EDI TO R I A L

Stop arguing and cut emissions

T

Jay Apt
is at the Tepper
School of Business
and Department of
Engineering and
Public Policy, Carnegie
Mellon University,
Pittsburgh, PA, USA.
apt@cmu.edu

y

“…a mixed
technology
portfolio…
offers the most
feasible route
forward…”

M. Granger Morgan
is in the Departments
of Engineering and
Public Policy and of
Electrical and
Computer Engineering,
Carnegie Mellon
University, Pittsburgh,
PA, USA. gm5d@
andrew.cmu.edu

y g
y
,
10.1126/science.adn9176

SCIENCE science.org

g

sonable way requires careful lifecycle accounting, but
one should avoid prescriptive standards.
As well, the US should avoid mandates that impose
high short-term costs. For example, as carbon dioxide
(CO2) emissions from electricity fall, switching home
heating to heat pumps becomes an attractive way to
reduce emissions. Some are now insisting that new
homes not use natural gas. However, on very cold
days, even the best of today’s heat pumps must switch
on electric resistance heat. Occasionally using a small
amount of natural gas in hybrid heat pumps, whose
expected life is only a couple of decades, would help
avoid huge short-term investments in further upgrading already stressed electricity distribution systems
while avoiding longer-term lock-in.
Rather than a ban on any use of gas,
it would be far better to implement
a performance standard that sets a
steadily diminishing upper bound on
emissions from new heating systems.
The Clean Heat Standards that Colorado and Vermont are considering
are modest steps in this direction.
The US should also continue to
ramp up investment in affordable
low-GHG technologies for electricity
and industrial processes while not
forbidding the use of specific fuels
if emissions can be controlled. The
biggest reason that the US electricity
system emits only about half as much
CO2 as it did in 2005, while generating 7% more power,
is that a lot of new natural gas plants have been built
and older, dirtier coal plants retired. Almost none of
these gas plants are equipped with CCS technology, but
many should be. Norway has had CCS at commercial
scale for almost 30 years. Because the US has no price
on CO2, it has been slower to adopt this technology, but
the 2022 tax incentives may unleash US innovation.
The US should also safely extend the life of as many
existing nuclear plants as possible. However, unless the
nuclear industry and regulators can achieve substantial
cost improvements, building new nuclear plants in the
US will not be economically viable.
Let’s not allow the perfect to become the enemy of
the good in responding to the climate emergency. A full
portfolio of low-carbon technology is the best way to do
that while minimizing the risks of impeding the adoption of even better solutions as they become viable and
affordable in the future.
–M. Granger Morgan and Jay Apt

p

he expert and policy communities have invested
enormous effort in debating what greenhouse
gas (GHG) emissions target to aim for, or which
decarbonization policies and technologies should
be mandated or banned. Because multiple trajectories can achieve similar targets and timelines,
some scenario analysis is useful. However, with
many players involved, it will be impossible to remain
on anyone’s optimal trajectory. As one of the world’s largest contributors to the climate crisis, the US should stop
arguing about perfect solutions and get on with reducing
emissions in ways that are feasible and affordable.
Renewables have made good progress in the US but
still produce only 21% of electric power. Arguments
that the US should ban all electricity
that is not made by wind or solar, or
that the only acceptable way to make
hydrogen is by using renewable electricity to split water, are costly and
delay action, when cost and time are
among our greatest challenges. Given
the urgency of the climate problem,
everyone should be focused on reducing the concentration of GHGs in the
atmosphere as rapidly as possible, using a broad portfolio of affordable and
achievable strategies.
In the long term, everyone needs
to stop burning all fossil fuels. In the
near term, multiple models show that
a mixed technology portfolio, which
includes some natural gas plants outfitted with carbon
capture and sequestration (CCS), offers the most feasible route forward for the US. We disagree with those
who argue that doing this is simply a smokescreen to
protect the fossil fuel industry. That said, we strongly
caution against those who champion natural gas as “a
bridge to the future” without simultaneously prioritizing the creation of the bridge abutment at the far end.
Beyond this there are three things the US should
be doing. For a start, because the country appears unable to implement a nationwide price on carbon or a
cap and trade regime, national performance standards
that become tighter over time should be specified for
each major category of emissions. Industry and users
can then innovate to meet or perhaps even exceed the
standards. The US has done this for control of conventional pollution and for automobile fuel economy;
California has done it for lighting. Rather than mandating specific solutions, the same approach should be
adopted for many other activities. Doing this in a rea-

1 MARCH 2024 • VOL 383 ISSUE 6686

933

NEWS
IN BRIEF

We warn about a risk of ‘scienticide,’ the systematic
destruction of Argentine science and technology.

“

”

Argentina’s Federal Science and Technology Roundtable, criticizing research budget cuts
and layoffs of science administrators under the new president, libertarian Javier Milei.

Edited by Jeffrey Brainard
PLANETARY SCIENCE

Early end for private Moon lander

| The governing body of the U.S.
National Science Foundation (NSF) plans
to recommend changes in the agency’s
grantsmaking process that it hopes will
prompt reviewers to pay more attention to a project’s potential benefits for
society. Since 1997, NSF has asked outside
reviewers to rate proposals based on both
intellectual merit and “broader impacts,”
or how a project could address societal
needs, such as public health and economic
development. But persistent concerns that
many reviewers give short shrift to these
potential effects prompted the presidentially appointed National Science Board in
2022 to name a committee to reexamine
the two metrics. Last week, the panel’s
chair, Stephen Willard, gave the board
a sneak peek at its report, due in May.
One recommendation would clarify the
intent of the broader impacts criterion by
changing its name to “societal benefits.”
The committee may also suggest reviewers
provide a separate score for that category.

L E G A L A F FA I R S

934

1 MARCH 2024 • VOL 383 ISSUE 6686

| Courts in Sweden and
Spain have rejected demands by a mysterious Polish company that scientists who
participated in its COVID-19 webinars pay
it tens of thousands of euros for “debate
fees”—charges the scientists called illegitimate. The company, Villa Europa, asked
researchers to sign licensing agreements
after the webinars, which contained long
clauses on the final page mentioning
charges for participation. The company
then billed the researchers for as much as
€80,000. Last month, a Swedish appeals
court announced it had dismissed claims
by the company against the three scientists; the decision cited “remarkable and
troubling circumstances” about the claims
and ordered it to pay about €68,000 to
partially cover their legal fees. The dismissal
followed a similar one in 2023 by a court in
Madrid, which backed a Spanish researcher’s challenge. This month, a court in Berlin
is set to hear another such request, from a
German scientist.
science.org SCIENCE

,

FUNDING

y

| Governments need to
step up oversight of research on microbes
that could cause a pandemic, says a
report this week from the Bulletin of the
Atomic Scientists. Its Pathogens Project,
a 28-member independent task force,
examined risks of an accidental or deliberate release from a lab studying “potential
pandemic pathogens” such as SARS-CoV,
SARS-CoV-2, H5N1 avian influenza, and
related viruses. Researchers should consider
using safer approaches for such studies
and weigh public health benefits against
risks, says the report, which also calls for
international rules on biosafety, mandatory reporting of lab accidents, and more
research on lab risks. The report stands
out because the panel included experts
from around the world, says co-chair David
Relman, a microbiologist at Stanford
University. It comes as U.S. officials are
considering new rules for research, known
as gain-of-function studies, that can make
dangerous pathogens riskier.

P U B L I C H E A LT H

y g

Courts quash webinar bills

y

NSF to seek ‘societal benefits’

g

Call for safer pathogen labs

p

T

he first private spacecraft to land on the Moon was
expected to shut down from a lack of power several
days ahead of schedule, as Science went to press. On
22 February, Intuitive Machines’s Odysseus lander,
built with $118 million from NASA, became the first
U.S. spacecraft (pictured during descent) since 1972
to touch down there, near the lunar south pole. Measuring
4.3 meters tall, Odysseus tipped on its side, which reduced
the light reaching its solar panels and blocked several antennas, limiting the operation of its scientific instruments. The
lander’s power supply dwindled, and it is unlikely to survive
the frigid, 2-week lunar night. Still, NASA hailed the mission as a successful start to its Commercial Lunar Payload
Services program, which aims to pay companies for low-cost
access to the Moon.

India genomes illuminate past

BIG GIFT A professor emeritus of pediatrics at the Albert Einstein College of
Medicine has donated $1 billion to the
school. Ruth Gottesman, who inherited a
fortune from her husband, a Wall Street
financier, stipulated that the college
use the gift, one of the largest to any
U.S. educational institution, to provide
future students free tuition. At Einstein,
Gottesman developed treatment methods for learning disabilities.

SCIENCE science.org

1 MARCH 2024 • VOL 383 ISSUE 6686

935

,

IN FOCUS Researchers last week reported finding more than 100 marine species
that may be new to science deep underwater off the coast of Chile. Using a robot,
they took images of the sponges, amphipods, urchins, crustaceans, and corals
(pictured) dwelling on four previously unmapped seamounts. The region is protected
by Chile as marine parks, which may help explain the biodiversity. The scientists’
continuing research cruise is funded by the Schmidt Ocean Institute.

MEASLES FREEDOM Public health
specialists criticized Florida’s surgeon
general for telling parents of unvaccinated students whose elementary
school had a measles outbreak they
are free to keep the potentially exposed
children in class. The stand by Joseph
Ladapo contradicts guidance from the
federal government that calls for such
students to remain home for 3 weeks.

y

NATURE LAW SURVIVES The European
Parliament this week narrowly approved
the Nature Restoration Law, which sets
targets to restore ecosystems. Last year,
lawmakers gave preliminary approval,
but the measure—a central piece of EU
leaders’ green agenda—was softened
after right-of-center parties criticized it
as unfairly burdening farmers.

y g

PHOTO: ROV SUBASTIAN/SCHMIDT OCEAN INSTITUTE

GENOMES AND RACE An uproar broke
out on X (formerly Twitter) after Nature
published a paper last week about a
massive U.S. research effort, the All of Us
project, that is studying links between
genes and health in people across
the country. Critics said a key figure,
which depicts patterns of relatedness
among nearly 250,000 study volunteers
whose genomes were sequenced, could
mislead some readers into thinking that
humans fall into distinct races defined
by genetics.

y

| A pilot program
will pay reviewers to check important published papers and preprints in psychology.
Reviewers will receive up to 3500 Swiss
francs (nearly $4000) depending on the
RESEARCH INTEGRITY

| More than one in 10 graduate students and postdocs in the health
sciences at Harvard University experiences food insecurity, a study has found.
Researchers surveyed 1745 early-career
researchers in 2023, finding that 17% of
graduate students and 13% of postdocs
consume food of inadequate quality or
quantity because of a lack of money or
other resources. Rates were even higher
—approximately 25%—among graduate
students who identified their race as other
than white or Asian and those who were
the first in their family to attend college.
The authors of the study, published last
week in JAMA Network Open and one of
the first of its kind, urged other institutions to gather similar information and
collectively develop policy interventions.
COMMUNITY

ANTARCTIC FLU Spanish researchers provided the first confirmation
that a highly pathogenic strain of
avian influenza has reached mainland
Antarctica. Tests showed that two
dead skuas found on 24 February near
Argentina’s Primavera research station
were infected with an H5 bird flu strain.
A government press release didn’t say
it was the H5N1 strain, but that seems
likely. H5N1, previously found on islands
in the Antarctic region, has killed millions
of wild birds and poultry worldwide since
2021 and poses a threat to Antarctica’s
dense penguin colonies.

g

Wanted: scientific bounty hunters

Harvard researchers face hunger

IN OTHER NEWS

p

| The largest genome
analysis of contemporary populations
living in India has revealed new details
about South Asia’s ancient genetic history. Indians’ genomes contain a striking
percentage of known Neanderthal and
Denisovan genes, researchers reported in a
preprint last week, based on an analysis of
2762 genomes from across India. Whereas
individual Indians derive 1% to 2% of their
ancestry from these archaic ancestors—a
rate similar to those of Europeans and East
Asians—the subcontinent’s populations
together hold a greater diversity of these
genes than seen in any other population,
comprising about 50% of all Neanderthal
genes known to have passed into modern
humans and 20% of known Denisovan
genes. The study also backed up previous
work suggesting modern Indians derive
primarily from three ancestral groups:
Iranian farmers, South Asian hunter-gatherers, and Eurasian Steppe pastoralists.

A N T H R O P O L O GY

severity of any errors uncovered. Authors
must consent in advance; they, too, will
receive compensation if they cooperate and
their work proves reliable. The program,
Estimating the Reliability and Robustness
of Research, is funded by the University of
Bern and modeled after payments by software companies to programmers who find
flaws in code. Backers say scientific authors
and reviewers lack incentives to identify
errors in the literature.

NE WS

IN DEP TH

p

REPRODUCTIVE HEALTH

g

Alabama IVF ruling may halt uterus transplants
Scientists fear wider research impacts if other states label frozen embryos “children”

T

science.org SCIENCE

,

1 MARCH 2024 • VOL 383 ISSUE 6686

y

936

A doctor removes frozen embryos created during in
vitro fertilization from a liquid nitrogen tank.

they are implanted in the uterus and a viable pregnancy can be detected—the impacts on would-be parents in Alabama were
immediate. Many IVF procedures have been
halted in a state where, in 2021, more than
1600 rounds of IVF treatment were completed, resulting in more than 400 babies.
But it also stands to affect UAB’s uterus
transplant program, led by immunologist
and transplant surgeon Paige Porrett, which
aims to help those who were born with defective or absent uteruses or had hysterectomies bear children.
For patients to qualify for a transplant,
they must first have embryos frozen for
implantation—an impossibility at UAB since
last week, when the university paused such
IVF procedures. “Thanks to the Supreme
Court ruling on Friday everything is at a full
stop,” Elizabeth Goldman, a UAB transplant
patient, wrote on Facebook. Goldman was
born without a uterus and received a uterus
transplant at UAB in 2022 after banking
frozen embryos; she gave birth to a daughter in November 2023 and had been scheduled to have another embryo transferred
in the coming months. “I’m on a timeline
with uterus transplant and this just makes
things way more complicated.”
UAB, which has touted the transplant program, declined to answer questions about it
or to make Porrett, who heads the program,
available for an interview. Porrett herself did
not respond to an emailed interview request.

y g

he Supreme Court of Alabama’s 16 February ruling declaring frozen embryos
at fertility clinics to be people has upended patient care there. The state’s
two leading private in vitro fertilization providers as well as the University
of Alabama at Birmingham (UAB) paused all
IVF procedures last week while officials assess the legal risks of continuing to create
and store embryos and impregnate patients.
But the impacts extend beyond wouldbe parents, to research. A uterus transplant
program in which women conceive by IVF—
one of only four in the country—is located
at UAB, and its leader is also co–principal
investigator on a related project studying
uterine immune cells. If other states follow
Alabama’s lead, other research, including
efforts to improve IVF outcomes and probe
developmental biology, could be imperiled
around the country. “There’s research that
is being done not just on uterus transplants,
but on IVF, on eggs, and on embryos,” says
Kate O’Neill, a reproductive medicine physician who directs the Uterus Transplant
Program at the University of Pennsylvania
(UPenn). Such studies, she says, “are going
to advance science, and if this spreads, [they]
would be very difficult to do.”
In its 8-1 decision, the Alabama Supreme
Court overturned a lower court decision
denying three couples the ability to recover

punitive damages for the accidental loss
of their frozen embryos. They sued a fertility clinic—the Center for Reproductive
Medicine—and an affiliated medical center
in Mobile because their embryos, created
through IVF, were dropped on the floor
and rendered unviable. The lower court had
held that frozen embryos were not children
under a 150-year-old law that allows parents
of a dead child to bring civil suits. The state
Supreme Court, however, declared “the law
applies to all unborn children, regardless of
their location.”
And although the future of the ruling in
Alabama is still uncertain—one state lawmaker, for instance, is working on a bill to
clarify that embryos are not “people” until

y

By Meredith Wadman

N E WS

Experiments suggest chemical reaction rates explain how
proteins came to be built from left-handed building blocks
By Robert F. Service

T

here’s a bias at the heart of life, and its
origin is an enduring mystery. Nearly
all the amino acid building blocks of
proteins today exist in mirror-image
forms, like right- and left-handed
gloves. But life only uses left-handed
ones, even though both forms should have
been equally abundant during the planet’s
early days and can readily link up in the lab.
Something must have tipped the balance
toward lefties in the primordial soup and
preserved the bias ever since.
Now, a trio of U.S. researchers proposes a
novel explanation. This week in Nature, they
report that by monitoring the formation

on one mirror-image form and not the other.
Several explanations have been advanced
in recent decades for life’s chirality, as the
bias toward a particular handedness is
known. For example, meteorites, which
could have seeded an early Earth, have
been shown to harbor amino acids with an
abundance of left-handed chirality, perhaps
because their contents were exposed to polarized light. Or magnetic fields on early
Earth could have given a twist to early biomolecules (Science, 16 June 2023, p. 1094).
But even if some external force imparted an
initial bias, what propagated it?
One clue comes from recent work by
Matthew Powner, an origin of life chemist
at University College London, and his col-

y
y g

rates of amino acid pairs, called dipeptides,
they’ve found multiple mechanisms that ultimately promote dipeptides whose two members share the same handedness.
“It’s quite convincing,” says Gerald Joyce,
a pioneering origin of life chemist and president of the Salk Institute for Biological Studies who was not involved with the study.
Researchers next hope to learn whether
the same mechanisms skew larger peptides
and proteins toward left-handedness—and
whether it can explain the opposite bias in
RNA and DNA, whose bases have sugars that
are inevitably right-handed. If so, the new
mechanisms could explain how life itself took

leagues. Over the past 5 years, Powner’s
group has discovered a set of sulfur-based
molecules that likely would have been
present on early Earth and shown how
they readily link individual amino acids to
amino acid precursors, called aminonitriles,
forming dipeptides. Because these reactions
take place in water and work with all the
amino acids found in living organisms, they
offer a plausible route to how the first proteins may have formed.
Powner’s team didn’t check whether its
sulfur-based catalysts had a chiral bias.
That’s where Donna Blackmond, an origin of life chemist at Scripps Research,
1 MARCH 2024 • VOL 383 ISSUE 6686

937

,

Biased chemical reactions within the primordial soup on early Earth may have led to amino acid pairs that are
fully left-handed (right) even though some with mixed handedness (left) were initially made faster.

y

SCIENCE science.org

New origin of life theory may
explain biomolecular handedness

g

Porrett and O’Neill are co–principal investigators on a research effort attached to
a similar transplant program at UPenn: a
$1.9 million U.S. National Institutes of
Health project to study key pregnancy-sustaining immune cells in the transplanted
uteruses. They aim to document the tissue
origins of uterine natural killer cells, which
are critical to successful implantation of the
embryo in the uterus. By taking biopsies of
the endometrium, or uterine lining, from
those who have received uterus transplants,
O’Neill says, the team is probing where in
the body the cells originate. “Understanding this will make huge impacts outside of
uterus transplants, in pregnancy in general,” and potentially in transplantation of
other organs.
Other avenues of experimental work—
for instance, comparing culture mediums
in fertility clinic labs to see whether one
produces better embryo survival rates
than another—will be affected in Alabama,
says Michael Allemand, a reproductive
endocrinologist at Alabama Fertility Specialists, an IVF clinic that has paused all
treatments. “No one in the state can do that
kind of work including the university practices if they can’t do IVF.”
Outside Alabama, other research relies
on frozen embryos, donated from IVF clinics with patient consent. Such studies have
illuminated how chromosomal abnormalities arise during the very earliest days of
embryonic development and how such
abnormalities are strongly associated with
IVF embryos ceasing to develop and then
dying in lab dishes. In other work, scientists
have compared the genetics of cells in the
outer layer of IVF embryos, which are often biopsied for genetic testing before the
embryos are implanted, with cells from the
rest of the embryo. The comparison showed
that in 90 out of 93 embryos, the outer cells
accurately mirrored chromosomal abnormalities in the rest of the cells.
Such research could be at risk if other
states explicitly declare frozen embryos to be
people. “It will be very tempting for states
that have very strong antiabortion leanings
to try to do copycat legislation … or litigation,” says Susan Crockin, an expert in reproductive technology law at Georgetown
University Law Center. She notes that several states—including Missouri, Arkansas,
Kentucky, and Oklahoma—already have statutes that define life as beginning at fertilization. “And so you’ve already got the stepping
stones if you want to argue that frozen embryos or [any] embryos should be swooped
into any protections for living people.” j

BIOCHEMISTRY

p

ILLUSTRATION: N. BURGESS/SCIENCE

Kristia Rumbley of Alabama holds her youngest child,
born via IVF. She has frozen two more embryos.

NE WS | I N D E P T H

Proposal would combat warming
by drying the stratosphere
Seeding clouds above the western Pacific would keep water
vapor, a greenhouse gas, out of the atmosphere’s rooftop
By Paul Voosen

y
,

science.org SCIENCE

y g

G

iven the alarm about rising levels of
carbon dioxide and methane, it’s easy
to forget that plain old water vapor is
a major greenhouse gas, too. It can linger for years in the stratosphere, for
example, absorbing heat from the surface and re-emitting it back down. According
to one study, a possible jump in stratospheric
water during the 1990s may have boosted
global warming by up to 30% during that
time. But what if you could stop water from
getting there in the first place?
That’s the idea behind a new geoengineering technique, proposed this week in Science
Advances. By targeting rising, moist air
and seeding it with cloud-forming particles
right before it crosses into the stratosphere,
geoengineers could cool the world with an
intervention far more delicate than other
schemes. Drying the stratosphere might take
as little as 2 kilograms of material a week,
says Shuka Schwarz, the study’s lead author
and a research physicist at the Chemical Sciences Lab of the National Oceanic and Atmospheric Administration (NOAA). “That’s an
amount of material that helps open the mind
to imagine a whole bunch of possibilities.”
“Intentional stratospheric dehydration,” as
it’s called, could only cool the climate moder-

ately, offsetting roughly 1.4% of the warming
caused by increased carbon dioxide over the
past few hundred years. But for geoengineers
who have talked about cooling the planet by
loading the stratosphere with thousands of
tons of reflective particles, “it’s clearly a new
idea,” says Ulrike Lohmann, an atmospheric
physicist at ETH Zürich. “This is something
that could work.”
The scheme relies on a key fact: Only a few
places in the world are hot enough to generate the powerful updrafts needed to lift air
into the stratosphere, which begins between
9 and 17 kilometers above the surface, depending on latitude. The most important
of these portals is found above the western
equatorial Pacific Ocean, in a region roughly
the size of Australia.
Along its upward journey, much of the water condenses into clouds and rains out of the
air. But in the past decade, NASA used a high
altitude, jet-powered drone to study the cold
layers just below the stratosphere and found
plenty of air masses moist enough to form
clouds, but lacking in particles that would
allow the moisture to condense into ice crystals and ultimately rain. “It’s a question of
chance, whether they get to this coldest spot
on their journey and there’s enough cloud
nuclei left to do anything,” Schwarz says. The
NASA studies also found that this moisture

y

1 MARCH 2024 • VOL 383 ISSUE 6686

GEOENGINEERING

g

938

Condensation nuclei
help moist updrafts form
towering storm clouds.

p

and her colleagues Min Deng and Jinhan Yu grabbed the baton. They tested
two of Powner’s sulfur compounds to see
whether the catalysts were sensitive to
chirality as they formed dipeptides. They
were, but not in the way Blackmond
had expected. The catalysts created
about four times as many “heterochiral”
dipeptides—those pairing a left-handed
amino acid (L) with a right-handed (D)
one—as fully chiral products. “We thought
it was bad news,” Blackmond says, because
it suggested that even if amino acids on
early Earth started with a bias, it would
have been scrambled as proteins formed.
But as Blackmond and her colleagues
looked more deeply, the news got better. In a
series of experiments, the Scripps researchers started with skewed proportions of L
and D amino acids—for example, 60% Ls
and 40% Ds. The L,D and D,L heterochiral
dipeptides formed most quickly, and
as they did they pulled equal numbers
of L and D amino acids out of the mix.
Because of the baseline bias, eventually a predominance of Ls remained in
the pool of unreacted amino acids, raising the likelihood of forming fully lefthanded dipeptides. “It’s like a domino
effect,” Powner says. The first heterochiral
reaction eventually encourages more
homochirals to form. “And it’s a general
process that works with all amino acids,”
Powner says. Joyce adds: “It’s just math.”
Follow-up experiments suggested a
second bias that amplifies the effect. The
team found that heterochiral dipeptides
precipitate out of a solution more quickly
than homochiral ones, speeding the way to
a relative abundance of either homochiral
L,L or D,D pairs, depending the starting
mix. Just why this precipitation bias occurs isn’t yet clear, Blackmond says. However, Joyce says, together with the other
effect, “it beautifully fits the [experimental] data.” Blackmond adds: “The wrong
answer turned out to be the right answer
to get us to homochirality.”
For now, this push toward a particular
handedness has only been shown with
dipeptides. But Blackmond says preliminary work suggests the same biasing process unfolds when the sulfur catalysts
stitch short peptides together into longer
peptide chains.
Joyce thinks it’s possible that the same
sort of math may also help explain how life’s
genetic molecules gained their handedness.
“This could happen with all kinds of other
things, like RNA,” he says. Perhaps it was just
a statistical coin flip that caused an original
bias toward building blocks of one handedness to form, Joyce says. “But once that coin
flipped it caused other coins to flip.” j

A dramatic shortage of the oral vaccine may ease in the
years ahead as more companies enter the market
By Pratik Pawar

T

y g

he world has run out of cholera
vaccines—just when the deadly disease is on a rampage not seen in
many years. Fifteen countries are
currently reporting active outbreaks,
with more than 40,900 cases and
775 deaths reported in January alone. But
all available doses of oral cholera vaccines
in the global stockpile have been allocated
until mid-March, Philippe Barboza, cholera
team lead at the World Health Organization
(WHO), said on 23 February. He said there
is now “no buffer for unforeseen outbreaks
or preventive campaigns.”
The catastrophic shortage is a result
not just of a surge in cases, but also of an
overdependence on a single vaccine manufacturer, EuBiologics in Seoul, South Korea, whose production capacity is limited.
“Unfortunately, there is no short-term solution to increase vaccine production,” says
Daniela Garone, international medical coordinator at Doctors Without Borders (MSF).
But hope is on the horizon. EuBiologics
is working to ramp up production of a simplified vaccine, and companies in South
Africa and India are preparing to enter the
market as well. The shortage “will lessen in
2024 and should be substantially addressed

by 2025,” says Julia Lynch, director of the
cholera program at the International Vaccine Institute (IVI), also in Seoul; by 2026 or
2027, it could even be a “crowded market,”
she predicts.
Cholera, caused by the Vibrio cholerae
bacterium, causes terrible diarrhea, with
the body flushing out as much as 1 liter of
fluid per hour. Left untreated, it can kill
in less than a day. Many public health experts attribute the current surge, which
began in late 2022, at least partially to climate change. Extreme weather events in
Pakistan, Malawi, and Mozambique have
destroyed health and sanitation infrastructure, allowing the bacterium to thrive.
Armed conflict and displacement of people
in the Democratic Republic of the Congo
and Yemen have also led to outbreaks.
The two-dose oral cholera vaccine, developed by IVI based on earlier research
in Vietnam, contains several strains of inactivated V. cholerae bacteria, along with a
part of the toxin secreted by the bacterium,
which boosts the immune response. Two
doses given 2 weeks apart can offer robust
protection for at least 3 years, and, when
deployed early enough, can prevent outbreaks from ballooning.
The global stockpile of the vaccine, established in 2012, is managed by an Inter-

y
y
,

PHOTO: NYASHA MUKAPIKO/EFE VIA ZUMA PRESS

World’s cholera vaccine stash is
empty—but relief is on its way

g

SCIENCE science.org

GLOBAL HEALTH

p

was concentrated: Just 1% of the air parcels
explored accounted for half of the water that
could end up in the stratosphere.
In a simple model, the team simulated
injecting bismuth triiodide, a nontoxic compound that has been used in lab studies of ice
nucleation, into the 1% areas most ripe for water harvesting. In an optimistic scenario, just
2 kilograms a week of seeds 10 nanometers
in diameter would be enough to convert
those moist air parcels into clouds, they
found. Such an amount could be sprayed by
balloons or drones, with no airplane needed.
Daniel Cziczo, an atmospheric chemist at
Purdue University, says the idea is interesting but could pose risks. If the seeds failed
to form clouds in the right place and spread
elsewhere, they could speed the formation of
the wrong kinds of clouds: thin, wispy cirrus
clouds, which reflect little sunlight but absorb infrared heat from the surface, Cziczo
says. “You’re basically exploring a technique
that could have a warming effect and not a
cooling effect.”
Mark Schoeberl, an atmospheric scientist
at the Science and Technology Corporation
who previously identified the stratospheric
gateway in the Pacific, agrees with the need
for further study. “You want to avoid unintended consequences and make a cleareyed assessment of implementation cost.”
The technique likely won’t be effective all
year round, he adds, because most water
reaches the stratosphere during the Asian
monsoon seasons. And just how much a reduction in stratospheric water would cool the
surface is uncertain, he says.
Schwarz sat on the idea for a while, wary of
the controversy that surrounds all proposals
for tinkering with the planet to offset humancaused warming. But now that the U.S. Congress has mandated that NOAA study solar
geoengineering, “the stigma around considering climate intervention is abating a bit,” he
says. “Two years ago, I for one would have really hesitated to consider these possibilities.”
The openness is spreading. For example,
the European Union is now supporting research into geoengineering governance.
Switzerland last week called on the United
Nations to support research in the area. And
Lohmann’s group last week won a grant from
the Simons Foundation to study another intervention: thinning heat-trapping clouds
above polar regions to mitigate warming.
“Things have changed on the science
agenda,” Lohmann says. She says climate
scientists have reservations about exploring
these schemes but feel there is no choice.
Emission cuts simply haven’t happened fast
enough, and carbon dioxide cannot yet be
sucked out of the air cheaply. “It’s clear we’re
looking for something else,” she says. “It’s our
failure as humans to avoid this.” j

A woman receives a dose of the oral cholera vaccine at a clinic in Kuwadzana, Zimbabwe, on 29 January.
1 MARCH 2024 • VOL 383 ISSUE 6686

939

NE WS | I N D E P T H

national Coordinating Group on Vaccine
Provision made up of experts from WHO,
UNICEF, MSF, and the International Federation of Red Cross and Red Crescent Societies. It can rapidly send vaccine to countries
in need. The number of doses available
for shipment reached 36 million in 2023
(see graphic, below) and could be close to
50 million this year.
But all will be needed to fight ongoing
outbreaks, and the shortage has forced
some difficult choices. In late 2022, the coordinating group announced it would stop
giving people second doses; even a single
dose, studies have found, can provide substantial protection against cholera for a few
years at least. The stockpile group must
also choose where to send the scarce doses,
and how to distribute them on the ground.
UNICEF, one of the partners managing the

35
30
25
20
15

Count as of
22 February
Februaryy

10

0
2016

2018

2020

2022

2024

1 MARCH 2024 • VOL 383 ISSUE 6686

science.org SCIENCE

,

940

Pratik Pawar is a science journalist in Bengaluru, India.

P

igs, cattle, and other livestock with
edited genes are still far from most
dinner plates, but a U.K. company
has taken a big step toward the
supermarket by engineering several
commercial breeds of pigs to be resistant to a virus that devastates the swine
industry. The firm, Genus plc, hopes that by
year’s end the U.S. Food and Drug Administration (FDA) will formally approve the pigs
for widespread human consumption, a first
for a gene-edited animal.
Alison Van Eenennaam, an animal geneticist at the University of California,
Davis, is cheering the news. “There’s no
point having a pig getting sick and dying if
there’s an approach to genetically prevent
it from doing so,” she says, adding that this
benefits farmers, the pigs, and, ultimately,
the consumer.
But Van Eenennaam laments the regulatory hoops the company is having to jump
through. FDA views the DNA change made
by the genome editor CRISPR as an “investigational new drug” that requires multiple
submissions from Genus to establish the
altered gene’s safety, heritability, and ability to protect against the virus. “You’re talking about a very, very expensive regulatory
pathway,” she says. It is unnecessary, she
argues, because unlike genetically modified organisms, which sometimes have DNA
from other species added, the gene editing involved the pigs’ own DNA, creating
changes that could happen naturally.
The gene edit made by Genus outwits a
virus that kills nearly all the suckling pigs
it infects and weakens older ones as well.
The virus, which causes a condition called
porcine reproductive and respiratory syndrome (PRRS), has spread worldwide
and costs the pork industry an estimated
$2.7 billion annually. Eight years ago, a

y

stockpile, says the aim is “to allocate supply
in the most impactful manner.” In the case
of Zimbabwe, for example, the 2.3 million
doses approved in late January were prioritized for use in the capital city Harare and
in Buhera district, the outbreak’s epicenter,
even though cholera has occurred in 63 of
the country’s 64 districts.
To boost supply, EuBiologics has simplified its original vaccine, with funding from
the Bill & Melinda Gates Foundation. The
new formulation, Euvichol-S, contains two
strains of the inactivated V. cholerae bacteria
instead of five, and the recipe drops a heat
inactivation step. That makes the vaccine
easier to produce and about 20% cheaper.
This will further expand EuBiologics
production capacity by 38%, to about
52 million doses annually, according to a De-

By Jon Cohen

y g

5

United States and other
countries may soon approve
virus-proof pigs

y

Doses approved for shipping (millions)

40

Meat from
gene-edited
pigs could hit
the market

g

As cholera outbreaks have surged, the number
of cholera vaccine doses approved for shipment
to outbreak areas has grown rapidly, depleting
the global stockpile. Close to 50 million doses are
expected to ship this year.

AGRICULTURE

p

A surge in demand

cember 2023 press release from IVI, which
has supported the development. A phase 3
trial in Nepal in 2022 showed the simplified
version protects as well as the original.
Euvichol-S is currently under review for
“prequalification,” a seal of approval from
WHO that the company expects to come
through in April, says EuBiologics Director Rachel Park. Then Gavi, the Vaccine Alliance will procure the vaccine at roughly
$1.5 per dose to replenish the stockpile.
EuBiologics is also building a new facility
that could expand its production capacity to
90 million doses annually.
To end the reliance on a single manufacturer, IVI in 2022 began to help Biovac,
a South African company, set up a facility
to produce the simplified vaccine. Biovac
tells Science it plans to start clinical trials in 2025 and hopes to produce 30 million vaccine doses annually in 5 or 6 years.
Biovac has been encouraged by a so-called
advanced market commitment from Gavi, a
pledge to buy many vaccine doses at established prices. So is the Indian manufacturer
Biological E, which plans to produce IVI’s
simplified vaccine.
Another Indian biotechnology company,
Bharat Biotech, is working on its own lowcost cholera vaccine, Hillchol. Bharat has
said little about its progress and did not respond to questions, but a phase 3 trial—at
various locations in India—finished in early
2023, Lynch says. Manufacturers in China,
Vietnam, and Bangladesh also make cholera vaccines, but their products have not
been prequalified by WHO.
IVI is developing an injectable conjugate
vaccine, in which an antigen from V. cholerae is attached to a protein carrier. Such
conjugated vaccines are known to induce
a stronger T cell immune response, generally leading to better and longer lasting
protection. They could be used to immunize infants, for whom the oral vaccine is
not licensed, through routine vaccination
programs. They may also offer better protection for young children, in whom the
oral vaccine is “least efficacious,” Lynch
says. IVI just finished phase 1 trails for a
conjugate vaccine in adults and is seeking funds for phase 2 studies in choleraendemic regions.
All of that is good news for cholera
control—but scientists and public health
workers keep stressing that the real solution
to preventing cholera is clean drinking water
and better sanitation and hygiene. That is a
tall order, however, says Abi Kebra Belaye, an
MSF coordinator working in Zimbabwe—so
for the foreseeable future, “we are pushing
for more vaccines.” j

By Elisabeth Pain

S

1 MARCH 2024 • VOL 383 ISSUE 6686

941

,

cientists in France are reeling after
the government announced last week
a €904 million cut to this year’s budget for research and higher education.
The cut is part of a broader €10 billion savings package, which the government says is necessary for the country to
limit its deficit in the face of dwindling economic growth since the budget was passed
in December 2023. But scientists say the research sector is bearing a disproportionate
share of the pain.
Under the planned cut—which represents
a 2.8% reduction in this year’s higher education and research budget—funding for national research organizations such as CNRS
and the National Research Agency, which
funds competitive research, will be reduced
by €383 million across all disciplines. Universities are set to lose €80 million for teaching
and research, and funding for student support
measures will be reduced by €125 million.
The higher education and research ministry has reassured scientists that public institutions would continue to receive funding
for their routine operations and that staff
salaries for researchers and existing commitments for student support will be preserved.
And despite the cuts, a spokesperson says the
ministry’s budget is still higher than in 2023.
Nonetheless, the move “comes as a surprise,” says Valérie Masson-Delmotte, a
climate scientist from the Pierre-Simon Laplace Institute, who worries the budget cuts
will reduce opportunities for early-career
scientists in particular. In a statement on
26 February, France Universités, which represents all universities in France, expressed
“deep concern” about the impact of the cuts.
Initiatives to protect the environment
and curb climate change are also facing a
€2.2 billion reduction. Despite the environmental crisis, climate policy and funding for
research and education “are used as mere
budgetary adjustment variables by our leaders,” says Julien Gossa, a political scientist at
the University of Strasbourg. “This leaves us
very little hope.” j

y

PHOTO: GENUS

Research and education to
lose almost €1 billion

y g

SCIENCE science.org

Budget cuts
alarm French
scientists

y

Pigs that were gene edited to make them resistant to a deadly virus look and taste like normal swine.

EUROPE

g

sugar, and Revivicor aims to sell meat from
GalSafe pigs to this specialized market.
FDA more recently gave less formal
endorsements to five pigs that had been
CRISPR-edited by a Washington State University team and a line of cattle edited by
Acceligen Inc. to have short hair to better
withstand heat. But neither received full approval for human consumption or is being
produced at a commercial scale: The pigs
received an “investigational food use authorization,” which took 2 years and more than
$200,000 to obtain, and two short-haired
cattle and their future offspring were given
a “low risk determination” for marketing.
In both cases, the introduced changes occur
naturally in the animals.
The CD163 modification used to protect
against PRRS could well occur naturally but
has never been observed in pigs, creating
higher hurdles for FDA clearance, says Clint
Nesbitt, a molecular biologist who oversees
regulatory affairs at Genus. As a result, he
says, “We have to go through the full, complete review system at FDA. There are no
shortcuts for us.” Still, he says, Genus has
made “good progress” with the agency.
The challenging U.S. regulatory environment for gene-edited food animals was on
the agenda of a National Academies workshop this week. Other countries are less restrictive. Regulators in Colombia in October
2023 indicated that because the Genus pigs
do not involve transgenics, they will treat
the swine the same as conventionally bred
animals. The firm is also seeking approval
in China, the largest consumer of pork.
Nesbitt says it will take time for producers to breed them into their herds. “Nowhere on the planet is it going to be a light
switch, where suddenly everybody’s got the
edited pigs,” he says. “It’s going to be much
more like a dimmer switch. … And we still
have to have a lot of conversations about
market acceptance.” j

p

team led by Randall Prather at the University of Missouri reported it could make pigs
resistant to PRRS by using CRISPR to disable a receptor, CD163, on pig cells that the
virus uses to establish an infection.
Now, Genus has translated “proof-ofconcept work to a commercial scale,” it reports in the February issue of The CRISPR
Journal. Scientists at the company used
CRISPR to modify early embryos in four
lines of pigs that are used in commercial
production of pork. By implanting the edited embryos into females, then breeding
the progeny, they created lines with both
their copies of the CD163 gene disabled.
Rodolphe Barrangou, a food scientist at
North Carolina State University who is also
editor-in-chief of The CRISPR Journal, says
the study is the “end of the beginning” of
bringing gene-edited livestock to the wide
market because so many farmers will likely
want PRRS-resistant pigs. “It’s not just a
nice study in a nice model,” Barrangou says.
“It’s actually doing it in the real world.”
Vaccines exist for PRRS but they lack the
100% protection seen with the gene edit.
Prather, whose university holds patents on
this modification and has a licensing agreement with Genus, says the CRISPR edit has
benefits beyond reducing financial losses in
the pork industry. The virus, he says, threatens food security and creates “psychological
and emotional issues” for producers that
have to euthanize the sick pigs. “CD163edited pigs are a solution.”
FDA has so far formally approved two genetically modified food animals, but neither
is widely consumed. One is a salmon that
has a gene from another fish species and
grows faster, but consumer concerns have
limited sales. The second, known as the GalSafe pig and made by Revivicor, had DNA
inserted to cripple a gene for a sugar molecule on the surface of its cells. Some people
have mild to severe allergic reactions to this

NE WS

FEATURES

p
g
y
y g

TAKING THE STAND

For scientists, going to court as an expert witness brings risks and rewards
y
,

L

ate last year, the sound of scientific
argument echoed through a New
York City courtroom packed with
legal and financial experts. Studies from top epidemiology journals
flashed onto large screens, as lawyers debated their statistical power
and whether their conclusions
rested on “cherry-picked” data.
Billions of dollars were at stake.
The scientists themselves were absent,
and attorneys argued on their behalf. But
the crucial issue was whether some of the
scientists would be allowed to appear at
a future trial, where they would tell jurors that children had developed autism
942

1 MARCH 2024 • VOL 383 ISSUE 6686

By Dan Charles
or attention-deficit/hyperactivity disorder
(ADHD) as a result of exposure to the painkiller acetaminophen, often sold as Tylenol,
while still in the womb.
Five researchers from Columbia University, the Baylor College of Medicine, the
Albert Einstein College of Medicine, and
other prominent institutions had submitted
reports arguing that acetaminophen’s links
to autism and ADHD are real. They’d been
paid by lawyers for the plaintiffs, who included parents alleging their children had
been harmed by the painkiller. But, “These
scientists are not professional witnesses,”

plaintiffs’ attorney Ashley Keller told the
court as he displayed their faces on a screen.
“They care deeply about public health.”
The opposing side had its own
scientists—an additional half-dozen of
them, with equally illustrious academic credentials, paid by companies that make or
sell acetaminophen.
U.S. District Judge Denise Cote played
the role of all-powerful peer reviewer. She
had to decide whether the plaintiffs’ expert
opinions were based on “reliable principles
and methods,” and thus admissible in court.
It was a pivotal issue; if Cote ruled against
admitting the experts for the plaintiffs,
their case could collapse. As the hearing
science.org SCIENCE

N E WS

FOR MANY academics, life as an expert witness

begins with a lawyer’s recruiting pitch, often unexpected and sometimes unwelcome.
Brent Wisner, a lawyer in Los Angeles who
often sues pharmaceutical or chemical companies, says about half of the scientists he
recruits ultimately turn him down.
That was Beate Ritz’s initial impulse.
When approached to testify against the
makers of glyphosate—on the side opposite
Mucci—the University of California (UC),
Los Angeles epidemiologist says her first
reaction was, “No, I don’t want to do this. I
have enough to do.”

“Attorneys, by their nature,
want to polarize things. …
And in my experience,
y.”
things are sort of gray.”

y g

THOSE WHO SAY yes to becoming an expert

1 MARCH 2024 • VOL 383 ISSUE 6686

943

,

witness describe a mix of motives.
Cristian Tomasetti, a mathematician at
the City of Hope Cancer Center, saw testifying on behalf of Bayer in the glyphosate
case “as another way to help change how
cancer causation is understood.” Tomasetti’s
research suggests cancer is caused not just
by genetic or environmental factors, but
also by mutations that occur randomly.
Fombonne says he ultimately decided to
take the stand in defense of vaccine safety
for the sake of public health. He saw vaccination rates declining because of unfounded
concerns that vaccines caused autism.
In some cases, serving as an expert witness can be professionally rewarding, says
Christopher Higgins, an environmental
engineer at the Colorado School of Mines
who has testified about the sources of con-

y

Eric Fombonne, a specialist on autism at
Oregon Health & Science University, also
initially resisted requests, more than a decade ago, to testify against claims that vaccines cause autism. “I had no experience
with litigation. And I’m a bit more reserved
as a person,” he says.
Frequently, there’s a fear that “you’ll be
seen as selling out,” Swan says. “It’s not
looked on very favorably by a lot of people”
in the scientific community. And some expert witnesses run into problems with their
employers, especially if their testimony
draws the ire of powerful political or business interests (see sidebar, p. 945).
Some scientists also worry taking sides
in litigation will force them to present onesided interpretations of the evidence. “The
more you study and construct arguments
and think through your arguments, it
reinforces the position you’ve taken,” says
David Eastmond, an environmental scientist who recently retired from UC Riverside. Eastmond worked briefly as an
expert witness, but found the adversarial
process uncomfortable.
“Attorneys, by their nature, want to polarize things, so that things become more black
and white. And in my experience, things are
sort of gray,” Eastmond says.

y

David Eastmond,
environmental scientistt

Social pressures can also come into play,
says David Sedlak, an environmental engineer at UC Berkeley. “You hang out with
the lawyers for many hours. You’re having
meals and socializing with them. You’re
part of that team. Your human nature is,
you’re going to want to please them, by telling them what they want to hear.”
Sedlak tells colleagues who are considering work as expert witnesses to keep some
distance from their legal team, for the sake
of their reputations. “Your job is not to spin
the facts,” he tells them. Lawyers care most
about how facts fit their legal arguments, he
says, but “for scientists, the facts are important, and that’s the end of the story.”
For some scientists, the rewards simply
aren’t worth the risks. Jane Hoppin,
an epidemiologist at North Carolina
State University who studies the
health effects of chemical exposures,
doesn’t work for litigants because it
might prevent her from serving on
government advisory panels. “If you
take money from anybody besides
the university or the federal government, basically you can’t play” because of concerns about conflicts of
interest, she says.
In other cases, scientists disagree
with the legal arguments they’ve
been asked to support. Hugh Taylor
of the Yale School of Medicine is the
co-author of a “call for precautionary action” on acetaminophen use during
pregnancy. But he turned down invitations
to work with the lawyers suing companies
that sell the drug. The evidence so far is
a “warning,” he says, “but to accuse somebody of doing something wrong, that we
can sue them over it, I think is wrong. I
don’t think the level of evidence ever rose
to that.”

g

SCIENCE science.org

I wasn’t going to say this, who was going
to say this?”

p

a prompt decision.
closed, Cote promised
p
academic firepower brandished in
The academ
2023 federal court hearthat 7 December
Decem
unusually impressive. But sciing was unu
regular guest in U.S. courtrooms.
ence is a regu
Hydrologists and toxicologists testify routinely about tthe sources and consequences
of groundwater
groundwat contamination. Structural
engineers assign blame for collapsed buildings. In criminal cases, scientists explain
DNA evidence, or the limitations of prosecutorial tools such as fingerprints or eyewitness identification.
Lawyers often hire technical specialists
from consulting firms who have made a
career out of serving as an expert witness.
But when the stakes are high, and the science is crucial—as in the acetaminophen case—they often prefer to
bring in university professors. It’s an
unusual, and often taxing, role for
most academics, and Science set out
to learn what it is like, interviewing
both scientists and the lawyers who
hire and question them.
Academics who have served as
expert witnesses say the experience
comes with a complicated mix of
rewards and risks. The work sometimes involves depositions and trials
that can last days, weeks, or months.
Cross-examinations can be hostile
and challenging.
The work can be lucrative: Some
experts in the acetaminophen case charged
more than $600 per hour. But it is haunted
by the specter of science for hire, of expertise distorted by the lure of easy money.
And those accusations, researchers say, can
be emotionally bruising and, potentially,
professionally damaging.
“It was a tough experience,” says Lorelei
Mucci, an epidemiologist at Harvard University’s T.H. Chan School of Public Health
(HSPH) who 6 years ago testified on behalf
of the company Monsanto, now owned
by Bayer. The firm was fighting claims
that its weed killer glyphosate, also called
Roundup, causes non-Hodgkin lymphoma.
Facing hostile questioning on the witness
stand, Mucci says, “was really one of the
most challenging things I’ve experienced
in my professional life.”
Still, some researchers feel an obligation
to share their expertise within the justice
system. “It was absolutely an ethical responsibility,” says epidemiologist Shanna
Swan of the Icahn School of Medicine at
Mount Sinai, who was among the first to
recognize the harmful effects of diethylstilbestrol (DES), a synthetic hormone once
commonly prescribed during pregnancy.
She decided to testify in cases against
DES manufacturers, she says, because “if

NE WS | F E AT U R E S

THERE ARE TRIALS where science speaks

with a single voice. Charles Weaver, an expert on human memory at Baylor University, often takes the stand to explain that the
memories of eyewitnesses are fallible. He’s
usually the only academic scientist testifying. “There’s really an emerged consensus”
in this field, Weaver says. “There’s really not
another side” that’s supported by data.
Even if the science isn’t clear cut, there’s
usually a dominant scientific view on one
side, says Andrew Jurs, an expert on legal
evidence at Drake University’s Law School.
“Then the other side has to get an outlier.”
Cases in which top scientists end up clashing are “unusual,” he says.
In those exceptional, but often important, cases, the opposing experts sometimes
reflect existing debates in the scientific
community. Ritz says, for instance, that
epidemiologists who specialize in the genetic causes of cancer are disinclined to attribute it to environmental exposures. They
often testify for the companies in cases involving exposure to chemicals. The opposite
is true for experts like Ritz, who focus on
environmental factors and tend to testify
for plaintiffs in those cases.
On the whole, though, Swan says the
dominant culture of science tends to discourage researchers from taking the side
of plaintiffs, because their claims are somescience.org SCIENCE

,

Mucci charged $350 per hour; Ritz, like
several experts for the plaintiffs, earned
$500 an hour for reviewing documents and
preparing, but $5000 per day when testifying at trial. Both spent hundreds of hours
on the case, testifying in both federal and
state courts.
They had plenty of company. Scientists
from Boston University, the City of Hope
Cancer Center, UC Berkeley, Michigan State
University, Columbia University, and McMaster University ended up testifying for
and against the claim that glyphosate causes
non-Hodgkin lymphoma.
Neither Ritz nor Mucci had worked as
expert witnesses before, and they had little
inkling of what awaited them.
Academic life does prepare a scientist
for some aspects of litigation. “We’re used
to having our science attacked,” says Jamie
DeWitt, director of the Environmental
Health Sciences Center at Oregon State
University, and an experienced expert
witness in cases involving per- and polyfluoroalkyl substances, long-lasting chemicals used to make nonstick coatings and
stain-resistant fabrics.
In litigation, though, the hostility is
overt, and questions quickly move past
data and methods to focus on a scientist’s
motives and biases. “What they’re trying
to do is show that you are a shoddy scientist,” Ritz says. “And that is really hard to
take. You have to have a very good sense of
yourself, in order to not feel denigrated as
a professional.”

y

1 MARCH 2024 • VOL 383 ISSUE 6686

Beate Ritz, University of California, Los Angeles

y g

944

“If they get really mad
at you because they cannot
trip you up on the science,
they try to paint you as a hack.”

y

lions of dollars hung in the balance, Ritz
and Mucci ended up on opposite sides.
Ritz ultimately agreed to work with the
plaintiffs who were suing Bayer, whereas
Mucci worked with the company. Each says
encounters with their respective lawyers
helped allay their fears.
As Ritz describes it, the plaintiffs’
lawyer “was very respectful and said,
‘We would never tell a scientist what to
say. That’s the worst thing you can do.’”
Mucci, for her part, says she came to believe Bayer’s lawyer “of course was working for the company, but I felt that I could
trust what he was telling me. And that
he valued the fact that I would approach
things independently.”
Both reviewed the scientific literature
on glyphosate before agreeing to take on
the case. Ritz found clear evidence of a
link between glyphosate and non-Hodgkin
lymphoma, whereas Mucci came to believe
the opposite.
Ritz relied on an array of “case-control”
studies that compared people who’d devel-

For Mucci, the stress reached its peak
while being cross-examined by Wisner,
the lawyer for the plaintiff. “I came at her
pretty hard,” Wisner recalls. “Because I
thought she was the most important witness for the defendants.”
“It felt very adversarial,” Mucci says. “I
tried to focus on my breathing, focus on
the science, and tried to disassociate from
what felt like a lot of anger being thrown at
me. But it was very hard.”
Wisner says Mucci’s discomfort was visible. “She came across as scared,” he
says. “She kind of fell apart.”
Mucci also failed to convince the
jury, which in that trial decided that
glyphosate had caused the plaintiff ’s
cancer and awarded hundreds of
millions of dollars in damages.
“I really took it personally,” Mucci
says. “I felt that I had failed.”
She also faced attacks outside the
courtroom. Activists sent out postcards carrying her photograph, condemning her for testifying on behalf
of Bayer. Mucci got one, and so did
colleagues at HSPH.
The glyphosate litigation continues. In 19 trials held so far, Bayer has prevailed in 10 of them, and the plaintiffs have
won nine times.

g

IN THE GLYPHOSATE litigation, in which bil-

oped the lymphoma with those who had not.
The studies showed the people with cancer were more likely to report exposure to
glyphosate than those in the control group.
Mucci, in contrast, gave more credence to
a large “prospective cohort” study that has
monitored the health of tens of thousands of
agricultural workers and pesticide applicators for more than 20 years, and has found
no association between glyphosate exposure
and cancer. Each dismissed the other’s preferred studies as flawed.

p

taminants in water. “It’s a great way to see
the real-world implications of your work,”
he says.
He adds that in academic life, intellectual
debates happen “in bits and pieces,” sandwiched between meetings and rewriting
papers. But, “When you get into a deposition, it’s a full-on, 8-hour—I don’t want to
say battle of wits, but you have to be paying attention. Having to be on your feet,
thinking, is, for me, quite intellectually
stimulating. That’s kind of the fun part.”
Plus, “The money is good,”
Higgins says. “Sometimes it becomes, like, man, I really could do
this full time if I wanted to.” The
thought passes quickly, though. “I
really do enjoy being a professor.”
A few hired experts, such as geochemist Avner Vengosh at Duke University, have worked for free in order
to show they aren’t just scientific
hired guns. Vengosh did, however,
accept research funding from environmental groups who successfully
sued the Tennessee Valley Authority
(TVA) over groundwater pollution
near one of the its coal ash ponds. He
used the money for studies that uncovered
evidence that contamination came from
the TVA site. Vengosh published his findings in a peer-reviewed journal. “So when
the trial happened and I took the stand,
basically I used the paper,” Vengosh says.
“I was looking at them and saying, ‘Hey, it’s
not me. … That’s the science talking to you.’
I think it was very powerful.”

A witness, then a target

O

ILLUSTRATION: A. MASTIN/SCIENCE

,

945

y

1 MARCH 2024 • VOL 383 ISSUE 6686

y g

hearing, Cote
opened the proceedings by noting that a
key federal rule on expert testimony “has
been revised recently to emphasize the
court’s gatekeeping role.”
The revised rule, which took effect in
December 2023, states that judges must
be persuaded by a “preponderance of evidence” that an expert’s testimony is based
on reliable scientific methods. Otherwise,
the testimony should be excluded.

IN THE ACETAMINOPHEN

Cote embraced that test. She posed detailed questions to Keller, the plaintiffs’
lawyer, as she presented studies showing
evidence for an acetaminophen-autism
link. When Keller accused the sellers of acetaminophen of trying to “keep pregnant
women in the dark” about potential dangers of their product, Cote interrupted and
told Keller to stick to facts. “Disinformation
is a big issue in our society,” she admonished him.
As he left the courtroom after the hearing, Keller looked a little shell-shocked. “I
think it’s time for tequila,” he said, to no
one in particular.
His worries were well placed. Cote took
less than 2 weeks to reject the scientific
case that the plaintiffs’ experts had laid
out. “Their analyses have not served to
enlighten but to obfuscate the weakness
of the evidence on which they purport
to rely,” she wrote. Her decision put an
abrupt end to the case, unless it’s appealed
and overturned.
For now, those scientists will not be appearing in court. j

y

SCIENCE science.org

Both Ritz and Mucci say they’d be willing to work as expert witnesses again. Ritz,
in fact, is currently working with attorneys
who are exploring a lawsuit targeting the
herbicide paraquat, which has been implicated as a cause of Parkinson’s disease.
But Ritz recently advised a younger colleague to reject work as an expert witness
in a different case. “I warned him,” she
says. “I said, ‘You’re too young to get your
name tainted.’ Because if they get really
mad at you because they cannot trip you
up on the science, they try to paint you as
a hack.” The colleague decided not to do it.

g

times hard to prove. “If you’re going out and
saying, ‘This causes that’—that’s a pretty
hard thing to say in science,” Swan says. “It’s
easier to say that more research is needed.
There are lots of safe things you can say that
nobody would object to.”
Personal sympathies and professional
loyalties can steer experts to one side or the
other. Swan, for example, notes she spent
the first 18 years of her career working for
California’s health department. “That really
instilled in me, ‘How do I protect people’s
health?’” As a result, Swan says she’s predisposed to work for plaintiffs who claim
harm, rather than companies accused of
causing it.
Mucci, for her part, says it was the
epidemiological data, not the client, that
convinced her to work for the company.
“It was, for me personally, kind of a tough
thing initially to say, ‘I’m working for Monsanto,’” she says. Some acquaintances, she
says, “definitely thought that I was on the
wrong side.” Professionally, though, she
hasn’t felt any negative repercussions. Since
the trial, she was promoted to full professor.

Carpenter wasn’t worried by the request.
“I thought, if they’re just trying to see if it’s
r
really
true that I don’t accept money myself,
it’ll take them about half a day to figure that
out. I gave them all my invoices,” he recalls.
For months, however, Carpenter remained in limbo. Finally, early in 2023, he
told his story to a reporter for an Albany
newspaper, and the ensuing front-page
story got immediate results. A state legislator demanded an explanation from UA,
and Ian Rosenblum, senior vice chancellor of
the State University of New York system, summoned Carpenter and university officials to his
office. Ac
According to Carpenter, Rosenblum “basically
said, ‘You will resolve this, and you’re not going home
until you do.’”
On 21 February 2023, 9 months after Carpenter was barred
from campus, university officials told him he could go back to
work. There was no further disciplinary action. Carpenter did
agree to sign an agreement laying out “safeguards” to prevent
his work as an expert witness from interfering with his academic
responsibilities.
Describing the experience, Carpenter sounds unperturbed.
And in a statement, university officials declared “our full commitment to unfettered academic freedom.”
But a report by the University Senate, which represents faculty
members, condemned UA’s actions. “The community at large
needs expert testimonies from reputable researchers,” it wrote,
adding that the university’s behavior “plays directly into the
hands of the external entities that would profit from silencing
them.” —D.C.

p

e Unin 27 May 2022, administrators at the
avid
versity at Albany (UA) summoned David
’s
Carpenter, director of the university’s
Institute for Health and the Environment, to an unscheduled meeting.
They told Carpenter he was the target
of a disciplinary investigation, and offered him the option to resign. Carpenter
refused. They then told him he would be
on
barred from campus while the investigation
continued.
University officials didn’t explain what
prompted the investigation, and they still
penter
refuse to comment on the case. But Carpenter
s work as
says the investigation was triggered by his
cademics, joinan expert witness—illustrating that for academics,
ing litigation comes with risks.
For the past 2 decades, Carpenter has testified frequently on
behalf of plaintiffs who have sued companies such as Monsanto, now owned by Bayer, for exposing them to polychlorinated biphenyls, carcinogenic compounds once widely
used in industrial and consumer products. “I was on the debating team in high school, and I loved it,” Carpenter says. His
background made him “feel pretty confident in myself,” and able
to hold his own in the courtroom.
Three months before Carpenter met with university administrators, he learned that a law firm representing Bayer had filed a
request for copies of financial records relating to his work as an
expert witness. Carpenter often tells jurors that he’s not doing
it for the money. The payments go instead to his staff or graduate students. In a typical year, his witness gigs bring in about
$200,000 for his research program.

I NS I GHTS

and AMPA receptor (AMPAR) accumulation at synapses. Importantly, SynGAP has
been recognized as a substrate for calcium/
NEUROSCIENCE
calmodulin–dependent protein kinase type
II (CAMKII), which is the central signaling
kinase in several forms of synaptic plasticity,
including LTP. Phosphorylation of SynGAP
by CAMKII enhances its enzymatic activity
and reduces its affinity for PSD95. Consequently, neuronal activity causing calcium
influx through NMDARs, as occurs during
LTP induction, leads to dispersion of Synmemory as well as neuronal development (9,
By Daniel Choquet1,2
GAP from the postsynaptic density (PSD)
10). These multiple roles have led to conflict(1, 11), which is a protein-dense region at
ing theories about the function of SynGAP.
he synaptic Ras/Rap guanosine trithe postsynaptic membrane where PSD95,
SynGAP influences excitatory synaptic
phosphatase (GTPase)–activating proNMDAR, and AMPAR accumulate in front of
transmission through N-methyl-D-aspartate
tein (SynGAP) plays substantial, albeit
neurotransmitter release sites (12).
(NMDA) receptors by modulating the mistill elusive, roles in synaptic function
The exact mechanism responsible for
togen-activated protein kinase (MAPK)–ex(1). SynGAP has attracted considerable
AMPAR recruitment at synapses during LTP
tracellular signal-regulated kinase (ERK)
attention owing to its pivotal role in
remains elusive but involves diffusion trapsignaling pathway (8). The GAP domain of
modulating excitatory glutamatergic synapping of surface AMPARs at synapses and
SynGAP stimulates the GTPase activity of
tic transmission and neuronal development
increased AMPAR exocytosis (12). Activitysmall GTPases, such as RAS and RAP. Howand because loss-of-function mutations in the
dependent trapping of AMPARs at synapses
ever, RAS and RAP have opposing roles in
SYNGAP1 gene account for up to 1% of genethas been proposed to arise from CAMKIIsynaptic function, blurring the understandically based intellectual disabilities. SynGAP
induced phosphorylation of transmembrane
ing of the role of the GAP domain (1).
comprises two primary functional domains:
AMPAR regulatory proteins (TARPs) and
Long-term potentiation (LTP), a form
the GAP domain and the C-terminal domain
their consequent increased binding to PSD95
of synaptic plasticity that is a core cellular
(CTD). The GAP domain negatively regu(12). This finding, together with the activitymechanism for learning and memory, inlates small G protein signaling, which may
dependent dispersal of SynGAP, has been
volves NMDA receptor (NMDAR) activation
be crucial for activity-dependent changes in
integrated into the “slot hypothesis”
synaptic strength, whereas the CTD
(13) of synaptic plasticity, in which
binds to postsynaptic density protein
the PSD harbors binding sites (slots)
95 (PSD95), but the functional conseSynGAP in synaptic potentiation
for AMPAR (presumably PSD95)
quences of this are unclear. On page
Wild-type SynGAP
that are in part occupied by SynGAP
963 of this issue, Araki et al. (2) reIn resting basal conditions, part of the PSD95 “slots” in the postsynaptic
when the synapse is at rest. The Synveal that contrary to common belief,
density are occupied by SynGAP. This prevents accumulation of AMPAR and
GAP-PSD95 complex undergoes liqGAP domain activity is dispensable
inhibits RAS activity, limiting synapse growth from actin polymerization.
Upon strong synaptic stimulation, SynGAP is dispersed, allowing diffusion
uid-liquid phase separation, which
for many SynGAP functions and that
trapping of AMPAR and relief of RAS inhibition, which results in synapse
is necessary for its synaptic localthe CTD is important for the plaspotentiation and growth.
ization (14). Upon neuronal activity
ticity of synaptic transmission. This
and CAMKII-dependent phosphorysuggests that potential therapeutics
Basal
Potentiated
lation of TARPs and SynGAP, some
targeting the GAP domain should be
TARP
slots are freed, allowing for diffusion
reconsidered.
trapping of AMPARs and increased
Discovered in 1998 (3, 4), SynGAP
AMPAR
Active
RAS
synaptic transmission (1, 13). This
is expressed in neurons and localized
PSD95
Actin
hypothesis has mixed acceptance;
at excitatory synapses. It has various
Inactive RAS
on the one hand, the GAP function
structural isoforms resulting from
SynGAP
of SynGAP has been proposed to be
different promoter usage and alterinstrumental for AMPAR exocytosis,
native splicing (5–7). These isoforms
whereas on the other hand, TARPpotentially contribute differently to
GAP-deficient SynGAP*
g8 phosphorylation is proposed to
SYNGAP1-associated cognitive disWhen GAP activity of SynGAP is blocked (SynGAP*), synapses are larger in
basal conditions owing to alleviation of RAS inhibition but are not functionally
disrupt PSD95 phase separation,
orders. Loss-of-function mutations
potentiated because SynGAP* still binds PSD95. Upon neuronal activity,
leading to reduced AMPAR clusterin the SYNGAP1 gene, which arise
SynGAP* disperses, resulting in diffusion trapping of AMPAR and functional
ing (15). This conflicts with the idea
de novo, are associated with intelsynaptic potentiation.
of TARP-g8 mediating AMPAR diflectual disability, autism, and epiBasal (potentiated size)
Potentiated
fusion trapping at synapses in an
lepsy, owing to haploinsufficiency (1).
activity-dependent manner.
The resulting disruption of SynGAP
The study of Araki et al. chalfunction impairs normal synaptic
lenges existing ideas about SynGAP’s
plasticity (8), affecting learning and
function and provides support for
1Interdisciplinary Institute for Neuroscience,
the slot hypothesis. The authors elGAP-deficient SynGAP*
University of Bordeaux, CNRS, UMR 5297,
egantly show, using in vitro and in
F-33000 Bordeaux, France. 2Bordeaux Imaging
vivo approaches in mice, that expresCenter, University of Bordeaux, CNRS, INSERM,
AMPAR, AMPA receptor; PSD95, postsynaptic density protein 95; SynGAP, synaptic Ras/Rap
sion of a SynGAP mutant devoid of
US4, UAR 3420, F-33000 Bordeaux, France.
guanosine triphosphatase (GTPase)–activating protein; TARP, transmembrane AMPAR
regulatory protein.
Email: daniel.choquet@u-bordeaux.fr
GAP function (SynGAP*) can sustain

PERSPECTIVES

Shifting rules in a brain disorder
The mode of action of a synaptic protein is challenged

y g
y
,

science.org SCIENCE

y

1 MARCH 2024 • VOL 383 ISSUE 6686

g

950

p

T

By Lauren Sumner-Rooney

T

D.C. receives support from European Research Council (ERC)
grant Dyn-Syn-Mem.
10.1126/science.adn8707
SCIENCE science.org

Leibniz Institute for Evolution and Biodiversity Research,
Museum für Naturkunde, Berlin, Germany. Email: lauren.
sumner-rooney@mfn.berlin

1 MARCH 2024 • VOL 383 ISSUE 6686

951

,

AC KN OWL EDG M ENTS

y

he extent to which evolution is predictable is a frequent flashpoint for
debate. Do organisms simply stumble
around the adaptive landscape, or are
they “channeled” along certain paths
or trajectories? The emergence of convergent traits, whereby the same phenotypes
evolve from different origins, spotlights this
question. However, it is notoriously difficult
to answer, particularly at macroevolutionary scales. On page 983 of this issue, Varney
et al. (1) characterize “critical junctions” as
the moments in evolutionary time when the
trajectory toward one of several observed
phenotypes is set, using the visual systems
of chitons, a class of benthic marine mollusk.
Their findings highlight an apparently crucial prerequisite to vision and a fundamental
divide that appears to predate the evolution
of either eyes or eyespots, which implicates a
strong role for path dependence in the origin
of complex traits over geological timescales.
The general body plan and ecology of chitons have remained relatively stable for more
than 300 million years, but they still exhibit
a broad range of adaptations and modifications. Some chitons have visual systems
comprising dozens, hundreds, or even up
to hundreds of thousands of units that are
spread across the surface of the dorsal shell
plates in a network that expands continuously over the life span of the animal (2, 3).
These visual units can take one of two forms
in different species: shell eyes, which have an
aragonite lens covering a retina that is capable of resolving images (4–6), or eyespots,
which are smaller, more numerous, and generally lack a distinct lens (7, 8). Whether the
eyespots (individually or as a network) are
capable of image formation remains unclear,
but they mediate defensive responses to
shadows and the alignment of the body with
darker areas of their environment (8).
Unlike most molluscan shells, chiton
valves are stuffed with nerves, which
thread through them in dense networks
of channels. At the surface sit tiny sensory
organs called aesthetes, which have been
implicated in chemo-, mechano-, and pho-

toreception (9–12). These aesthetes are
thought to be a precursor for the evolution of the more elaborate eyespots and
shell eyes, through both enlargement and
the addition of components that confer
new functional abilities, such as pigmentation and lenses. Previous work suggested
that eyespots represent an intermediate
step between aesthetes and shell eyes (8).
Although this at first appears to be a parsimonious explanation in line with theoretical expectations for eye evolution (13),
the phylogenetic distribution and composition of the eyes and eyespots do not lend
strong support to this stepwise hypothesis.
Evidence presented by Varney et al. further demonstrates multiple, separate, and
recent origins of eyes and eyespots (14),
raising the fascinating question, Why do
some lineages evolve eyes and others eyespots? Varney et al. explored two characteristics of the aesthete networks: aesthete
density at the shell surface and the number of slits in the anteriormost shell valve,
where efferent nerves coalesce and connect to the visceral nervous system. They
found that although increased aesthete
density always predated or co-occurred
with the elaboration of aesthetes to visual
organs, this was true for both shell eyes
and eyespots. However, the number of
slits in the anterior valves was consistently
higher in lineages that lead to the simpler
eyespots than in those that lead to eyes. In
a reconstructed evolutionary history (phylomorphospace), the pathways to the two
visual systems appear clearly separated,
highlighting increased slit numbers as an
empirical example of a critical junction in
their evolutionary path (see the figure).
These results are intriguing. The insertion slits represent the connection of the
shell pore network to the surrounding
body, and their complexity may be related
to signal organization and efficiency. One
tempting possibility is that the increase in
insertion slits provides capacity for directionality or even spatial resolution across
the network of eyespots, even if eyespots cannot achieve this individually. An increased
number of slits could reflect more potential
sites of projection from these nerves to the
chiton medullary cords, possibly providing
a basis for the preservation of coarse spatial
information in very rough analogy to the in-

y g

9.
10.
11.
12.
13.
14.
15.

T. R. Gamache et al., J. Neurosci. 40, 1596 (2020).
Y. Araki et al., Science 383, eadk1291 (2024).
J. H. Kim et al., Neuron 20, 683 (1998).
H. J. Chen et al., Neuron 20, 895 (1998).
Y. Araki et al., eLife 9, e56273 (2020).
M. Kilinc et al., eLife 11, e75707 (2022).
A. C. McMahon et al., Nat. Commun. 3, 900 (2012).
G. Rumbaugh et al., Proc. Natl. Acad. Sci. U.S.A. 103,
4344 (2006).
J. P. Clement et al., Cell 151, 709 (2012).
M. Birtele et al., Nat. Neurosci. 26, 2090 (2023).
Y. Araki et al., Neuron 85, 173 (2015).
L. Groc, D. Choquet, Science 368, eaay4631 (2020).
W. G. Walkup IV et al., eLife 5, e16813 (2016).
M. Zeng et al., Cell 166, 1163 (2016).
M. Zeng et al., Neuron 104, 529 (2019).

Key traits set the course of de novo
visual system evolution in marine mollusks

y

1.
2.
3.
4.
5.
6.
7.
8.

Critical junctions in evolution

g

RE F ER E NC ES AND NOTES

EVOLUTION

p

normal activity-dependent synaptic dispersal, AMPAR recruitment at synapses, LTP,
and learning—even in homozygous mouse
Syngap1* mutants (see the figure). This finding contrasts with the lethal phenotype observed when Syngap1 is deleted in mice.
Notably, Araki et al. observed that
SynGAP* does not support activity-dependent spine enlargement. On the contrary, the
basal spine size in rat neurons expressing
SynGAP* or in homozygous Syngap1* mice
is enlarged, but the spontaneous synaptic
current amplitude remains unchanged. This
supports the idea that the GAP activity may
be important for changes in spine size but
does not affect AMPAR content in synapses.
This is also important because spine size is
commonly used as a surrogate measure of
synapse strength. Indeed, structural synaptic
plasticity is often confounded with functional
synaptic plasticity. The study of Araki et al. is
another example of the dissociation between
functional and structural synaptic plasticity.
It is fascinating that the authors discovered five carriers of GAP-disabling SYNGAP1
mutations that were not associated with any
diagnosed neurological or cognitive disorder.
Altogether, this probably calls for reevaluation of the understanding of SynGAP in synaptic plasticity and cognition. These findings
shift the focus from the enzymatic activity of
SynGAP to its structural properties and interactions at the PSD. There are also implications for developing treatments for SYNGAP1
mutation–associated disorders. Future therapies might need to focus on preserving SynGAP’s structural functions rather than solely
targeting its GAP activity. Conversely, the role
of SynGAP in early neuronal development
(10) indicates that SYNGAP1 mutation–associated brain disorders may also arise through
nonsynaptic mechanisms. Thus, the roles of
SynGAP need to be deciphered by using tools
that can acutely regulate its functions to distinguish its involvement in neuronal development and synaptic function. j

I NS I GHTS | P E R S P E C T I V E S

why chitons evolve vision at all; given the
overlap of the eyeless chiton morphospace
with that of both visual system types, the
observed fates are accessible rather than
inevitable. These outstanding questions do
not diminish but rather emphasize the potential impact of identifying critical junctions in directing future research on the
evolution of complex traits. The findings
of Varney et al. represent a promising step
toward understanding the enigmatic visual
systems of these understudied organisms and
consolidate chitons as an important evolutionary model. j
R E F E R E N C ES A N D N OT ES

L.S.-R. receives support from the Deutsche
Forschungsgemeinschaft Emmy Noether Programme
(SU1336/1-1).
10.1126/science.ado1700

In chiton shells, the number of insertion slits, where efferent nerves coalesce and connect to the visceral nervous
system, acts as a critical junction in the evolution of either eyes or eyespots. Chiton lineages that evolved visual
systems exhibit a higher density of sensory organs called aesthetes, but the type of visual system (eyespots or
shell eyes) that evolved was constrained by the number of slits on the anteriormost shell valve.
Critical junction

Shell valve
Girdle

Insertion slit

Eyespot

Aesthete
Posterior

952

1 MARCH 2024 • VOL 383 ISSUE 6686

Increased slit number

science.org SCIENCE

,

Eye

Anterior

T

he biological activity of a living cell is
reflected in the dissipation of heat to
its surroundings. However, quantifying heat and cellular activity at the
nanoscale has been a challenge (1).
The recent development of stochastic
thermodynamics (2) has brought access to
measuring the energetics and efficiency of
microscopic systems, such as cells, through
a combination of stochastic theory and
high-resolution experimental techniques.
On page 971 of this issue, Di Terlizzi et al.
(3) report a thermodynamic constraint
applicable to nonequilibrium, stationary
fluctuations and apply it to determining
the heat dissipated by living cells at the nanoscale. Their analysis reveals nonuniform
heat dissipation along the equatorial cell
contour of red blood cells. The approach
may lead to more accurate measurements
and a deeper understanding of energy efficiency in living matter, from single cells
to whole organisms.
The distinction between the motion
of living (active) and dead (passive) matter at the microscale is at the core of the
genesis of statistical mechanics. Having
examined the motion of pollen grains
under the microscope, botanist Robert
Brown concluded that their lively, zigzag-like motion was not attributable to
biological processes (4). His observations
triggered Einstein’s microscopic theory
describing the statistics of Brownian motion, which is now considered one of the
building blocks of statistical mechanics.
Most active systems, however, do not move
like Brownian particles. For example, the
bacterium Escherichia coli exhibits the socalled run-and-tumble motion in which
periods of ballistic-like motion alternate
with sudden changes in direction of motion (5). Similarly, differentiated living
cells are often not optimized to swim like
bacteria but execute specific functions,
such as metabolism, that result in nonBrownian fluctuating motion. A hallmark
of cells with metabolic activity is the consumption of chemical fuel resources, such

y

Prerequisite
Increased aesthete density

By Édgar Roldán

y g

Eyes or eyespots?

Nonequilibrium fluctuations
reveal nonuniform heat
dissipation in living cells

y

AC K N OW L E D G M E N TS

Thermodynamic
probes of life

g

1. R. M. Varney et al., Science 383, 983 (2024).
2. D. R. Chappell, D. I. Speiser, D. J. Eernisse, A. C. N. Kingston,
in Distributed Vision: From Simple Sensors to Sophisticated
Combination Eyes, Springer Series in Vision Research, E.
Buschbeck, M. Bok, Eds. (Springer, 2023), pp. 147–167.
3. J. D. Sigwart, L. Sumner-Rooney, Biol. Bull. 240, 23 (2021).
4. D. I. Speiser, D. J. Eernisse, S. Johnsen, Curr. Biol. 21, 665
(2011).
5. L. Li et al., Science 350, 952 (2015).
6. N. T. Moseley, Q. J. Microsc. Sci. 25, 37 (1885).
7. M. G. Sturrock, J. M. Baxter, J. Zool. 235, 127 (1995).
8. A. C. N. Kingston, D. R. Chappell, D. I. Speiser, J. Exp. Biol.,
221, jeb.183632 (2018).
9. P. Omelich, Veliger 10, 77 (1967).
10. P. R. Boyle, Cell Tissue Res. 172, 379 (1976).
11. L. B. Arey, W. J. Crozier, J. Exp. Zool. 29, 157 (1919).
12. F. P. Fischer, Spixiana 1, 209 (1978).
13. D.-E. Nilsson, Vis. Neurosci. 30, 5 (2013).
14. X. Liu, J. D. Sigwart, J. Sun, Mar. Life Sci. Technol. 5, 525
(2023).
15. D. R. Chappell, D. I. Speiser, J. Exp. Biol. 226, jeb244710
(2023).

BIOPHYSICS

p

sect central complex. This would represent
a starkly different approach to achieving
spatial resolution compared with species
with shell eyes, where the arrangement of
neural projections from individual eyes to
the medullary cords directly reflects their
sampling distributions (15).
The findings of Varney et al. suggest a
powerful role for evolutionary path dependence, where slit number can irreversibly
restrict chitons to one of two fixed visual
system types and has done so in four separate origins of vision. If so, this unlocks
the exciting possibility of understanding
the aesthete and visual systems in fossil
chitons and offers a tractable new model
for studying constraint and convergence
at macroevolutionary timescales. Future
work, particularly on the function and
distribution of eyespots, will be crucial to
testing this assertion. Notably, unlike shell
eyes, eyespots are not readily preserved
in fixed or fossilized material and their
phylogenetic, and phylomorphospace, distribution may therefore be greater than
presently appreciated. Moreover, the particular relevance of the anterior valves
must be explored to determine whether,
and how, the fate of a vast distributed visual network can be so tightly constrained
by variation in only one region.
Finally, the possible existence and nature of evolutionary triggers that surround
this proposed critical junction remain unknown. Other factors that affect anterior
slit numbers, and improved data on the
sensory and behavioral ecology of different chiton lineages, will help to clarify

y
,

PHOTO: STEVE GSCHMEISSNER/SCIENCE PHOTO LIBRARY UR

y g

953

y

1 MARCH 2024 • VOL 383 ISSUE 6686

g

SCIENCE science.org

p

in the position of a sensing probe (9, 10).
perimentally inaccessible, as is the case of
as adenosine triphosphate. Although cells
Estimating the heat dissipation of a livprobes sensing active fluctuations (such as
can convert most chemical energy input
ing cell by recording spontaneous fluctuamembrane vibrations of red blood cells).
into mechanical work, it is never 100% eftions in the motion of a sensing probe is a
Data from three different red blood cell
ficient, resulting in energy dissipation to
major challenge. One problem is that often
assays were studied: optical-tweezer sensthe environment in the form of heat. Unonly partial, coarse-grained information is
ing, optical-tweezer stretching, and optiderstanding energy efficiency in biologiavailable (such as the position of a probe),
cal microscopy of membrane vibrations.
cal processes thus requires extending the
whereas several nonequilibrium variables
The three platforms yielded consistent
theory of nonequilibrium thermodynamics
(such as intracellular active forces) remain
estimates for the heat dissipation of up to
to the nanoscale.
inaccessible to experimental measurement.
106 kBT/s for red blood cells (where kB is
Research on stochastic thermodynamBoltzmann’s constant, T is the
ics has led to universal laws
temperature of the environconstraining the performance
ment, and s is seconds). This is
of small systems (2). For exaround a thousand to a million
ample, the second law of thertimes higher than estimates for
modynamics was extended to
red blood cells depleted of an
Langevin stochastic dynamics,
energy source, and of the same
which describe the fluctuating
order of magnitude as the heat
motion of cells. For a system
dissipation measured from bulk
that is not in thermodynamic
calorimetry. The variance sum
equilibrium, there exists an
rule allowed access to uneximbalance between the change
plored local heat flux density
of entropy—the overall amount
along the equatorial cell conof molecular disorder—and
tour. Spatially resolved heat disthe entropy that flows to the
sipation maps revealed a finite
system’s environment. The avcorrelation length of around
erage of such an imbalance is
The rate of heat dissipation varies along the surface of a human red blood cell.
half a micrometer, paving the
often called entropy producway toward topographic calorition inasmuch as it is always
Some approaches have established lower
metric considerations (14) in active matter.
positive and thus generated within the
bounds for heat dissipation of biological
Besides estimating heat dissipation,
system. In most experiments with living
systems in terms of quantitative measures
applications of the variance sum rule are
cells, nonequilibrium steady states arise in
of the progression (so-called arrow) of time
envisaged in the field of inference. It can
isothermal (constant temperature) condi(11–13). However, quantifying the arrow of
be combined with artificial intelligence
tions. For most isothermal steady states,
time from noisy biological data is a hercuand machine learning algorithms to exthe rate of entropy production is well aplean task involving large uncertainties.
tract unknown parameters from stochastic
proximated by the steady-state rate of heat
Di Terlizzi et al. show that the steadymodels described by Langevin equations.
dissipation divided by the temperature of
state rate of heat dissipation can be accuDeveloping reliable estimates in nonstathe environment.
rately estimated by simply measuring the
tionary processes could provide better unAdvances in experimental techniques
variance of the fluctuations of a position
derstanding of biological processes such
have enabled high-resolution tracking of
(such as the center of mass of a Brownian
as embryo development, cell differentiation,
passive probes interacting with living sysparticle) and a force (such as that exerted
and cell division (15). j
tems, such as colloidal particles interacting
by an optical tweezer). The main result is a
with metabolically active cells. Single-cell
RE FE REN C ES AN D N OT ES
proposed variance sum rule equality. This
studies on living auditory hair-cell bundles
1. X. Yang et al., Proc. Natl. Acad. Sci. U.S.A. 118,
relates variance of the position, variance of
(6), oocytes (7), and red blood cells (8) ree2026786118 (2021).
vealed that cellular activity affects fluctuathe impulse (the integral of the force with
2. L. Peliti, S. Pigolotti, Stochastic Thermodynamics: An
Introduction (Princeton Univ. Press, 2021).
tions in the position of the sensing probe.
respect to time), and excess variance that
3. I. Di Terlizzi et al., Science 383, 971 (2024).
By comparing the spontaneous motion of
quantifies dissimilarities between position
4. R. Brown, Philos. Mag. 4, 161 (1828).
the probe with its linear response funcand force fluctuations over time. From
5. H. C. Berg, Sci. Am. 233, 36 (1975).
tion to weak external stimuli, these studies
their variance sum rule, the authors de6. P. Martin, A. J. Hudspeth, F. Jülicher, Proc. Natl. Acad. Sci.
U.S.A. 98, 14380 (2001).
revealed a violation of a thermodynamic
rived an estimate for the heat dissipation
7. É. Fodor et al., Europhys. Lett. 116, 30008 (2016).
principle—Kubo’s fluctuation-dissipation
rate in terms of the variance of the force
8. H. Turlier et al., Nat. Phys. 12, 513 (2016).
theorem. When in contact with a living
and the short-time curvature of the mean
9. E. Roldán, J. M. R. Parrondo, Phys. Rev. Lett. 105, 150607
(2010).
cell, the probe’s motion could not be desquared displacement of the position as a
10. É. Fodor et al., Phys. Rev. Lett. 117, 038103 (2016).
scribed as thermal, Brownian fluctuations.
function of time. In other words, the vari11. É. Roldán, J. Barral, P. Martin, J. M. R. Parrondo, F.
Thus, to further determine the in vivo nonance sum rule establishes that deviations
Jülicher, New J. Phys. 23, 083013 (2021).
equilibrium thermodynamic properties of
from Einstein’s theory for diffusion are a
12. A. Bacanu, J. F. Pelletier, Y. Jung, N. Fakhri, Nat.
Nanotechnol. 18, 905 (2023).
living cells, it is essential to develop tools
signature of nonequilibrium.
13. I. A. Martínez, G. Bisker, J. M. Horowitz, J. M. R. Parrondo,
that allow one to quantify how much acDi Terlizzi et al. verified their findings
Nat. Commun. 10, 3542 (2019).
tive matter is away from equilibrium by
with theoretical stochastic models, using
14. P. Dolai, C. Maes, K. Netočný, SciPost Phys. 14, 126
measuring only spontaneous fluctuations
experimental data from optically trapped
(2023).
15. J. Rodenfels, K. M. Neugebauer, J. Howard, Dev. Cell 48,
colloidal particles dragged through wa646 (2019).
ter. Moreover, they extended the variance
Quantitative Life Sciences Section, The Abdus Salam
sum
rule
to
situations
where
one
or
few
International Centre for Theoretical Physics (ICTP),
Trieste, Italy. Email: edgar@ictp.it
system-probe interaction forces are ex10.1126/science.adn9799

I NS I GHTS | P E R S P E C T I V E S

CHEMISTRY

Nitrogen cuts in during C–C cross-coupling
A catalyst system diverts traditional C–C bond coupling into desired C–N bond formation
By Kevin H. Shaughnessy

T

science.org SCIENCE

,

10.1126/science.ado0068

y

1. A. Biffis, P. Centomo, A. Del Zotto, M. Zecca, Chem. Rev.
118, 2249 (2018).
2. A. B. Pagett, G. C. Lloyd-Jones, in Organic Reactions, vol.
100, S. E. Denmark, Ed. (Wiley, 2019), chap. 9, pp. 547.
3. J. F. Hartwig et al., in Organic Reactions, vol. 100, S. E.
Denmark, Ed. (Wiley, 2019), chap. 14, pp. 853.
4. D. G. Brown, J. Boström, J. Med. Chem. 59, 4443 (2016).
5. P. Onnuch et al., Science 383, 1019 (2024).
6. L.-C. Campeau, N. Hazari, Organometallics 38, 3 (2019).
7. D. Bhattacherjee et al., Adv. Synth. Catal. 363, 1597
(2021).
8. C. M. Abdulla Afsina, R. M. Philip, P. V. Saranya,
G. Anilkumar, Curr. Org. Synth. 20, 308 (2023).

y g

1 MARCH 2024 • VOL 383 ISSUE 6686

R E F E R E N C ES A N D N OT ES

y

954

Onnuch et al. extended this atom insertion concept by achieving an aminative SM
coupling through the three-component coupling of a nitrene (NH) precursor, an aryl
halide or triflate (–O3SCF3), and an arylboronic acid. In this reaction, the nitrene unit
is introduced during the SM catalytic cycle,
resulting in the formation of two new C–N
bonds. Successful development of this reaction required overcoming several potential
challenges. Modern SM catalysts afford
high rates for C–C bond formation. Selective formation of the diarylamine product
requires efficient nitrogen insertion before
the C–C reductive elimination step. In addition, the nitrene reagent must efficiently
react with the arylpalladium(II) intermediate but not with the palladium(0) species
responsible for insertion into the carbonleaving group bond.
These challenges were overcome by Onnuch et al. through the appropriate choice
of the nitrene precursor and palladium
catalyst. O-Diphenylphosphinyl hydroxylamine (DPPH) was the optimal nitrene
precursor, whereas less electrophilic nitrogen sources were less selective for nitrogen
insertion. Sterically demanding, electronrich phosphine ligands were also critical to
achieving selective formation of the desired
unsymmetric diarylamine product.

g

Department of Chemistry and Biochemistry, University of
Alabama, Tuscaloosa, AL, USA. Email: kshaughn@ua.edu

“The nitrogen insertion approach
can be used to create new
potential drugs from existing
drug compounds...”

This methodology opens new avenues for
the synthesis of pharmaceuticals and other
fine chemicals through late-stage functionalization of halogen-containing drug molecules to incorporate arylamine moieties.
The nitrogen insertion approach can be
used to create new potential drugs from existing drug compounds containing a biaryl
structure. No other changes are required to
the other steps in the synthesis to introduce
nitrogen in this way.
Onnuch et al. further demonstrated the
potential utility of this approach with a
four-component coupling of an aromatic
bromide, CO, DPPH, and an arylboronic
acid to give an N-aryl benzamide derivative. The nitrogen insertion approach was
also applied to the coupling of allyl acetates
and arylboronic acids to give N-allylaniline
derivatives in modest yield.
The nitrogen insertion approach developed by Onnuch et al. represents a groundbreaking new avenue in metal-catalyzed
cross-coupling in which heteroatoms can be
introduced in traditional C–C bond–forming reactions. There is the potential to apply
this approach to other classes of coupling
reactions beyond the SM coupling. Although
an exciting development, the method must
overcome some challenges to become widely
applicable. Yields range from modest to
high, which likely stems from undesired
side reactions. In some cases, the arylboron
electrophile and the nitrogen reagent react,
leading to undesirable aniline side products.
In addition, further optimization is needed
to afford consistently high selectivity for the
diarylamine product over the biaryl product
of SM coupling. Additional development of
the catalyst system will open the door for
wider application of this method for latestage introduction of nitrogen and other heteroatoms into target molecules. j

p

ransition metal–catalyzed cross-coupling, which creates a bond between
two target molecules, is a workhorse
method for synthesizing pharmaceuticals, agricultural chemicals, electronic
materials, and other fine chemicals
(1). The generation of biaryl compounds by
the Suzuki–Miyaura (SM) coupling (2) and of
aryl amines by the Buchwald–Hartwig (BH)
coupling (3) reactions are two of the most
used transformations in the pharmaceutical
industry (4). Their prevalence has resulted
in biaryl and arylamine structures as common motifs in drug candidates. New methods to prepare these structures will expand
the chemical space that can be accessed in
cross-coupling reactions. On page 1019 of this
issue, Onnuch et al. (5) report a palladiumcatalyzed aminative SM coupling reaction in
which the traditional SM coupling of an aryl
(pseudo)halide and an arylboron compound
is interrupted by the insertion of nitrogen.
This results in the formation of two new C–N
bonds in one reaction.
Transition metal–catalyzed coupling reactions are the most widely used methods for
C–C and C–heteroatom bond formation. In
these reactions, the metal promotes nucleophilic substitution at an electrophilic carbon
bearing a leaving group (6). The nucleophiles
are typically organometallic reagents, such as
organomagnesium (Grignard compounds),
organozinc, or organoboron reagents, and
the leaving group is typically a halide or
sulfonate. These reactions are highly useful
synthetic methods because they can typically
be carried out under mild conditions and are
tolerant of a wide range of functional groups.
The mechanism for these bond-forming
reactions involves three basic catalytic steps:
oxidative addition of the carbon electrophile,
typically a haloaromatic compound, to the
metal center; substitution of the nucleophilic
coupling partner for the leaving group on the
metal; and reductive elimination to form the
product. In SM coupling, the nucleophile is
an organoboron reagent, which results in the
formation of a new C–C bond. In the case of
the BH coupling, a C–N bond is formed from
a nitrogen nucleophile.
One method to expand the scope of these

reactions is to introduce additional catalytic
steps before the bond-forming reductive
elimination step, creating multicomponent
reactions. For example, addition of carbon
monoxide (CO) to cross-coupling reactions
leads to CO incorporation between the electrophilic and nucleophilic coupling partners.
In SM coupling, the result is a ketone product
(7), whereas amides are prepared through the
palladium-catalyzed coupling of aromatic halides, CO, and amines (8). These reactions are
highly selective for the carbonylative product
because CO insertion into the metal-carbon
bond occurs much faster than the subsequent steps of the catalytic cycle.

INSIGHTS
p
g
y

P OLICY FORUM
y g

ENERGY

Leverage demand-side policies for energy security
y

Conventional supply-side approaches overlook potential benefits

,

By Nuno Bento1, Arnulf Grubler2,3,
Benigna Boza-Kiss2, Simon De Stercke4,
Volker Krey2,5, David L. McCollum2,6,7,
Caroline Zimm2, Tiago Alves1

E

nergy security is a top priority for
governments, companies, and households because energy systems and the
critical functions that they support
are threatened by disruptions from
wars, pandemics, climate change, and
other shocks (1). More often than not, governments rely on policies focused on energy
supply to enhance energy security while
generally ignoring demand-side possibili-

946

1 MARCH 2024 • VOL 383 ISSUE 6686

ties. Further, the indicators traditionally
used to measure energy security are also
tilted toward the supply side; this fails to
capture the full spectrum of vulnerability
to energy crises. Energy security assessments need to reflect the wider benefits of
security-related interventions more accurately. To that end, we develop a systematic
approach to measuring the energy security
impacts of policy interventions that explicitly considers energy demand (buildings,
transport, and industry). We determine that
demand-side actions outperform conventional supply-side approaches at making
countries more resilient.

Energy demand links more directly than
supply to the satisfaction of critical social
functions and human well-being that are
at the core of energy security. Yet, demandside perspectives tend to be neglected or
underrepresented in analysis and policy
debates on energy security. Factors that
contribute to this supply-side bias include
the traditional sectoral organization of industries and policy institutions along fuels
(coal, oil, and gas) and energy forms (electric utilities) as well as the decentralized
and multivaried activities characteristic of
energy demand (from vehicles to household
appliances to manufacturing and more),
science.org SCIENCE

An infrared scan of a residential building is used
to check thermal insulation and energy efficiency.
Building efficiency is key to reducing energy demand.

which leads to a multitude of actors and
institutional fragmentation. The basic fundamentals of energy systems and markets,
where demand and supply are intricately
linked, have also not yet risen from vague
awareness to a central organizing principle
among policy-makers for structuring the
energy security discourse.

1 MARCH 2024 • VOL 383 ISSUE 6686

947

,

SCIENCE science.org

y

Leoben, Austria. 4Department of Civil and Environmental Engineering, Imperial College London, London, UK. 5Industrial Ecology Programme and Energy Transitions Initiative, Norwegian
University of Science and Technology (NTNU), Trondheim, Norway. 6Energy Science and Technology Directorate, Oak Ridge National Laboratory, Oak Ridge, TN, USA. 7Baker School of Public
Policy and Public Affairs, University of Tennessee, Knoxville, TN, USA. Email: nuno.bento@iscte.pt

y g

1Instituto Universitário de Lisboa (ISCTE-IUL), DINÂMIA’CET, Lisbon, Portugal. 2International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria. 3Montanuniversität Leoben,

y

COMPARING ENERGY SECURITY POLICIES
Four stylized policy interventions aimed at
enhancing energy security are devised and

g

FACTORS ENHANCING ENERGY SECURITY
Indicators used to measure energy security
tend to emphasize supply diversification
and substitution, costly storage, and redundancies in energy infrastructure. This
overlooks the scale of vulnerability to energy crises; the benefits of energy demand
reduction; and the energy cost burden to
countries, firms, and households. Meanwhile, demand-side indicators more directly
measure the satisfaction of individual needs
and well-being because the focus is on access to and use of energy services. Examples
of demand-based indicators of energy security include energy intensity of the economy
(energy needed per unit of GDP), energy efficiency (the inverse of intensity), energy
expenditures, and access to critical energy
services (10). However, data availability can
present a challenge to operationalizing the
use of these indicators (11).
Illustrating the importance of including energy demand in energy security
analyses, our analysis of Organisation
for Economic Co-operation and Development (OECD) countries found a positive
relationship [coefficient of determination
(R2) = 11%, P = 0.08] between the ratio of
energy expenditures to GDP and energy
intensity (see fig. S2 in the SM). In other
words, countries with higher energy intensity (less energy efficient) tend to face
higher energy cost burdens. We also found
a robust negative correlation between energy intensity and the Shannon diversity
index—a supply-side measure based on the
diversity of a country’s energy sources—for
a set of 20 countries and macro regions (R2
= 48%, P < 0.0001, in 2014; R2 = 56%, P
< 0.001, in 2019; see fig. S3, A and B, respectively, in the SM) (12). Put differently,
a discernible trend between higher efficiency and diversification is evident across
several countries over the past century (see
fig. S3C in the SM).
Such analyses, possible only by including demand-side indicators, help in gaining
a deeper understanding of the factors that
reinforce or jeopardize a country’s supply
security. Yet, to our knowledge, such analyses have not been central to the scientific or
policy discourses for many years.

their effectiveness is compared across a
range of security indicators. Each of these
aim at the same target, an ~10% change
(reduction or reallocation) in total primary
energy (i.e., unconverted natural resource
inputs, such as coal, oil or gas, uranium,
wind, or solar). These interventions are
targeted at one of several points along the
energy conversion chain: primary (PE), final (FE) (energy converted and delivered to
end users—e.g., electricity or refined petroleum in the form of gasoline for vehicles),
and useful (UE) (energy actually put to the
intended use—e.g., the light resulting from
the electricity used by a light-emitting diode
light bulb).
We use an accounting model based on
physical energy flows that first calculates
conversion efficiencies throughout the energy system from primary to useful energy
for specific end uses (mobility, thermal
comfort, etc.). The model then calculates
changes backward from useful to final to
primary—e.g., reflecting how reduced gas
use to provide low-temperature heat (useful energy) would affect the primary energy
balance for gas and the corresponding potential to reduce gas imports (see box S2 in
the SM for more details about the method
and data). The four policies analyzed are
import diversification (PE 1), fuel substitution (FE 1), reduction of low-temperature
heat demand in buildings (UE 1), and transport electrification (UE 2) (see box S3 in the
SM for more details about the policies).
An additional, somewhat extreme scenario estimates the impact on energy security of ensuring only the most basic energy
services that guarantee critical social functions (CSFs) (minimum thermal comfort
in buildings, transport, illumination, etc.).
The CSF scenario assesses the ultimate
social vulnerability of countries to energy
crises and illustrates the upper potential of
demand-side policies that is far larger than
in the four other scenarios examined.
Historical analogies help to put the stylized policy measures into context and to
demonstrate the order of magnitude of
interventions (see boxes S3 and S4 in the
SM). On the energy supply side, Germany’s
reduction of imports from Russia since
the start of the war in Ukraine is an example of rapid import diversification that
involved around a third of primary energy
(PE 1). The Brazilian ethanol program illustrates fuel substitution (FE 1) that reached
an impact equivalent to 10% of primary energy. On the demand side, Germany’s reduc-

p

SUPPLY AND DEMAND IMBALANCE
The energy security literature describes a
plethora of indicators and multi-indicator
indexes. However, two-thirds of the indicators and more than 80% of the indexes focus on the energy supply side [see fig. S1 in
the supplementary materials (SM)], aligned
with the International Energy Agency (IEA)
approach to define energy security solely
as the security of energy supply (2). In addition, energy security analyses often rely
on a small number of indicators, such as
import dependency, diversity of energy
sources (both produced and imported), or
cost of energy imports as a proportion of
gross domestic product (GDP) (3–7). In the
few studies that also consider demand indicators of energy security [a noteworthy exception being (8)], these metrics are rarely
quantified (see box S1 in the SM).
Not only is the structure of measures of
energy security imbalanced, but the predominant indicators do not reflect the full
picture. For example, a reduction in energy
demand may leave a country’s import dependence ratio unchanged—by simultaneously reducing both the volume of imports
in the numerator and total energy consumption in the denominator. In this case,
the unchanged ratio masks a marked reduction in the country’s energy vulnerability—a
smaller system being more flexibly satisfied
from different sources—as well as benefits
for the environment and trade balance.
Moreover, important gaps remain in the
usage of indicators for energy security assessments. Some studies use scenarios for
assessing future energy security (9). Others
analyze the evolution of energy security in
retrospect (3, 7). To our knowledge, no assessment has yet combined scenario-based
and historical analyses to determine the impact on energy security for different policy
options. To be sure, demand-side indicators
of energy security are neither perfect nor
all-encompassing; still, they merit greater

consideration for comprehensive energy security assessment.

I NS I GHTS | P O L I C Y F O RU M

Impact of the policy interventions on enhancing energy security

separately (see the figure) for the average of
the 14 countries in the assessment. Fuel substitution leads to the highest impact on only
two indicators—the share of nonfossil fuels
and import independency—confirming
the bias against energy efficiency of these
UE 1—Low-temp. heat demand reduction
PE 1—Import diversification
Index values relative to 200
a preintervention
two popular energy security indicators.
FE 1—Fuel substitution
UE 2—Transport electrification
baseline of 100
100 80
Demand measures score best in 8 of
the 12 indicators, with transport electrification the most impactful (seven
0
Shannon diversity
Compound Shannon
indicators).
Import diversification is
index PE
diversity index PE
consistently the least effective intervenCompound Shannon
tion, despite being the most applied in
diversity index PE with
Europe
as a consequence of the Russian
import diversification
Import
invasion of Ukraine.
independency
Supply
National contexts influence the effects of the policy interventions (see the
Demand
table and see table S1 and fig. S5 in the
SM). Transport electrification enhances
Share of
Compound
nonfossil fuels
energy security indicators the most (150
Shannon FE
times) and only leads to worsening in 12
cases, whereas fuel substitution more often (29 times) deteriorates energy security,
Compound
Savings in
although it also improves it in 117 cases.
Shannon FE
primary
including electricity
energy demand
Low-temperature heat demand leads to
by source
improvements in 137 cases and worsening in 23. Import diversification, although
a commonly applied approach by governments, has a relatively muted effect—imSavings in energy
Final energy
provement in only 19 cases and worsening
expenditures/GDP
efficiency
in nine.
Looking at four representative coun% of energy
Savings in energy
tries more closely (see fig. S6 in the SM),
expenditures/GDP
expenditures/fossil
Japan and Australia (both high-income
total PE/GDP
economies) show a similar pattern, with
FE, final energy; GDP, gross domestic product; PE, primary energy; UE, useful energy.
transport electrification being the most effective policy. Meanwhile, Japan and China
tions of gas demand—mainly from buildings
policy for each country. The choice of indi(both large energy importers) share more
(UE 1)—saved 5% of primary energy in 2022.
cators followed three criteria: They must
similar results than energy exporters such
Sustained promotion of electric vehicles in
be representative, feasible to calculate, and
as Australia and Nigeria.
Norway enabled a 10% reduction of primary
complementary to each other. For example,
energy in 2021, an example of the benefits of
four diversity indexes are included to asROBUSTNESS OF DEMAND ACTIONS
enhancing transport efficiency through elecsess different configurations of the energy
An extensive sensitivity analysis that calcutrification (UE 2). Similarly, the Corporate
system (e.g., different importing regions,
lates every combination of the 12 security
Average Fuel Economy (CAFE) standards in
different primary energy carriers, and difindicators (from 1 to 12 indicators in each
the US have saved more than 10% of primary
ferent structures of energy demand).
combination; see fig. S7 in the SM) strongly
energy over time (13). Finally, as an example
To reach the same goal of changing 10%
supports the robustness of our concluof energy demand reduction from reduced
of primary energy, the four policies entail
sions. For example, there are only 6 of the
activities, the COVID-19 response measures
very different changes in energy flows at
792 combinations (0.76%) that are posin the US led to a 7.5% decrease in primary
different points along the energy conversible to create with five indicators where
energy in 2020 (14) (followed eventually by
sion chain (see the figure). A 10% change in
fuel substitution is the most effective of
a rebound back to prepandemic amounts).
primary energy (PE 1) required only a 9%
the policies in improving energy security.
We use statistical data for a represenchange in final energy (FE 1) and only beDemand-side options and particularly
tative sample of 14 countries in 2019 to
tween a 5% and 3% change in useful energy
transport electrification rank first in the
simulate the impact of the various policy
level (UE 1 and UE 2, respectively), reflectremaining combinations.
interventions on national energy security.
ing the corresponding conversion losses in
Our multidimensional indicator and
These countries account for two-thirds of
energy systems. The difference is higher for
policy modeling framework also allows for
global energy use in 2019 and include a diexporting countries (see box S6 and fig. S4
testing alternative policies beyond the four
verse mix of high-income and low-income
in the SM). Overall, interventions at moredescribed above. For example, when connations as well as energy importers and
downstream levels benefit from a leverage
sidering a representative energy importer
exporters (see SM). We quantify a set of
effect by avoiding cascading losses in the
nation, such as Japan, fuel substitution
12 indicators of energy security from both
successive stages of energy conversions.
(FE 1) with hydrogen (also domestically
demand and supply perspectives (see box
The effects of the four stylized policy inproduced) instead of biofuels leads to a
S5 in the SM) to assess the impacts of each
terventions for all indicators are presented
similar level of impacts on energy security

Twelve indicators, five supply-side and seven demand-side, reflect impacts of four policy interventions,
each aimed at achieving a 10% reduction in primary energy. Impacts are shown in index values relative
to preintervention baseline normalized to 100, with larger values reflecting greater security.

y
y g
y
,

science.org SCIENCE

g

1 MARCH 2024 • VOL 383 ISSUE 6686

p

948

Effect of policy interventions on 12 energy security indicators,
in number of countries (n = 14)
Shading indicates the number of countries where energy security improved ( ) or worsened ( ) based
on the given indicator due to the intervention. Policy interventions appear in decreasing order of performance.
SUPPLY INDICATORS

UE 2 - TRANSPORT
ELECTRIFICATION

UE 1 - LOW-TEMP.
HEAT DEMAND REDUCTION

FE 1 - FUEL
SUBSTITUTION

PE 1 - IMPORT
DIVERSIFICATION

14

–1

Import independency

13

9

13

0

Shannon diversity index PE

8

6

5

0

Compound Shannon diversity
index PE

8

6

5

1

Compound Shannon diversity
index PE with import diversification

10

8

6

10

Compound Shannon FE

11

3

9

–1

Compound Shannon FE including
electricity by source

7

7

9

1

Final energy efficiency

13

14

0

1

% of energy expenditures/GDP

14

13

13

–1

Savings in energy expenditures/
fossil total PE/GDP

12

11

10

–2

Savings in energy expenditures/
GDP

14

14

12

1

Savings in primary energy demand

14

14

–8

1

Total (improved net of worsened)

138

114

88

10

DEMAND INDICATORS

Total worsened

12

23

29

9

Total improved

150

137

117

19

FE, final energy; GDP, gross domestic product; PE, primary energy; UE, useful energy.

ACKNOWLEDGMENTS

The authors thank anonymous reviewers for constructive
feedback that substantially improved the manuscript.
They also thank M. J. Machado of DINÂMIA’CET-ISCTE for
administrative assistance. This research was supported by
the EDITS project, which is a collaborative initiative coordinated by the Research Institute of Innovative Technology for
the Earth (RITE) and the International Institute for Applied
Systems Analysis (IIASA) and funded by the Ministry of
Economy, Trade, and Industry (METI), Japan. N.B. and T.A.
acknowledge funding from the Sus2Trans project, supported
by the Fundação para a Ciência e a Tecnologia (PTDC/GESAMB/0934/2020). D.L.M. acknowledges support from the
Laboratory Directed Research and Development Program
of Oak Ridge National Laboratory (ORNL), managed by
UT-Battelle, LLC, for the US Department of Energy (DOE). The
views expressed do not represent those of ORNL/UT-Battelle
or US DOE. For data and code, see Zenodo (15).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adj6150

10.1126/science.adj6150
1 MARCH 2024 • VOL 383 ISSUE 6686

949

,

CONCLUSIONS
Demand-side policies offer clear advantages
for energy security improvement across
many dimensions, including continuity, affordability, and sustainability. They also
have advantages in terms of flexibility. Demand-side policies give more opportunities

y

SCIENCE science.org

assumptions that the needed electricity for
transportation comes from domestic renewable sources. If, instead, this electricity is
generated from, for example, imported natural gas, the policy would be less efficient compared with other policies.
The overall ranking of policy options also
holds when looking at key indicators beyond
energy security, namely carbon dioxide emissions reductions (see fig. S9 in the SM). For
the four archetypal policy cases analyzed, demand-side policies again outperform supplyside policies (0% for PE 1 and –11% for FE 1
against –12% for UE 2 and –13% for UE 1),
even if the ranking order of transport electrification over heating demand reduction
reverses. The ultimate potential of demandside policies is illustrated in the (extreme demand reduction) CSF scenario (–67%).

1. G7 Ministers of Climate, Energy and the Environment,
“G7 Climate, Energy and Environment Ministers’
Communiqué,” joint press release (G7 Ministers’
Meeting on Climate, Energy and Environment, 17 April
2023).
2. IEA, “Emergency response and energy security:
Ensuring the uninterrupted availability of energy
sources at an affordable price” (2023); https://www.iea.
org/areas-of-work/energy-security.
3. P. Gasser, Energy Policy 139, 111339 (2020).
4. B. Kruyt, D. P. van Vuuren, H. J. M. de Vries, H.
Groenenberg, Energy Policy 37, 2166 (2009).
5. B. K. Sovacool, I. Mukherjee, Energy 36, 5343 (2011).
6. C. Winzer, Energy Policy 46, 36 (2012).
7. B. W. Ang, W. L. Choong, T. S. Ng, Renew. Sustain. Energy
Rev. 42, 1077 (2015).
8. F. Creutzig, Nature 606, 460 (2022).
9. J. Jewell, A. Cherp, K. Riahi, Energy Policy 65, 743 (2014).
10. T. B. Johansson, A. Patwardhan, N. Nakicenovic, L.
Gomez-Echeverri, Eds., Global Energy Assessment:
Toward a Sustainable Future (Cambridge Univ. Press,
2012).
11. E. Kisel, A. Hamburg, M. Härm, A. Leppiman, M. Ots,
Energy Policy 95, 1 (2016).
12. S. De Stercke, “Primary, Final and Useful Energy
Database (PFUDB)” (IIASA Models and Databases,
2023); https://iiasa.ac.at/models-tools-data/pfudb.
13. D. L. Greene, J. M. Greenwald, R. E. Ciez, Energy Policy
146, 111783 (2020).
14. US Energy Information Administration (EIA),
International Energy Statistics online (2023); https://
www.eia.gov/international/overview/world.
15. N. Bento et al., Auxiliary Supplementary Materials
(SM), Zenodo (2024); https://doi.org/10.5281/
zenodo.10573539.

y g

for all indicators but lower final energy efficiency and lower savings in energy expenditures (i.e., hydrogen increases primary
energy use and energy costs). Similarly,
rolling out heat pumps instead of insulation
(UE 1)—in this case, an active measure in
place of a passive one—to reduce energy demand for low-temperature heat for buildings
in Japan would not ameliorate the energy security indicators (see fig. S8 in the SM). The
variations of the policies tested in the sensitivity analysis confirm the robustness of the
order of merit of the interventions: Demandside policies generally have more and higher
positive impacts on improving energy security compared with supply-side measures.
Assumptions for fuel substitution with
biofuels are quite optimistic. For example,
not every country has enough available biomass, not to mention the potential land-use
conflicts with agriculture and environmental
concerns. Yet, even under these assumptions,
demand-side policies remain the top choice
in most cases. It is also worth noting that
the higher scores of transport electrification
compared with reducing low-temperature
heat demand in buildings benefit from our

RE FE REN C ES AN D N OT ES

y

9

g

14

p

Share of nonfossil fuels

for intervention at the national level, relative to fuel substitution, for example, which
requires international coordination. Energy
security is more than security of supply
because there are other economic, social,
and environmental aspects that are also
relevant. Future studies should compare
the benefits of different energy security
policies more systematically by including a
demand-side perspective instead of relying
on partial assessments and problematic indicators, such as import dependency. They
could also expand to encompass a more
comprehensive assessment of the social and
environmental effects. This approach would
contribute to a more nuanced understanding of energy security and inform more effective policy decisions on both a national and a
global scale. j

B O OKS et al .

PHILOSOPHY OF SCIENCE

Embracing our role as active participants in the Universe
should be a vital part of science, contend a trio of authors

955

,

1 MARCH 2024 • VOL 383 ISSUE 6686

y

SCIENCE science.org

10.1126/science.adn6303

y g

PHOTO: BRIGHTSTARS/ISTOCK PHOTO

W

y

isms betray what the British philosopher
Alfred North Whitehead referred to as “the
estern science was founded on the
bifurcation of nature.”
premise of divorcing objective and
Frank, Gleiser, and Thompson explicate
subjective aspects of nature—an
the nature and origin of this deficiency at
approach to understanding the
the core of science. They then navigate across
world that has proven very successdisciplines that deal with the greatest scienful indeed. And yet, such a strategy
tific mysteries. Time, matter, and cosmology
has major shortcomings. In The Blind Spot,
are covered in the second section of the book;
astronomer Adam Frank, theoretical physilife, cognition, and consciousness in the
cist Marcelo Gleiser, and philosopher Evan
third. The Blind Spot, readers learn, is hidThompson set out to reclaim the central
den in plain sight everywhere.
place of human experience in the scientific
When it comes to time and Einstein’s relaenterprise by invoking the image
tivity, the authors cite the French
of a “Blind Spot.” “At the heart
philosopher Henri Bergson’s inof science lies something we do
tuition that the experience of the
not see that makes science pospassage of time is alien to clocks.
sible, just as the blind spot lies at
They argue, however, that “noththe heart of our visual field and
ing illustrates the Blind Spot as
makes seeing possible,” they dedramatically as the emergence
clare. “In the visual blind spot sits
of quantum physics.” The weirdthe optic nerve; in the scientific
ness of superposition, entangleThe Blind Spot:
blind spot sits direct experience.”
ment, and the measurement
Why Science Cannot
Seeking to identify and corproblem results, in part, from
Ignore Human Experience
rect what the Austrian German Adam Frank, Marcelo Gleiser, insisting on a God’s-eye view of
philosopher Edmund Husserl
reality, they propound.
and Evan Thompson
MIT Press, 2024. 328 pp.
called “the surreptitious substiThe
French
philosopher
tution,” the authors have taken
Georges Canguilhem’s insightful
on a formidable challenge because the
realization that “there is no distinction beBlind Spot is a conceptual Frankenstein, an
tween normal and pathological in physics”
amalgam of views that includes materialand the German philosopher Hans Jonas’s
ism, reductionism, objectivism, instrumenremark that “only life can know life” preface
talism, and epiphenomenalism. All these
the return of the primacy of the organism
currently underway in biology, where agency,
purpose, and freedom are being entertained
The reviewer is at the Instituto de Neurociencias,
again after a long hiatus. In cognitive science,
Consejo Superior de Investigaciones Científicas–
the computational Blind Spot is epitomized
Universidad Miguel Hernández de Elche, Alicante, Spain.
Email: agomezmarin@gmail.com
in the imminent perils of artificially intelli-

g

By Alex Gomez-Marin

gent systems devoid of human wisdom. But
beyond quanta, chaos, and complexity, the
greatest opportunity to spot science’s foundational scotoma is consciousness.
“The phenomenologist brackets the everyday positing of the world as existing outside
consciousness in order to examine the world
strictly as it is disclosed to consciousness,”
write the authors. The primacy of consciousness is then brought to the fore. We cannot
step outside consciousness, and it is not simply another object of knowledge “but also,
and more fundamental[ly], that by which any
object is knowable.” Their conclusion is unflinching: “the hard problem [of consciousness] is an artifact of the Blind Spot.”
To ask “how” the brain gives rise to experience begs the question of “whether” it actually does so. Here, the authors reject not only
physicalism and illusionism but also panpsychism and idealism. They contend instead
that the real problem of consciousness is
“how the brain as a perceptual object within
consciousness relates to the brain as part of
the embodied conditions for consciousness.”
The Blind Spot is not just endemic in
science, the authors maintain, it has also
percolated to education, journalism, culture,
and society writ large. Touching on political
economics, the final chapter reimagines our
relationship with planet Earth.
This is a very important book that has the
potential to become a classic text. I wish to
note, however, three qualms. First, its diagnosis is much stronger than its prognosis.
Having claimed at the start that “we need
nothing less than a new kind of scientific
worldview,” the authors ultimately leave
readers with suggestions for “best practices.”
Second, there is a whiff of disdain for speculation, particularly in mathematics and metaphysics. Phenomenology can feel a bit like being in an elevator that is stuck between two
floors (science and philosophy): One can see
what is wrong in both and yet is unable to
contribute much to either. And finally, the
book is simultaneously daring and yet mellowly heterodox—the authors could have
been bolder in entertaining anomalous experiences at the edges of consciousness.
Science is indeed a strange loop: “a highly
refined form of experience” whose bounty
lies, in part, in its ability to distill “objects
of public knowledge” from experience, in
ever-ascending “cycles of abstraction.” But,
like a kite, it cannot properly fly if it loses its
grounding. Being aware of the Blind Spot is
a necessary step toward reinscribing human
experience back into science’s core. j

p

Experiencing science

Often overlooked, direct human experience is a
central factor at play in understanding reality.

I N SI G H TS | B O O K S

ANTHROPOLOGY

Knowing the Neanderthal

The Naked Neanderthal:
A New Understanding of
the Human Creature
Ludovic Slimak
Pegasus, 2024. 208 pp.

An archaeologist seeks to strip away modern
misconceptions about our extinct relatives
By April Nowell

y g
y
,

science.org SCIENCE

y

1 MARCH 2024 • VOL 383 ISSUE 6686

g

956

p

to look carefully. He argues, on the basis of
the lack of standardization of Neanderthal
stone tools (another point many in the field
ver the course of 300,000 years,
would debate), that Neanderthals were
Homo sapiens shared Earth with
more creative than Homo sapiens. Whereas
potentially as many as seven other
Neanderthal creativity is apparent in the
species of ancient humans. But of all
uniqueness of every stone tool, Homo sapiof these species, it is the Neanderens, he reasons, are only superficially crethals who have captured the imagiative, and yet their “material rationalization
nation of both scientists and the public
of the world,” as seen in their efficient and
alike. When they went extinct some 40,000
highly standardized tools, allowed them to
years ago, we were without a close relative
drive Neanderthals to extinction within a
on this planet for the very first time. As
few seasons wherever they encountered each
a result, Neanderthals have become the
other. Here, again, the lack of nuance in the
quintessential “other.”
modeling of Neanderthal extinction is out
In his newly translated book, The Naked
of step with current understandings
Neanderthal, archaeologist Ludovic
in the field, as is Slimak’s contention
Slimak seeks to strip away the modthat it was the more restricted creativern trappings he argues that scientists
ity of Homo sapiens that gave them an
have projected onto our distant cousadaptive advantage.
ins in order to understand the true
There are other aspects of the book
nature of the Neanderthal “soul.” The
that are likely to be stumbling blocks
book focuses primarily on the author’s
for the reader. For example, because
own field projects over the past three
the book is structured around Slimak’s
decades, including detailed descripcareer in the field, it does not engage
tions of his pioneering research at
with many other pertinent behaviors
the French site of Mandrin investigator sources of data to a meaningful
ing the nature of Neanderthal–Homo
extent even though they stand to insapiens interactions and the timing
fluence the reader’s understanding of
and cause of Neanderthal extinction.
who the Neanderthals were. These inThe book is filled with evocative
clude the species’ medicinal plant use,
imagery. Slimak’s descriptions of
pyrotechnic knowledge, and cooking,
traveling by train to the Arctic Circle
as well as the growth and development
and digging through sediment ridof Neanderthal “minibrains” (2, 3).
dled with shark teeth to find traces
Slimak (foreground) and colleagues work in Grotte Mandrin, France.
Another issue is that the
of Neanderthals below leave a lasting impression on the reader and provide
Can he be? He maintains that other scienNeanderthal in The Naked Neanderthal is
compelling details of discovery and explortists are incapable of imagining a humanity
decidedly male. Beyond Slimak’s use of male
ation. Given that it was written primarily
that is different from their own and instead
pronouns throughout the book and his emfor a popular audience, it is surprising,
remake Neanderthals in their own likeness.
phasis on hunting as an exclusively male achowever, that the book contains no maps
For his part, Slimak dismisses all evitivity, population dynamics are described as
or timelines, no glossary of terms, and no
dence that Neanderthals made cave art and
“I give you my sister, you give me your sister.”
images of stone tools, personal ornaments,
items of personal adornment, behaviors seen
No female agency here.
or even Neanderthals themselves, other
as hallmarks of “humanness.” He further
Although he omits any consideration
than one well-known drawing of what a
argues that Neanderthal burials simply link
of the lives of Levantine Neanderthals,
Neanderthal might look like in modern
them to other animals who feel loss at the
Slimak’s suggestion that the last refuges of
clothing. Without these contextualizing and
passing of conspecifics. Although new eviNeanderthals may have been in polar regions
explanatory elements, it may be difficult for
dence has come to light since the book was
and not just the warmer climes of Gibraltar
those with little to no previous knowledge
published—for example, the discovery of
is interesting. Despite its shortcomings, such
of Neanderthal biological and cultural evoNeanderthal cave art in the form of digital
musings are among the many things this
lution to find their footing.
tracings at La Roche-Cotard, France (1)—it
book will leave the reader thinking about. j
The Naked Neanderthal takes a deeply
seems unlikely to have swayed him. In this,
RE FE REN CES A ND N OT ES
he is at odds with the vast majority of scien1. J.-C. Marquet et al., PLOS ONE 18, e0286568 (2023).
tists in the field.
2. L. S. Weyrich et al., Nature 544, 357 (2017).
The reviewer is at the Department of Anthropology,
For Slimak, the archaeological record
3. C. A. Trujillo et al., Science 371, eaax2537 (2021).
University of Victoria, Victoria, BC V8W 2Y2, Canada.
lays bare the Neanderthal soul, if we choose
10.1126/science.adn6093
Email: anowell@uvic.ca

O

philosophical approach to the study of
Neanderthals and is written in a poetic
style that is at times cumbersome. Similarly,
the chapters can be difficult to follow, and
Slimak’s choice of topics can seem somewhat
random (for example, why discuss cannibalism but not language?) except that they are
linked to each other through the author’s
field projects. The Naked Neanderthal is thus
almost best thought of as a memoir—its meandering style a reflection of the nature of
memory, its philosophical approach a product of the author’s soul-searching.
After a lifetime of “hunting Neanderthals,”
is Slimak any closer to really knowing them?

Peru’s new law removes obstacles to converting
diversity-rich forests to farmland.

LET TERS

Eric G. Cosio2, Cony Decock6, William Farfan-Rios7,
Kenneth Feeley8,9, Eurídice Honorio Coronado10,
Isau Huamantupa11, Alfredo J. Ibañez2, Juliane
Koepcke de Diller12, Blanca León13, Reynaldo
Linares-Palomino14, José L. Marcelo Peña15, Betty
Millán5, Justin F. Moat1, R. Toby Pennington16,17,
Nigel Pitman18, Norma Salinas2, Roxana RojasVeraPinto19, Philip C. Stevenson1,20, Carolina
Tovar1, Oliver Q. Whaley1, Kenneth R. Young13
1Royal

Botanic Gardens Kew, Surrey TW9 3AB, UK.
Universidad Católica del Perú, 15088
Lima, Peru. 3Philipps-Universität Marburg, 35032
Marburg, Germany. 4Duke University, Durham,
NC 27705, USA. 5Universidad Nacional Mayor
de San Marcos, 15081 Lima, Peru. 6Université
Catholique de Louvain, B-1348 Louvain-la-Neuve,
Belgium. 7Department of Biology and Center for
Energy, Environment, and Sustainability, Wake
Forest University, Winston-Salem, NC 27106,
USA. 8Department of Biology, University of Miami,
Coral Gables, FL 33146, USA. 9Fairchild Tropical
Botanic Garden, Coral Gables, FL 33156, USA.
10University of St. Andrews, St. Andrews KY16
9AL, UK. 11Universidad Nacional Amazónica de
Madre de Dios, Puerto Maldonado, Peru. 12Area de
Conservación Privada Panguana, Huánuco, Peru.
13Geography and the Environment, University of
Texas at Austin, Austin, TX 78712, USA. 14Center
for Conservation Education and Sustainability,
Smithsonian’s National Zoo & Conservation
Biology Institute, Washington, DC 20008, USA.
15Universidad Nacional de Jaén, Jaén, Peru.
16Royal Botanic Garden Edinburgh, Edinburgh EH3
5LR, UK. 17Department of Geography, University
of Exeter, Exeter EX4 4RJ, UK. 18Science and
Education, The Field Museum, Chicago, IL 60605,
USA. 19University of Reading, Reading RG6 6EX,
UK. 20Natural Resources Institute, University of
Greenwich, Kent ME4 4TB, UK.
*Corresponding author. Email: c.martel@kew.org
2Pontificia

10.1126/science.ado0050
1 MARCH 2024 • VOL 383 ISSUE 6686

957

,

1. N. Myers, R. A. Mittermeier, C. G. Mittermeier, G. A. B. da
Fonseca, J. Kent, Nature 403, 853 (2000).
2. Ministerio del Ambiente, Government of Peru, “Sexto
informe nacional sobre diversidad biológica” (2019) [in
Spanish].
3. K. Brandon, “Ecosystem services from tropical forests–Review of current science,” Center for Global
Development Working Paper 380 (2014).
4. O. L. Phillips, R. J. W. Brienen; RAINFOR collaboration,
Carbon Balance Manag. 12, 1 (2017).
5. R. Gómez et al., Sustainability (Basel) 15, 4788 (2023).
6. “Ley N° 31973,” El Peruano (2023); https://busquedas.
elperuano.pe/dispositivo/NL/2251964-1 [in Spanish].
7. Ministerio del Ambiente, Government of Peru,
“Cobertura y pérdida de bosque húmedo amazónico
2021” (2022); https://geobosques.minam.gob.pe/geobosque/descargas_geobosque/perdida/documentos/
Reporte_Cobertura_y_Perdida_de_Bosque_Humedo_
Amazonico_2021.pdf [in Spanish].
8. Servicio Nacional Forestal y de Fauna Silvestre,
Government of Peru (Cuenta de Bosques del Perú, 2021)
[in Spanish].
9. C. A. Nunes et al., Proc. Natl. Acad. Sci. U.S.A. 119,
e2202310119 (2022).
10. R. Beuchle, F. Achard, C. Bourgoin, C. Vancutsem,
“Deforestation and forest degradation in the Amazon
– Updated status and trends for the year 2021”
(Publications Office of the European Union, 2022).
11. N. Giardino, “Narco violence surge in Peru’s Amazon
sends Indigenous leader into hiding” (2023).
12. D. Valdivia Blume, “Ley Forestal: ¿quiénes estuvieron
detrás de la modificación de la norma que ahora permitirá la deforestación en la Amazonía?” (2023) [in
Spanish].

y

PHOTO: ARMIN NIESSNER (PANGUANA BIOLOGICAL STATION)

RE FE REN CES A ND N OT ES

y g

SCIENCE science.org

Carlos Martel1,2*, Glenda Mendieta-Leiva3,
Patricia C. Alvarez-Loayza4, Asunción Cano5,

y

Peru, one of the most biodiverse countries
in the world (1), contains many endemic
species that are at risk of extinction (2). The
Peruvian Amazon biome harbors most of
this diversity and provides globally important ecosystem services and benefits to
all people (3, 4) as well as economic and
cultural value for Indigenous communities (5). However, recent amendments to
Peru’s Forestry and Wildlife Law No. 29763
threaten these important forest ecosystems.
On 11 January, the Peruvian Congress
amended Law No. 29763 by enacting Law
No. 31973 (6), which removes obstacles to
deforestation by changing zoning laws and
regulatory bodies. Previously, exploitation
could only take place in areas zoned as
“permanent production forests.” No area
could be rezoned as permanent production
forests without an evaluation study and
approval from the Ministry of Environment,
the regulatory body for land use in forested
areas. Law No. 31973 removes the evaluation
requirement and allows zoning changes with
permission from the Ministry of Agricultural
Development and Irrigation, which replaces
the Ministry of Environment as Peru’s forest
regulatory body. Given that a priority of
the Ministry of Agricultural Development
and Irrigation is to increase agricultural
production, it will likely facilitate zoning
changes to allow forest exploitation despite

g

Peru’s zoning amendment
endangers forests

the threat they pose to diversity-rich areas.
These changes allow private agricultural
companies that already own forested land to
freely convert it to farms, facilitating rapid
land-use change.
Between 2015 and 2017, Peru lost more
than 4770 km2 of forest, comprising 0.7% of
Peru’s total forest area (7), 83% of which was
transformed for agriculture and livestock
(8). Such land conversions lead to biodiversity loss, alter soil properties, and reduce
aboveground carbon pools (9). In addition
to releasing substantial amounts of carbon
into the environment, deforestation affects
the hydrological cycle and other natural
processes (10). The intrusion of industrial
activities in the Amazon could also lead to
increased crime against Indigenous communities [e.g., (11)].
Private business groups, such as the
National Confederation of Private Business
Institutions (CONFIEP), lobby the government in support of land-use change in the
Amazon rainforest to establish large-scale
intensive agriculture (12). Their demands
place profits above long-term environmental
and human health. Instead of capitulating to
industry, the Peruvian Congress should protect the country’s land and people by ensuring that its legislation serves to preserve and
promote the sustainability of the Peruvian
forests as well as protect the country’s natural ecosystems and biodiversity. Peruvian
citizens and scientists can fight the business
lobby by calling on their congressional representatives to act accordingly.

p

Edited by Jennifer Sills

I N SI G H TS | L E T T E R S

Nicaraguan government
puts mining over justice

de Pós-Graduação em Biodiversidade e
Conservação, Universidade Federal do Maranhão,
65080-805, São Luís, MA, Brazil. 2Programa de
Pós Graduação em Biodiversidade e Biotecnologia
da Amazônia Legal, Universidade Federal do
Maranhão, 65085-580, São Luís, MA, Brazil.
3Facultad de Gobierno, Universidad de Chile,
8320000 Santiago, Región Metropolitana, Chile.
4Internet Interdisciplinary Institute, Universitat
Oberta de Catalunya, 08018 Barcelona, CAT, Spain.
*Corresponding author.
Email: raphaelmais12@gmail.com
R E F E R E N C ES A N D N OT ES

y g

Michael J. Staplevan and Faisal I. Hai*

R E F E R E N C ES A N D N OT ES

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

P. Bhatt et al., Environ. Res. 200, 111762 (2021).
M. J. Stapleton et al., Water Res. 233, 119790 (2023).
S. Acharya et al., Text. Res. J. 91, 2136 (2021).
G. Guerranti et al., Environ. Toxicol. Pharmacol. 68, 75
(2019).
M. M. Haque et al., J. Hazard. Mater. 8, 100166 (2022).
G. Suzuki et al., Environ. Pollut. 303, 119114 (2022).
Y. Guo et al., Sci. Tot. Environ. 838, 156038 (2022).
E. Brown et al., J. Hazard. Mater. 10, 100309 (2023).
M.J. Stapleton et al., Sci.Total Environ. 902, 166090 (2023).
R. Geyer et al., Sci. Adv. 3, 1700782 (2017).
H. B. Sharma et al., Sci. Total Environ. 800, 149605 (2021).
European Commission, “Microplastics” (2024);
https://environment.ec.europa.eu/topics/plastics/
microplastics_en.
10.1126/science.ado1473

science.org SCIENCE

,

Strategic Water Infrastructure Laboratory, School
of Civil, Mining, Environmental, and Architectural
Engineering, University of Wollongong,
Wollongong, NSW 2522, Australia.
*Corresponding author. Email: faisal@uow.edu.au

y

1. A. Marena, “Guía para el Manejo de la Biodiversidad”
(2020).
2. F. Galão, “Como a Nicarágua está devastando a segunda
maior floresta tropical das Américas,” Gazeta do Povo, 31
December 2023 [in Portuguese].
3. International Union for Conservation of Nature, Red
List, “Red List Category, Critically Endangered and
Endangered in Nicaragua” (2024);
https://archive.vn/J6d7m.
4. “Acuerdo de Escazú en papel: Régimen de Daniel Ortega
niega información en materia ambiental,” Onda Local,
26 December 2022 [in Spanish].
5. Nicaragua, Presidente de la República, “Ley Creadora
de la Empresa Nicaragüense de Minas (ENIMINAS)”
(2017); https://archive.vn/OSkiQ [in Spanish].
6. S. Quartucci, “Illegal Mining and Its Impacts on Human
Rights-Nicaragua,” Latina Republic, 6 December 2023.
7. “Ortega ha concesionado la cuarta parte de Nicaragua
a empresas mineras,” Divergentes, 16 July 2022 [in
Spanish].
8. A. Hines, “Decade of defiance,” Global Witness, 29
September 2022.
9. The Global Organized Crime Index, “Nicaragua” (2023);
https://archive.vn/AWNNo.
10. M. Civillini, “UN climate fund suspends project in
Nicaragua over human rights concerns,” Climate Home
News, 26 July 2023.
11. S. Cortés, R. Sáenz, in Serie América Central en
Perspectiva Ístmica Vol. 2: Pueblos, Movimientos,
Saberes y Migraciones en el Istmo Centroamericano
(Edições EACH, 2023), pp. 108–146 [in Spanish].
12. T. Clifford, S. Heavey, “U.S. mining sanctions take aim at
Nicaragua’s Ortega,” Reuters, 24 October 2022.
10.1126/science.ado1264

y

1 MARCH 2024 • VOL 383 ISSUE 6686

1Programa

Potentially harmful microscopic plastics
(microplastics) have been identified in
flora, fauna, and humans (1, 2), and their
volume and impact in the environment
are difficult to quantify. The most effective
microplastics mitigation strategy is to
pinpoint their sources and prevent their
release. Many industries, including textiles,
cosmetics, and pharmaceuticals, have
been linked to the release of microplastics
into the environment (3–5). However, one
counterintuitive source has been overlooked:
the plastic recycling industry.
The size-reduction units at plastic
recycling facilities can generate substantial
amounts of microplastics, which can be
released into the environment during the
subsequent washing process (6–9). In
one case, 200,000 microplastic particles
were released per liter of effluent (6). The
commercial process for plastic recycling
may have been emitting microplastics since
its first use nearly half a century ago (10).
Plastic recycling is integral to the transition
from a linear to circular economy (11), but
to ensure that the process is a net benefit,
measures need to be put in place to prevent
microplastic contamination.
Preventing the release of microplastics
from the recycling sector will require
cooperation among scientists, industry, and
governments. Researchers need to work
with the recycling industry to find ways to
effectively contain the microplastics that
facilities emit. In addition, environmental
regulatory agencies should implement and
enforce wastewater emission standards
that specifically target microplastics as
a contaminant of concern, similar to the
policies the European Commission has
proposed (12).

g

958

Rafael F. Oliveira1*, Felipe P. Ottoni1,2, Lucas O.
Vieira2, Grettel N. Obando3, Ronald Sáenz4, Diego
S. Campos2

Recycling process
produces microplastics

p

Nicaragua contains 7% of global biodiversity
and encompasses 60% of Central America’s
ecosystem types (1). In recent decades, corruption, crime, and government negligence
have threatened the country’s biodiversity,
especially in areas such as the Bosawás
Reserve, the second largest tropical forest in the Americas, and the Indio Maíz
Biological Reserve, the second largest forest
in Nicaragua (2), both protected areas that
shelter endemic and threatened species
(3). In 2019, Nicaragua signed the Escazú
Agreement, a transnational pact pledging
public participation and access to information to ensure environmental justice in Latin
America and the Caribbean (4). However,
the Nicaraguan government’s prioritization of mining activities, at the expense of
the environment and Indigenous peoples,
undermines the agreement.
In 2017, President Daniel Ortega’s
administration approved Law No. 953 (5),
which established the Nicaraguan Mining
Company (ENIMINAS) and thus mandated
state participation in mineral exploitation.
Law No. 953 does not specify the percentage
of land to be allocated to mining activities,
but since it took effect, there has been a
substantial increase. In 2021, mining was
taking place on approximately 835,000 hectares (about 7% of Nicaragua’s total land),
with the possibility of expanding to 4.2
million hectares (approximately 36% of the
country) (6). Some of that land is located
in natural reserves or traditional territories, where previous laws prohibit mineral
extraction (2, 6, 7).
Government approval of mining in areas
where mining is illegal has led to increasing conflict as traditional communities face
persecution from land grabbers (2, 6–8). The
Global Organized Crime Index, which tracks
environmental offenses such as timber trafficking and illegal mining, rates 2023 crime
in Nicaragua a 5.72 out of 10 and Nicaragua’s
ability to address crime-related challenges
(resilience) in 2023 only a 2.08 out of 10
(9). In response to the escalating violence
against Indigenous peoples, the Green
Climate Fund, a United Nations linked
funder with the goal of reducing deforestation in reserves, suspended US$117 million
from Nicaragua in 2023 (2, 10). Because
of government-driven mineral extraction,
Nicaragua faces increased environmental
degradation and decreased access to funding
to address it.
In the past two decades, opposition to
the government has primarily stemmed
from an alliance between grassroots

farmworker and environmentalist movements who are concerned about resource
depletion, including land, water, and biodiversity in protected areas (11). The international community has also taken steps
to help protect Nicaragua’s biodiversity
and Indigenous communities. For example,
in 2022, the US government increased
economic pressure on Nicaragua by banning business with the gold industry and
placing sanctions on the national mining
authority (12). Increased diplomatic pressure is needed, along with support for local
nongovernmental organizations and independent investigations. International organizations should foster dialogue among
stakeholders to find joint solutions to this
socio-environmental crisis and ensure
that Nicaragua complies with the Escazú
Agreement. Most importantly, the root
cause of the problem must be addressed:
Nicaragua’s government must honor its
protective laws instead of passing legislation that overrides and contradicts them.

RESEARCH
IN S CIENCE JOU R NA L S
Edited by Michael Funk

p

MARINE CONSERVATION

Protecting pelagic species

SCIENCE science.org

PHOTO: MANU SAN FELIX

Science p. 967, 10.1126/science.adk3863

NEUROSCIENCE

Touch receptor diversity
Touch sensation is mediated
by mechanically activated
ion channels. Although some
of these channels have been
identified, the mechanosensitivity of sensory neurons

is likely to involve multiple
types of mechanically gated
channels. Chakrabarti et al.
identified a putative ion channel, ELKIN1, that is activated
by mechanical force and is
necessary for normal touch
sensation in mice. Deletion of
Elkin1 reduced the activation
of low-threshold, mechanically activated currents. A
similar reduction was observed
in induced human sensory
neurons upon small interfering RNA–mediated reduction
in ELKIN1. The identification of
ELKIN1’s contribution to touch
sensation expands our understanding of the molecular basis
of cutaneous sensation. —MMa
Science p. 992, 1 0.1126/science.adl0495

RIVER FLOW

Changing seasonal
changes
Patterns of river flow vary
seasonally, which has important
effects on the occurrence of
floods and droughts, degrees
of water security, and ecology.
What is anthropogenic climate
change doing to these seasonal
cycles? Wang et al. used in
situ observations of monthly
average river flow from 1965 to
2014, combined with modeling,
to show that human effects on
climate have already caused a
reduction of river flow seasonality at latitudes above 50° N.
Understanding these changes
is necessary for ensuring that

1 MARCH 2024 • VOL 383 ISSUE 6686

959

,

In vitro work has shown that
endocannabinoids mediate
a form of synaptic plasticity
called depolarization-induced
suppression of inhibition (DSI).
However, whether DSI occurs
in vivo and if it contributes to
physiological function have
needed more investigation.
Dudok et al. now demonstrate
that a subpopulation of cells in
the hippocampus, which fire in
specific locations called place
cells, trigger endocannabinoid
signaling in their place field
that can be detected both in
the postsynaptic membrane

and the presynaptic inhibitory axons. The authors show
that inhibiting the endocannabinoid signaling alters place
cell firing. These results reveal
that a form of DSI-like plasticity occurs in vivo and plays
an important role in shaping
hippocampal spatial representation. —MMa

y

Location-dependent
signaling

y g

NEUROSCIENCE

y

Researchers deploy a remote underwater video station in French Polynesia
for unbiased surveys of ocean fish species.

g

M

arine megafauna are increasingly threatened and are difficult to protect. Understanding the influence of humans
on body size in fishes is also challenging given that data on
marine species often come from fishery-based activities.
Letessier et al. deployed more than 17,000 remotely operated baited devices to collect data on fish size and abundance as
related to habitat (pelagic or benthic), human activities, and marine
protected areas. Pelagic species were strongly influenced by human
pressures and protection. The authors concluded that benthic species could be effectively protected even near markets, whereas only
more remote protected areas will effectively safeguard large pelagic
species. —SNV Science p. 976,, 10.1126/science.adi7562

R ES EA RCH | I N S C I E N C E J O U R NA L S

freshwater ecosystems maintain
their essential functions, for
securing sustainable water
resources, and for determining allocations for irrigation or
hydropower generation. —HJS
Science p.1009, 10.1126/science.adi9501

safe passage out of the body.
INSPIRE could offer a direct
treatment for postoperative
ileus, but future studies will
need to focus on optimizing
electrical stimulation, as well
as on safety and efficacy for
human translation. —MY

IN OTHER JOURNALS

Edited by Caroline Ash
and Jesse Smith

Sci. Robot. (2024) 10.1126/scirobot.adh8170

STAR FORMATION

Ultraviolet light erodes
protoplanetary disks

Growing larger
crystals faster
The production of large crystals
of porous covalent organic frameworks (COFs) usually requires
slow growth over weeks to avoid
precursor assembly that results
in defects. Han et al. found that
large imine-linked single-crystal
COFs (15 to 100 micrometers)
can be grown in 1 or 2 days using
trifluoroacetic acid as a catalyst
and trifluoroethylamine as an
intermediate reactant that is
displaced by the reactant amine.
This approach grew a wide
variety of large COF crystals with
x-ray diffraction resolutions up to
0.8 angstroms. —PDS

p
g
y

During their formation process,
young stars are surrounded by
a protoplanetary disk of gas
and dust within which planets
can form. Stars mostly form in
clusters, and bright, high-mass
stars irradiate the disks around
low-mass stars with ultraviolet
light. Berné et al. combined infrared, submillimeter, and optical
observations of a protoplanetary
disk in the Orion Nebula to
determine the effect of ultraviolet irradiation. The authors found
that the heating and ionization
induced by the ultraviolet photons caused gas to be lost. They
measured the loss rate and discuss the implications for planet
formation in the disk. —KTS

FRAMEWORK MATERIALS

Science p. 1014, 10.1126/science.adk8680

Science p. 988, 10.1126/science.adh2861

CANCER
MEDICAL ROBOTS

Restoring intestinal
peristalsis

The world has seen unprecedented fertility declines across
advanced, industrialized countries. Although childless families
are rarely idealized, does the
number of children still shape
our perceptions of a successful family? Previous research
tended to narrowly examine
individual attributes (e.g., wealth,
gender roles, communication
quality, or number of children)
in isolation instead of holistically accounting for the several
dimensions that affect contemporary family ideals. Aassve
et al. used family vignettes
to investigate perceptions of
ideal family characteristics
across eight culturally diverse
industrialized countries. The
countries consistently prioritized
good communication, wealth,

Proc. Natl. Acad. Sci. U.S.A. (2024)
10.1073/pnas.2311847121

,

1 MARCH 2024 • VOL 383 ISSUE 6686

Sci. Transl. Med. (2024)
10.1126/scitranslmed.adf9874

Is a two-child
family still ideal?

egalitarian gender roles, and
parenthood involving at least
one child. However, having only
one child (versus more than one)
did not matter, suggesting that
policies supporting satisfying
marriages should target dimensions other than fertility. —EEU

y

960

Aromatase inhibitors prevent
estrogen production and can
be effective against estrogen
receptor–positive breast cancer,
but subsequent tumor metabolic adaptation often thwarts
treatment efficacy. Bacci et
al. report that acetyl-CoAcarboxylase-1 (ACC1) promotes
lipid mobilization in estrogendeprived breast cancer cells,
leading to anti–estrogen therapy
resistance. Pharmacologically
targeting ACC1 in patientderived, treatment-resistant
xenograft models reduced tumor
growth and increased mouse
survival. This work indicates
that targeting ACC1 may be an
avenue to resensitizing estrogen
receptor–positive breast cancer
to endocrine-based therapies.
—CAC

FAMILY IDEALS

y g

Paralysis of the intestinal track
is a common postoperative
complication that not only
causes several painful symptoms, but also often results
in prolonged hospital stays.
Srinivasan et al. developed
an ingestible robotic device
designed to reanimate the
intestines through electrical
stimulation. The S-shaped
device, named INSPIRE, contains electrode contacts on its
outer surface. An expansion
mechanism triggered in the
small intestine creates contact
between the lumen and the
electrodes. INSPIRE improved
intestinal contractions by 44%
in anesthetized swine and by
140% in a model of induced
ileus. Made of biodegradable
polymers, the device degrades
within 24 hours, enabling

Repotentiating
aromatase inhibitors

IMMUNOLOGY

How PD-1 influences
T cell signaling
T lymphocytes are immune cells
that, when properly activated,
attack and kill cancer cells.
Inhibitory receptors such as
programmed cell death 1 (PD-1)
impair the activation of T cells.
Therefore, cancer immunotherapy that blocks PD-1 can be
used to restore tumor-specific T
cell responses. Despite the success of PD-1 inhibitor therapy
science.org SCIENCE

Genetic basis of color
is a real hoot

C

oloration is vital for predator
avoidance, attracting mates,
and many other keys to survival. However, color is difficult
to investigate mechanistically
because it is often both polygenic
and environmentally influenced.
Cumer et al. use whole-genome
data from 75 barn owls to identify
genetic variation underlying plumage coloration. They replicated
previously known associations with
MC1R, a gene involved in pigmentation in many species, and identified
two new variants, including one
with effects that are only seen in
the presence of the MC1R allele that
confers white pigmentation. These
results highlight the oligogenic and
complex interplays of epistasis and
dominance among genetic variants
underlying pigmentation. —CNS

Plant Cell (2024)
10.1093/plcell/koae033

g

Proc. Biol. Soc. (2024) 10.1098/rspb.2023.1995

SPINTRONICS

In many settings, particularly lowand middle-income countries,
challenges faced by both lenders
SCIENCE science.org

TEXTILES

Clean water from cloth
Water resources are inadequate
in many parts of the world. Li et
al. developed a fabric-based solar
steam generator to produce clean

PLANT SCIENCE

Leaf shape and
development
Leaf shapes exhibit remarkable diversity. Leaflet initiation
and boundary formation are
key developmental events that
control compound leaf morphogenesis. He et al. identified a

Phys. Rev. Lett. (2024)
10.1103/PhysRevLett.132.056704

1 MARCH 2024 • VOL 383 ISSUE 6686

961

,

Accessing education with
digital collateral

Adv. Funct. Mater. (2024)
10.1002/adfm.202312613

y

ECONOMICS

Q. J. Econ. (2024)
10.1093/qje/qjae003

water. They were able to accomplish relatively efficient water
production by designing a fabric
that absorbs sunlight but is structured to efficiently separate salt
from water. As a proof of concept,
the authors constructed a floating
outdoor water purification device.
Using fabrics as a basis for solar
steam generation may be helpful
for scaling in a cost-efficient way,
providing a different pathway for
clean water generation. —BG

The vast majority of today’s
computing technologies
are based on the transport
of charged carriers around
electrical circuits. However, the
speed at which electrons can
be moved around, together
with ohmic losses, restricts
how far these technologies
can progress. The next generation of devices based on wave
technologies such as acoustics, optics, or spin waves (the
collective magnetic properties
of electrons) could overcome
these limitations. Hwang et al.
have shown that magnons, the
quanta of spin waves, propagating in a thin magnetic film can
be strongly coupled to another
wave excitation, phonons
(surface acoustic waves), propagating across the surface of
the film. This effect may provide
opportunities to develop hybrid
wave–based devices in which
information (classical and possibly quantum) can be stored,
manipulated, and carried in a
variety of different ways. —ISO

y g

J. Exp. Med. (2023)
10.1084/jem.20231242

and borrowers in securing loans
with physical collateral can affect
households’ access to resources.
Describing their field experiment
in Uganda, Gertler et al. show
how securing loans with digital
collateral, a home solar system
that the lender doesn’t physically
repossess but can temporarily deactivate using lock-out
technology until loan payments
are made, increased the rate of
return to the lender and reduced
loan default rates. The loans
allowed borrowers to pay school
fees, increasing enrollment in, and
expenditures on, schooling. —BW

y

Strongly coupling
magnons to phonons

European barn owls exhibit continuous color
variation that is primarily generated by
interactions between three genomic regions.

in cancer patients, there is
still much to learn about how
inhibitory receptors affect
immune cells. Chan et al. developed an imaging technology
called FILMSTAR (Fluorescent
Intracellular Live Multiplex Signal
Transduction Activity Reporter)
to study activation signals in
T cells. The system allows for
real-time, single-cell tracking of
multiple signaling pathways and
may improve our understanding of how PD-1 and related
molecules influence the T cell
anticancer response. —PNK

p

transcription factor, PINNATELIKE PENTAFOLIATA2 (PINNA2),
in Medicago truncatula (a small
annual legume) that controls
compound leaf development.
They found that PINNA2 is
specifically expressed at organ
boundaries, and its loss-of-function mutations convert trifoliate
leaves into a pentafoliate form.
PINNA2 synergistically acts
with other genes to down-regulate the expression of SINGLE
LEAFLET1 (SGL1), a positive
regulator of leaflet initiation.
Precise SGL1 expression by
PINNA2 in compound leaf
primordia maintains a proper
pattern of leaflet initiation.
Therefore, regulatory pathways
are intrinsically coordinated
in time and space to regulate
compound leaf morphogenesis.
—AWa

GENETICS

R ES E ARCH

ALSO IN SCIENCE JOURNALS
NEUROSCIENCE

Closing the (Syn)GAP
on plasticity

Science p. 964, 10.1126/science.adh0755

GROUNDWATER

A changing dynamic

Science p. 962, 10.1126/science.adf0630

INNATE IMMUNITY

A bacteria-killing coat
Multiplying crop
improvement

1 MARCH 2024 • VOL 383 ISSUE 6686

The pre-T cell receptor a
(PTCRA) chain is critical for ab
T cell development in mice, but
whether this also holds true for
humans is unclear. Materna et al.
examined 10 patients with rare
biallelic loss-of-function PTCRA
variants. Despite having small
thymi and low circulating naïve
ab T cell counts, the memory ab
T cell counts in these patients
were normal, suggesting that the
pre-TCRa may not be absolutely
required for ab T cell development in humans. The authors
also identified two common
hypomorphic PTCRA variants
that were responsible for partial
pre-TCRa deficiency in homozygotes in about one in 4000
individuals from the Middle East
and South Asia, resulting in high
circulating naïve gd T cell counts
and a significantly increased
incidence of autoimmunity.
—STS
Science p. 966, 10.1126/science.adh4059

THERMODYNAMICS

Measuring the changes
in entropy
Entropy in a closed system is
a measure of the disorder or
randomness and represents the
energy unavailable to do work.
Di Terlizzi et al. propose a
method for evaluating the
steady-state entropy production
in nonequilibrium stochastic
systems (see the Perspective by
Roldán). This method is achieved
using a variance sum rule that
connects changes in positions with the forces required
to restore those positions. The
approach was verified using
high-resolution experimental data on optically trapped
Brownian particles and red blood
cells, including stretching of the

ATMOSPHERIC DYNAMICS

Building up flow
How does random turbulence
organize to form large-scale
structures in planetary atmospheres? Such a process implies
the existence of an inverse
energy cascade, an idea that
has been suggested but not yet
demonstrated for Earth’s atmosphere. Alexakis et al. conducted
numerical simulations at high
spatial resolutions to show that
rotating and stratified flows can
support a three-dimensional,
bidirectional cascade of energy
under conditions applicable to
those on Earth. These results
explain how spontaneous order
can arise in a dry atmosphere
through an inverse cascade of
energy to large spatial scales.
—HJS
Science p. 1005, 10.1126/science.adg8269

EVOLUTION

Follow the path
Established morphological traits
can direct trait evolution along
particular trajectories in a process known as path dependence.
Varney et al. explored this process in two lineages of chitons
that have evolved two different
visual systems, eye spots and
shell eyes (see the Perspective
by Sumner-Rooney). They found
that lineages with more nerve
openings in their shell evolved
eye spots, whereas those with
fewer openings evolved shell
eyes. —SNV
Science p. 983, 10.1126/science.adg2689;
see also p. 947, 10.1126/science.ado1700

ORGANIC CHEMISTRY

C–C coupling diverted
to form C–N bonds
Two of the most common
reactions in pharmaceutical
science.org SCIENCE

,

961-B

A spectrum of pre-TCRa
deficiency

Science p. 971, 10.1126/science.adh1823;
see also p. 948, 10.1126/science.adn9799

y

Plants often throw out offspring
with different chromosome numbers (ploidy). This phenomenon
can create bigger flowers and
fruits or provide environmental
resilience, and it has become a
target in plant breeding for crop
improvement. Westermann et
al. investigated why de novo
polyploids show reduced fertility compared with established
descendant polyploid plants.
Fertility loss was originally
thought to result primarily from
issues relating to segregation of
multiple chromosomes during
meiosis, but this is not the only
obstacle. In a survey of genes
under selection, the authors
identified two: AGC1, which

Human cells have many mechanisms to detect and respond
to bacterial and viral intruders.
Proteins that recognize bacterial
cell wall components can coat
the surface of invading bacteria
and serve as a scaffold for the
assembly of signaling proteins
and antimicrobial enzymes.
Zhu et al. performed genetic,
biochemical, and structural
experiments to reveal how
this large protein complex is
formed and to understand how
it functions to protect cells from
intracellular infection. Cryo–
electron tomography revealed
that monomers and dimers of
GBP1 form an even coating on
bacteria, and medium-resolution
reconstructions suggested
that an extended conformation
allows the protein to insert into
and disrupt the bacterial outer

IMMUNOLOGY

cells and contour fluctuations, a
measure of the cells’ metabolic
activity. —MSL

y g

PLANT SCIENCE

Science p. 965, 10.1126/science.abm9903

y

The availability of fresh groundwater is vital for agriculture,
industry, people, and ecosystems, but its quality and
quantity have been significantly
affected by climate change and
anthropogenic activities. Kuang
et al. review the changes that
groundwater is experiencing
now and will experience in the
near future and discuss future
challenges to groundwater supplies. Considering these changes
is important for managing this
critical resource in an ever-more
challenging environment. —HJS

membrane. —MAF and SMH

g

Science p. 963, 10.1126/science.adk1291;
see also p. 946, 10.1126/science.adn8707

is important for pollen tube
growth, and ACA8, a calcium
transporter, both of which are
known to affect pollen tube
structure. Experiments verified
the role that adaptations in the
expression of these genes play in
correcting defective pollen tube
growth after ploidy generation.
—CA and AWa

p

Synaptic plasticity is the critical
mechanism supporting learning, memory, and many other
neurophysiological processes
during brain development and
in adulthood. The GTPaseactivating protein SynGAP has
been shown to be necessary for
synaptic plasticity, and mutations have been associated with
autism and other intellectual
disabilities. Araki et al. found
that the GAP activity of SynGAP
is not required for synaptic
plasticity (see the Perspective by
Choquet). Instead, the protein
modulates synaptic plasticity
by competing with the AMPA
receptor–TARP complex at
excitatory synapses, influencing the formation of molecular
condensates and ultimately
regulating the recruitment of
AMPA receptors during plasticity. These results will help in the
development of treatments for
SynGAP-mediated neurological
disorders. —MMa

Edited by Michael Funk

Science p. 1019, 10.1126/science.adl5359;
see also p. 950, 10.1126/science.ado0068

CATALYSIS

A protective layer
of innate-like T cells
Although T cells are primarily
activated in response to antigen
recognition through their T cell
receptor (TCR), they can also
undergo TCR-independent
“bystander” activation in
response to certain cytokines.
Using single-cell transcriptomics
and chromatin accessibility
profiling, Watson et al. found that
neonatal CD8+ T cells undergo a
distinct and robust program of
innate-like bystander activation
controlled by a balance of Bach2
and AP-1 transcription factor
activity. Neonatal CD8+ T cells
could protect against bacterial,
viral, and parasitic pathogens
in a TCR-independent manner.
Innate-like CD8+ T cells were
also present in both adult mice
and humans, indicating that
multiple layers of defense may
exist within the adult naïve T cell
pool. —CO

y g
y

Sci. Immunol. (2024)
10.1126/sciimmunol.adf8776

,

Science p. 998, 10.1126/science.adk5195

T CELLS

y

An effective catalyst for propane
dehydrogenation must avoid
unwanted carbon buildup and
metal agglomeration by stabilizing rhodium atoms in silicalite-1
zeolite. Rhodium catalysts for
this reaction tend to be unstable
at the high temperatures need to
drive high conversion. Zeng et al.
found that alloying with indium
formed RhIn4 groups attached
to the zeolite through an In–O
linkage. This catalyst was stable
for more than 1200 hours at
600°C, exhibited high propane
conversion (~65%) and propylene selectivity (98%), and was
also highly active for ethane and
butane dehydrogenation. —PDS

Sci. Signal. (2024)
10.1126/scisignal.adh1178

g

Stabilizing rhodium
atoms in zeolites

clinical severity of RA. The US
Food and Drug Administration–
approved ALOX5 inhibitor
zileuton suppressed CD4+ T cell
pyroptosis and reduced joint
inflammation in rodent models
of RA, suggesting ALOX5 as a
potential therapeutic target.
—AMV

p

chemistry entail the
Pd-catalyzed formation of C–C
and C–N bonds. Onnuch et al.
report that conditions under
which the two reactants primed
to form C–C bonds through
Suzuki coupling can instead
both be coupled to a common N
center to form an amine (see the
Perspective by Shaughnessy).
The intervention, which hinges
on a bulky phosphine ligand on
Pd and a P-based electrophilic
N source, offers a simple means
of diversifying existing Suzuki
reactant libraries. —JSY

IMMUNOLOGY

ALOX5 stokes
rheumatoid arthritis
Pyroptosis of CD4+ T cells
is associated with synovial
inflammation in rheumatoid
arthritis (RA). Cai et al. found
that increased abundance of
the leukotriene biosynthetic
enzyme ALOX5 in circulating
and synovium-infiltrating CD4+
T cells drove pyroptosis in these
cells, which correlated with the
SCIENCE science.org

1 MARCH 2024 • VOL 383 ISSUE 6686

961-C

RES EARCH
◥

GROUNDWATER

The changing nature of groundwater
in the global water cycle
Xingxing Kuang, Junguo Liu*, Bridget R. Scanlon, Jiu Jimmy Jiao, Scott Jasechko, Michele Lancia,
Boris K. Biskaborn, Yoshihide Wada, Hailong Li, Zhenzhong Zeng, Zhilin Guo, Yingying Yao,
Tom Gleeson, Jean-Philippe Nicot, Xin Luo, Yiguang Zou, Chunmiao Zheng*

5

2

12
4

1
2
3
4
5

7

Unconfined
aquifer
Seawater
Shale
intrusion
formation

Flux (×1000 km3/yr)
Ocean evaporation (420 ± 20%)
Precipitation on ocean (380 ± 20%)
Net water vapor flux transport (46 ± 20%)
Rainfall (98.5 ± 10%)
Snowfall (12.5)

11
8

Storage (×1000 km3)
6
7
8
9
10

Terrestrial evapotranspiration (69 ± 10%)
Runoff (46 ± 10%)
Groundwater discharge (4.5 ± 70%)
Groundwater withdrawal (~1.0)
Managed aquifer recharge (MAR) 0.01

11
12
13
14

Ocean 1,338,000
Glaciers and snow 24,064
Permafrost 300
Groundwater 23,400

Simplified global water cycle with its components. Groundwater is becoming increasingly more dynamic
in the global water cycle.
Kuang et al., Science 383, 962 (2024)

1 March 2024

▪

The list of author affiliations is available in the full article online.
*Corresponding author. Email: liujg@sustech.edu.cn (J. L.);
zhengcm@sustech.edu.cn (C. Z.)
Cite this article as X. Kuang et al., Science 383, eadf0630
(2024). DOI: 10.1126/science.adf0630

READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.adf0630
1 of 1

,

Co
nfin
ed
aqu
ifer

Industry

9

y

6
14

Bedrock

1

10

13

y g

3

OUTLOOK: The role of groundwater in the global water cycle is becoming increasingly dynamic
and complex while the security of groundwater
resources faces considerable threats worldwide in terms of both quantity and quality. The
sustainable use of groundwater resources has
become a crucial global concern. In planning
for a more sustainable future, groundwater
resources should be considered from both
regional and global perspectives, especially for
large, transboundary groundwater systems.
As global changes continue to affect these resources, it is imperative to manage groundwater and surface water as a single resource.
Additionally, ensuring food and water security and maintaining ecosystem health must be
addressed concurrently. Various management
strategies, including forest and wetland conservation, desalination, wastewater recycling,
managed aquifer recharge, water diversion projects, and green infrastructure development
may be employed to bolster the resilience of
groundwater. Major research gaps exist that
warrant further exploration, including detailed
studies of groundwater in high-latitude and
mountainous regions, more accurate predictions
of groundwater recharge, quantitative assessments of injected and discharged groundwater
volumes, and accurate modeling of the global
water balance. To address these gaps effectively, comprehensive observational datasets
are essential, as they enable a thorough evaluation of the current state and future changes
in groundwater resources.

y

pogenic activities have led to regional and global transformations in groundwater dynamics.
Climate-driven modifications include shifts in
groundwater recharge rate across continents,
increased groundwater contributions to streamflow in glacierized catchments, and profound
alterations in groundwater flow patterns within permafrost areas. Glacial meltwater infiltrates into the subsurface, sustaining a stable
groundwater discharge to streams during dry
seasons. Permafrost thaw fosters increased
rainfall infiltration, amplifies groundwater storage, creates new subsurface flow pathways, and
increases groundwater discharge to streamflow. Direct anthropogenic activities include
groundwater withdrawal, unconventional oil and
gas production, geothermal energy exploration,
managed aquifer recharge, afforestation, land
reclamation, urbanization, and international food
trade. These undertakings engender groundwater withdrawal and injection, reshaping regional groundwater flow regimes, impacting water

g

ADVANCES: Climate change and other anthro-

able freshwater resource and forms an active
component of the global water cycle. It serves
as the primary source of fresh water for billions of people and provides drinking water
to numerous communities. Moreover, groundwater supplies over 40% of global irrigation
demand and is becoming increasingly important in mitigating water scarcity induced by
climate change. In the past few decades, climate change and other anthropogenic activities have substantially altered groundwater
recharge, discharge, flow, storage, and distribution. Climate warming–induced glacier retreat and permafrost thaw have led to changes
in groundwater in glacierized and permafrost
areas. In the interest of fostering a more comprehensive understanding of the state of global groundwater, we present a synthesis of its
changing nature in the global water cycle over
the recent decades, shaped by the impacts of
climate change and other various anthropogenic activities.

p

BACKGROUND: Groundwater is the largest avail-

tables and groundwater storage, and redistributing embedded groundwater in foods globally. Groundwater depletion occurs across the
globe and has intensified over recent decades.
Groundwater pumped from aquifers participates in the global water cycle by contributing to river discharge and evapotranspiration.
Groundwater withdrawal transfers fresh water
from long-term storage to the active water cycle
at the Earth’s surface. Moreover, nonrenewable
groundwater withdrawal from deep aquifers
integrates deep ancient fossil groundwater
into the active contemporary water cycle, ultimately contributing to rising sea levels. The risks
of saltwater intrusion and groundwater inundation in coastal regions are exacerbated by sea
level rise. The importance of groundwater for
drinking and irrigation is poised to increase in
response to climate change. Consequently, the
effects of groundwater depletion on sea level rise
are expected to become magnified in the future.

ILLUSTRATION ADAPTED FROM EREBORMOUNTAIN/SHUTTERSTOCK

REVIEW SUMMARY

RES EARCH

REVIEW

◥

GROUNDWATER

The changing nature of groundwater
in the global water cycle
Xingxing Kuang1, Junguo Liu1,2*, Bridget R. Scanlon3, Jiu Jimmy Jiao4, Scott Jasechko5,
Michele Lancia1, Boris K. Biskaborn6, Yoshihide Wada7, Hailong Li1, Zhenzhong Zeng1, Zhilin Guo1,
Yingying Yao8, Tom Gleeson9, Jean-Philippe Nicot3, Xin Luo4, Yiguang Zou1, Chunmiao Zheng10,1*

1 March 2024

1 of 13

,

Kuang et al., Science 383, eadf0630 (2024)

y

*Corresponding author. Email: liujg@sustech.edu.cn (J.L.);
zhengcm@sustech.edu.cn (C.Z.)

y g

State Environmental Protection Key Laboratory of
Integrated Surface Water-Groundwater Pollution Control,
Guangdong Provincial Key Laboratory of Soil and
Groundwater Pollution Control, School of Environmental
Science and Engineering, Southern University of Science and
Technology, Shenzhen, China. 2Henan Provincial Key Lab of
Hydrosphere and Watershed Water Security, North China
University of Water Resources and Electric Power,
Zhengzhou, China. 3Bureau of Economic Geology, Jackson
School of Geosciences, University of Texas at Austin, Austin,
TX 78758, USA. 4Department of Earth Sciences, The
University of Hong Kong, Pokfulam Road, Hong Kong, China.
5
Bren School of Environmental Science and Management,
University of California, Santa Barbara, CA 93106, USA.
6
Alfred Wegener Institute Helmholtz Centre for Polar and
Marine Research, 14473 Potsdam Germany. 7Biological and
Environmental Science and Engineering Division, King
Abdullah University of Science and Technology, Thuwal,
Saudi Arabia. 8Department of Earth and Environmental
Science, School of Human Settlements and Civil Engineering,
Xi’an Jiaotong University, Xi’an, China. 9Department of Civil
Engineering and School of Earth and Ocean Sciences,
University of Victoria, 3800 Finnerty Road, Victoria, BC V8P
5C2, Canada. 10Eastern Institute for Advanced Study, Eastern
Institute of Technology, Ningbo, China.

y

1

water could mitigate the impacts of climate
extremes on water resources (11, 12) and is already widely used as a buffer against water
scarcity during droughts (7). The importance
of groundwater as a drinking and irrigation
source is expected to increase as a result of climate change (13). Global warming may cause
shifts in groundwater recharge rates (7, 14);
warming also causes accelerated glacier retreat (15, 16) and permafrost degradation in
high-latitude and high-altitude areas (17). Glacier retreat and permafrost degradation in turn
lead to changes in groundwater in glacierized
and permafrost areas (18, 19). Lowland populations are often dependent on water resources
derived from mountain headwaters as irrigation sources (20).
Many different anthropogenic activities have
changed groundwater flow, storage, and distribution during past decades. Groundwater
overexploitation occurs in many regions globally and groundwater depletion has grown in
past decades (21). Other anthropogenic activities that can lead to changes in groundwater
flow and storage include unconventional oil
and gas production (22), geothermal energy
exploration (23), managed aquifer recharge
(24), afforestation (25), land reclamation and
urbanization (26), and international food trade
(27). Much of the withdrawn groundwater
eventually enters the oceans and contributes
to sea level rise (28). Rising sea levels increase
the water table in coastal areas, which may
cause flooding through groundwater inundation (29).
In the interest of developing a more comprehensive understanding of the state of global groundwater, we synthesize aspects of the
changing nature of groundwater in the global
water cycle over recent decades resulting from

Groundwater recharge is affected by climate
variability and change (30, 31). Climate change
affects groundwater resources by changing
precipitation, evapotranspiration (ET), recharge,
and pumpage (7, 32). On a global scale, modern global mean groundwater recharge fluxes are
estimated to be at least ~12,000 to ~17,000 km3
per year (33–36). However, recharge rates vary
substantially across different regions. Fig. 2A
shows simulated mean annual groundwater
recharge between 1960 and 2010 modeled by
PCR-GLOBWB and considering lateral groundwater flow (37). A nonlinear relationship is
found between precipitation and groundwater
recharge in some regions, with wetter regions
having higher recharge than drier areas (38).
At the global scale, the effects of precipitation
change on global average groundwater recharge may be insignificant. Higher precipitation (and recharge) in some areas may be
offset by lower precipitation (and recharge) in
other areas, leading to relatively small changes
in interannual groundwater recharge rates at
the global scale but large changes at the local
scale (31). Both increasing and decreasing
trends in groundwater recharge have been
found in response to climate change (14).
Increases in recharge projected in some areas
have been attributed to projected increases
in precipitation in regions such as the Upper
Colorado River Basin in the United States (39)
and to increasing intensity of precipitation in
regions such as Indonesia and East Africa
(14, 38). Increases in induced recharge may
also be caused by groundwater overexploitation (30). Groundwater withdrawals vary over
time with climate extremes, with more withdrawals during droughts and less withdrawals
during wet periods (30). Declines in groundwater recharge are projected in some tropical
and temperate climate regions (14, 40), such as
much of the western United States (41). An
average decline of 10 to 20% in total recharge
is estimated for some aquifers in the southwestern United States (41). Climate models

g

G

roundwater is the largest available freshwater resource and constitutes a major
component of the global hydrological
cycle (1). Groundwater also provides
drinking water for billions of people
(2) and supplies ~40% of global irrigation demand (3), in which it is becoming increasingly important (4–6). As a key component of
the global water cycle (Fig. 1), groundwater
sustains river discharge, lakes, wetlands, crops,
forests, and ecosystems (7). The global water
cycle is being modified by climate change and
other anthropogenic activities at an unprecedented rate (8), the effects of which need
to be better understood to meet the challenges that these changes present.
Climate change is expected to fundamentally
alter the global water cycle (9, 10). Ground-

Groundwater changes driven by climate change
Effects on groundwater recharge variability
p

In recent decades, climate change and other anthropogenic activities have substantially affected
groundwater systems worldwide. These impacts include changes in groundwater recharge, discharge,
flow, storage, and distribution. Climate-induced shifts are evident in altered recharge rates, greater
groundwater contribution to streamflow in glacierized catchments, and enhanced groundwater flow
in permafrost areas. Direct anthropogenic changes include groundwater withdrawal and injection,
regional flow regime modification, water table and storage alterations, and redistribution of embedded
groundwater in foods globally. Notably, groundwater extraction contributes to sea level rise, increasing
the risk of groundwater inundation in coastal areas. The role of groundwater in the global water cycle
is becoming more dynamic and complex. Quantifying these changes is essential to ensure sustainable supply
of fresh groundwater resources for people and ecosystems.

climate change and other anthropogenic activities. First, we discuss alterations to groundwater systems driven by climate change,
including shifts in groundwater recharge
and variations in groundwater flow systems
in glacierized and permafrost areas. Then,
we review other anthropogenic activities that
lead to changes in groundwater levels, storage,
and regional groundwater flow regimes. Finally, we evaluate the contribution of groundwater
to sea level rise and groundwater inundation in
coastal areas induced by sea level rise. We
acknowledge that human activities also affect
groundwater quality but a thorough discussion of groundwater quality changes is beyond
the scope of this Review.

RES EARCH | R E V I E W

Snowfall
(12.5 ± 10%)

Net water vapor flux transport
(46 ± 20%)

Glaciers and snow
24,064

Precipitation on ocean
(380 ± 20%)

Rainfall
(98.5 ± 10%)

Permafrost
300

Terrestrial
evaporation
(69 ± 10%)

MAR
(0.01)

Ocean
evaporation
(420 ± 20%)
Industry

Groundwater
withdrawal
(~1.0)

Groundwater
23,400

Runoff (46 ± 10%)
Co

nf

in

ed

Unconfined
aquifer

fe

Sh

Be

dro

ck

ale

for

ma

r

Ocean
1,338,000

p

aq
ui

Seawater Groundwate
r
intrusion
discharge
(4.5 ± 70%)

tio

n

g
y

Fig. 1. Global hydrological cycle with its components. The global water fluxes (×1000 km3 per year) in brackets and water storage (×1000 km3) were obtained from
previous studies (9, 36). The upward arrows show annual evaporation from the ocean and terrestrial evapotranspiration. Global groundwater withdrawal is set at 1000 km3 per
year based on data from 2010 in the literature (21). Antarctica was not included in the terrestrial water balance. [Adapted from EreborMountain/Shutterstock]

y g
y
,
Fig. 2. Groundwater recharge, withdrawal, water level decline, and storage changes. (A) Mean annual groundwater recharge from 1960 to 2010 modeled by PCR-GLOBWB
coupled with MODFLOW (37). Positive values indicate groundwater recharge and negative values indicate capillary rise. (B) Mean annual potential net groundwater withdrawal
from 1980 to 2016 simulated by WaterGAP 2.2d (44). Negative values indicate an increase in groundwater storage caused by surface water irrigation whereas positive values
indicate a net removal of groundwater from aquifers due to human water use. (C) Annual groundwater storage change rate from 1980 to 2016 modeled by WaterGAP 2.2d (44).
(D) Annual averaged decline in the groundwater level in the world’s major aquifers from 1990 to 2014 simulated by PCR-GLOBWB 2 run coupled with MODFLOW (101).
Kuang et al., Science 383, eadf0630 (2024)

1 March 2024

2 of 13

RES EARCH | R E V I E W

precipitation result in large uncertainties in
projected groundwater recharge (7). Groundwater recharge is affected by rainfall amount
and intensity (38, 40). Regions that experience
increases in rainfall intensity may experience
increases in groundwater recharge (12, 40).
However, many predictions of future changes
in precipitation frequency and intensity are
highly uncertain. Current representations of
hydrological processes and groundwater in
global hydrological models may also lead to
large uncertainties in the projected groundwater recharge (14). Incorporation of the impact of the changing climate and atmospheric
CO2 levels on vegetation in global hydrological
models can lead to variations of 100 mm per
year in simulated groundwater recharge (14).
Regionally, the predominant sources of uncertainty may stem from selection of global climate models and emissions scenarios (49).
Increases in groundwater contribution to
streamflow due to glacier retreat

y

A

g

Glacial meltwater has been identified as an
important source of aquifer recharge in glacierized catchments (50). A portion of glacial
meltwater infiltrates and recharges groundwater; groundwater then discharges farther
down-gradient to streams (Fig. 3A) (51). For
example, in the rapidly retreating Virkisjökull
glacier in southeastern Iceland, >25% and
often >50% of the groundwater is recharged
from glacial meltwater in summer (52). In the

Upper Indus River Basin, ~44% of annual
groundwater recharge is derived from glacial meltwater (53).
Groundwater in glacierized catchments contributes substantially to river discharge. In
Nepal, groundwater flowing through fractured basement aquifers contributes ~66% of
annual river discharge, which is six times
higher than the contribution from glaciers
and snow melt (54). The percentage of river
discharge derived from groundwater can be
>90% (55). During dry periods and winter,
groundwater may be the main source of river
discharge, with contributions of 50 to 90%
(56). In the Shullcas watershed in central Peru,
a typical proglacial watershed, groundwater
provides ~70% of the dry season streamflow
(57). These examples highlight the importance
of groundwater in sustaining streamflow in
mountainous areas.
Accelerating glacier retreat may threaten
the sustainability of water resources in mountainous areas (57); however, groundwater in
high mountain areas may provide some resilience to glacier retreat (19). Groundwater
storage in glacier forelands can buffer streamflow changes (52). The stored groundwater is
released during dry seasons and compensates
for high variability in glacial meltwater and
sustains streamflow (58). Climate change has
induced substantial glacier retreat in recent
decades, with glaciers retreating in High
Mountain Asia and many of the world's other

p

project that droughts will become more frequent and intense in California, decreasing
recharge and increasing demand for groundwater (42). However, considerable uncertainty
exists in some of these climate projections.
Surface water irrigation can increase groundwater recharge and replenish aquifers from
irrigation return flows (40, 43, 44). Inefficient
surface water irrigation will increase groundwater recharge and storage (45). Canal leakage
and return flow are the main pathways for
increased groundwater recharge from surface
water irrigation. Groundwater storage in the
Indo-Gangetic Basin increased by ~420 km3
during the 20th century before large-scale
groundwater withdrawal began in the late
1990s and early 2000s (46). Leakage from surface water irrigation increased groundwater
storage by ~20 km3 in the Columbia Plateau in
the northwestern United States between ~1940
and ~1970 (45). Previous studies estimated that
10 to 50% of total irrigation becomes irrigation return flow (47); the latter can be reduced
with more efficient irrigation schemes such as
drip rather than flood irrigation (48).
Uncertainty in recharge projections arises
from several sources, including uncertainty in
changes in future precipitation rates and, critically, intensities (7, 40, 41). Annual and seasonal
precipitation and temperature are identified
as some of the most important factors in predicting spatial variation in groundwater recharge (31). Considerable uncertainties in future

C
River

Active layer

Lake

Permafrost

y g

Glacier
Runoff
Groundwater
flow paths

River

Sub-permafrost aquifer

Bedrock

y
,

D

B

River

Active layer
Lake

Permafrost
Glacier
Runoff
Groundwater
flow paths

Talik

River
Sub-permafrost aquifer

Bedrock

Fig. 3. Schematics of groundwater flow systems in glacierized catchments and permafrost areas. (A) and (B) Groundwater flow system in a glacierized
catchment before and after glacier retreat (A) and (B), respectively (61). The blue curves with arrows show the groundwater flow from subglacial meltwater recharge.
(C) and (D) Groundwater flow system in a permafrost area before and after climate warming (C) and (D), respectively (78, 79). The blue arrows in (D) show the
enhanced groundwater flow.
Kuang et al., Science 383, eadf0630 (2024)

1 March 2024

3 of 13

RES EARCH | R E V I E W

,

4 of 13

y

1 March 2024

Groundwater is pumped out of many aquifers
globally (89, 90). As an essential water source
for humans, groundwater withdrawal accounted for an estimated ~22% of total water
withdrawal in 2000 according to global hydrological models (34) and ~26% in 2010 according to national and international databases
(91). Groundwater is withdrawn from both
unconfined and confined aquifers (Fig. 4, A
and B). Global groundwater withdrawal increased from ~310 ± 37 to 460 km3 per year
in 1960 (4, 92, 93), to ~570 to 790 ± 30 km3
per year in 2000 (21, 33, 34), and then to
~1000 km3 per year in 2010 (21). Fig. 2B shows
the mean annual potential net groundwater
withdrawal from 1980 to 2016 simulated by
the global hydrological model WaterGAP 2.2d
(44). Although global groundwater withdrawal
has increased from 1960 to the present, groundwater withdrawal has stabilized during recent
decades in countries such as the United States,
China, Pakistan, and Iran (91, 94). Large groundwater withdrawals have caused substantial
declines in global aquifer storage (Fig. 2C)
(6, 95, 96) and groundwater depletion may
account for ~15% of total groundwater withdrawal (97, 98). The remaining 85% of groundwater withdrawal is linked to surface water
capture, reduced evapotranspiration, and decreased discharge (97, 98). Groundwater withdrawal has resulted in substantial groundwater
level declines in many areas in recent decades,
such as parts of the US High Plains aquifer, the

y g

Kuang et al., Science 383, eadf0630 (2024)

Groundwater changes driven by other
anthropogenic activities
Groundwater withdrawal

y

Permafrost underlies 14 to 16 million km2 of
the Earth's exposed land surface (66), with a
mean active layer thickness of ~0.8 m in the
High Arctic and ~2.3 m in the alpine and highplateau regions (67). The observed permafrost
temperatures have increased continuously over
the past few decades (17, 68). Permafrost thaw
and active layer thickening occur throughout
cold regions globally (Fig. 3, C and D); the
latter has been observed since the 1990s (68),

charge to streams, providing stable baseflow
during winter and dry periods (84). Enhanced
infiltration, groundwater storage, and groundwater flow indicate an expanding role for
groundwater in the high-latitude hydrological cycle (85). Continued permafrost degradation may exacerbate regional ecological
challenges, including a reduction in soil water
availability, vegetation degradation or greening, and land desertification (86). It is crucial
to recognize the intricate interdependencies
between the permafrost thermal regime and
vegetation, as the impact of vegetation on
permafrost degradation is complex (68, 86).
Substantial increases in groundwater discharge to streams induced by permafrost thaw
are likely to occur in the next few centuries
(73, 87). An increase of 2°C in the mean annual
surface temperature of the Tibetan Plateau
could increase groundwater discharge to streams
by a factor of three (88). The increase in runoff
is caused by infiltrated water flowing through
the subsurface and discharging to rivers during
periods of flow recession (71). In catchments
with ice-rich permafrost, excess ground ice
provides large quantities of potential meltwater
for groundwater flow (87).

g

Groundwater flow enhancement by
permafrost thaw

and in the Russian Arctic the active layer thickness increased by 0.4 m between 1999 and 2019
(68). In the Tibetan Plateau, the active layer
thickness increased at 19.5 cm per decade from
1980 to 2018 (69), and the permafrost area
decreased by ~1500 × 103 km2 during the past
half century (70).
Thawing permafrost increases groundwater
storage, deepens groundwater flow pathways,
and augments groundwater discharge to streams
(Fig. 3, C and D) (18, 71, 72), especially during
low-flow seasons (73). The permafrost area in
the Yangtze River source region decreased by
~8000 km2 between 1962 and 2012, increasing
groundwater storage at a rate of 1.6 km3 per
year (74). In the Yukon River basin, long-term
(>30 years) observations indicate a 7 to 9%
increase in groundwater discharge to streamflow per decade (18). Thawing permafrost and
thickening of the active layer can augment
baseflow by enhancing groundwater flow pathways and releasing groundwater from storage
to streams (75–77). Thickening of the active
layer can eventually lead to the formation of
large taliks (unfrozen zones in permafrost),
plausibly increasing infiltration rates, subsurface storage volumes, and flow depths that
alter groundwater flow pathways (Fig. 3, C and
D) (78, 79), increasing groundwater discharge
to streams through baseflow (72). Progressive
permafrost thaw facilitates shallow groundwater flow systems whereas complete permafrost thaw creates new deep groundwater flow
systems (73). Thawing permafrost also increases
hydrologic connectivity and linkages between
surface water and groundwater (77).
Vertical talik expansion enhances regional
groundwater circulation (76, 79). When a closed
talik degrades to an open talik (i.e., a talik
completely penetrates the permafrost), a pathway is created for groundwater flow (78). Open
taliks connect shallow groundwater in the active layer to the aquifer below the permafrost,
serving as vertical conduits for groundwater
flow (Fig. 3D), thus enhancing regional groundwater circulation and discharge (79). Open taliks
enhance surface water–groundwater interactions and groundwater flow converges at the
talik (78, 80). Open taliks allow migration of
relatively warm groundwater from above or
geothermally warmed groundwater from below, thus accelerating permafrost thaw and
expanding the talik network (81).
As global warming persists in the coming
decades, permafrost is projected to continue
thawing (82). Over 40% of permafrost area
may disappear if the climate is stabilized at
2°C above preindustrial levels (82). The low
permeability of permafrost generally provides
a hydraulic barrier that reduces rainfall and
snowmelt infiltration (83). Where permafrost
is discontinuous, rainfall and snowmelt can
infiltrate and recharge groundwater, flow within the groundwater system, and finally dis-

p

high mountain areas (15). Global projections
suggest that glaciers will lose ~20 to ~50% of
their mass by 2100 relative to 2015 (59). In the
Shullcas watershed in Peru, glaciers are projected to disappear entirely by 2100; however,
the relatively consistent groundwater discharge
to rivers is expected to compensate for the
reduction in glacial meltwater (57). Continued
future climate change may further decrease
glacial meltwater contributions to rivers and
some stream sources may undergo a progressive shift toward snow melt and groundwater
(Fig. 3B) (60).
As climate warming continues, many debriscovered glaciers will transform into rock glaciers, which are poorly sorted, angular rock
debris with ice (61). Groundwater stored in
rock glaciers discharges to streams through
springs (61, 62). In the Canadian Rockies,
groundwater discharge from one rock glacier spring accounts for 50% of streamflow
during summer and up to 100% during winter (63). Continued climate warming may lead
to the thawing of ice in rock glaciers. Although
streamflow may initially increase as a result
of ice melting in rock glaciers (62), after the
volume of stored ice has declined, snowmelt
and rainwater formerly flowing on the ice surface may infiltrate into the rock glacier matrix
and flow out of the basin as groundwater (62)
(Fig. 3B).
Rock glaciers, talus, moraines, and alpine
meadows are typical to alpine aquifers (64).
Alpine aquifers can store large volumes of
groundwater, which is vital in sustaining baseflow in rivers during low-flow seasons (64). In
the Canadian Cordillera, which is experiencing glacier retreat, minimal reductions in
winter streamflow have been observed in 17
rivers, indicating that groundwater storage in
alpine headwater aquifers supports streamflow during the low-flow season (64). The high
hydraulic conductivities of these alpine aquifers
allow rapid infiltration of rainwater and snowmelt to unconfined aquifers above bedrock
surfaces (63); these aquifers can then provide
steady discharge to rivers for many months
(64, 65). A large amount of groundwater stored
in these aquifers sustains river runoff and
stabilizes catchment outflow, which may affect
catchment responses to climate change (65).

RES EARCH | R E V I E W

A

Withdrawal

B

Ground surface

0

Withdrawal

C
0

Ground surface

0

Before withdrawal
10

10

After withdrawal

15

Unconfined aquifer

F

1

3
4

2 km

Direct discharge
Cold
water

Reinjection

Production
well

Injection
well

2
3
4

5

Withdrawal

Geothermal
energy
utilization

Hot
water

Injection
well

Production
well

2

5

1 km

H

1 km

I

ET

Infiltration pond
0
5

After MAR

10

Reinjection

0

1

Ground surface

5

Direct discharge
Cold
water

Before MAR

15

After MAR
10

0

Before MAR

Depth (m)

Injection

Geothermal
energy
utilization

0

Shale formation
Deep formation

Hot
water

Depth (km)

Surface Pumping
Injection
water
well
well
Shallow aquifer

E

y

0

10 m

125

g

G

Depth (m)

Produced water

2

4

Deep confined
aquifer

50

Shell well

3

10 m

Confined aquifer

Depth (km)

Depth (km)

40

p

1

Confined aquifer

75

Aquitard

10 m

Hydraulic fracturing
recycling
0

30

Depth (m)

D

Aquitard

Aquitard

50

100

20

25

After withdrawal

After withdrawal

20

Before withdrawal

25

Depth (m)

Before withdrawal

Depth (m)

Depth (m)

5

Withdrawal
Ground surface

15

Aquifer

Aquifer

100 m

100 m

enhanced geothermal system (23, 129). (G and H) Schematics of MAR: (G) Aquifer
storage and recovery, in which water is injected into the aquifer for storage
and recovery using the same well; (H) Infiltration ponds, in which water infiltrates
from a constructed pond into an unconfined aquifer for storage and recovery (140).
(I) Water table before and after afforestation. ET, evapotranspiration.

137 km3 per year from 1960 to 2010 according
to PCR-GLOBWB (93). Estimates of cumulative global groundwater depletion between
1960 and the early 21st century range from
2000 to as much as ~27,000 km3 (93, 103),
highlighting the substantial uncertainty in
cumulative groundwater depletion estimates.
Groundwater depletion varies substantially
across different regions (21); depletion estimates include 8 ± 3 km3 per year from 2000
to 2012 in the transboundary Indo-Gangetic
Basin (104), ~4 km3 per year from 2003 to

1 March 2024

Aquifer

After afforestation

2010 in the California Central Valley (105), and
~6 km3 per year from 1945 to 2020 in the
North China Plain (106).
Groundwater with mean renewal times surpassing human timescales (i.e., 100 years) is
globally widespread and has been termed “nonrenewable” in some works (21, 107). Pumpage
of this old groundwater is especially common when wells tap deep aquifers (Fig. 4C).
An estimated ~20% of global gross irrigation water demand was derived from this old
groundwater in 2000 (4). Groundwater that
5 of 13

,

North China Plain, and the Indo-Gangetic Basin
(Fig. 2D) (32, 46, 99–101).
Groundwater depletion is often caused by
withdrawals for irrigation (5, 99). Global annual irrigation water use was estimated to be
960 ± 130 km3 per year from 2011 to 2018 (102).
Groundwater accounts for 45 to 50 and 60%
of irrigation in India and the United States,
respectively (5, 99). Global groundwater depletion was estimated to be 56 km3 per year
from 1960 to 2000 and 113 km3 per year from
2000 to 2009 according to WaterGAP (43) and

Before
afforestation

y

Fig. 4. Schematics of different types of groundwater withdrawal and
recharge. Groundwater withdrawal in an (A) unconfined aquifer, (B) confined
aquifer, and (C) deep confined aquifer. (D) Schematics of shale gas development
with hydraulic fracturing of a horizontal well (117, 118). (E and F) Schematics
of different geothermal systems: (E) two-well circulation system and (F)

8
10

25

25

Kuang et al., Science 383, eadf0630 (2024)

100 m

4
6

y g

20

20

2

RES EARCH | R E V I E W

Unconventional oil and gas production

Geothermal energy exploration

,

6 of 13

y

Geothermal energy can be used for either
geothermal power generation or direct utilization (129). Geothermal power generation has
increased significantly worldwide in recent
decades (23, 130). From 2010 to 2014, at least
2200 wells were drilled in 42 countries for
both direct utilization and power generation, a 6.2% increase compared with 2005 to
2009 (131). From 2015 to 2019, at least 2647
wells were drilled by 42 countries for both
direct utilization and power generation, with
an additional ~20,000 shallow heat pump wells
up to 100 m deep (132). Geothermal direct
utilization worldwide increased from 71 GWt
in 2014 to 108 GWt in 2019 (131, 132). The number of countries with direct utilization of geothermal energy increased from 28 in 1995 to
88 in 2019, including China, the United States,
Sweden, Germany, and Turkey (132).
The utilization of geothermal energy can be
realized by pumping hot groundwater out of
a hydrothermal system. Intensive withdrawal
of deep thermal groundwater is needed during
geothermal energy production. After geothermal

y g

1 March 2024

y

To enable unconventional oil and gas production from low-permeability source rocks such
as shale, coal, or tight sandstone formations,
hydraulic fracturing is widely used (Fig. 4D)
(117, 118). During hydraulic fracturing, highvolume, high-pressure fluids, chemical additives, and proppants are injected into the
low-permeability shale and tight rocks to
fracture and maintain open fractures in the
rocks (117, 119, 120). Horizontal drilling and
hydraulic fracturing allows for large quantities
of gas or oil to be extracted from these rocks
(Fig. 4D). From 2009 to 2017, 1.8 km3 of water
was used to fracture ~73,000 wells with a total
lateral length of 134,000 km in the United States
(22). Annual water use for hydraulic fracturing
for major plays in the United States has increased rapidly since 2009 (22, 121). Groundwater was the primary water source (~13,000 wells)
for hydraulic fracturing from 2010 to 2019 for the
Permian Basin in the United States (121). Future
increases in unconventional oil and gas production would require larger quantities of water for
hydraulic fracturing, which would also lead to
larger volumes of produced water (120).
Produced water is coproduced with oil and
gas over the life of a well and is mostly comprised of formation water (122, 123). The volume of produced water is estimated to vary
widely from 1720 to 50,000 m3 per well, with
1720 to 14,320 m3 per well for the major US
unconventional plays, 10,000 to 20,000 m3
per well during the first year of production
for China, and 10,000 to 50,000 m3 per well
for Canada (118, 120). Common methods for
produced water management include deep
underground injection, reuse for hydraulic
fracturing, and surface discharge. Part of the
produced water is reused for hydraulic fracturing and the percentage of water used for
the latter derived from recycling tends to increase over time (Fig. 4D) (22, 118, 122). For
the Marcellus Shale in the United States, 13%
of the produced water was recycled from 2000
to 2010; this percentage had increased to 56%
by 2011 (122). Deep underground injection is
the primary method for produced water management (Fig. 4D) (119, 122). Most of the produced water in the United States is managed
by deep underground injection (22, 122).
In semiarid regions and/or areas with high
groundwater consumption, the use of groundwater for drilling and hydraulic fracturing
may change local water availability or lead
to water stress (22, 124, 125). Globally, 20% of
shale deposits are located in regions with
groundwater depletion (125). In the United
States, nearly half of shale wells are distrib-

uted in water-scarce basins, in which unconventional wells increased water use (22, 124).
The withdrawal of groundwater for shale oil
or gas development may also lead to declining water levels and decrease the contribution of baseflow to streams (22, 121, 126). For
the Eagle Ford play and Permian basin in the
United States, a total of ~11,000 water wells
were drilled to meet water demands for hydraulic fracturing from 2009 to 2017 (22). Water
levels declined more considerably in confined
aquifers in the Eagle Ford play (6 to 18 m per
year over a ~5-year period) than in unconfined
aquifers (22). From 2009 to 2013, the use of
groundwater for hydraulic fracturing in the
Eagle Ford play resulted in an estimated local
drawdown of ~30 to 60 m in ~6% of the western play area (126).
Regional groundwater flow regimes may be
modified by unconventional oil or gas production. When groundwater is used for hydraulic
fracturing, large volumes of groundwater are
generally pumped out from shallow aquifers.
Shallow groundwater is injected into shale
layers during hydraulic fracturing and part
of it remains in the shale layer. The produced
water is then injected into deep underground
geologic formations. The withdrawal and injection of groundwater leads to the redistribution of groundwater at different depths.
Upward hydraulic gradients may be caused by
injection that could potentially result in upward fluid leakage into shallow aquifers (127).
Hydraulic fracturing also provides additional
pathways for groundwater flow. Additionally,
abandoned wells can provide potential conduits for produced water, and groundwater
may flow from one aquifer to another (121, 128).

g

Kuang et al., Science 383, eadf0630 (2024)

changes to global hydrological cycling induced
by increased groundwater withdrawals as well
as to assess the role of capture in groundwater
resources in different regions.

p

was recharged by precipitation that fell before
the Holocene (~12,000 years ago) is termed
fossil groundwater (7, 9, 108). A synthesis of
~6500 wells globally shows that fossil groundwater dominates storage at depths of ≥ ~250 m
(108). In the US High Plains Aquifer, the estimated depletion of fossil groundwater—much
of which was recharged during the past 13,000
years—was 330 km3 from the 1950s to 2007 (5).
Groundwater overexploitation in some aquifers leads to permanent depletion of water resources, sometimes referred to as groundwater
mining (109). In the United States, the proportion of newly drilled wells that are sufficiently deep (200 ± 100 m) to tap fossil aquifers
has grown in recent decades, although this deep
drilling is not necessarily associated with depletion (110).
Groundwater withdrawal is expected to increase under different future climate change
scenarios (21, 111, 112). By 2050, the estimated
global groundwater withdrawal rate is projected to be ~1250 ± 118 km3 per year, and
the depletion rate is estimated to be ~300 ±
50 km3 per year (21, 111). By 2099, the projected
global groundwater withdrawal is ~1600 ±
130 km3 per year, and the depletion is ~600 ±
85 km3 per year (92). Declining water levels
may result in wells drying up, meaning deeper
wells must be drilled to supply water (113).
If global groundwater levels were to decline
by only a few meters, millions of wells would
be at risk of running dry (114). Deep fresh
groundwater will become a strategic resource
in areas with high extraction and low recharge
rates (113). However, drilling deeper wells is an
unsustainable stopgap measure for addressing
groundwater depletion (113).
Groundwater pumped from aquifers participates in the global water cycle by discharging
to rivers and providing water for evapotranspiration (33, 99). In regions with groundwaterfed irrigation, increased groundwater use may
cause higher evapotranspiration (102), potentially leading to higher precipitation downwind and thus augmenting river discharge
(40). Irrigation in California's Central Valley
strengthens the regional water cycle by an
estimated ~15% increase in summer precipitation and a nontrivial increase in Colorado
River streamflow (115). Large groundwater
withdrawals can modify natural groundwater
flow systems. Groundwater discharges to
streams—which are vital to sustaining streamflow especially during dry seasons and
droughts—may decline or even stop flowing,
springs may dry up, and streamflow may decrease (2, 112). With greater groundwater
withdrawals, particularly in areas with dry
climates, it is likely that there will be more
ephemeral and losing rivers (streams with
water levels higher than those in adjacent
wells) that can seep into underlying aquifers
(100, 116). More studies are needed to evaluate

RES EARCH | R E V I E W

y

Coastal groundwater flow systems can be
modified by land reclamation and urbanization. During urbanization of coastal areas, land
reclamation from the sea and high-rise building
construction with deep foundations are two
common measures implemented to meet the
growing demand for land (26, 163). Land reclamation in coastal areas is practiced worldwide (164). Large-scale land reclamation can
change the regional groundwater regime by
increasing groundwater levels and altering
or slowing seaward groundwater discharge
(Fig. 5A) (164, 165). Locally, seaward groundwater discharge may increase as a result of
additional recharge in reclaimed land (163).
The saltwater-freshwater interface may also
move seaward after land reclamation. The
7 of 13

,

1 March 2024

Land reclamation and urbanization

y g

Kuang et al., Science 383, eadf0630 (2024)

Afforestation can potentially increase annual
evapotranspiration (Fig. 4I) while reducing
annual streamflow. Global tree cover increased
by 2.2 million km2 from 1982 to 2016 (153). Climate simulations suggest that tree plantations
can increase summer evapotranspiration by
more than 0.3 mm per day (154). Large-scale
tree restoration has been found to increase
terrestrial evapotranspiration by 1.2% and increase terrestrial precipitation by 0.7% due to
recycling of increased evaporation (155). Largescale tree plantations may lead to groundwater
declines where the enhanced evapotranspiration rates reduce recharge (25, 156). However,
divergent impacts of tree restoration on streamflow have been found (155, 156). Some rivers
experienced a decrease in streamflow by 6% as
a result of enhanced evapotranspiration whereas for other rivers, the greater evapotranspiration is counterbalanced by enhanced moisture
recycling (155).
Afforestation can cause declines in the water
table (Fig. 4I) as well as reductions in groundwater recharge, effective infiltration, soil moisture, and baseflow to streams. Afforestation
may lead to water table declines in arid and
semiarid areas of 0.5 to 3.0 m from 1952 to 2011
(25, 157). Compared with grasslands, groundwater recharge decreased by 3 to 7% for deeprooted forests (158). Much greater reductions
in recharge of 33 to >90% were found in forests
related to surrounding bare sandy soil in semiarid areas (159). Groundwater recharge is reduced as a result of increased transpiration
and interception (160). Increased tree cover
reduces soil moisture (161); for example, revegetation of a 16,000 km2 area in the Loess
Plateau in China decreased soil moisture by
~2.4 mm per year and reduced runoff by
~0.5 mm per year from 2000 to 2010 (162).
As forest plantations increase evapotranspiration (162), groundwater discharge to streams
(baseflow) tends to decline, especially in drylands or during dry seasons (25, 156).

y

Managed aquifer recharge (MAR) refers to intentionally recharging and storing water in
aquifers for subsequent recovery and various
beneficial uses (24, 140–142). As a means of
adapting to climate change and land use change
and realizing sustainable water management,
MAR has been implemented in many regions
globally (143), including Europe, Australia,
North and South America, Africa, and South

Afforestation

g

Managed aquifer recharge

Asia (30, 141, 144). MAR has also been implemented in mines to preserve aquifers, manage
surplus water, or adhere to licensing (145).
Effective MAR means that both water quantity and quality are managed effectively and
is a water management strategy that is becoming increasingly important (24). There
are ~1200 MAR sites in 62 countries (143).
MAR has increased by 5% per year since the
1960s (24). The average MAR volume increased
from 1.0 km3 per year in 1965 to 10 km3 per year
in 2015, representing ~1% of global groundwater
withdrawal in 2015 (24); however, it can be
important for alleviating regional water stress
(98). MAR is projected to exceed 10% of global
groundwater extraction as MAR techniques
become more advanced (24). MAR as a percentage of groundwater use varies significantly for different continents, from 0.4% in
Africa to 9% in the Middle East (24, 142).
MAR refers to a suite of methods that can be
used to maintain and enhance groundwater
systems under climate change and groundwater overexploitation (24, 146). MAR projects
have various goals, such as raising groundwater levels, increasing groundwater storage,
improving groundwater quality, preventing
saltwater intrusion, and meeting irrigation demand (24, 143, 146). Different water sources
have been used for MAR, including surface
water (rivers and lakes), stormwater, treated
wastewater, desalinated water, rainwater, and
fresh and brackish groundwater from other
aquifers (141, 144, 147). Surface waters such as
river and lake water are the dominant sources
of MAR (142, 144). There are many types of MAR
methods, including infiltration basins, percolation tanks, bank filtration, recharge wells, and
agricultural MARs (140, 141, 148) (Fig. 4, G and
H). Depleted aquifers provide additional subsurface reservoir storage capacity for MAR in
many regions, estimated at ~1000 km3 in the
United States, exceeding the surface reservoir
storage capacity (98).
MAR buffers against the adverse impacts of
climate extremes or change (142) and groundwater overexploitation (149). MAR can be used
to enhance resilience to drought by storing
excess surface waters and recycled water (146).
Additional water is recharged during flooding
or wet periods for subsequent abstraction during drought or dry periods (142, 146, 149). Depleted aquifers can be used to store water by
recharging groundwater with surface water
through MAR (147, 150). In semiarid areas where
groundwater is either overexploited or saline,
MAR has the potential to store excess runoff in
aquifers (140). Stormwater or floodwater can
drain into aquifers through infiltration basins,
wells, or sumps to reduce flood and drought
risks and then reuse this water for drinking or
irrigation purposes (140, 151). Guidelines and
regulations are vital to implementing MAR safely and sustainably (152).

p

energy utilization, the cold water is either reinjected deep underground or discharged directly to the surface (Fig. 4E). For a single-well
extraction system, hot groundwater was pumped
out and then released on the surface after use
(133). The ratio of reinjected mass to produced
fluid can vary from 5 to 100% (23). The distance between the production and reinjection
zones ranges from 0.1 to 6.0 km, with an average value of 1.3 km (23). Sources of water used
for reinjection include produced and surface
water such as waste-, rain-, stream-, lake-, and
groundwater (23). Enhanced geothermal systems use engineering strategies to enhance
geothermal energy production, in which hydraulic fracturing is utilized to improve rock permeability and the injected water is heated by the
rock (129, 134). The heated water is pumped
out by the production well and the cold water
is reinjected (Fig. 4F). Throughout many stages
of enhanced geothermal systems, a substantial
amount of water is introduced into the deep
subsurface, including water from well drilling, hydraulic fracturing, and fluid circulation,
in addition to water lost during the recovery
process (129).
As geothermal production and injection wells
are generally several kilometers deep, groundwater withdrawal and injection may cause
deep groundwater redistribution among different formations, perturbing local water cycling to some degree (135). Injection can also
lead to elevated pore pressure, which may reactivate faults and cause new thermal fractures
(134, 136, 137), thus providing new paths for
deep groundwater flow. At the Geysers geothermal field in the United States, injected
water can migrate >3 km below the injection
point due to a hydraulically conductive fracture network (136). At the Nesjavellir geothermal field in Iceland, the injected fluid flowed
through faults from the injection zone to the
northeast (138). Intensive withdrawal of deep
thermal groundwater can decrease the artesian
pressure in deep fractures and allow shallow
groundwater to flow into these deep fractures
(135). In some geothermal fields, excessive
withdrawal of geothermal fluid has resulted
in sea or lake water intrusion when a sea or
lake is nearby (23). Additional water may also
be injected into geothermal systems to sustain
production rates and maintain reservoir pressures (23, 139).

RES EARCH | R E V I E W

A

C
Gr

ou

nd

After reclamation

su
rf

5m

ac
e

100 m

500 m

500 m

Reclamation

Before reclamation

Spring

Sea

Sea

-SW e
FWerfac
t
in

Bedrock

B

D
100 m

Gr

ou

nd

High-rise buildings

su

rfa

After reclamation

10 m

ce

2 km

Spring

Sea

Before reclamation

-SW e
FWerfac
t
n
i

Land
reclamation

p

Reclamation

500 m

Sea
Sea
Bedrock

Deep foundations

Fresh groundwater lens

before land reclamation and urbanization (26, 173). (D) Coastal groundwater flow
system after land reclamation and urbanization with deep foundations (26, 173).
Curves with arrows show the groundwater flow paths. FW-SW interface,
freshwater-saline water interface.

g

Fig. 5. Groundwater flow system changes caused by land reclamation and
urbanization. (A) Land reclamation beside a coastal hillside with an unconfined
aquifer (164, 165). (B) Land reclamation beside an elongated oceanic island
with an unconfined aquifer (164–166). (C) Coastal groundwater flow system

y

A substantial share of groundwater depletion
has primarily resulted from irrigation and
an estimated ~11% of groundwater depletion
is linked to the international food trade (176).
The global groundwater depletion embedded in
international food trade increased from 18 km3
per year in 2000 to 26 km3 per year in 2010
(176). “Virtual water trade” refers to exchanges
of virtual water (amount of water embedded
in a commodity) between different regions or
nations through the exchange of physical com-

1 March 2024

8 of 13

,

International food trade

modities such as food (177). Enhanced international food trade is the main reason for increasing virtual water trade (177). At a national
scale, groundwater depletion in three major
US aquifers (Central Valley, High Plains, and
Mississippi Embayment) related to food trade
in the United States is linked primarily to
domestic food transfers (31 km3), accounting
for 90% of trade-related groundwater depletion, with the remaining 10% accounted for
by international exports (178). Groundwater
depletion linked to domestic trade grew from
26 km3 in 2002 to 35 km3 over a decade (2002
to 2012), with a similar rise in international
trade (2.7 to 3.7 km3) (179).
Embedded green water (soil moisture) and
blue water (surface- and groundwater) exports
are projected to more than triple from 2010 to
2100 from ~905 to 3200 km3 for green water
and ~56 to ~170 km3 for blue water (27). To meet
future crop demands, international food trade
is projected to nearly triple by 2050, including
virtual water transfers from water-abundant
regions to water-scarce regions (180).
Large groundwater volumes embedded in
international food trade redistribute groundwater demand globally. Virtual water trading
generates a virtual water flux that links water
resources used physically in the production area
to the consumption area (181). Unsustainable
irrigation embedded in virtual water trade globally demonstrates a redistribution of irrigation

y

Kuang et al., Science 383, eadf0630 (2024)

gradient (169, 172). Deep foundations limit
groundwater flow, raising the water table
and leading to upward groundwater flow in
the transition zone between the natural slope
and urbanized areas (26, 172, 173) (Fig. 5, C and
D). When permeable natural soils are replaced
by much less permeable deep foundations, the
hydraulic conductivity of the aquifer system is
reduced locally (26, 173).
During urbanization, native surface soils are
replaced by impervious surfaces, including
roads, foundations, and pavement. These impervious surfaces prevent infiltration, leading
to more surface runoff and less groundwater
recharge (171). However, other studies show
that urbanization leads to increased recharge
due to rain and runoff infiltration and losses
from water supply systems and sewer systems
(174, 175).

y g

response of a given groundwater system to
land reclamation can be a slow process, requiring several decades to reach a new equilibrium
(163). Land reclamation around an oceanic
island can change the groundwater system
on the entire island, raising the water table on
the island, shifting the water divide toward
the reclaimed side, and increasing submarine groundwater discharge on the other side
(164, 165). The saltwater-freshwater interface
at the reclamation side may also move seaward
after land reclamation (Fig. 5B) (164, 165). Lab
experiments and numerical modeling indicate that land reclamation can enlarge fresh
groundwater lenses by up to 85% in tropical
islands (166).
Groundwater systems can also be modified
by dewatering and underground structures.
Construction of high-rise buildings or underground infrastructures usually includes dewatering, deep excavations, and diaphragm
walls (167). In areas with shallow water tables,
dewatering requires pumping large quantities
of groundwater (168). Artificial recharge beyond
the excavation site can mitigate the impacts
of dewatering on the foundation stability of
neighboring buildings (167). Underground
structures below the water table impact the
groundwater flow system by acting as barriers
to flow and altering the groundwater budget
(168–171). The water table rises up-gradient of
the underground structures and falls down-

RES EARCH | R E V I E W

water demand, including groundwater demand
(182). From 2000 to 2015, an estimated 15% of
global unsustainable irrigation was virtually exported (182). Studies on changes in the global
water cycle should consider both the physical
and virtual water cycles (181).
Groundwater and sea level rise
Contribution of groundwater to sea level rise

Globally, groundwater resources face substantial threats in terms of both their quantity and
quality (21, 30). Excessive groundwater withdrawals continue to drive substantial groundwater depletion and the demand for groundwater
is projected to rise. Climate warming has led
to a diverse array of changes in groundwater
recharge across different regions of the world.
Other anthropogenic activities are reshaping
regional groundwater flow regimes, complicating groundwater storage dynamics, altering
groundwater discharge to streams, and redistributing embedded groundwater in the global food supply chain. Groundwater depletion
transfers fresh water from long-term storage
to the active water cycle, thereby contributing
to sea level rise. Moreover, pollution from anthropogenic sources and interactions between
surface water and groundwater have led to
deterioration in groundwater quality (195).
Groundwater-dependent ecosystems and geological environments have been severely affected
by water table changes or poor groundwater
quality (195).
Given these challenges, the sustainable use
of groundwater resources is a crucial global

C
d surfa

ce

Gro

Water table

und
surf
a

y g

Groun

ce

Water table

Sea

Sea

FW-SW
interface

FW-SW
interface
1m

Aquifer

100 m

B

Saline water

1m

100 m

,

Saline water

y

Aquifer

D
Withdrawal

y

Rising sea levels can cause water tables to rise
in unconfined coastal aquifers (29, 191, 192).
This rise can then cause groundwater discharge
to surface drainage networks and flooding from
below in low-lying coastal areas (Fig. 6, C and
D), which is referred to as groundwater inundation (29, 192). In California, areas flooded in
this manner are projected to expand ~50 to
130 m inland in response to a sea level rise of
1 m (192). In northern California's coastal
plains, a 1- to 2-m sea level rise may cause
widespread groundwater emergence (193). In
urban Honolulu, Hawaii, a 1-m sea level rise
may inundate an estimated 10% of a 1-km
wide coastal zone that is heavily urbanized
(29). Groundwater inundation alone may increase the area flooded by seawater inundation by a factor of two (29). In urbanized coastal

Sustainable use of groundwater resources

g

A

Groundwater inundation induced by sea level rise

areas, dense networks of buried and low-lying
infrastructure may lead to thinning and loss
of unsaturated subsurface space, which may
further magnify the risk of groundwater inundation (194). Groundwater inundation caused
by sea level rise enlarges the likelihood of groundwater discharge at the surface and accelerates
groundwater circulation within the water cycle
in coastal areas.

p

Groundwater withdrawal transfers fresh water
from long-term groundwater storage to the
active water cycle at the Earth's surface (7).
Much of the groundwater ultimately returns
to the ocean and causes sea level rise, which is
particularly important in coastal areas (Fig. 6,
A and B) (28, 183). Groundwater withdrawal
also causes land subsidence, and coastal land
subsidence contributes to relative sea level
rise (184). From 1900 to 2008, the estimated
contribution of cumulative global groundwater
depletion to sea level rise was 13 mm (103) and
ranged from 13 to 19 mm from 1948 to 2016
(185). The rate of global mean sea level rise
increased from 1.56 ± 0.33 mm per year from
1900 to 2018 to 3.35 ± 0.47 mm per year from
1993 to 2018 (183). Similar increasing rates
were reported in other studies from 1.7 ± 0.3 mm
per year since 1950 to 3.3 ± 0.4 mm per year
from 1993 to 2009 (186). Estimated contributions of past groundwater depletion to rates of
sea level rise range from 0.2 to 0.9 mm per year
(187, 188). Global groundwater depletion was
estimated to contribute 0.31 mm per year (2000
to 2009) to sea level rise based on WaterGAP
(43) and 0.40 ± 0.11 mm per year (2000 to 2008)
based on in situ measurements (103), accounting for ~10% of global mean sea level rise.

Te global mean sea level has been predicted
to rise by 0.5 to 1.4 m by 2100 (29, 186), with
the contribution of groundwater depletion to
sea level rise projected to increase in the future
(111). By 2050, groundwater depletion has been
projected to contribute 0.82 ± 0.13 mm per
year to sea level rise (111) and the percentages
of cumulative contribution of groundwater depletion to global sea level rise range from ~10
to ~27% (29, 111). Groundwater depletion and
sea level rise may lead to seawater intrusion into
coastal freshwater aquifers which is becoming
a critical environmental issue, with ~500 coastal
cities experiencing seawater intrusion crises
globally (189, 190). Seawater intrusion may become even more challenging to manage because
of climate change (190).

Groundwater
utilization
Gro
und
sur

Discharge
fac

e

FW-SW
interface
Aquifer

Gro
und

surf

After sea level rise
Before sea level rise
1m

Saline water

100 m

ace

Groundwater
inundation

FW-SW
interface
Aquifer

After sea level rise
Before sea level rise

1m

Saline water

100 m

Fig. 6. Schematics of groundwater withdrawal, sea level rise, and inundation. (A to B) Contribution of groundwater withdrawal to sea level rise (28, 187).
(C to D) Groundwater inundation caused by sea level rise (29). FW-SW interface, freshwater-saline water interface.
Kuang et al., Science 383, eadf0630 (2024)

1 March 2024

9 of 13

RES EARCH | R E V I E W

RE FE RENCES AND N OT ES

1.

3.

5.

6.

7.

9.

10.

13.

14.

15.

16.

Kuang et al., Science 383, eadf0630 (2024)

1 March 2024

44.

45.

46.

47.

48.

49.

50.

51.

52.

53.

54.

55.

56.

57.

58.

59.

60.

10 of 13

,

12.

43.

y

11.

42.

T. Meixner et al., Implications of projected climate
change for groundwater recharge in the western United
States. J. Hydrol. 534, 124–138 (2016). doi: 10.1016/
j.jhydrol.2015.12.027
N. S. Diffenbaugh, D. L. Swain, D. Touma, Anthropogenic
warming has increased drought risk in California. Proc. Natl.
Acad. Sci. U.S.A. 112, 3931–3936 (2015). doi: 10.1073/
pnas.1422385112; pmid: 25733875
P. Döll, H. M. Schmied, C. Schuh, F. T. Portmann, A. Eicker,
Global-scale assessment of groundwater depletion and
related groundwater abstractions: Combining hydrological
modeling with information from well observations and GRACE
satellites. Water Resour. Res. 50, 5698–5720 (2014).
doi: 10.1002/2014WR015595
H. M. Schmied et al., The global water resources and use
model WaterGAP v2.2d: Model description and evaluation.
Geosci. Model Dev. 14, 1037–1079 (2021). doi: 10.5194/
gmd-14-1037-2021
B. R. Scanlon et al., Effects of climate and irrigation on
GRACE-based estimates of water storage changes in major
US aquifers. Environ. Res. Lett. 16, 094009 (2021).
doi: 10.1088/1748-9326/ac16ff
D. J. MacAllister, G. Krishan, M. Basharat, D. Cuba,
A. M. MacDonald, A century of groundwater accumulation
in Pakistan and northwest India. Nat. Geosci. 15, 390–396
(2022). doi: 10.1038/s41561-022-00926-1
B. Dewandel, J.-M. Gandolfi, D. de Condappa, S. Ahmed,
An efficient methodology for estimating irrigation return
flow coefficients of irrigated crops at watershed and
seasonal scale. Hydrol. Processes 22, 1700–1712 (2008).
doi: 10.1002/hyp.6738
S. Zuidema et al., Interplay of changing irrigation
technologies and water reuse: Example from the upper Snake
River basin, Idaho, USA. Hydrol. Earth Syst. Sci. 24,
5231–5249 (2020). doi: 10.5194/hess-24-5231-2020
S. B. Ajjur, S. G. Al-Ghamdi, Quantifying the uncertainty in
future groundwater recharge simulations from regional
climate models. Hydrol. Processes 36, e14645 (2022).
doi: 10.1002/hyp.14645
A. Vincent, S. Violette, G. Aðalgeirsdóttir, Groundwater
in catchments headed by temperate glaciers: A review.
Earth Sci. Rev. 188, 59–76 (2019). doi: 10.1016/
j.earscirev.2018.10.017
W. W. Immerzeel, F. Pellicciotti, M. F. P. Bierkens, Rising river
flows throughout the twenty-first century in two Himalayan
glacierized watersheds. Nat. Geosci. 6, 742–745 (2013).
doi: 10.1038/ngeo1896
B. É. Ó Dochartaigh et al., Groundwater-glacier meltwater
interaction in proglacial aquifers. Hydrol. Earth Syst. Sci. 23,
4527–4539 (2019). doi: 10.5194/hess-23-4527-2019
S. A. Lone et al., Meltwaters dominate groundwater recharge
in cold arid desert of Upper Indus River Basin (UIRB),
western Himalayas. Sci. Total Environ. 786, 147514 (2021).
doi: 10.1016/j.scitotenv.2021.147514
C. Andermann et al., Impact of transient groundwater
storage on the discharge of Himalayan rivers. Nat. Geosci. 5,
127–132 (2012). doi: 10.1038/ngeo1356
D. Penna, M. Engel, G. Bertoldi, F. Comiti, Towards a
tracer-based conceptualization of meltwater dynamics
and streamflow response in a glacierized catchment.
Hydrol. Earth Syst. Sci. 21, 23–41 (2017). doi: 10.5194/
hess-21-23-2017
A. H. Laskar, S. K. Bhattacharya, D. K. Rao, R. A. Jani,
N. Gandhi, Seasonal variation in stable isotope compositions
of waters from a Himalayan river: Estimation of glacier
melt contribution. Hydrol. Processes 32, 3866–3880 (2018).
doi: 10.1002/hyp.13295
L. D. Somers et al., Groundwater buffers decreasing
glacier melt in an Andean watershed-but not forever.
Geophys. Res. Lett. 46, 13,016–13,026 (2019).
doi: 10.1029/2019GL084730
J. Pourrier, H. Jourde, C. Kinnard, S. Gascoin, S. Monnier, Glacier
meltwater flow paths and storage in a geomorphologically
complex glacial foreland: The case of the Tapado glacier, dry
Andes of Chile (30°S). J. Hydrol. 519, 1068–1083 (2014).
doi: 10.1016/j.jhydrol.2014.08.023
D. R. Rounce et al., Global glacier change in the 21st
century: Every increase in temperature matters. Science
379, 78–83 (2023). doi: 10.1126/science.abo1324;
pmid: 36603094
P. J. Blaen, D. M. Hannah, L. E. Brown, A. M. Milner, Water
temperature dynamics in High Arctic river basins. Hydrol.
Processes 27, 2958–2972 (2013). doi: 10.1002/hyp.9431

y g

8.

41.

y

4.

B. K. Biskaborn et al., Permafrost is warming at a global
scale. Nat. Commun. 10, 264 (2019). doi: 10.1038/
s41467-018-08240-4; pmid: 30651568
18. M. A. Walvoord, R. G. Striegl, Increased groundwater to
stream discharge from permafrost thawing in the Yukon
River basin: Potential impacts on lateral export of carbon
and nitrogen. Geophys. Res. Lett. 34, L12402 (2007).
doi: 10.1029/2007GL030216
19. L. D. Somers, J. M. McKenzie, J. M. MeKenzie, A review of
groundwater in high mountain environments. WIREs. Water 7,
e1475 (2020). doi: 10.1002/wat2.1475
20. D. Viviroli, M. Kummu, M. Meybeck, M. Kallio, Y. Wada,
Increasing dependence of lowland populations on mountain
water resources. Nat. Sustain. 3, 917–928 (2020).
doi: 10.1038/s41893-020-0559-9
21. M. F. P. Bierkens, Y. Wada, Non-renewable groundwater use
and groundwater depletion: A review. Environ. Res. Lett. 14,
063002 (2019). doi: 10.1088/1748-9326/ab1a5f
22. B. R. Scanlon, S. Ikonnikova, Q. Yang, R. C. Reedy, Will
water issues constrain oil and gas production in the United
States? Environ. Sci. Technol. 54, 3510–3519 (2020).
doi: 10.1021/acs.est.9b06390; pmid: 32062972
23. Z. Kamila, E. Kaya, S. J. Zarrouk, Reinjection in geothermal
fields: An updated worldwide review 2020. Geothermics 89,
101970 (2021). doi: 10.1016/j.geothermics.2020.101970
24. P. Dillon et al., Sixty years of global progress in managed
aquifer recharge. Hydrogeol. J. 27, 1–30 (2019).
doi: 10.1007/s10040-018-1841-z
25. B. A. Bryan et al., China’s response to a national land-system
sustainability emergency. Nature 559, 193–204 (2018).
doi: 10.1038/s41586-018-0280-2; pmid: 29995865
26. J. J. Jiao, X.-S. Wang, S. Nandy, Preliminary assessment
of the impacts of deep foundations and land reclamation
on groundwater flow in a coastal area in Hong Kong,
China. Hydrogeol. J. 14, 100–114 (2006). doi: 10.1007/
s10040-004-0393-6
27. N. T. Graham et al., Future changes in the trading of
virtual water. Nat. Commun. 11, 3632 (2020). doi: 10.1038/
s41467-020-17400-4; pmid: 32686671
28. Y. N. Pokhrel et al., Model estimates of sea-level change
due to anthropogenic impacts on terrestrial water storage.
Nat. Geosci. 5, 389–392 (2012). doi: 10.1038/ngeo1476
29. K. Rotzoll, C. H. Fletcher, Assessment of groundwater
inundation as a consequence of sea-level rise. Nat. Clim.
Chang. 3, 477–481 (2013). doi: 10.1038/nclimate1725
30. U. Lall, L. Josset, T. Russo, A snapshot of the world’s
groundwater challenges. Annu. Rev. Environ. Resour. 45,
171–194 (2020). doi: 10.1146/annurev-environ-102017-025800
31. C. Moeck et al., A global-scale dataset of direct natural
groundwater recharge rates: A review of variables, processes
and relationships. Sci. Total Environ. 717, 137042 (2020).
doi: 10.1016/j.scitotenv.2020.137042; pmid: 32062252
32. T. A. Russo, U. Lall, Depletion and response of deep
groundwater to climate-induced pumping variability.
Nat. Geosci. 10, 105–108 (2017). doi: 10.1038/ngeo2883
33. Y. Wada et al., Global depletion of groundwater resources.
Geophys. Res. Lett. 37, L20402 (2010). doi: 10.1029/
2010GL044571
34. N. Hanasaki, S. Yoshikawa, Y. Pokhrel, S. Kanae, A global
hydrological simulation to specify the sources of water used
by humans. Hydrol. Earth Syst. Sci. 22, 789–817 (2018).
doi: 10.5194/hess-22-789-2018
35. V. Mohan, A. W. Western, Y. Wei, M. Saft, Predicting
groundwater recharge for varying land cover and climate
conditions – a global meta-study. Hydrol. Earth Syst. Sci. 22,
2689–2703 (2018). doi: 10.5194/hess-22-2689-2018
36. B. W. Abbott et al., Human domination of the global
water cycle absent from depictions and perceptions.
Nat. Geosci. 12, 533–540 (2019). doi: 10.1038/
s41561-019-0374-y
37. I. E. M. de Graaf, K. Stahl, A model comparison assessing
the importance of lateral groundwater flows at the global
scale. Environ. Res. Lett. 17, 044020 (2022). doi: 10.1088/
1748-9326/ac50d2
38. R. G. Taylor et al., Evidence of the dependence of
groundwater resources on extreme rainfall in East Africa.
Nat. Clim. Chang. 3, 374–378 (2013). doi: 10.1038/nclimate1731
39. F. D. Tillman, S. Gangopadhyay, T. Pruitt, Changes in
groundwater recharge under projected climate in the upper
Colorado River basin. Geophys. Res. Lett. 43, 6968–6974
(2016). doi: 10.1002/2016GL069714
40. A. C. Amanambu et al., Groundwater system and climate
change: Present status and future considerations. J. Hydrol.
589, 125163 (2020). doi: 10.1016/j.jhydrol.2020.125163

g

2.

W. Aeschbach-Hertig, T. Gleeson, Regional strategies
for the accelerating global problem of groundwater depletion.
Nat. Geosci. 5, 853–861 (2012). doi: 10.1038/ngeo1617
T. Gleeson et al., Groundwater sustainability strategies.
Nat. Geosci. 3, 378–379 (2010). doi: 10.1038/ngeo881
S. Siebert et al., Groundwater use for irrigation-a global
inventory. Hydrol. Earth Syst. Sci. 14, 1863–1880 (2010).
doi: 10.5194/hess-14-1863-2010
Y. Wada, L. P. H. van Beek, M. F. P. Bierkens, Nonsustainable
groundwater sustaining irrigation: A global assessment.
Water Resour. Res. 48, W00L06 (2012). doi: 10.1029/
2011WR010562
B. R. Scanlon et al., Groundwater depletion and sustainability
of irrigation in the US High Plains and Central Valley. Proc.
Natl. Acad. Sci. U.S.A. 109, 9320–9325 (2012). doi: 10.1073/
pnas.1200311109; pmid: 22645352
M. Rodell et al., Emerging trends in global freshwater
availability. Nature 557, 651–659 (2018). doi: 10.1038/
s41586-018-0123-1; pmid: 29769728
R. G. Taylor et al., Ground water and climate change. Nat.
Clim. Chang. 3, 322–329 (2013). doi: 10.1038/nclimate1744
T. Gleeson, et al., Illuminating water cycle modifications and
Earth system resilience in the Anthropocene. Water Resour. Res.
56, e2019WR024957 (2020). doi: 10.1038/nclimate1744
T. Oki, S. Kanae, Global hydrological cycles and world water
resources. Science 313, 1068–1072 (2006). doi: 10.1126/
science.1128845; pmid: 16931749
A. M. DeAngelis, X. Qu, M. D. Zelinka, A. Hall, An
observational radiative constraint on hydrologic cycle
intensification. Nature 528, 249–253 (2015). doi: 10.1038/
nature15770; pmid: 26659186
M. O. Cuthbert et al., Global patterns and dynamics of
climate-groundwater interactions. Nat. Clim. Chang. 9,
137–141 (2019). doi: 10.1038/s41558-018-0386-4
M. O. Cuthbert et al., Observed controls on resilience of
groundwater to climate variability in sub-Saharan
Africa. Nature 572, 230–234 (2019). doi: 10.1038/
s41586-019-1441-7; pmid: 31391559
A. F. Lutz et al., South Asian agriculture increasingly
dependent on meltwater and groundwater. Nat. Clim. Chang.
12, 566–573 (2022). doi: 10.1038/s41558-022-01355-z
R. Reinecke et al., Uncertainty of simulated groundwater
recharge at different global warming levels: A global-scale
multi-model ensemble study. Hydrol. Earth Syst. Sci. 25,
787–810 (2021). doi: 10.5194/hess-25-787-2021
H. D. Pritchard, Asia’s shrinking glaciers protect large
populations from drought stress. Nature 569, 649–654
(2019). doi: 10.1038/s41586-019-1240-1; pmid: 31142854
R. Hugonnet et al., Accelerated global glacier mass loss in the
early twenty-first century. Nature 592, 726–731 (2021).
doi: 10.1038/s41586-021-03436-z; pmid: 33911269

17.

p

concern. In planning for a more sustainable
future for water in an increasingly interconnected world, groundwater resources should
be considered from both regional and global
perspectives, especially in the case of large,
transboundary groundwater systems (196).
As global changes continue to affect these resources, it is imperative to manage groundwater
and surface water as a single resource (98). Additionally, ensuring food and water security
and maintaining ecosystem health must be
addressed concurrently (197). There has been
a growing global trend toward incorporating
sustainability into groundwater laws, regulations, and policies (198, 199). Various management strategies, including forest and wetland
conservation, desalination, wastewater recycling (98), managed aquifer recharge, water
diversion projects, and green infrastructure
development (200) are already being employed
to bolster groundwater resilience, and will be
critical to combat the growing problem of
groundwater depletion globally.

RES EARCH | R E V I E W

61.

62.

63.

64.

65.

66.

67.

69.

70.

73.

74.

76.

79.

80.

81.

Kuang et al., Science 383, eadf0630 (2024)

87.

88.

89.

90.

91.
92.

93.

94.
95.
96.

97.
98.

99.

100.

101.

102.

103.

104.

105.

106.

107.

1 March 2024

11 of 13

,

78.

86.

y

77.

85.

y g

75.

84.

108. S. Jasechko et al., Global aquifers dominated by fossil
groundwaters but wells vulnerable to modern contamination.
Nat. Geosci. 10, 425–429 (2017). doi: 10.1038/ngeo2943
109. P. D’Odorico et al., The global food-energy-water nexus. Rev.
Geophys. 56, 456–531 (2018). doi: 10.1029/2017RG000591
110. M. GebreEgziabher, S. Jasechko, D. Perrone, M. GebreEgziabher,
Widespread and increased drilling of wells into fossil aquifers
in the USA. Nat. Commun. 13, 2129 (2022). doi: 10.1038/
s41467-022-29678-7; pmid: 35440593
111. Y. Wada et al., Past and future contribution of global
groundwater depletion to sea-level rise. Geophys. Res. Lett.
39, L09402 (2012). doi: 10.1029/2012GL051230
112. I. E. M. de Graaf, T. Gleeson, L. P. H. Rens van Beek,
E. H. Sutanudjaja, M. F. P. Bierkens, Environmental flow limits
to global groundwater pumping. Nature 574, 90–94 (2019).
doi: 10.1038/s41586-019-1594-4; pmid: 31578485
113. D. Perrone, S. Jasechko, Deeper well drilling an unsustainable
stopgap to groundwater depletion. Nat. Sustain. 2, 773–782
(2019). doi: 10.1038/s41893-019-0325-z
114. S. Jasechko, D. Perrone, Global groundwater wells at risk of
running dry. Science 372, 418–421 (2021). doi: 10.1126/
science.abc2755; pmid: 33888642
115. M.-H. Lo, J. S. Famiglietti, Irrigation in California’s Central
Valley strengthens the southwestern U.S. water cycle.
Geophys. Res. Lett. 40, 301–306 (2013). doi: 10.1002/
grl.50108
116. S. Jasechko, H. Seybold, D. Perrone, Y. Fan, J. W. Kirchner,
Widespread potential loss of streamflow into underlying
aquifers across the USA. Nature 591, 391–395 (2021).
doi: 10.1038/s41586-021-03311-x; pmid: 33731949
117. J. C. McIntosh et al., A critical review of state-of-the-art
and emerging approaches to identify fracking-derived
gases and associated contaminants in aquifers. Environ. Sci.
Technol. 53, 1063–1077 (2019). doi: 10.1021/acs.
est.8b05807; pmid: 30585065
118. C. Zhong et al., Comparison of the hydraulic fracturing
water cycle in China and North America. Environ. Sci.
Technol. 55, 7167–7185 (2021). doi: 10.1021/acs.est.0c06119;
pmid: 33970611
119. R. D. Vidic, S. L. Brantley, J. M. Vandenbossche,
D. Yoxtheimer, J. D. Abad, Impact of shale gas development
on regional water quality. Science 340, 1235009 (2013).
doi: 10.1126/science.1235009; pmid: 23687049
120. A. J. Kondash, N. E. Lauer, A. Vengosh, The intensification
of the water footprint of hydraulic fracturing. Sci. Adv. 4,
eaar5982 (2018). doi: 10.1126/sciadv.aar5982; pmid: 30116777
121. B. R. Scanlon, R. C. Reedy, B. D. Wolaver, Assessing
cumulative water impacts from shale oil and gas
production: Permian Basin case study. Sci. Total Environ. 811,
152306 (2022). doi: 10.1016/j.scitotenv.2021.152306;
pmid: 34906580
122. R. B. Jackson et al., The environmental costs and benefits
of fracking. Annu. Rev. Environ. Resour. 39, 327–362 (2014).
doi: 10.1146/annurev-environ-031113-144051
123. A. Kondash, A. Vengosh, Water footprint of hydraulic
fracturing. Environ. Sci. Technol. Lett. 2, 276–280 (2015).
doi: 10.1021/acs.estlett.5b00211
124. A. Vengosh, R. B. Jackson, N. Warner, T. H. Darrah,
A. Kondash, A critical review of the risks to water
resources from unconventional shale gas development and
hydraulic fracturing in the United States. Environ. Sci.
Technol. 48, 8334–8348 (2014). doi: 10.1021/es405118y;
pmid: 24606408
125. L. Rosa, M. C. Rulli, K. F. Davis, P. D’Odorico, The waterenergy nexus of hydraulic fracturing: A global hydrologic
analysis for shale oil and gas extraction. Earths Futur. 6,
745–756 (2018). doi: 10.1002/2018EF000809
126. B. R. Scanlon, R. C. Reedy, J. P. Nicot, Will water scarcity in
semiarid regions limit hydraulic fracturing of shale plays?
Environ. Res. Lett. 9, 124011 (2014). doi: 10.1088/1748-9326/
9/12/124011
127. K. Jellicoe, J. C. McIntosh, G. Ferguson, Changes in
deep groundwater flow patterns related to oil and gas
activities. Groundwater 60, 47–63 (2022). doi: 10.1111/
gwat.13136; pmid: 34519028
128. C. Perra, J. C. McIntosh, T. Watson, G. Ferguson,
Commingled fluids in abandoned boreholes: Proximity
analysis of a hidden liability. Groundwater 60, 210–224
(2022). doi: 10.1111/gwat.13140; pmid: 34617284
129. W. G. P. Kumari, P. G. Ranjith, Sustainable development of
enhanced geothermal systems based on geotechnical
research-A review. Earth Sci. Rev. 199, 102955 (2019).
doi: 10.1016/j.earscirev.2019.102955

y

72.

83.

S. E. Chadburn et al., An observation-based constraint on
permafrost loss as a function of global warming. Nat. Clim.
Chang. 7, 340–343 (2017). doi: 10.1038/nclimate3262
G. Cheng, H. Jin, Permafrost and groundwater on the
Qinghai-Tibet Plateau and in northeast China. Hydrogeol. J.
21, 5–23 (2013). doi: 10.1007/s10040-012-0927-2
M.-K. Woo, D. L. Kane, S. K. Carey, D. Yang, Progress in
permafrost hydrology in the new millennium. Permafr.
Periglac. Process. 19, 237–254 (2008). doi: 10.1002/ppp.613
L. C. Smith, T. M. Pavelsky, G. M. MacDonald,
A. I. Shiklomanov, R. B. Lammers, Rising minimum daily flows
in northern Eurasian rivers: A growing influence of
groundwater in the high-latitude hydrologic cycle. J. Geophys.
Res. 112, G04S47 (2007). doi: 10.1029/2006JG000327
X.-Y. Jin et al., Impacts of climate-induced permafrost
degradation on vegetation: A review. Adv. Clim. Chang. Res.
12, 29–47 (2021). doi: 10.1016/j.accre.2020.07.002
P. Wang et al., Potential role of permafrost thaw on
increasing Siberian river discharge. Environ. Res. Lett. 16,
034046 (2021). doi: 10.1088/1748-9326/abe326
S. G. Evans, S. Ge, S. Liang, Analysis of groundwater flow in
mountainous, headwater catchments with permafrost. Water
Resour. Res. 51, 9564–9576 (2015). doi: 10.1002/
2015WR017732
T. Gleeson, Y. Wada, M. F. P. Bierkens, L. P. H. van Beek,
Water balance of global aquifers revealed by groundwater
footprint. Nature 488, 197–200 (2012). doi: 10.1038/
nature11295; pmid: 22874965
A. S. Richey et al., Quantifying renewable groundwater stress
with GRACE. Water Resour. Res. 51, 5217–5238 (2015).
doi: 10.1002/2015WR017349; pmid: 26900185
J. Margat, J. van der Gun, Groundwater Around the World
(CRC Press, 2013). doi: 10.1201/b13977
Y. Wada, M. F. P. Bierkens, Sustainability of global water use:
Past reconstruction and future projections. Environ. Res. Lett.
9, 104003 (2014). doi: 10.1088/1748-9326/9/10/104003
I. E. M. de Graaf et al., A global-scale two-layer transient
groundwater model: Development and application to
groundwater depletion. Adv. Water Resour. 102, 53–67
(2017). doi: 10.1016/j.advwatres.2017.01.011
UNESCO. Groundwater making the invisible visible (The
United Nations World Water Development Report, 2022)
J. S. Famiglietti, The global groundwater crisis. Nat. Clim.
Chang. 4, 945–948 (2014). doi: 10.1038/nclimate2425
L. E. Condon, R. M. Maxwell, Simulating the sensitivity of
evapotranspiration and streamflow to large-scale
groundwater depletion. Sci. Adv. 5, eaav4574 (2019).
doi: 10.1126/sciadv.aav4574; pmid: 31223647
L. Konikow, Overestimated water storage. Nat. Geosci. 6, 3
(2013). doi: 10.1038/ngeo1659
B. R. Scanlon et al., Global water resources and the role of
groundwater in a resilient water future. Nat. Rev. Earth
Environ. 4, 87–101 (2023). doi: 10.1038/s43017-022-00378-6
M. Rodell, I. Velicogna, J. S. Famiglietti, Satellite-based
estimates of groundwater depletion in India. Nature 460,
999–1002 (2009). doi: 10.1038/nature08238;
pmid: 19675570
M. R. Khan et al., Megacity pumping and preferential flow
threaten groundwater quality. Nat. Commun. 7, 12833 (2016).
doi: 10.1038/ncomms12833; pmid: 27673729
WRI (World Resources Institute), Aqueduct: Using cuttingedge data to identify and evaluate water risks around the
world. www.wri.org/aqueduct.
K. Zhang, X. Li, D. Zheng, L. Zhang, G. Zhu, Estimation of
global irrigation water use by the integration of multiple
satellite observations. Water Resour. Res. 58,
e2021WR030031 (2022). doi: 10.1038/ncomms12833;
pmid: 27673729
L. F. Konikow, Contribution of global groundwater depletion
since 1900 to sea-level rise. Geophys. Res. Lett. 38, L17401
(2011). doi: 10.1029/2011GL048604
A. M. MacDonald et al., Groundwater quality and depletion in
the Indo-Gangetic Basin mapped from in situ observations.
Nat. Geosci. 9, 762–766 (2016). doi: 10.1038/ngeo2791
J. S. Famiglietti et al., Satellites measure recent rates of
groundwater depletion in California’s Central Valley. Geophys.
Res. Lett. 38, L03403 (2011). doi: 10.1029/2010GL046442
M. Lancia et al., The China groundwater crisis: A mechanistic
analysis with implications for global sustainability. Sustain.
Horiz. 4, 100042 (2022). doi: 10.1016/j.horiz.2022.100042
G. Ferguson, M. O. Cuthbert, K. Befus, T. Gleeson,
J. C. McIntosh, Rethinking groundwater age. Nat. Geosci. 13,
592–594 (2020). doi: 10.1038/s41561-020-0629-7

g

71.

82.

p

68.

D. B. Jones, S. Harrison, K. Anderson, W. B. Whalley, Rock
glaciers and mountain hydrology: A review. Earth Sci. Rev.
193, 66–90 (2019). doi: 10.1016/j.earscirev.2019.04.001
S. T. Geiger, J. M. Daniels, S. N. Miller, J. W. Nicholas,
Influence of rock glaciers on stream hydrology in the
La Sal Mountains, Utah. Arct. Antarct. Alp. Res. 46, 645–658
(2014). doi: 10.1657/1938-4246-46.3.645
J. S. Harrington, A. Mozil, M. Hayashi, L. R. Bentley,
Groundwater flow and storage processes in an inactive rock
glacier. Hydrol. Processes 32, 3070–3088 (2018).
doi: 10.1002/hyp.13248
M. Hayashi, Alpine hydrogeology: The critical role of
groundwater in sourcing the headwaters of the world.
Ground Water 58, 498–510 (2020). doi: 10.1111/gwat.12965;
pmid: 31762021
M. Cochand, P. Christe, P. Ornstein, D. Hunkeler,
Groundwater storage in high alpine catchments and its
contribution to streamflow. Water Resour. Res. 55,
2613–2630 (2019). doi: 10.1029/2018WR022989
J. Obu, How much of the Earth's surface is underlain by
permafrost? J. Geophys. Res.-Earth Surf. 126,
e2021JF006123 (2021). doi: 10.1029/2018WR022989
Y. Ran et al., New high-resolution estimates of the permafrost
thermal state and hydrothermal conditions over the
Northern Hemisphere. Earth Syst. Sci. Data 14, 865–884
(2022). doi: 10.5194/essd-14-865-2022
S. L. Smith, H. B. O’Neill, K. Isaksen, J. Noetzli, V. E. Romanovsky,
The changing thermal state of permafrost. Nat. Rev. Earth
Environ. 3, 10–23 (2022). doi: 10.1038/s43017-021-00240-1
D. Zhao et al., Changing climate and the permafrost
environment on the Qinghai-Tibet (Xizang) plateau. Permafr.
Periglac. Process. 31, 396–405 (2020). doi: 10.1002/
ppp.2056
Y. Ran, X. Li, G. Cheng, Climate warming over the past
half century has led to thermal degradation of permafrost on
the Qinghai–Tibet Plateau. Cryosphere 12, 595–608 (2018).
doi: 10.5194/tc-12-595-2018
M. Rogger et al., Impact of mountain permafrost on flow path
and runoff response in a high alpine catchment. Water
Resour. Res. 53, 1288–1308 (2017). doi: 10.1002/
2016WR019341
M. A. Walvoord, C. I. Voss, B. A. Ebel, B. J. Minsley,
Development of perennial thaw zones in boreal hillslopes
enhances potential mobilization of permafrost carbon.
Environ. Res. Lett. 14, 015003 (2019). doi: 10.1088/
1748-9326/aaf0cc
V. F. Bense, G. Ferguson, H. Kooi, Evolution of shallow
groundwater flow systems in areas of degrading permafrost.
Geophys. Res. Lett. 36, L22401 (2009). doi: 10.1029/
2009GL039225
W. Yi et al., Increasing annual streamflow and groundwater
storage in response to climate warming in the Yangtze River
source region. Environ. Res. Lett. 16, 084011 (2021).
doi: 10.1088/1748-9326/ac0f27
S. Ge, J. McKenzie, C. Voss, Q. Wu, Exchange of groundwater
and surface-water mediated by permafrost response to
seasonal and long term air temperature variation.
Geophys. Res. Lett. 38, L14402 (2011). doi: 10.1029/
2011GL047911
S. G. Evans, B. Yokeley, C. Stephens, B. Brewer, Potential
mechanistic causes of increased baseflow across northern
Eurasia catchments underlain by permafrost. Hydrol.
Processes 34, 2676–2690 (2020). doi: 10.1002/hyp.13759
S. N. Wright et al., Thaw-induced impacts on land and
water in discontinuous permafrost: A review of the
Taiga Plains and Taiga Shield, northwestern Canada.
Earth Sci. Rev. 232, 104104 (2022). doi: 10.1016/
j.earscirev.2022.104104
B. L. Kurylyk, K. T. B. MacQuarrie, J. M. McKenzie, Climate
change impacts on groundwater and soil temperatures in
cold and temperate regions: Implications, mathematical
theory, and emerging simulation tools. Earth Sci. Rev. 138,
313–334 (2014). doi: 10.1016/j.earscirev.2014.06.006
M. A. Walvoord, B. L. Kurylyk, Hydrologic impacts of thawing
permafrost: A review. Vadose Zone J. 15, 1–20 (2016).
doi: 10.2136/vzj2016.01.0010
J. M. Scheidegger, V. F. Bense, Impacts of glacially recharged
groundwater flow systems on talik evolution. J. Geophys.
Res. Earth Surf. 119, 758–778 (2014). doi: 10.1002/
2013JF002894
J. M. McKenzie, C. I. Voss, Permafrost thaw in a nested
groundwater-flow system. Hydrogeol. J. 21, 299–316 (2013).
doi: 10.1007/s10040-012-0942-3

RES EARCH | R E V I E W

174.

175.

176.

177.

178.

179.

180.

181.

182.

183.

184.

185.

186.

188.

191.

192.

193.

194.

12 of 13

,

190.

y

189.

y g

187.

y

1 March 2024

173.

Hydrol. Process. 22, 1857–1865 (2008). doi: 10.1002/
hyp.6768
J. J. Jiao, C.-M. Leung, G. Ding, Changes to the groundwater
system, from 1888 to present, in a highly-urbanized coastal
area in Hong Kong, China. Hydrogeol. J. 16, 1527–1539
(2008). doi: 10.1007/s10040-008-0332-z
I. Tubau, E. Vázquez-Suñé, J. Carrera, C. Valhondo, R. Criollo,
Quantification of groundwater recharge in urban
environments. Sci. Total Environ. 592, 391–402 (2017).
doi: 10.1016/j.scitotenv.2017.03.118; pmid: 28324856
F. La Vigna, Review: Urban groundwater issues and resource
management, and their roles in the resilience of cities.
Hydrogeol. J. 30, 1657–1683 (2022). doi: 10.1007/
s10040-022-02517-1
C. Dalin, Y. Wada, T. Kastner, M. J. Puma, Groundwater
depletion embedded in international food trade. Nature
543, 700–704 (2017). doi: 10.1038/nature21403;
pmid: 28358074
C. Dalin, M. Taniguchi, T. R. Green, Unsustainable
groundwater use for global food production and related
international trade. Glob. Sustain. 2, e12 (2019).
doi: 10.1017/sus.2019.7
L. Marston, M. Konar, X. Cai, T. J. Troy, Virtual
groundwater transfers from overexploited aquifers in
the United States. Proc. Natl. Acad. Sci. U.S.A. 112,
8561–8566 (2015). doi: 10.1073/pnas.1500457112;
pmid: 26124137
S. Gumidyala et al., Groundwater depletion embedded in
domestic transfers and international exports of the United
States. Water Resour. Res. 56, e2019WR024986 (2020).
doi: 10.1038/s41893-019-0287-1
A. V. Pastor et al., The global nexus of food-trade-water
sustaining environmental flows by 2050. Nat. Sustain. 2,
499–507 (2019). doi: 10.1038/s41893-019-0287-1
P. D’Odorico et al., Global virtual water trade and the
hydrological cycle: Patterns, drivers, and socio-environmental
impacts. Environ. Res. Lett. 14, 053001 (2019). doi: 10.1088/
1748-9326/ab05f4
L. Rosa, D. D. Chiarelli, C. Tu, M. C. Rulli, P. D’Odorico,
Global unsustainable virtual water flows in agriculture
trade. Environ. Res. Lett. 14, 114001 (2019). doi: 10.1088/
1748-9326/ab4bfc
T. Frederikse et al., The causes of sea-level rise since
1900. Nature 584, 393–397 (2020). doi: 10.1038/
s41586-020-2591-3; pmid: 32814886
M. Shirzaei et al., Measuring, modelling and projecting
coastal land subsidence. Nat. Rev. Earth Environ. 2, 40–58
(2021). doi: 10.1038/s43017-020-00115-x
D. Cáceres et al., Assessing global water mass transfers from
continents to oceans over the period 1948-2016. Hydrol.
Earth Syst. Sci. 24, 4831–4851 (2020). doi: 10.5194/
hess-24-4831-2020
R. J. Nicholls, A. Cazenave, Sea-level rise and its impact on
coastal zones. Science 328, 1517–1520 (2010). doi: 10.1126/
science.1185782; pmid: 20558707
Y. Wada et al., Fate of water pumped from underground and
contributions to sea-level rise. Nat. Clim. Chang. 6, 777–780
(2016). doi: 10.1038/nclimate3001
Y. Wada et al., Recent changes in land water storage
and its contribution to sea level variations. Surv. Geophys.
38, 131–152 (2017). doi: 10.1007/s10712-016-9399-6;
pmid: 32269399
H. Ketabchi, D. Mahmoodzadeh, B. Ataie-Ashtiani,
C. T. Simmons, Sea-level rise impacts on seawater intrusion
in costal aquifers: Review and integration. J. Hydrol. 535,
235–255 (2016). doi: 10.1016/j.jhydrol.2016.01.083
T. Cao, D. Han, X. Song, Past, present, and future of
global seawater intrusion research: A bibliometric
analysis. J. Hydrol. 603, 126844 (2021). doi: 10.1016/
j.jhydrol.2021.126844
G. Ferguson, T. Gleeson, Vulnerability of coastal aquifers to
groundwater use and climate change. Nat. Clim. Chang. 2,
342–345 (2012). doi: 10.1038/nclimate1413
K. M. Befus, P. L. Barnard, D. J. Hoover, J. A. F. Hart,
C. I. Voss, Increasing threat of coastal groundwater
hazards from sea-level rise in California. Nat. Clim. Chang.
10, 946–952 (2020). doi: 10.1038/s41558-020-0874-1
D. J. Hoover, K. O. Odigie, P. W. Swarzenski, P. Barnard,
Sea-level rise and coastal groundwater inundation and
shoaling at select sites in California, USA. J. Hydrol. Reg. Stud.
11, 234–249 (2017). doi: 10.1016/j.ejrh.2015.12.055
S. Habel, C. H. Fletcher, K. Rotzoll, A. I. El-Kadi, Development
of a model to simulate groundwater inundation induced by

g

Kuang et al., Science 383, eadf0630 (2024)

151. X. He et al., Climate-informed hydrologic modeling and
policy typology to guide managed aquifer recharge. Sci. Adv.
7, eabe6025 (2021). doi: 10.1126/sciadv.abe6025;
pmid: 33883132
152. E. F. Escalante, J. D. H. Casas, J. S. S. Sauto, R. C. Gil,
Monitored and intentional recharge (MIR): A model for
managed aquifer recharge (MAR) guideline and regulation
formulation. Water 14, 3405 (2022). doi: 10.3390/w14213405
153. X.-P. Song et al., Global land change from 1982 to 2016.
Nature 560, 639–643 (2018). doi: 10.1038/s41586-018-0411-9;
pmid: 30089903
154. R. B. Jackson et al., Trading water for carbon with biological
carbon sequestration. Science 310, 1944–1947 (2005).
doi: 10.1126/science.1119282; pmid: 16373572
155. A. J. H. van Dijke et al., Shifts in regional water availability
due to global tree restoration. Nat. Geosci. 15, 363–368
(2022). doi: 10.1038/s41561-022-00935-0
156. J. Jones et al., Forest restoration and hydrology.
For. Ecol. Manage. 520, 120342 (2022). doi: 10.1016/
j.foreco.2022.120342
157. C. Lu, T. Zhao, X. Shi, S. Cao, Ecological restoration by
afforestation may increase groundwater depth and create
potentially large ecological and water opportunity costs in
arid and semiarid China. J. Clean. Prod. 176, 1213–1222
(2018). doi: 10.1016/j.jclepro.2016.03.046
158. G. Chen et al., Evaluating potential groundwater recharge
in the unsteady state for deep-rooted afforestation in
deep loess deposits. Sci. Total Environ. 858, 159837
(2023). doi: 10.1016/j.scitotenv.2022.159837;
pmid: 36411672
159. T. Huang, Z. Pang, S. Yang, L. Yin, Impact of afforestation on
atmospheric recharge to groundwater in a semiarid area.
J. Geophys. Res.-Atmos. 125, e2019JD032185 (2020).
doi: 10.1016/j.scitotenv.2021.146336; pmid: 34030299
160. Z. Zhang et al., Salix psammophila afforestations can cause
a decline of the water table, prevent groundwater recharge
and reduce effective infiltration. Sci. Total Environ. 780,
146336 (2021). doi: 10.1016/j.scitotenv.2021.146336;
pmid: 34030299
161. Y. Li et al., Divergent hydrological response to large-scale
afforestation and vegetation greening in China. Sci. Adv. 4,
eaar4182 (2018). doi: 10.1126/sciadv.aar4182; pmid: 29750196
162. X. Feng et al., Revegetation in China’s Loess Plateau is
approaching sustainable water resource limits. Nat. Clim.
Chang. 6, 1019–1022 (2016). doi: 10.1038/nclimate3092
163. L. Hu, J. J. Jiao, Modeling the influences of land reclamation
on groundwater systems: A case study in Shekou
peninsula, Shenzhen, China. Eng. Geol. 114, 144–153 (2010).
doi: 10.1016/j.enggeo.2010.04.011
164. L. Hu, J. J. Jiao, H. Guo, Analytical studies on transient
groundwater flow induced by land reclamation. Water Resour.
Res. 44, W11427 (2008). doi: 10.1029/2008WR006926
165. H. Guo, J. J. Jiao, Impact of coastal land reclamation on
ground water level and the sea water interface. Ground Water
45, 362–367 (2007). doi: 10.1111/j.1745-6584.2006.00290.x;
pmid: 17470125
166. C. Sheng, J. J. Jiao, H. Xu, Y. Liu, X. Luo, Influence of land
reclamation on fresh groundwater lenses in oceanic islands:
Laboratory and numerical validation. Water Resour. Res. 57,
e2021WR030238 (2021). doi: 10.1111/j.1745-6584.2006.00290.x;
pmid: 17470125
167. E. Pujades, A. Jurado, Groundwater-related aspects during
the development of deep excavations below the water table:
A short review. Undergr. Space 6, 35–45 (2021). doi: 10.1016/
j.undsp.2019.10.002
168. Y.-X. Wu, S.-L. Shen, D.-J. Yuan, Characteristics of dewatering
induced drawdown curve under blocking effect of retaining
wall in aquifer. J. Hydrol. 539, 554–566 (2016). doi: 10.1016/
j.jhydrol.2016.05.065
169. E. Pujades, A. López, J. Carrera, E. Vázquez-Suñé,
A. Jurado, Barrirer effect of underground structures on
aquifers. Eng. Geol. 145-146, 41–49 (2012). doi: 10.1016/
j.enggeo.2012.07.004
170. G. Attard, T. Winiarski, Y. Rossier, L. Eisenlohr, Review:
Impact of underground structures on the flow of urban
groundwater. Hydrogeol. J. 24, 5–19 (2016). doi: 10.1007/
s10040-015-1317-3
171. J. Bonneau, T. D. Fletcher, J. F. Costelloe, M. J. Burns,
Stormwater infiltration and the ‘urban karst’-A review.
J. Hydrol. 552, 141–150 (2017). doi: 10.1016/
j.jhydrol.2017.06.043
172. G. Ding, J. J. Jiao, D. Zhang, Modelling study on the impact of
deep building foundations on the groundwater system.

p

130. R. Bertani, Geothermal power generation in the world
2010-2014 update report. Geothermics 60, 31–43 (2016).
doi: 10.1016/j.geothermics.2015.11.003
131. J. W. Lund, T. L. Boyd, Direct utilization of geothermal energy
2015 worldwide review. Geothermics 60, 66–93 (2016).
doi: 10.1016/j.geothermics.2015.11.004
132. J. W. Lund, A. N. Toth, Direct utilization of geothermal
energy 2020 worldwide review. Geothermics 90, 101915
(2021). doi: 10.1016/j.geothermics.2020.101915
133. S. Chen, Q. Zhang, P. Andrews-Speed, B. Mclellan,
Quantitative assessment of the environmental risks of
geothermal energy: A review. J. Environ. Manage. 276,
111287 (2020). doi: 10.1016/j.jenvman.2020.111287;
pmid: 32877889
134. F. Parisio, V. Vilarrasa, W. Wang, O. Kolditz, T. Nagel, The
risks of long-term re-injection in supercritical geothermal
systems. Nat. Commun. 10, 4391 (2019). doi: 10.1038/
s41467-019-12146-0; pmid: 31558715
135. L. Jiang, L. Zhu, E. Hiltunen, Large-scale geo-energy
development: Sustainability impacts. Front. Energy 13,
757–763 (2019). doi: 10.1007/s11708-017-0455-9
136. C. W. Johnson, E. J. Totten, R. Bürgmann, Depth migration
of seasonally induced seismicity at The Geysers geothermal
field. Geophys. Res. Lett. 43, 6196–6204 (2016).
doi: 10.1002/2016GL069546
137. W. Luo, A. Kottsova, P. J. Vardon, A. C. Dieudonne,
M. Brehme, Mechanisms causing injectivity decline and
enhancement in geothermal projects. Renew. Sustain. Energy
Rev. 185, 113623 (2023). doi: 10.1016/j.rser.2023.113623
138. E. Gómez-Díaz, S. Scott, T. Ratouis, J. Newson, Numerical
modeling of reinjection and tracer transport in a
shallow aquifer, Nesjavellir Geothermal System, Iceland.
Geotherm. Energy 10, 7 (2022). doi: 10.1186/s40517022-00217-3
139. R. Duggal et al., A comprehensive review of energy extraction
from low-temperature geothermal resources in hydrocarbon
fields. Renew. Sustain. Energy Rev. 154, 111865 (2022).
doi: 10.1016/j.rser.2021.111865
140. P. Dillon, Future management of aquifer recharge.
Hydrogeol. J. 13, 313–316 (2005). doi: 10.1007/s10040004-0413-6
141. C. Sprenger et al., Inventory of managed aquifer recharge
sites in Europe: Historical development, current situation and
perspectives. Hydrogeol. J. 25, 1909–1922 (2017).
doi: 10.1007/s10040-017-1554-8
142. G. Ebrahim, J. F. Lautze, K. G. Villholth, Managed aquifer
recharge in Africa: Taking stock and looking forward. Water
12, 1844 (2020). doi: 10.3390/w12071844
143. C. Stefan, N. Ansems, Web-based global inventory of
managed aquifer recharge applications. Sustain. Water
Resour. Manag. 4, 153–162 (2018). doi: 10.1007/
s40899-017-0212-6
144. S. Alam, A. Borthakur, S. Ravi, M. Gebremichael,
S. K. Mohanty, Managed aquifer recharge implementation
criteria to achieve water sustainability. Sci. Total Environ.
768, 144992 (2021). doi: 10.1016/j.scitotenv.2021.144992;
pmid: 33736333
145. S. Sloan, P. G. Cook, I. Wallis, Managed aquifer recharge in
mining: A review. Groundwater 61, 305–317 (2023).
doi: 10.1111/gwat.13311; pmid: 36950867
146. N. Ulibarri, N. E. Garcia, R. L. Nelson, A. E. Cravens,
R. J. McCarty, Assessing the feasibility of managed aquifer
recharge in California. Water Resour. Res. 57,
e2020WR029292 (2021). doi: 10.1111/gwat.13311;
pmid: 36950867
147. B. R. Scanlon, R. C. Reedy, C. C. Faunt, D. Pool, K. Uhlman,
Enhancing drought resilience with conjunctive use and
managed aquifer recharge in California and Arizona.
Environ. Res. Lett. 11, 035013 (2016). doi: 10.1088/
1748-9326/11/3/035013
148. E. Levintal et al., Agricultural managed aquifer recharge
(Ag-MAR)—a method for sustainable groundwater
management: A review. Crit. Rev. Environ. Sci. Technol. 53,
291–314 (2023). doi: 10.1080/10643389.2022.2050160
149. S. Alam, M. Gebremichael, R. Li, J. Dozier, D. P. Lettenmaier,
Can managed aquifer recharge mitigate the groundwater
overdraft in California's Central Valley? Water Resour. Res.
56, e2020WR027244 (2020). doi: 10.1088/1748-9326/
abcfe1
150. D. E. Wendt, A. F. Van Loon, B. R. Scanlon, D. M. Hannah,
Managed aquifer recharge as a drought mitigation strategy in
heavily-stressed aquifers. Environ. Res. Lett. 16, 014046
(2021). doi: 10.1088/1748-9326/abcfe1

RES EARCH | R E V I E W

195.

196.

197.

198.

sea-level rise and high tides in Honolulu, Hawaii. Water Res.
114, 122–134 (2017). doi: 10.1016/j.watres.2017.02.035;
pmid: 28229950
Y. Wang, C. Zheng, R. Ma, Review: Safe and sustainable
groundwater supply in China. Hydrogeol. J. 26, 1301–1324
(2018). doi: 10.1007/s10040-018-1795-1
T. Gleeson, M. Cuthbert, G. Ferguson, D. Perrone,
Global groundwater sustainability, resources, and systems
in the Anthropocene. Annu. Rev. Earth Planet. Sci. 48,
431–463 (2020). doi: 10.1146/annurev-earth-071719-055251
S. M. Gorelick, C. Zheng, Global change and the groundwater
management challenge. Water Resour. Res. 51, 3031–3051
(2015). doi: 10.1002/2014WR016825
A. S. Elshall et al., Groundwater sustainability: A review
of the interactions between science and policy.
Environ. Res. Lett. 15, 093004 (2020). doi: 10.1088/
1748-9326/ab8e8c

199. E. Stokstad, Deep deficit. Science 368, 230–233 (2020).
doi: 10.1126/science.368.6488.230; pmid: 32299932
200. M. A. Palmer, J. Liu, J. H. Matthews, M. Mumba, P. D’Odorico,
Manage water in a green way. Science 349, 584–585
(2015). doi: 10.1126/science.aac7778; pmid: 26250670
ACKN OWLED GMEN TS

We thank three anonymous reviewers for their constructive
comments and suggestions that have led to significant
improvement of this work. Funding: This work was supported by
the following: National Natural Science Foundation of China
grants 92047202 and 91747204 (to X.K.), Guangdong Provincial
Key Laboratory of Soil and Groundwater Pollution Control
(2023B1212060002) and Ministry of Education of the People's
Republic of China (D20020) (to C.Z.), Strategic Priority Research
Program of the Chinese Academy of Sciences grant XDA20060402
and Shenzhen Science and Technology Program grant

KCXFZ20201221173601003 (to J.L.) Author contributions: C.Z,
X.K., and J.L. conceptualized the study. X.K. wrote the initial
manuscript; J.J.J., C.Z., J.L., B.K.B., Y.W., B.R.S., S.J. T.G., and J.-P. N.
reviewed and edited the manuscript. All authors made substantial
contributions to discussions of the content of the manuscript.
Competing interests: Authors declare that they have no
competing interests. Data availability: All data presented here are
based on previously published studies that are cited in the review.
License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https: per per www.sciencemag.org per about per sciencelicenses-journal-article-reuse

Submitted 27 September 2022; accepted 5 January 2024
10.1126/science.adf0630

p
g
y
y g
y
,

Kuang et al., Science 383, eadf0630 (2024)

1 March 2024

13 of 13

RES EARCH

RESEARCH ARTICLE SUMMARY

◥

INNATE IMMUNITY

Native architecture of a human GBP1 defense
complex for cell-autonomous immunity to infection
Shiwei Zhu†, Clinton J. Bradfield†, Agnieszka Maminska†, Eui-Soon Park†, Bae-Hoon Kim†,
Pradeep Kumar, Shuai Huang, Minjeong Kim, Yongdeng Zhang, Joerg Bewersdorf, John D. MacMicking*

complex assembled directly on Stm inside human cells. This defense complex comprised

RATIONALE: Since their discovery in the phys-

ical assembly of antimicrobial and innate im-

Extended GBP1
conformers

Human GBP1
defense complex

Bacterial
LPS

Architecture of a human GBP1 defense complex. (Left) 3D reconstruction of human GBP1 on the outer
membrane (OM) of Salmonella bacteria from cryo-ET. IM, inner membrane. Size of scale bar shown in Angstroms.
(Right) Pseudomodel showing the extended upright GBP1 conformer inserting into the OM LPS layer. Release of
LPS by GBP1 insertion subsequently activated caspase-4 following its coassembly on the same bacterial surface.
Zhu et al., Science 383, 965 (2024)

1 March 2024

▪

The list of author affiliations is available in the full article online.
*Corresponding author. Email: john.macmicking@yale.edu
†These authors contributed equally to this work.
Cite this article as S. Zhu et al., Science 383, eabm9903
(2024). DOI: 10.1126/science.abm9903

READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abm9903
1 of 1

,

Bacterial IM

y

Bacterial
ribosomes

nate immune signaling cascades is the
higher-order assembly of repetitive protein
units that generate large polymers capable
of amplifying signal transduction. Our results identify human GBP1 as the principal
repetitive unit, numbering thousands of proteins per bacillus, that undergoes dramatic
conformational opening to establish a host
defense platform directly on the surface of
gram-negative bacteria. This platform enabled the recruitment of other immune partners, including GBP family members and
components of the inflammasome pathway,
that initiate protective responses downstream
of activating cytokines such as interferon-g.
Elucidating this giant molecular structure not
only expands our understanding of how human cells recognize and combat infection but
may also have implication for antibacterial
approaches within the human population.

y g

200 Å

200 Å

Bacterial OM

y

CONCLUSION: An emerging paradigm for inRESULTS: We identified a multiprotein defense

Salmonella

g

mune signaling complexes over a decade ago,
guanylate-binding proteins (GBPs) have emerged as major organizers of intracellular host
defense to a broad array of bacteria, viruses, or
parasites in animals and plants. In mammals,
these large dynamin-like guanosine triphosphatases (GTPases) relocate to intracellular
pathogens, where they can establish macromolecular assemblies on the microbial outer
membrane (OM) that serve as interactive hubs
for inflammatory proteins or bactericidal effectors as part of the cell-autonomous innate
immune response. To better understand the
mechanistic details underlying these distinct
hybrid structures, we enlisted host and bacterial genetics plus single-particle nanoscopy and
cryo–electron tomography (cryo-ET) to visualize GBP defense complexes on the surface of a
gram-negative bacterial pathogen, Salmonella
enterica serovar Typhimurium (Stm), within
the cytosol of human cells.

p

INTRODUCTION: The compartmentalized landscape of eukaryotic cells offers a wide variety
of intracellular niches for microbial pathogens
to hide and replicate. Consequently, eukaryotes have evolved compartment-specific immune surveillance mechanisms that alert the
host to infection and recruit antimicrobial
proteins that help bring microbial replication
under control. Many of these host defense proteins form giant macromolecular complexes
when encountering either pathogens or their
products to amplify innate immune signaling
and spatially localize protein partners at the
site of microbial recognition. In some cases,
complete signaling cascades are built directly
upon the invading pathogen itself, a distinctive situation in which a large foreign object
acts as the anchoring platform for assembling
the entire host defense machinery. How these
massive host-pathogen platforms are initiated
and structurally organized at the molecular
level remains unknown.

four members of the human GBP family (GBP1,
GBP2, GBP3, and GBP4) together with human
caspase-4 and one of its natural substrates,
full-length Gasdermin D (GSDMD). It triggered
innate immune signaling through caspase-4
cleavage of GSDMD into its pore-forming subunits, resulting in the extracellular release of
an immune cytokine, interleukin-18 (IL-18),
and pyroptotic cell death needed for protection against this bacterial pathogen. Notably, human GBP1 was obligate for initiating
the entire signaling cascade; its genetic removal prevented the remaining immune proteins being recruited onto the gram-negative
bacterial surface. C-terminal anchorage and
GTPase-driven self-assembly enabled GBP1
to bind to and polymerize over the surface
of cytosolic Stm to establish the recruitment
platform. Nearly 30,000 GBP1 molecules
were assembled in just a few minutes. Reconstitution of this massive GBP1 defense
complex with a bacterial minicell system allowed cryo-ET visualization of the entire
coat structure in its native state. Within this
native platform, individual GBP1 dimers
were found to adopt an open conformation
for vertical insertion into the bacterial OM
through their extended C-terminal lipidated
tail. Anchorage of the upright GBP1 conformer led to OM disruption, which released
the gram-negative cell wall component, lipopolysaccharide (LPS), to activate coassembled caspase-4.

RES EARCH

RESEARCH ARTICLE

◥

INNATE IMMUNITY

Native architecture of a human GBP1 defense
complex for cell-autonomous immunity to infection
Shiwei Zhu1,2,3,4†‡, Clinton J. Bradfield1,2,3,4†§, Agnieszka Maminska1,2,3,4†, Eui-Soon Park1,2,3,4†,
Bae-Hoon Kim1,2,3,4†¶#, Pradeep Kumar1,2,3,4, Shuai Huang1,2,3,4**, Minjeong Kim1,2,3,4,
Yongdeng Zhang5††, Joerg Bewersdorf5,6, John D. MacMicking1,2,3,4*

1 March 2024

1 of 18

,

Zhu et al., Science 383, eabm9903 (2024)

y

*Corresponding author. Email: john.macmicking@yale.edu
†These authors contributed equally to this work.
‡Present address: Department of Physiology & Biophysics, Case
Western Reserve University School of Medicine, Cleveland, OH,
USA. §Present address: Signaling Systems Section, Laboratory of
Immune System Biology, National Institute of Allergy and
Infectious Diseases, National Institutes of Health, MD, USA.
¶Present address: Rare Disease R&D Center, PRG Science and
Technology Co. Ltd., Pusan, Republic of Korea.
#Present address: Department of Molecular Biology, College of Natural
Sciences, Pusan National University, Pusan, Republic of Korea.
**Present address: Department of Molecular Genetics, The Ohio State
University, Columbus, OH, USA.
††Present address: School of Life Sciences, Westlake University,
Hangzhou, Zhejiang, PR China.

y g

Howard Hughes Medical Institute, Yale University School of
Medicine, New Haven, CT 06510, USA. 2Yale Systems
Biology Institute, West Haven, CT 06477, USA. 3Department
of Microbial Pathogenesis, Yale University School of
Medicine, New Haven, CT 06510, USA. 4Department of
Immunobiology, Yale University School of Medicine, New
Haven, CT 06510, USA. 5Department of Cell Biology, Yale
University School of Medicine, New Haven, CT 06510, USA.
6
Yale Nanobiology Institute, West Haven, CT 06477, USA.

y

1

This ability to generate extended, filamentous signaling platforms stems in part from
the modularity of the proteins involved. Signalosome proteins often harbor leucine-rich
repeat domains, caspase activation and recruitment domains (CARDs), or death effector domains that concentrate receptors, adaptors,
and effectors through cooperativity (3). As a
result, they yield some of the most iconic and
visible structures inside immune-activated cells.
These include retinoic acid–inducible gene I
filaments and mitochondrial antiviral-signaling
protein prion-like structures that control RNA
sensing; CARD and nucleotide-binding domain,
leucine-rich (NLR)–containing protein filaments
as part of the inflammasome machinery;
Myddosomes orchestrating nuclear factor kB
and interferon (IFN) regulatory factor signaling;
and NLR HOPZ-ACTIVATED RESISTANCE
1 pentamers that underpin the plant resistosome (3–6). Collectively, these large polymeric
structures represent an increasingly pervasive paradigm for cell-autonomous innate
immunity throughout metazoan evolution.
Their mode of assembly differs from that of
classical antigen or immunoglobulin signaling
at the plasma membrane, which are typically
transmitted through clustered synapses (3).
To this list can now be added members of a
dynamin-like guanosine triphosphatase (GTPase)
family termed guanylate-binding proteins (GBPs),
which undergo polymerization for generating
innate immune signaling platforms and coordinating local antimicrobial activity. GBPs assemble into large multimeric structures inside
Arabidopsis, zebrafish, mouse, and human cells,
in some cases using phase-separation to further concentrate homotypic complexes (7–12).
In plants, the 69- to 122-kDa GBP-like proteins

g

I

n biological systems, environmental cues
are often sensed by ligand-induced allosteric changes in cell surface receptors that
rapidly transmit signals to the interior for
mobilizing the desired physiological response. Within the cell-autonomous defense
systems of plants and animals (1, 2), additional
modalities are used to decode the outside world,
including intracellular protein assemblies formed
through helical symmetry (3, 4). These higherorder assemblies amplify innate immune signals
through protein polymerization or “prionization” events to facilitate proximity-induced autoactivation of latent zymogens, caspases, and
kinases. The benefits to such repetitive design are manifold: lowered signaling thresholds,
all-or-none responsivity, and stable signalosome
platforms that can recruit and accommodate
numerous protein partners (3).

p

All living organisms deploy cell-autonomous defenses to combat infection. In plants and animals,
large supramolecular complexes often activate immune proteins for protection. In this work, we resolved
the native structure of a massive host-defense complex that polymerizes 30,000 guanylate-binding
proteins (GBPs) over the surface of gram-negative bacteria inside human cells. Construction of this
giant nanomachine took several minutes and remained stable for hours, required guanosine triphosphate
hydrolysis, and recruited four GBPs plus caspase-4 and Gasdermin D as a cytokine and cell death
immune signaling platform. Cryo–electron tomography suggests that GBP1 can adopt an extended
conformation for bacterial membrane insertion to establish this platform, triggering lipopolysaccharide
release that activated coassembled caspase-4. Our “open conformer” model provides a dynamic view
into how the human GBP1 defense complex mobilizes innate immunity to infection.

(GBPLs) respond to inducible immune signals,
including salicylic acid and pipecolic acid, to
assemble large nuclear RNA polymerase II hubs
that transcribe host defense genes during infection (7, 8). In animals, immune cytokines
such as IFNs induce 65- to 73-kDa GBP
expression to control microbicidal or inflammasome responses within the cytosol of both
immune and nonimmune cells (9–13). These
activities often coincide with the relocation
of GBPs to the site of microbial replication
where they completely “coat” targeted pathogens to build mesoscale signaling or killing
platforms (9, 12–16). GBP-coated pathogens can
range in size from ~750 nm in diameter for
Salmonella enterica serovar Typhimurium (Stm)
to >5 mm for Toxoplasma gondii tachyzoites (17).
Assembling such a large coat complex must
enlist biochemical properties synonymous
with dynamin-like proteins (DLPs). DLPs
typically exhibit robust GTPase activity (kcat,
~2 to 100 min−1), low-micromolar substrate
affinity and nucleotide-dependent self-assembly
to generate >0.5- to 1-MDa complexes within
cells (18, 19). Hence, they are “large” GTPases,
which often function as mechanoenzymes to
deform or tubulate membranes during vesicular trafficking, organelle division, or cytokinesis (18, 19). Human GBPs also exhibit high
intrinsic catalytic activity (kcat, ~80 min−1)
to produce guanosine diphosphate (GDP) or
monophosphate (GMP) from guanosine triphosphate (GTP) in a two-step hydrolysis reaction (20, 21). GTP hydrolysis likely initiates
conformational changes leading to cooperative self-assembly of GBP dimers. Structurally, all human GBPs possess a large globular
(LG) N-terminal catalytic domain and extended
a-helical C-terminal tail (19); the latter spans a
middle domain (MD) comprising helices a7 to
a11 and GTPase effector domain (GED) encompassing the final a12 and a13 helices (20, 22, 23).
In human GBP1 and GBP5 crystal structures,
the GED folds back tightly onto the LG and
MD regions (20, 22, 23). This could represent a
closed, autoinhibited state, as substrate binding results in different geometries for isolated
GBP1 when viewed by conventional electron
microscopy (24). Indeed, substrate catalysis
could theoretically release the GED to yield
an open, active dimer that undergoes further
multimerization, although high-resolution GBP
structures captured directly on the pathogen
surface after GTP hydrolysis have yet to be
reported. In addition, GBP1, GBP2, and GBP5
each possess C-terminal CaaX motifs for isoprenyl modification to facilitate membrane binding (19, 24). Isoprenylation could offer not only
anchorage but also serve as a nucleating template to deposit more GBPs on the microbial
surface to build a signaling platform during
innate immunity (12–16).
In this study, we characterized a massive
immune defense complex comprising nearly

RES EARCH | R E S E A R C H A R T I C L E

30,000 GBP1 molecules assembled on the
bacterial surface using cryo–electron tomography (cryo-ET). Notably, despite their central role for host defense across plant and
animal kingdoms (25–27), the ultrastructural
organization of these mesoscale GBP coat complexes on an intact pathogen membrane is
currently unknown. Visualizing such structures below the light diffraction limit would
enhance our understanding of how eukaryotic cells recognize and combat infection. It
would also yield information on GBP1 coat
assembly under native conditions, which trigger innate immune signaling. Each has important implications for anti-infective therapy
as well as the basic biology of immune recognition and host defense within the human
population.

y
,

2 of 18

y g

1 March 2024

Coat construction required GBP1 C-terminal
attachment and N-terminal catalytic activities
to polymerize over the entire Salmonella surface. Whether microbial activities or physical
features of bacteria also influence this process
was examined by engineering 13 Stm strains
differing in size, shape, motility, and OM
composition. Stm isolates that were longer
(StmpBAD-ftsZ, up to 20 mm), wider (StmMreB K27E
mutants, ~2 mm in diameter), or smaller
(StmDminD minicells 250 to 300 nm in diameter
arising from aberrant septation) still recruited
endogenous GBP1 in IFN-g–primed cells to
activate IL-18 release and pyroptotic cell death
(Fig. 1, G and H). Bent StmMreB D78V mutants
likewise mobilized this pathway (34, 35) (Fig.
1, G and H). Hence, microbial cell size, division,
or curvature did not seem to influence GBP1
coat formation to generate an innate immune
signaling platform. Flagellin-expressing (36)
and flagellin-deficient Stm (StmDflhD) were
both targeted by GBP1, ruling out motility or
bacterial immobilization as a cue to begin coat
complex assembly.
We therefore turned to the OM itself. Gramnegative bacteria harbor long polysaccharide LPS
chains that form a divalent cation-crosslinked
barrier that is impermeable to hydrophobic
solutes (37). The LPS moiety consists of O-antigen
polysaccharide, outer core galactose- and
inner core heptose- and 3-Deoxy-D-mannooctulosonic acid (KDO)–enriched saccharides,
and a lipid A module with multiple acyl chains
embedded at the base by electrostatic and hydrophobic interactions (Fig. 1G). Isogenic Stm
mutants with progressively shorter LPS chains
[generated by inactivating enzymes at successive steps of the LPS biosynthetic pathway

y

Zhu et al., Science 383, eabm9903 (2024)

Host and bacterial determinants of GBP1 coat
complex assembly
g

We first examined the human GBP1 coat complex surrounding cytosolic bacteria using 4Pi
single-molecule switching (4Pi-SMS) nanoscopy
and fast, live three-dimensional structured illumination microscopy (3D-SIM) with an OMXSR Blaze imaging instrument equipped with
high-speed galvanometers. 4Pi-SMS is a dualobjective, single-molecule localization, superresolution microscope that resolves 3D structures
to ~20 nm (200 Å) isotropically throughout
entire mammalian cells (28, 29); it enabled
us to detect single GBP molecules on the surface
of virulent Stm inside human cells. 3D-SIM
imaging with the OMX-SR instrument is capable
of ~180 frames per second, ensuring that coat
complex assembly could be followed throughout the entire bacterial encapsulation process.
We tracked GBP1 as the forerunner of this
seven-member DLP family in humans. Recent
work discovered that human GBP1 recruitment
onto the outer membrane of cytosol-invasive
pathogens, including Stm and Shigella flexneri,
enables bactericidal activity by apolipoprotein
L3 (APOL3) in IFN-g–activated primary human intestinal epithelia, myofibroblasts, and
endothelium, as well as in HeLa CCL2 cells
(13). We and others also reported that GBP1
recruits additional GBP family members plus
endogenous human caspase-4 to stimulate
cytokine release [interleukin-18 (IL-18)] and
pyroptosis in primary intestinal human organoids, human macrophages, and human
epithelial cell lines (12, 14–16). The GBP coat
complex has thus emerged as a central hub
for intracellular host defense and innate immune signaling in humans.
To visualize real-time GBP1 coat complex
formation by live 3D-SIM imaging, we deleted endogenous GBP1 in human HeLa CCL2
cells with CRISPR-Cas9 and replaced it with a
functional monomeric red fluorescent protein
(mRFP)–GBP1 reporter (GBP1–/–mRFP-GBP1) detected at physiological levels to avoid over-

age and postprenyl processing (32) (fig. S3, A
and B). LPS is the major constituent of gramnegative OMs (~75% in Stm) (33) and was recently reported to bind human GBP1 (15, 16).
We found that Stm LPS captured farnesylated
FLAG-tagged GBP1 at physiological pH and
temperature in fluorescence anisotropy assays
[dissociation constant (Kd), ~3.971 mM], whereas it failed to capture nonfarnesylated FLAGGBP1C589S, nondimerized or catalytic mutants,
or farnesylated FLAG-GBP1R584–586A (fig. S3B).
GBP1 assembly together with the farnesyl moiety
therefore appears critical for Stm LPS engagement with the arginine patch strengthening
these interactions electrostatically to maintain a stable coat (15, 16). This stability was in
some cases very long-lived: Live imaging revealed that a single GBP1 coat can persist for
up to 2 to 3 hours within the human cytosol
(fig. S3C). Thus, thousands of GBP1 molecules
generate a highly durable signaling platform
once anchored through initial farnesyl and
polybasic contacts to the bacterial OM.

p

Human GBP1 coat size, kinetics, and stability
in cellulo

expression artefacts (fig. S1A). When infected
with Stm expressing enhanced green fluorescent protein (StmEGFP), mRFP-GBP1 completely
encapsulated individual bacilli over a ~1- to
6-min time period (Fig. 1A and movies S1 and
S2). Comprehensive coating was also observed
in single-molecule 4Pi-SMS imaging of endogenous GBP1 and GBP2 on the bacterial surface
in IFN-g–activated GBP1+/+ cells, which in some
cases depicted colocalization (Fig. 1B and movie
S3). Volumetric and kinetic measurements
of GBP1 yielded 29,542 ± 5156 molecules per
bacterium assembled at a rate of 103 ± 11.6
molecules per second (fig. S1, B and C).
Such rapid kinetics required massive GBP1
cooperativity involving sequential hydrolysis
of GTP and GDP for nucleotide-dependent
self-assembly of GBP1 dimers on the bacterial
surface, as revealed by loss-of-function mutants.
GTPase (GBP1S52N), GDPase (GBP1DD103,108NN),
or dimerization (GBP1D184N) mutants blocked
coat formation and downstream IL-18 release
plus pyroptotic cell death (as LDH release) in
stably reconstituted GBP1–/– cells (10) (Fig. 1, C
to H). Notably, these N-terminal mutants were
still posttranslationally prenylated at a C-terminal
CaaX motif (Fig. 1F and fig. S2, A and B),
which may otherwise help anchor GBP1 to
the bacterial outer membrane (OM). Indeed,
mutating the CaaX box (GBP1C589S) prevented
both C15 farnesylation and coat attachment
inside human cells (Fig. 1, F to H). It did not,
however, interfere with nucleotide-dependent
dimer self-assembly as shown through size
exclusion chromatography by using a transition state analog, GDP plus aluminum fluoride
(AIF3–), in recombinant protein assays (Fig. 1E).
Thus, GBP1 mutants uncoupled OM attachment from subsequent polymerization, revealing distinct steps in coat-complex formation
during immunity to gram-negative infection.
OM anchorage also required a polybasic
patch (amino acids 584 to 586) that resembled
lipid-binding motifs in small H-Ras GTPases
(30) within the GBP1 C-terminal a-13 helix
(20) (fig. S2, A and B). Alanine-scanning mutagenesis of all three arginines (GBP1R584–586A)
(31) ablated coat formation and impaired downstream cytokine plus cell death signaling (Fig.
1H and fig. S2, C and D). Because GBP1R584–586A
was heavily farnesylated inside human cells,
the loss of coat complex assembly was not due
to R584-to-586A substitution interfering with
lipidation of the nearby CaaX motif (Fig. 1F).
Instead, it appears that GBP1 farnesylation is
necessary but not sufficient for OM anchorage,
requiring a second site to stably engage the
OM and help retain it on the bacterial surface.
The bivalent nature of this C-terminal anchor was reinforced in lipopolysaccharide (LPS)–
binding profiles for the GBP1R584–586A and
GBP1C589S mutants purified from human embryonic kidney (HEK) cells lacking endogenous
GBP expression to ensure correct farnesyl link-

RES EARCH | R E S E A R C H A R T I C L E

A

Fig. 1. Human GBP1 coat kinetics and functional
determinants in cellulo. (A) Live 3D-SIM showing
full encapsulation of EGFP-expressing Stm by
mRFP-GBP1 in IFN-g–activated HeLa cells. GBP1 coat
complex initiation, expansion, and completion
are noted. Volume rendering through Imaris software.
Maximum intensity projection of one of six similar
3D-SIM videos shown. Scale bar, 2 mm. (B) 4Pi-SMS
nanoscopy of cytosolic Stm coated by endogenous
human GBP1 and GBP2, detected by using rat anti-GBP1
and mouse anti-GBP2 antibodies, respectively, at
2 hours postinfection. Maximum intensity projection
of one of eight similar 4Pi-SMS images shown.
Scale bar, 1 mm. (C) Domain structure of farnesylated
human GBP1 (PDB 6K1Z) depicting catalytic LG,
MD, and GED. The polybasic patch and CaaX motif in
the C-13 a helix are denoted. RRR, polybasic triplearginine patch. (D) TLC of 32[P]-GTP hydrolysis
products for recombinant GBP1 and its mutants.
Representative of four independent experiments.
(E) Size exclusion chromatography profiles depicting
GBP1 and its mutants in the absence or presence
of the transition state analog, GDP plus AIF4–.
Representative of three independent experiments.
(F) Prenylation profile of EGFP-GBP1 and its
mutants in HEK-293E cells, detected by using
FPP-azide-biotin CLICK chemistry coupled to anti-GFP
immunoprecipitation. SA, streptavidin–horseradish
peroxidase (HRP); IgG, immunoglobulin heavy
chains. One of three independent biological
experiments. (G) Salmonella and GBP1 mutants
used to examine determinants of coat complex
function. Acyl chains removed through each mutation
are depicted by colored brackets. c11088,
StmDlpxRDPagLDPagP triple mutant. (H) Bacterial and
host determinants in human GBP1 coat complex
formation on Stm coating and downstream IL-18 release
or cell death in IFN-g–activated wild-type or mutant
HeLa cells infected with different bacterial strains.
Mean ± standard deviations is shown for triplicates.
Significant one-way analysis of variance (ANOVA) values
with Bonferroni post hoc test are shown. NS, not
significant. Representative of three to five independent
experiments.

B

Live 3D-SIM
Expansion

Initiation
17:57:845

18:42:746

20:12:831

19:27:713

4Pi-SMS
Cytosolic Stm

GBP1
GBP2

Bacterial
surface

~10-12nm
20:57:881

21:43:148

24:42:716

25:28:218

Colocalized
GBPs
Expansion

C

Completion
317

1

LG
LG

E

479 486 588
MD
GED CaaX

Buffer

F

GDP/AIFx

kDa: 670 158 44

FPP-azide-biotin CLICK chemistry

17

Cell lysate

MD ( 7- 11)

IP:GFP
4-5
86
A

GF
GF P -G
B
GFP -GB P1
GF P -G P1 S5
GFP -G BP1 D 2N
GF P-GBP1 D1D103.1
P- BP 84N 08N
GB 1 C5
N
P1 R 89S
GF
58
4-5
P
86
GF -G
A
B
GFP -GB P1
GF P -G P1 S5
GFP -G BP1 D 2N
GF P -GBP1 D1D103.1
P- BP 84N 08N
GB 1 C5
N
P1 R 89S
58

rGBP1

rGBP1S52N
GED ( 12, 13)

D

GBP1 mutants

[32P]-GMP
[32P]-GDP
[32P]-GTP

rGBP1DD103.108NN
IB:
SA

Prenylated
GBP1
GFP-GBP1
IgG

rGBP1D184N

rGBP1C589S

IB:
GFP

rG
r
B
BP GBP P1
1 DD 1 S52
10
N
rG 3.108
NN
BP
rG 1 D18
4N
rG B P
BP 1 C58
1 R58 9S

4-5
86
A

GFP

(GBP1 prenylation in cellulo)

rGBP1R584-586A

rG

Retention volume (mL)

G

g
y

P < 0.0005

P < 0.0005

P < 0.0002

P < 0.0004

P < 0.0012
P < 0.0003

P < 0.0001
A
86

P < 0.0007

P < 0.0002
P < 0.0001

P < 0.0006
P < 0.0002
03

4 -5

08
11

Zhu et al., Science 383, eabm9903 (2024)

proteins LpxR, PagL, or PagP (that remove a
3′-acyloxyacyl moiety, remove a single R-3hydroxymyristate chain, or palmitoylate the
hydroxymyristate chain, respectively) (40) failed
to prevent GBP1 coating and caspase-4 activation (Fig. 1, G and H). CRISPR-Cas9 deletion of
human acyloxyacyl hydrolase (AOAH–/–), which
is expressed at low levels in HeLa CCL2 cells
and removes secondary acyl chains from lipid

1 March 2024

A as a deactivation mechanism (41), or human
E3 ubiquitin ligase ring finger protein 213
(RNF213–/–), which modifies lipid A through
ubiquitinyation (42), also had no effect on coatdependent signaling (Fig. 1, G and H). Thus,
functional coat formation on Salmonella was
primarily governed by host GBP1 activities
rather than lipid A modifications or other
microbial determinants in cellulo (i.e., within
3 of 18

,

P < 0.0002

NS

P < 0.0001

NS

NS

30
25
20
15
10
5
0
60
50
40
30
20
10
0
450
400
350
300
250
200
150
100
50
0

,1

NS

NS

8

NS

NS

08
NN
GB
P1 D
GB 184N
P
1C
GB
P1 R 589S
58

NS

30
25
20
15
10
5
0
60
50
40
30
20
10
0
450
400
350
300
250
200
150
100
50
0

GB
P
GB GBP 1
P1 D 1 S 52N
D1

NS

y

30
25
20
15
10
5
0
60
50
40
30
20
10
0
450
400
350
300
250
200
150
100
50
0

P < 0.0001

H

y g

(13, 38)] revealed that OM truncations in StmDwzy,
StmDwaaL, StmDwaaJ, StmDwaaI, or StmDwaaG did
not block GBP1 coat formation and downstream
innate immune signaling (Fig. 1, G and H).
Beneath these truncations, we modified the
lipid A module positioned at the base of the
OM where it interacts with the phospholipid
inner leaflet; lipid A is recognized and directly
bound by caspase-4 (39). Mutations in Stm

p

Absorbance (100%)

RRR CaaX

RES EARCH | R E S E A R C H A R T I C L E

a living cell). Human GBP1 still targeted cytosolic Stm irrespective of bacterial size, shape,
motility, or OM composition; the latter spanned
LPS chains of different length, charge, and
chemical structure. Such broad ligand promiscuity may help GBP1 combat gram-negative
pathogens that modify their LPS moiety in
an attempt to evade innate immune recognition and antimicrobial killing.
Human GBP1 coat initiates a multiprotein
platform for bacterial recognition

y
,

4 of 18

y g

1 March 2024

y

How the GBP1 coat complex promotes LPS recognition by caspase-4, especially given that lipid
A is buried at the base of the OM, remains a
question. We initially probed LPS release by
live bacteria in cellulo (Fig. 3, A and B). First,
copper (Cu2+)–free CLICK chemistry was used
to label Stm LPS with Alexa Fluor attached
through KDO-azide derivatives adjacent to
lipid A within the inner core (Fig. 3A). AntiSalmonella O-antigen antibody was used in
conjunction to verify the KDO–Alexa Fluor
signal, which decreases during transit to the
cytosol. We likewise incorporated fluorescent
D-alanine into the L-Ala-D-meso-diaminopimelateD-Ala-O-Ala pentapeptide through metabolic la-

g

Zhu et al., Science 383, eabm9903 (2024)

GBP1 defense complex triggers bacterial LPS
release in cellulo and in reconstituted systems

beling of the underlying peptidoglycan (PG)
scaffold in active bacteria (44) (Fig. 3A). 3D-SIM
found that complete GBP1 coating triggered
release of LPS but did not seem to disturb
the Stm PG layer within the cytosol of IFN-g–
activated HeLa cells (Fig. 3B), confirming that
GBP1 can promote caspase-4 ligand availability in cellulo. Other exteriorized structures such as flagella were still evident on
GBP1-coated bacteria (Fig. 3C); hence, GBP1 primarily affected OM disruption, leaving the
underlying PG scaffold and flagellar apparatus
intact.
We next directly assayed lipid A release to
test if GBP1 was indeed sufficient for LPS
liberation. We used a cell-free reconstitution
system to measure soluble lipid A, which cannot be undertaken in situ owing to contamination of the host cell cytosol by whole
bacteria. Incubation of farnesylated recombinant RFP (rRFP)–GBP1 with axenic Stm found
that >95% of bacteria were fully coated within
60 min after addition of GTP substrate (Fig.
3D). Bacterial encapsulation by rRFP-GBP1
followed a highly accelerated sigmoidal curve
that yielded a half maximal value (“coat Km”)
of 225 nM and a steep Hill slope of 5.122 (Fig.
3E). Notably, this all-or-none behavior did not
arise from crossing a phase-transition boundary because rRFP-GBP1 does not phase separate, unlike plant GBPLs, which possess a
C-terminal intrinsically disordered region to
generate biomolecular condensates during
infection (6, 7) (fig. S5A). Pronounced coating
was evident irrespective of Stm size or LPS status;
comparison with gram-negative Pseuodomonas
aeruginosa and gram-positive Listeria monocytogenes
showed that the latter two bacteria had lower
levels of rRFP-GBP1 encapsulation (fig. S5,
B and C).
Notably, reconstituting the GBP1 coat complex triggered Stm LPS release, which was
measured by limulus amebocyte lysate (LAL)
assay that detects the soluble lipid A moiety.
Unfarnesylated rRFP-GBP1C589S and farnesylated catalytic or assembly mutants that could
not coat Stm failed to release LPS even after
addition of GTP, mimicking the results seen
inside human cells (Fig. 3F). Omission of GTP or
substitution with nonhydrolyzable GTP analogs (GTP-g-S or GMPPNP) or transition-state
mimics (GDP.AIF3– or GMP.AIF4–) also failed
to trigger rGBP1-dependent LPS release because these analogs could not support hydrolysisdriven GBP1 coating on bacteria (Fig. 3F). The
amount of lipid A released by GTP-dependent
rGBP1 assembly on Stm greatly exceeded the
Kd range of caspase-4 (39) yet comprised <1%
of the total LPS present [based on 2 × 106
molecules of LPS per Stm (31)]. Hence, the
GBP1 coat complex disrupts the LPS outer
leaflet; depending on the extent of this disruption, it may coincide with inflammasome
activation or even bacterial killing (13, 15). Indeed,

p

Broad multivalent GBP1 interactions with the
bacterial surface provide a stable platform to
recruit downstream protein partners for innate immune signaling (12, 13, 16). Previous
work from us together with other groups found
that GBP2 to GBP4 and caspase-4 form part
of this GBP1 signaling platform in primary
human intestinal organoids and cervical epithelial cell lines (12, 16). In this study, stable
CRISPR-Cas9 deletions corroborated their importance with endogenous caspase-4 autoproteolysis (denoted by the active p30 subunit),
IL-18 release, and pyroptotic cell death (as
LDH release) significantly diminished in IFN-g–
activated GBP1–/– and GBP2–/– single-knockout
and GBP1–/–2–/– double-knockout cells (fig. S4,
A to C). Single deletions of GBP3–/– and GBP4–/–
had a lesser effect, however, en bloc removal of
the entire 335-kb human GBP1 to GBP7 cluster
through genome engineering on chromosome
1q22.2 (GBPD1q22.2) yielded almost complete
loss of downstream signaling, underscoring
concerted GBP action and phenocopying
CASP4–/– and GSDMD–/– (Gasdermin D) cells
(fig. S4, A to C).
Complete GBP gene cluster deletion in the
septuple GBPD1q22.2 mutant provided an ideal
tool to reconstitute the entire coat complex on
an empty background and test if GBP1 is the
critical organizer of this hierarchical complex
on the same bacilli in situ (fig. S4D). We included full-length GSDMD as a natural caspase-4
substrate with potential bactericidal activities
(43) to see if it is brought into this new supramolecular platform as well. Each coat component was fused to one of seven fluorescent
proteins [mAzurite, mSapphire, pmTurqouise2,
pmEmerald, pmVenus, pmOrange, pmCardinal,
or pmIFP24; fluorescence excitation and emission (Ex/Em) range, 384/450 to 684/708 nm] to
identify compatible combinations for reconstituting GBPD1q22.2 cells, as exchange-PAINT
lacked appropriate antibodies for detection of
endogenous proteins (fig. S4, D and E). Our
color-coded orthogonal matrix (GBP-COAT450–708)
successfully resolved five-color objects to yield
a complete signaling platform with GBP1 to
GBP4 plus caspase-4 or GSDMD all coating
the same bacilli after 90 to 120 min of infection (Fig. 2, A to C, and fig. S4, F and G).
Notably, this multicolored coat was completely lost when GBP1 was omitted, indicat-

ing that GBP1 establishes the entire signaling
cascade at the outset (12, 14) (Fig. 2, A and D).
Indeed, the GBP-COAT450–708 assay found human GBP1 was obligate for GBP2, GBP3, GBP4,
and caspase-4 or GSDMD simultaneously sharing the same bacterial surface (Fig. 2D). Excluding caspase-4 had no effect on GBP1 to
GBP4, placing them upstream, but it largely
blocked full-length GSDMD targeting, positioning it downstream in this six-member signaling cascade as part of two-step hierarchical
model (Fig. 2, C to F). Notably, N- or C-terminal
GSDMD fragments mimicking the processed
substrate failed to be recruited (Fig. 2E). Hence,
caspase-4 appears to bring only its full-length
GSDMD substrate to this location for cleavage
once the protease is activated by lipid A after
recruitment by GBP1; this was also corroborated in individually deleted CASP4–/– or GBP1–/–
cells (Fig. 2F). By tracking the other major
caspase-4 substrate in this pathway, pro–IL-18,
we found that it largely failed to be recruited
onto the coat. Hence, GSDMD is a new component of the GBP-CASP4 signaling complex
as part of a two-step hierarchical model (Fig. 2F).
Its interaction with caspase-4 likely occurs
before reaching the GBP platform, as shown
by coimmunoprecipitation assays in GBPD1q22.2
cells (Fig. 2G).
This hierarchical model was further supported
by LPS binding profiles (Fig. 2H). FLAG-GBP1,
-GBP2, -GBP3, -GBP4, or –caspase-4C258A (preventing autoproteolysis) (39) were purified from
human HEK cells to avoid bacterial contaminants and ensure proteins were posttranslationally modified; GBP1 (Kd, ~3.776 mM) and
caspase-4C258A (Kd, ~313 nM) strongly interacted with Stm LPS as major OM binding
proteins (13, 14, 39) (Fig. 2H). GBP1 therefore
serves as the principal organizer of this multiprotein complex atop the coated bacterium
in situ (12, 16). Its assembly facilitates bacterial recognition of the LPS lipid A moiety
by caspase-4 that can recruit GSDMD as a
subsequent step to activate cell death and
cytokine release downstream.

RES EARCH | R E S E A R C H A R T I C L E

Coat reconstitution in GBP

pc
D
mI NA3
FP .1
mO
24
mV rang-GBP
mEenuse-GB1
mT mer -GB P2
a P
mT-Sap ld-G 3
-Sa phi BP
pp re-C 4
hir A
e-G SP
SD 4
MD
-CASP4 All
78 -100
-75
91 78 -100
-75
70

97

E
Stm
FL-GSDMD Inset

P < 0.004

P < 0.002
P < 0.006

NS
NS

- GSDMD

- CASP4
NS
NS
NS
NS

NS

NS
NS
NS

NS
NS

NS
NS

- GBP4

CTGSDM
D

NTGSDM
D

Stm FL-GSDMD
Inset

chr.1q22.2

NS
NS

All
-GSDMD -GBP4 -GBP3 -GBP2 -GBP1
NS

P < 0.0003

- GBP3

chr.1q22.2

IFP24- Orange- Venus- Emerald Sapphire Merge
GBP1 GSDMD GBP3 -GBP4 -CASP4

-GAPDH

NS
NS
NS

NS

P < 0.0003

-100
-75

70

-GSDMD

- GBP2

GBP1 -/-

CASP4 -/-

15

Stm coating (%)
FL
-G
N T SD M
D
-G
C T SD M
-G
D
P < 0.0042
SD
MD
FL
P < 0.0042
-G
SD
NT
-G MD
C T SD M
NS
-G
SD D
NS
MD
FL
-G
N T SD M
-G
D
C T SD M
NS
-G
D
SD
NS
MD

P < 0.00007
P < 0.0005
P < 0.0004
P < 0.0005
P < 0.004
P < 0.004

-100
-75

92

-CASP4

-GFP

Coat reconstitution in GBP

-100
-75

100

-GBP2

P < 0.0002

All

12
9

6
3
0

Recruitment onto bacteria
GBP2
GBP3
GBP4
[FL-GSDMD]-CASP4
(Prebound)

+- +
++
IgG
-37
-50
-37
-50

Stm LPS Binding

H
Relative mP (%)

.2
22
1q

s9

GB
P

W
T Ca
Flag + Flag-GSDMDL192D - +
++
C258A
Myc-CASP4
IP: IB: -Myc
-Flag IB: -Flag
Input IB: -Myc
IB: -Flag

CASP4
NT-GSDMD
Activation and cleavage

rGBP1
LPS only

120
90
60

Kd = 3.776 M

30
0
-3 -2 -1 0 1 2 3 4

120

0

rGBP2
LPS only

120

rCASP4C258A
LPS only

90

90

90

600 K = >10 M
d
300

60 Kd = >10 M
30

60 Kd = >10 M
30

60 Kd = 0.313 M
30

00
0
0
-3 -2 -1 0 1 2 3 4
-3 -2 -1 0 1 2 3 4
-3 -2 -1 0 1 2 3 4
Log 10 protein concentration (nM)

0
-3 -2 -1 0 1 2 3 4

5 of 18

,

mediated killing (Fig. 3G). Thus, insertion of
human GBP1 seems to disrupt lateral LPS-LPS
interactions to compromise OM integrity. This
not only activates the caspase-4 inflammasome

y

GSDMD fused to GFP reintroduced into GSDMD–/– cells activated with IFN-g and
infected with Stm for 2 hours. The GFP channel is pseudocolored turquoise, and
the targeting of FL-GSDMD to bacteria is indicated by a dashed circle. Nuclear
staining was done with 4′,6-diamidino-2-phenylindole. Scale bar, 5 mm. (Bottom)
Loss of full-length GSDMD targeting in GBP1–/– and CASP4–/– cells. Quantitation
of GSDMD targeting in IFN-g–treated knockout cell lines. n = 74 to 124 events
for each group. Micrographs and quantitation were from at least three independent
experiments. NT, N-terminal fragment; CT, C-terminal fragment. (F) Two-step
hierarchical model showing GBP1 dependency and recruitment of full-length
GSDMD by caspase-4. (G) Coimmunoprecipitation of FLAG-GSDMDL192D by
Myc-CASP4C258A [to reduce cytotoxicity (39)] in wild-type and GBPD1q22.2 mutant
HeLa cells. One of two similar experiments is shown. (H) Binding curves for recombinant
coat proteins to Salmonella LPS–Alexa Fluor 488 in fluorescence anisotropy assays
under physiological pH and temperature. mP, maximal polarization. Mean ± SD
determined in triplicate for each protein concentration. Representative of three
independent experiments. Significant one-way ANOVA values with Bonferroni
post hoc test shown in (D) and (E).

Such dependency was confirmed by using nonhydrolyzable GTP-g-S or GDP.AIF3– that not
only prevented rGBP1 coat formation and LPS
release (Fig. 3F) but also subsequent rAPOL3-

1 March 2024

120

rGBP4
LPS only

y g

coincubating rGBP1 with human rAPOL3 led
to loss of bacterial viability (Fig. 3G) because
OM disruption by GBP1 allows APOL3 access
to the Stm inner membrane underneath (13).

120

rGBP3
LPS only

900

Fig. 2. Human GBP1 initiates a multiprotein platform for cytosolic LPS recognition. (A) Multicolor confocal imaging of the reconstituted coat complex (GBPCOAT450–708) in GBPD1q22.2 mutant cells (chr, chromosome). Stepwise omission of
each coat component revealed GBP1 dependence. Circles depict Stm targeting. IFN-g
was added to ensure that sufficient caspase-4 was expressed endogenously when
using mT-Sapphire-GSDMD. Micrographs are representative of at least three
independent experiments. (B) Immunoblot of GBP1, GBP2, caspase-4, and GSDMD by
specific antibodies. Anti-GFP was used to detect related fluorescent proteins fused to
GBP3, GBP4, caspase-4, or GSDMD in reconstituted GBPD1q22.2 cells. (C) Caspase-4–
dependent GSDMD recruitment shown by exchanging mOrange-GBP2 with mOrangeGSDMD because GBP2 is dispensable for reconstituted coat formation in nonactivated
GBPD1q22.2 cells. Scale bar, 2 mm. One of three independent experiments is shown.
(D) Quantitation of Stm targeted by each coat protein when all are present or after
individual omission in reconstitution assays. Sapphire-GSDMD used throughout except
when it was coexpressed with CASP4, where it was used as in panel (A). n = 156 to
208 events for each group from four to five independent experiments. (E) (Top)
Widefield imaging of full-length (FL), N-terminal, or C-terminal fragments of human

Zhu et al., Science 383, eabm9903 (2024)

y

G

LPS binding
GBP1

g

GB
GB P1
GBP2
GB P3
CA P4
GS S P 4
DM
D
GB
GB P1
GBP2
G P3
CABP4
GS S P 4
DM
D
GB
P
GB 1
GBP2
G P3
CABP4
GS S P 4
DM
D
GB
GB P1
GBP2
GB P3
CA P4
GS S P 4
DM
D
GB
GB P1
GBP2
GB P3
CA P4
GS S P 4
DM
D
GB
GB P1
GBP2
G P3
CABP4
GS SP4
DM
D
GB
GB P1
GBP2
GB P3
CA P4
GS S P 4
DM
D

Stm coating (%)

30
25
20
15
10
5
0

C

chr.1q22.2

p

F

-GBP1

Coat reconstitution in GBP
- GBP1

GBP

IFP24- Orange- Venus- Emerald Sapphire Merge
GBP1 GBP2 GBP3 -GBP4 -GSDMD

D
All present

B

chr.1q22.2

NS
NS
NS

Coat reconstitution in GBP chr.1q22.2
IFP24- Orange- Venus- Emerald Sapphire Merge
GBP1 GBP2 GBP3 -GBP4 -CASP4

-CASP4 -GBP4 -GBP3 -GBP2 -GBP1

A

RES EARCH | R E S E A R C H A R T I C L E

B

A

In cellulo 3D-SIM
GBP1

LPS

Merge

PG

Uncoated

Coated

C

D

rRFP-GBP1F (+ GTP)

rRFP-GBP1F (- GTP)

DNA

rGBP1 (+ GTP)

rGBP1 (- GTP)

20 m

20

800
600
400
200

2

-7

2

-6

2-5 2 2-3 2 2-1 2 21
-4

-2

0

rRFP-GBP1 F ( M) + GTP

0

-

+ -

Zhu et al., Science 383, eabm9903 (2024)

+ + + + +

60
40
20
0

rAPOL3
GTP

+ + + +
- + - -

GTP (2 mM) imaged by confocal microscopy. Representative of 10 to 12 micrographs
shown. (Bottom) rRFP-GBP1F coating across increasing diluents imaged by
confocal microscopy. (E) Best-fit interpolation curve of means ± SD together with
regression analysis revealed half maximal coat Km and Hill slope values. One of three
similar experiments is shown. (F) Soluble lipid A release by LAL in rRFP-GBP1F
coat reconstitution assays (triplicate ± SD) on live unfixed bacilli. Substrate analogs
and GBP mutants are denoted. One of six independent biological experiments is
shown. EU, Endotoxin Unit. (G) Stm killing assay by rAPOL3 requires LPS disruption by
the GBP1 coat complex. Significant one-way ANOVA values with Bonferroni post hoc
test are shown for (F) and (G). One of three independent experiments is shown.

and insertion into the bacterial outer leaflet
for triggering LPS disruption. GBP1 conformers could be delineated with powerful imaging tools such as cryo–electron microscopy
(cryo-EM) and cryo-ET. We engineered bacterial minicells and outer membrane vesicles
(OMVs) because they are considerably smaller than isogenic rod-shaped bacilli but retain

1 March 2024

80

NS

intact features of the pathogen OM, unlike
artificial liposomes or soluble LPS (Fig. 4A).
More notably, the reduced sample thickness
of minicells improves resolution in cryo-ET
samples (34), and negative-stain EM initially
confirmed that Stm minicells and OMVs are
coated by rRFP-GBP1 like the parental Stm
1344 strain (fig. S6, A and B).
6 of 18

,

Our reconstituted coat complex provided us
with the opportunity to view GBP1 assembly

-

100

y

A bacterial minicell system enables cryo-EM
and cryo-ET of the native GBP1 coat

-

120

y g

Fig. 3. The human GBP1 coat complex triggers LPS release in cellulo and in
cell-free systems. (A) Fluorescent labeling strategy for LPS with copper-free CLICK
chemistry with a DIBO alkyne intermediate. (B) 3D-SIM imaging of cytokine-activated
HeLa cells (1000U IFN-g, 18 h). Images were collected at 2 hours postinfection with
prelabeled Stm at an MOI of 20. Endogenous GBP1 was detected by anti-GBP1,
and LPS was detected with anti–Sal-O antibody for verification (pseudocolored
magenta). Maximum intensity projection of one of five similar 3D-SIM images. Scale
bar, 2 mm. (C) GBP1-coated cytosolic bacteria harbor flagellin detected by anti–Fli-C
antibody in 3D-SIM imaging. Maximum intensity projection of one of four similar
3D-SIM images. (D) (Top) Farnesylated rRFP-GBP1F (2 mM) coat assembly on Stm ±

pathway but allows the passage of small
antimicrobial proteins such as APOL3 to directly kill pathogenic bacteria (13).

-

rRF
rRF rR P-G
B
rRF P-G FP-G P1
P-G BP1 BP
rRF
/GT 1
P BP
P
S
rRF -GB 1/GM
P
P-G 1/G PP
BP DP NP
1
.A
rRF rR /GMP IF3 .A
P-G FPBP GBP IF4 1 SD D 1 S52
rRF
N
10
3 ,1
P
0
rRF -GBP 8N N
rRF P-G 1 D 184
N
P-G BP
BP 1 C 589
1 R58 S

GTP
0

Stm viability (% control)

“Coat Km ”
(225nM)

P < 0.0001

1000

Substrate
analogs

rRF
P-G
rR
rRF FP-GrRFP BP1
P-G BP -GB
BP 1/G P1
1/G TP
DP S
.AI
F3 -

40

1200

4 -5

Stm Lipid A (EU.mL -1)

60

G

rGBP1
mutants

y

rRFP-GBP1 F coated Stm (%)

Hill slope
(5.122)

21

g

rGBP Substrate
GTP analogs

GBP1 coat complex assembly

20

2-1

P < 0.002

2-2

p

F

E

80

2 m

rRFP-GBP1F ( M) + GTP
2-3

100

rGBP1 (+ GTP)

20 m

Endogenous
coat

A

Flagellin

86

GBP1

RES EARCH | R E S E A R C H A R T I C L E

A

B

Stm

waaG::pBAD-ftsZ

OMVs

rRFP-GBP1 (+ GTP)

rRFP-GBP1 (- GTP)

Outer Leaflet of
Outer Membrane
(OLOM)
Inner Leaflet of
Outer Membrane
(ILOM)

GBP1 coat

OLOM
ILOM

50 nm

50 nm

1. Stm

minicell

2. Stm

(n = 11,889)

minD

minicell

(n = 12,487)

3. Stm

waaG::pBAD-ftsZ

OMV

(n = 6,107)

G

E

waaG::pBAD-ftsZ

minicell

H

35
30
25
20

1 2 3

p

Stm

G

rGBP1 length (nm)

D
waaG::pBAD-ftsZ

28
.0
6
27 ±1.
0
.9
6± 4
1
27
.8 .53
6±
1.
24

C

NS

IM
OM

F

g

20 nm
Inset
Human GBP1 coat
complex

y

OM
IM
20 nm

1 March 2024

measurements found that GBP1 spanned ~25
to 27 nm, tightly juxtaposed over the entire
bacterial surface (fig. S6E). In a recent crystal
structure of farnesylated human GBP1 bound
to the nonhydrolyzable GTP analog, GMPPNP
(PDB ID 6K1Z), the protein is half this length
(12.89 nm) (22), with the final a13 helix folded
back onto the a12 C-terminal segment in a
“closed” conformation (fig. S6, E and F). To
span the ~25 to 27 nm measured with cryoEM, at least two “closed” GBP1 molecules
vertically positioned on top of one another
would be needed, or four to six molecules
7 of 18

,

Zhu et al., Science 383, eabm9903 (2024)

dimer near the point of OM insertion. Both
O-antigen and outer core segments are also
dispensable for GBP1 coat attachment and
downstream signaling inside human cells because the lipid A region is still intact (Fig. 1, G
and H). Our dual genetic strategy was therefore devised to limit nonspecific sample noise
during collection of tilt images by cryo-ET.
Cryo-EM at 200 kV revealed successful reconstitution of the GBP1 coat complex on
StmDwaaG::pBAD-ftsZ minicells and OMVs in
comparative samples with rRFP-GBP1 ± GTP
(Fig. 4B and fig. S6, D and E). Preliminary EM

y

Minicells arise from abnormal asymmetric
cell division (Fig. 4A); we found that overexpression of a septation gene, ftsZ, yielded
the smallest Stm minicells (~150-300 nm) that
could still be isolated intact by differential
centrifugation along with OMVs (fig. S6C).
We deliberately introduced ftsZ overexpression into a waaG-deficient background lacking the LPS O-antigen and outer core segment
(StmDwaaG::pBAD-ftsZ) because these segments
have unstructured density that interfere with
subtomogram averaging (45) and removing it
would allow us to see more of the native GBP1

(3. StmDwaaG::pBAD-ftsZ). (D) Violin plots of GBP1 conformer length (nanometers, mean ± SD) for each measured sample. NS across all groups;
found using one-way ANOVA with Bonferroni post hoc test from three to
four independent experiments. (E) Representative tomographic slice of
StmDwaaG::pBAD-ftsZ minicell completely coated with rRFP-GBP1F after GTP
hydrolysis. (F) Enlarged view of the boxed inset from (E). (G and H) 3D
segmentation of the same rRFP-GBP1F–coated StmDwaaG::pBAD-ftsZ minicell
shown in (E) and (F). Tomographic series collected over 102° tilt range. OM,
outer membrane. IM, inner membrane.

y g

Fig. 4. Native GBP1 coat complex on the Salmonella surface observed by
cryo-EM and cryo-ET. (A) Strategy for constructing StmDwaaG::pBAD-ftsZ minicells and
OMVs missing LPS O-antigen and outer core for reduced noise detection of the
native GBP1 coat complex. (B) GTP dependency of rRFP-GBP1F coat formation on
the StmDwaaG::pBAD-ftsZ OMV surface shown in 200-kV cryo-EM images. Black dots
are 6-nm diameter fiducial beads. (C) Measurement of coat length from 300 kV
cryo-ET images by using computer script quantitation (yellow) for LPS-O
antigen and outer core truncated minicell (1. StmDwaaG::pBAD-ftsZ), LPS wild-type
minicell (2. StmDminD), and LPS-O antigen and outer core truncated OMV

RES EARCH | R E S E A R C H A R T I C L E

arrayed horizontally and perpendicular to the
outer leaflet (46). Alternatively, AlphaFold2
modeling (47) found that GBP1 may undergo
dynamic C-terminal extension of its GED to
present a fully unhinged C15 farnesyl group
to the OM (fig. S6F). This “open” extended
conformer had a predicted length of 278 Å,
which also fits the EM measurements. Each
potential configuration was probed by cryoET along with biochemical evidence to discern
how GBP1 directly operates on the bacterial
surface under native conditions in the presence of its bona fide substrate, GTP.
Cryo-ET reveals thousands of open GBP1
conformers insert into the bacterial OM

y
,

8 of 18

y g

1 March 2024

y

The conformationally dynamic model of human GBP1 was further validated by three separate approaches. Firstly, we devised sequential
wash conditions to reduce protein crowding
within the coat complex itself. It enabled us to
observe isolated GBP1 dimers still attached to
the surface of Stm OMVs and minicells in vitro
(Fig. 6, A and B, and fig. S12, A and B). Tomographic images of 66 isolated GBP1 proteins
found that all had the globular LG positioned
at the perimeter with the extended C terminus
anchored underneath at the base; these isolated dimers spanned 25.08 to 27.44 nm, sug-

g

Zhu et al., Science 383, eabm9903 (2024)

A conformationally dynamic model for GBP1
defense complex assembly

gesting an extended conformation (Fig. 5, D to
F, and fig. S12, C and D). Secondly, nanogold
labeling validated this orientation. We chose
1.8-nm Ni-NTA-nanogold particles to bind a
histidine tag (His6) fused to the N-terminal
LG of GBP1 for detecting its location within
the mature coat. Labeling was observed exclusively at the outer border (Fig. 6C), which was
consistent with the LG being positioned at the
top of the coat, as seen for isolated dimers (Fig.
6, A and B). Notably, the exposed bacterial
OM did not bind Nanogold particles in the
absence of His6-GBP1, indicating that the labeling signal originated from the protein itself (Fig.
6C). Attempts to introduce a functional His6
sequence near the GBP1 C terminus proved
impractical owing to the presence of the CaaX
motif used for farnesylation and the polybasic R584 to 586A sequence needed for
OM anchorage. Nonetheless, cryo-ET of both
washed OMVs and nanogold labeling validated both the length and the orientation of
the GBP1 dimer on the gram-negative bacterial surface.
Lastly, a requirement for dynamic opening
of the GBP1 GED was assessed through a
cysteine replacement assay. Seven paired cysteine (Cys) mutants (I369C-E533C, E366C-G530C,
I365C-H527C, R362C-E526C, I365C-G526C,
G361C-S523C, and G389C-K520C) were generated spanning the C-terminal GED a12 helix
and its spatially adjacent a7 MD helix to enable disulfide cross-linking of the “closed”
conformer (fig. S13A); this prevented the
C-terminal a12,13 region from opening to
~280 Å unless exposed to a reducing agent
such as dithiothreitol (DTT). To ensure that
the introduction of Cys-Cys pairs did not lead
to gross structural alterations having effects
beyond opening the C-terminal hinge, we
first tested if they could still target Stm within the reducing environment of the human
cytosol. All seven Cys-Cys pairs coated cytosolic Stm in HeLa cells (fig. S13B). Of these
mutants, we chose the middle R362C-E526C
mutant with the shortest Cys-Cys interbond
length (2 Å) to make recombinant RFP-GBP1
for direct cell-free coating assays after confirming its normal activity in situ (Fig. 6, D
and E and fig. S13, A and B). In the absence
of DTT, the “closed” disulfide-linked R362CE526C conformer completely failed to coat
Stm even when given its GTP substrate, unlike its wild-type rRFP-GBP1 control (Fig. 6F).
Addition of DTT, however, fully restored bacterial encapsulation by the R362C-E526C
mutant (Fig. 6F). Thus, biochemical evidence
supports the requirement for dynamic opening of the human GBP1 dimer to ~280 Å for
assembly on the bacterial surface. The C15
farnesylated tail anchors GBP1 to the OM
with the catalytic LG domain positioned at
the perimeter to generate this distinctive host
defense platform.

p

We next used 300-kV cryo-ET to acquire highcontrast 3D images in dose-fractionated mode
of a human GBP1 coat complex fully assembled on the bacterial surface (Fig. 4, C to H,
and fig. S7, A to L). rRFP-GBP1 fluorescence
initially helped us to locate coated bacilli
within vitrified samples, and the linkage of
RFP did not alter the overall length of GBP1
(fig. S6F). Mean conformer lengths of ~280 Å
were found across 30,483 measurements of
StmDwaaG::pBAD-ftsZ and StmDminD minicells plus
OMVs (Fig. 4, C and D, and movies S4 and
S5). Elongated GBP1 conformers radiating
from the bacterial surface were discernible
within multiple tomographs (Fig. 4, E to H,
and fig. S7G). The 44,891 particles collected
during coarse classification yielded 15,683
particles for 3D segmentation of this native
host defense complex (Fig. 4, G and H). It
resolved a massive coat fully surrounding
StmDwaaG::pBAD-ftsZ minicells up to 384 nm
in diameter (movie S6). Zoomed-in views revealed what appear to be vertically upright
GBP1 conformers often aligned in register when
attached to the OM (Fig. 5A).
To verify individual protein configurations
within this native coat, we next enlisted tagless rGBP1 to ensure that RFP was not interfering with lateral packing and applied tighter
final masks on 12,858 particles for subtomogram averaging (STA) (Fig. 5B and fig. S8, A
to C). Such refinement necessitated user-side
scripting to accommodate the small 68-kDa
size of untagged GBP1, its flexibility, and the
dense array on the bacterial OM. Using this
strategy, we were able to resolve the larger
GBP1 dimer and its smaller monomeric subunit to ~9 to 17 Å directly on the bacterial surface (Fig. 5, C and D, and fig. S9, A to D). At
these resolutions, the LG, MD, and GED regions could be delineated (Fig. 5D and fig. S9,
A to C). Native GBP1 adopts an asymmetric
dimer with the LGs twisted and tilted relative
to each other when attached to the bacterial
OM (Fig. 5D). This may resemble the tilted
LGs found in the crystallized dimer of human
GBP5 (PDB ID 7E5A) (23), although the a12
and a13 GED helices appear extended into the

LPS inner core and lipid A regions for native
GBP1, yielding a minimum length of 22.4 nm
(Fig. 5, D and E). Hence, one possibility arising from our cryo-ET analysis is that human
GBP1 operates as an “open” conformer in its
functional state (Fig. 5E); this contrasts with
all monomer or dimer GBP crystal structures
reported so far, which position the GED in a
“closed” conformation (Fig. 5E).
Enumerating these native conformers and
the bacterial surface area present within each
subtomogram segment, along with overall size
measurements of StmDwaaG::pBAD-ftsZ minicells,
we found 11,760 ± 735 GBP1 proteins assembled
on the OM, yielding a >1-GDa structure. For a
mature rod-shaped Stm, this extrapolates to
~22,000 to 32,000 GBP1 molecules per bacterium, which is consistent with earlier light
microscopy estimates in situ. Cryo-ET thus
provided a new structural view of this giant
immune complex in which thousands of GBP1
dimers may stretch their C-terminal GED domain to 280 Å for anchoring the farnesyl tail
and insertion of the nearby polybasic patch.
This mesoscale coat also appeared evident
in situ. Focus ion beam (FIB)–milled lamella
together with correlative light and electron
microscopy (CLEM) found GBP1-coated bacteria inside human cells (fig. S10, A to D).
Longer StmpBAD-ftsZ bacilli were needed to detect rare coating events in thin ~150- to 300-nm
FIB-milled lamella, and EGFP-GBP1 expression in a GBPD1q22.2 background ensured
that it was the only GBP family member targeting bacilli (fig. S10A). This dual strategy
succeeded in resolving what appear to be GBP1
coat complexes on CLEM-validated bacteria
that had escaped into the cytosol; vacuolar
Stm were devoid of EGFP-GBP1 as a negative
control (fig. S10, B to D, and fig. S11, A and B).
Individual GBP1 dimers could not be delineated at this resolution; however, the ~30-nm
EGFP-GBP1 boundary surrounding Stm closely resembled the length of reconstituted coats
in cell-free assays (fig. S11A). Hence, the conformational changes seen in vitro probably
operate in cellulo to generate this massive
GBP1 defense complex.

RES EARCH | R E S E A R C H A R T I C L E

A

B

2D tomographic slice

2D tomographic slice

C

x,z slice

x,y slice

Dimeric mask

LG

MD

20 nm

GED

OM

Human GBP1 coat complex

Monomeric mask
on dimer

LG

MD

GED

Human GBP1 coat complex
20 nm

D
Salmonella OM
20 nm

OM

STA dimeric mask
C1 symmetry

STA monomeric mask
C1 symmetry

9.8 nm

5.7 nm

p

LG
4.5 nm

Native GBP1 conformers

Salmonella OM

MD
7- 11
7.9 nm

22.4 nm

22.4 nm

20 nm

g

GED
12- 13
10.0 nm

Native GBP1 conformers

y

Salmonella OM

20 nm

Inner core
Lipid A
Salmonella OM

E
Crystallized
monomer
(+ GMPPNP)

Crystallized
dimer
(+ GDP.AIF3-)

GBP1 monomer
subunit model
(+ GTP)

147°
rotation

STA dock
monomeric-dimeric
masks of native GBP1

Dynamic GBP1
dimer model
(+ GTP)

LG

13

y g

Fig. 5. Human GBP1 could adopt a dynamically “open” conformer when assembled on
the bacterial OM following GTP hydrolysis.
(A) 3D segmentation of the human GBP1
coat complex with multiple upright GBP1 conformers attached to the bacterial OM, shown
at 31.2 Angstroms by cryo-ET in the presence
of 2-mM GTP. A 2D tomographic slice is shown
at the top. (B) Representative 2D tomographic
slice of StmDwaaG::pBAD-ftsZ OMVs coated with
tag-free GBP1 in the presence of GTP. (Inset)
The yellow box highlights elongated GBP1 conformers containing dimers (yellow arrows,
white boxes). Scale bar, 20 nm. (C) (Top)
Asymmetric GBP1 dimer on the bacterial OM
captured through a larger mask. Image shows
the 182nd of 256 slices used for generating
a 3D volume of the GBP1 dimer STA within
256×256×256 voxels. The tomogram was
rotated counterclockwise at 45° to reveal both
LGs of the dimer. (Bottom) Smaller mask on
one monomeric subunit of the native GBP1 dimer,
which yielded higher resolution. Image shows
the 155th of 256 slices used for generating a 3D
volume of the GBP1 STA within 256×256×256
voxels. (Right) Cross-sections (x,z slices) of the
GBP1 STA at the LG, MD, and C-terminal GED
in the original orientation. (D) STA of native
human GBP1 directly on the bacterial outer
membrane. Native dimer (17-Å final resolution)
and monomeric subunit of the dimer (9.7-Å final
resolution) show a12 and a13 helical domains
extending down into the bacteria outer membrane. (E) (Left) Monomer and dimer of models
of crystallographic GBP structures in the presence of substrate analogs. (Right) Computational
GBP1 monomer subunit and dimer models that
incorporate a tilted LG domain of the GBP5 dimer
(PDB ID 7E5A) together with an extended GED
on the bacterial membrane following hydrolysis of
its natural substrate, GTP. Monomeric STA
docked onto its dimeric counterpart from cryoET studies is placed in between for comparison.
The positions of LG, MD, and GED are noted,
along with the CaaX motif for C15-farnesyl
attachment.

CaaX
MD
12

PDB:7E5A
(Human GBP5)

y

PDB:6K1Z
(Human GBP1)

12

,

GED

+ GTP
hydrolysis

13

CaaX
Salmonella OM

Discussion

Higher-order protein assemblies help amplify
innate immune signaling and spatially regulate signal propagation by localizing partners at the site of ligand recognition (1, 3).
Such assemblies form on the host plasma
membrane, mitochondria, peroxisomes, and
chloroplasts (1, 48). They also occur in the
cytosol, nucleus, endoplasmic reticulum, and
Zhu et al., Science 383, eabm9903 (2024)

endosomal network, in some cases yielding
membraneless condensates through liquidliquid phase separation (49), as recently discovered for plant GBPLs during cell-autonomous
defense against bacterial phytopathogens
(7, 8).
By contrast, human GBP1 builds a multiprotein complex upon a completely foreign
object, the gram-negative bacterial OM. This

1 March 2024

huge nanomachine solicits GBP2 to GBP4,
caspase-4, and full-length GSDMD as part of a
six-member platform to propagate cytokine
and cell death signaling in multiple human
cell types (10, 16). We found that human GBP1
is the central organizer of this platform and
enlists nucleotide-dependent hydrolysis to selfassemble like other members of the DLP superfamily of large GTPases (18). GTP-driven GBP1
9 of 18

RES EARCH | R E S E A R C H A R T I C L E

Fig. 6. Validation of the dynamically “open”
conformer model for the GBP1 coat complex.
(A) Direct visualization of isolated GBP1 dimers
on StmDMinD minicell after GTP hydrolysis followed
by the sequential wash protocol to remove
crowdedness. (Inset) A zoomed-in view showing
isolated GBP1 conformers (yellow arrows).
(B) 3D segmentation of the StmDMinD minicell
after wash treatment. (Inset) Zoomed-in view
of isolated GBP1 depicting the LG domain (dashed
circles) at the periphery with helical stalk
underneath. (C) Topological evidence of the
His6-GBP1 upright conformer. Representative
tomographic slice of StmDMinD minicell (top) with
and (bottom) without the His6-GBP1 coat complex
in the presence of GTP. (Insets) Zoomed-in
views of top (red) and bottom (blue) images.
His6-GBP1 GD is labeled at the outer perimeter
of the coat with 1.8-nm Ni-NTA-nanogold particles
are shown in red arrow. Scale bars, 100 nm
(top and bottom), 20 nm (insets). One of three
independent experiments. Tomographs
denoised with cryoCARE software to delineate
nanogold particles. (D) Cross-link design between
a7 in the MD and a12 in the C-terminal GED.
(Inset) Residues selected for cysteine substitution
for forming a disulfide linkage. (E) In cellulo
examination of mRFP-GBP1 with Cys replacements
targeting onto the Salmonella surface, indicating
that cysteine mutations do not grossly alter
GBP1 function inside human cells. GFP-expressing
bacteria targeted with RFP-GBP1R362C–E526C,
indicated by arrows. The GFP channel is pseudocolored turquoise. One of two similar experiments
is shown. (F) (Left) Release of the covalent
a7 to a12 crosslinked cysteines (recombinant
RFP-GBP1R362C–E526C) by DTT allows GTP-dependent
assembly on the bacterial OM in reconstitution
assays. (Right) Wild-type recombinant RFP-GBP1
is unaffected by the presence of DTT in
GTP-dependent coat assays. Scale bar, 2 mm.
One of three similar experiments is shown.
Single-letter abbreviations for all the amino acid
residues tested are as follows: A, Ala; C, Cys;
E, Glu; G, Gly; H, His; I, Ile; K, Lys; R, Arg; and S, Ser.

A

B

C

p
g

D

E
y

7

F
y g

-DTT

y
,

+DTT

cooperativity gave a sigmoidal Stm coating
curve on bacteria that steeply accelerated above
125 nM, resembling other “prionizing” proteins in which all-or-none responsivity occurs
once a concentration threshold is reached (3).
Anchorage to LPS may help accelerate GBP1
catalysis and oligomerization (16). Both GTPase
and GDPase activities contributed to GBP1 reZhu et al., Science 383, eabm9903 (2024)

sponsivity through GTP and GDP turnover as
part of a two-step enzymatic process (15, 21).
Because hydrolysis of GTP and GDP liberates
large amounts of Gibbs free energy, the GBP1
coat conforms to a nanomachine by performing
“work” in establishing this massive signaling
platform. It may explain why transition-state
analogs such as GDP.AIF3– fail to recapitulate

1 March 2024

+DTT

the native coat on intact bacteria (15) despite
helping GBP1 form dimers (20, 21). Instead,
multiple rounds of GTP and GDP hydrolysis
are needed to continually drive assembly of
adjacent dimers that probably undergo lateral
interactions to stabilize the coat in register.
Cryo-ET suggests that these lateral interactions could result from GTP-induced changes
10 of 18

RES EARCH | R E S E A R C H A R T I C L E

1 March 2024

HeLa (CCL-2) and 293T (CRL-3216) cells were
purchased from the American Type Culture
Collection (ATCC). Cells were grown in DMEM
supplemented with 10% (v/v) heat-inactivated
fetal bovine serum (FBS) at 37°C in a 5% CO2
incubator. Lentiviral (LentiCrisprV2; Addgene
Plasmid #52961) or retroviral (pMSCV-puro;
Takara 634401) transductions were done by
incubating dilutions of 0.45 mm filtered supernatants from transfected 293T cells with 8 mg/
mL polybrene for 24 hours. For selection of
stable transductants, 1 mg/ml puromycin was
included. For transient transfections, TransIT®LT1 (MIRUS; MIR 2300) was used according
to manufacturer’s instructions. To minimize
toxicity in microscopy experiments, 200 ng
of DNA was transfected per 24-well cover
slip. HeLa cells were stimulated with 500 to
1000 U/mL IFN-g for 18 hours.
11 of 18

,

GTP sodium salt (G8877), GTP-g-S (tetralithium
salt; G8634), GDP sodium salt (GDP; G7127),
GMP sodium salt (GMP; G8377), aluminum
trifluoride (449628), chloramphenicol (220551),
Ficoll (F5415), and biotin (B4501) (Sigma/
Millipore); Guanosine-5′-[(b,g)-imido]triphosphate
(tetralithium salt)/GMPPNP/GppNHp (Jena
Biosciences; NU-401-50); recombinant human
IFN-g (285-IF/CF; R & D Systems; 285-IF);
Salmonella minnesota LPS–Alexa Fluor 488
(ThermoFisher; L23356).

Cell culture and transfection
y

Reagents

y g

Antibodies used were anti-Flag M2 (F1804,
Sigma), anti-HA (16B12, Biolegend), anti-Myc
(9E10, ThermoFisher), anti-GFP (11814460001,
Roche; A0174, Genscript), anti-GST (1E5,
Origene), anti-DsRed (sc-390909, SCBT), antiGBP1 (sc-53857, SCBT), anti-GBP2 (sc-10581,
SCBT), anti-Salmonella O Group B antiserum
(240984, BD), anti-flagellin (FliC-1, BioLegend),
anti-b-actin (ab6276, Abcam), anti-GAPDH
41335; SCBT), anti-IL-18 (PM014; MBL), anticaspase-4 (clone 4B9; Enzo), and anti-GSDMD
(NBP2-33422, Novus Biologicals). See table S1
for applications.

Bacterial strains were generated in-house
or kindly provided by the following groups:
S. enterica serovar Typhimurium (Stm) strain
1344 and flagellin-deficient StmDflhD (Dr. Jorge
Galan); Stm UK-1 wildtype, Dwzy, DwaaL,
DwaaJ, DwaaI, DwaaG, DlpxR, DpagL, DpagP,
and c11088 (StmDlpxRDpagLDpagP triple mutant)
(Dr. Roy Curtiss III, Dr. Soo-Young Wanda)
(38, 40); P. aeruginosa L2 strain (Dr. Barbara
Kazmierczak), Bacillus subtilis (Dr. Farren
Isaacs), and L. monocytogenes 140203S (Dr.
Herve Agaisse).
The following Stm strains were constructed
on a 1344 isogenic background: StmDminD,
StmpBAD::ftsZ, StmmreB(K27E), StmmreB(D78V),
Stm mScarlet-I , and Stm eGFP . In addition,
StmDWaaG::pBAD-ftsZ was generated on the UK-1
background for cryo-ET. See table S2 for details.
To generate Stm MinD deletion and MinD/
waaG double deletion mutants in UK-1, the
lambda red recombinase system was used. A
kanamycin cassette with minD flanking sequences was amplified by primers minDKO-L
(GTTTACGATTTTGTAAACGTCATTCAGGGCGATG CGACtgtgtaggctggagctgcttcg) and
minDKO-R (GGAGATGTTCTTTAATCGGTTCTTCGCC ATTTTCTcatgggaattagccatggtcca)
with pKD4 vector as the template. The polymerase chain reaction (PCR) product was gelextracted and electroporated into wild-type
UK-1 and waaG-deletion UK-1 Stm competent
cells, which were expressing lambda red recombinase. Kanamycin (Km)–resistant strains
were plate selected, and kanamycin cassette
insertion into the minD gene was checked by
minD Km insert validation primers, minDKm-L (ATTTTGTAAACGTCATTCAGGGCG) and
minD-Km-R (gcagttcattcagggcaccg). To check
the deleted region of minD, primers minDWT-L (GCTGATCAAAGATAAGCGTACTGA)
and minD-WT-R (CGATGCCAGAATACCCAG
AATACG) were used.

y

Materials and Methods
Antibodies and reagents
Antibodies

Bacterial strains

g

Zhu et al., Science 383, eabm9903 (2024)

directly from human cells using fast protein
liquid chromatography (FPLC) also ensured
proper OM anchorage and insertion. C15
lipidation requires sequential addition by human farnesyltransferase, tripeptide removal
by CAAX carboxypeptidase, and carboxy-group
methylation by isoprenylcysteine carboxymethyltransferase (32). Postprenyl processing
thus brings the fully modified farnesyl tail almost adjacent (three amino acids apart) to
the triple-arginine patch, creating a powerful
bipartite anchor. OM insertion of the polybasic
motif likely undergoes electrostatic interactions with the negatively charged PO4– groups
of lipid A and inner core saccharides, whereas
farnesylation makes the tail more hydrophobic
(15, 16). Together, these modifications enabled
STA of human GBP1 bound to gram-negative
bacteria. It should help annotate tomographic
densities of FIB-milled human cells now that
we have established initial conditions to detect
GBP1-coated bacteria in cellulo.
Our 3D reconstruction from cryo-ET elucidates the mesoscale architecture of a distinctive host defense structure, the massive
GBP1 coat complex, that cooperatively functions
on the surface of microbial pathogens inside
the human cytosol. Our findings reinforce the
importance of higher-order protein assemblies
within the innate immune systems of animals
and plants. These nanomachines concentrate
signaling and effector proteins for rapid mobilization of cell-autonomous resistance to
infection.

p

to the twisted LG itself, given that GBP1 dimers
appeared slightly asymmetric when embedded
within the native coat complex.
The functional importance of GTP hydrolysis was further reinforced by its ability to trigger GBP1-dependent LPS release, which activates
caspase-4 and sensitizes bacteria to APOL3mediated killing (13). Again, transition-state
(GDP.AIF3–, GMP.AIF4–) or nonhydrolyzable analogs (GTPgS or GMPPNP/GppNHp) failed to
elicit GBP1-dependent disruption of the OM
to liberate LPS. These results highlight the
limitation of using substrate mimics to probe
GBP1 defense complex formation and functionality on the pathogen surface, despite their
usefulness in earlier studies of isolated GBP1
(20, 21, 24).
Cryo-ET provided us with structural information about GBP1 and its supramolecular
architecture on the gram-negative Salmonella
surface in its native state. Previous crystal
structures of GBP1 used full-length recombinant protein produced in Escherichia coli to
capture the monomeric (apo) state ± GMPPNP
(20, 22). The N-terminal G-domain also produced
in bacteria was crystallized as a homodimer in
the presence of multiple nonhydrolyzable
nucleotides (GMPPNP, GDP.AIF3–, GMP.AIF4–,
and GMP) (21). Both full-length human GBP1
crystal structures position a12 and a13 GED
helices tucked up against the a7 to a11 MD in
a folded conformation, whereas under native
conditions, we found that recombinant GBP1
produced in human cells is fully stretched with
the a12 and farnesylated a13 helices inserting
vertically into the bacterial outer leaflet.
Identifying an “open” GBP1 conformer as the
principal repetitive unit of the mature coat
complex reinforces the capacity of cryo-ET to
yield insights into the behavior of assembled
proteins on their natural membrane targets
and in the presence of their natural substrate,
which is in this case GTP (50).
Our attempts to resolve the GBP1 defense
complex assembled on its physiological surface proved challenging across two scales:
first, the small size of GBP1 dimers (~140 kDa),
and second, the giant ~1.5-GDa size of the
final polymer. Determining the length and
orientation of individual GBP1 molecules to
9.7 Å among thousands of identical proteins
benefitted not only from post-acquisition
masking but also from bacterial genetics plus
recombinant protein preparation. Purified 150to 300-nm minicells and OMVs with reduced
thickness helped improve the quality of our
tilt images. It was further aided by genetic
removal of the O-antigen and outer core to
prevent unstructured LPS density from interfering with GBP1 C-terminal resolution (45).
Part of the GED of GBP1 otherwise hidden
within the inner leaflet could be reconstructed
from tomographs to help confirm the dynamically open model. Farnesylated GBP1 purified

RES EARCH | R E S E A R C H A R T I C L E

CRISPR-Cas9 cell lines and stable complementation

1 March 2024

HeLa cells were grown on 12-mM high performance cover glass #1.5h (Thorlabs CG15KH1)
for microscopy of fixed samples. For live imaging, cells were seeded on four-well chambers
(Cellvis C4-1.5H-N) with 1.5 high performance
cover glass. Here seeding occurred 48 hours
prior to imaging to reach 80% confluency on
the day of infection. They were treated with
IFN-g for 18 to 24 hours prior to imaging.
Bacteria were added to cells as described for
infections at an MOI of 20. Images were analyzed on a DeltaVisionTM OMX SR Blaze microscopy system (GE Healthcare) or a laser
scanning confocal model SP8 (Leica).
For ultrafast live imaging, GBP1–/– HeLa
cells were transfected with RFP-GBP1, induced
with 1000-U/mL IFN-g for 18 hours, and infected with EGFP-Stm at MOI 20. After 40 min,
the media was changed to DMEM with 30 ug/
mL gentamycin and the sample was imaged
starting 60 min postinfection at 37°C, 5% CO2.
Images were collected every 45 s using OMXSR Blaze microscope (GE) in 3D-SIM 512 ×
512-pixel mode at ~180 frames sec−1. Images
represent maximum intensity projections of
deconvolved z-stacks of Moire fringe patterns
(Softworx, GE). Postacquisition calculations
for real-time voxel (boxed) assembly events
used Imaris (Oxford instruments) software.
When combined with 3D atomic structure
volumes of GBP1 (PDB ID 1F5N) and RFP
12 of 18

,

The KDO of log-phase Stm was labeled by
using CLICK-mediated according to the manufacturer’s instructions (Jena Bioscience). Briefly, an azide modification of the C8-position of
KDO with a biorthogonal azido group was
introduced to prevent reverse metabolism by
KDO-8-P phosphatase. This 8-azido-8-deoxyKDO modification enabled a biotin group
within a DIBO alkyne dye intermediate (Jena
Bioscience CLK-A105P4) to be introduced via
Cu(I)-free CLICK chemistry for addition of the
Alexa Fluor 594 tag (Jena Bioscience CLK-1295).
For infection experiments, we conducted click
chemistry reactions on bacteria that had already

Microscopy
3D-SIM and multicolor confocal microscopy

y

KDO-azide Cu2+-free CLICK chemistry

Fluorescent blue or red D-alanine analogs
[HCC-amino-D-alanine, HADA (MCE HY131045); TAMRA 5-amino-D-alanine using 5Carboxytetramethylrhodamine (ab145438)]
were incorporated as described (46). Infection
of HeLa cells at MOI 20. After 40 min of infection the media was changed to DMEM with
100 mg/mL gentamycin after next 1 hour to
30 mg/mL gentamycin. Cells were fixed with
PFA at 2 hours postinfection and immunestained for GBP1 or LPS. At least 10 GBP1positive and 10 GBP1-negative fields of view
were collected using the DeltaVisionTM OMX
SR Blaze microscopy system (GE Healthcare)
in the 3D-SIM mode (512 × 512 pixels, 1-ms
exposure, 125-nm step, 8 z slices, 15 images
per slice). All images were subjected to processing to widefield image, deconvolution,
and maximum intensity projection for semiautomatic analysis in CellProfiler (Broad Institute, Open Scholar, 2021).

y g

Zhu et al., Science 383, eabm9903 (2024)

293E cells (ATCC HEK CRL-1573) expressing
GFP or GFP-GBP1 mutant constructs were
plated at 2 × 105 cells in each well of a sixwell plate. Cells were then treated with 20-mM
Azido farnesyl pyrophosphate (Cayman C10248)
for 18 hours followed by cell lysis in 500 mL of
20-mM Tris (7.5), 100-mM NaCl, 1% TX-100
buffer containing Roche protease inhibitors.
Cells were further sonicated (30% power Virtis
Virsonic 600, three times for 1 min each) on
ice to liberate membrane bound proteins.
Lysates were centrifuged at 21,000g for
10 min. Supernatants were transferred to a
new Eppendorf tube and were treated with
100-mM Biotin dibenzocyclooctynol (DIBO)
(Thermo C10412) overnight at room temperature in the dark. 1 mg Roche monoclonal
anti-GFP antibody was added to each IP followed by a 2 hours incubation at 4°C. 20 mL
of preequilibrated Protein G Sepharose (Cytiva
17-0756-01) was added to each reaction for an
additional 2 hours. Beads were pelleted at
4000 g for 5 min followed by eight washes
in lysis buffer. Samples were subsequently
eluted by 100 mL addition of 2X SDS-Sample
Buffer followed by heating at 100°C for
20 min for immunoblotting. Immunoblots
were performed with streptavidin-HRP and
Roche anti-GFP.

Metabolic labeling of Stm peptidoglycan
and LPS release in cellulo

y

For Stm infections, overnight bacterial cultures were diluted 1:33 in fresh Luria broth
(LB; Thermo 12780029), grown for 3 hours
before being washed once in PBS, and used to
infect HeLa cells at 80% confluence with a
multiplicity of infection (MOI) of 20 to 50 as
indicated. Plates were centrifuged for 10 min
at 1000x g and incubated for 30 min at 37°C

CLICK chemistry and metabolic labeling
Farnesylpyrophosphate (FPP)–azide-biotin
CLICK chemistry

incorporated fluorescently labeled D-alanine
into the underlying peptidoglycan scaffold
using a pulse of 500 mM HCC-amino-D-alanine
(HADA; MCE HY-131045). Unincorporated
DIBO was removed through extensive washing
in PBS.

g

Bacterial infections, LDH assay, and IL-18 ELISA

to allow invasion. Extracellular bacteria were
killed by replacing media with fresh DMEM
(Thermo 11965092) containing 100 mg/ml gentamicin (Thermo 15710064) for 30 min. Cells
were washed 3 times and incubated with
20 mg/ml gentamicin for the duration. To
enumerate live bacteria, cells were lysed in
PBS + 0.5% Triton X-100 and serial dilutions
plated on LB agar. For LDH assay, cell death
was measured by CytoTox 96® Non-Radioactive
Cytotoxicity Assay (Promega G1780). IL-18 release in supernatants was detected via human
IL-18 ELISA kit (Abcam; ab215539) per the
manufacturer’s instructions with a detection
sensitivity of 8.3 pg/mL.

p

To generate stable gene knockouts in HeLa
CCL2 cells, single guide RNAs (sgRNAs) were
cloned into pX459 (Addgene Plasmid #62988)
per established protocols. Two to four sgRNAs
targeting each gene (200 ng total DNA) were
transfected in 24 well plates for 24 hours, followed by selection with 1 mg/mL puromycin
for 48 hours. Surviving cells were expanded
into media lacking puromycin for 48 hours,
then subject to limiting dilution to obtain
single colonies. Colonies were screened first
by PCR, then by Western blot, and the genotype of each positive clone was determined by
Sanger sequencing. The following 10 CRISPRCas9 mutants were generated: GBP1–/–, GBP2–/–,
GBP3–/–, GBP4–/–, GBP1–/–GBP2–/– double mutant, GBPD1q22.2 septuple mutant (GBP1–/–
GBP2–/–GBP3–/–GBP4–/–GBP5–/–GBP6–/–GBP7–/–),
CASP4–/–, GSDMD–/–, AOAH–/–, and RNF213–/–.
A complete list of guide RNAs (gRNAs) used to
generate these CRISPR deletions are provided
in Table S3.
In addition, we generated a series of cell
lines stably or transiently complemented with
GBP mutants, affinity probes or reporters. These
included GBP1–/– clonal lines complemented
with either of the following: EGFP-GBP1, RFPGBP1, mNG-GBP1, EGFP-GBP1S52N, EGFPGBP1DD103,108NN, EGFP-GBP1D184N, EGFP-GBP1C589S,
EGFP-GBP1a13ARR, EGFP-GBP1a13RAR, EGFPGBP1a13RRA, EGFP-GBP1a13ARA, EGFP-GBP1a13AAR,
EGFP-GBP1a13AAA (R584-586A). Here alanine
scanning mutations in the C-terminal polybasic patch (amino acid 584 to 586) of GBP1
were introduced according to Stratagene
Quickchange (Agilent 200523) protocol and
confirmed by DNA sequencing. Transfections
of plasmids for complementation into GBP1–/–
HeLa cells were performed using Mirus LTI
according to manufacturer protocol. Briefly,
1 mL Mirus LTI was added to 50 mL serumfree DMEM followed by the addition of 500 ng
of respective construct. The mixture was allowed to sit for 30 min followed by gentle
mixing through pipetting. Cells were selected
for hygromycin resistance on the puromycinresistant GBP1–/– HeLa cell background.
We also complemented CASP4–/– with EGFPcaspase-4 and GSDMD–/– with either EGFP- FLGSDM EGFP-NT-GSDMD, or EGFP-CT-GSDMD.
The latter plasmids were also introduced into
CASP4–/– and GBP1–/– clonal lines as above for
fluorescent microscopy.

RES EARCH | R E S E A R C H A R T I C L E

4Pi-SMS nanoscopy

,

13 of 18

y

The coding sequences of human GBP1 and its
mutants were cloned into a customized vector
pCMV-His10-Halo-HRV-mRFP-TEV for coating
assays. HEK 293F suspension cells (a gift from
Dr. James Rothman; mycoplasma-negative)
was maintained at a concentration of 0.4 ×
106~4 × 106 cells/m in Expi293 expression
medium (ThermoFisher A1435101). 24 hours
prior to transfection, cells were seeded at a
concentration of 1.2 × 106 cells/ml. For transfection, cells were harvested and resuspended
in fresh medium at a concentration of 2.5 ×
106 cells/ml. Cells were transfected by adding
pCMV-His10-Halo-HRV-mRFP-TEV–containing
clones to a final concentration of 1 mg/ml in
media containing PEI at a concentration of
5 mg/ml. 24 hours after transfection, cells
were diluted 1:1 (v/v) with fresh medium containing 4-mM valproic acid and cultured for
an additional two days. 2 × 109 cells were harvested via centrifugation (500 x g, 10 min),
washed once in cold PBS, resuspended in
lysis buffer (50 mM HEPES, pH 7.5, 500 mM
NaCl, 1 mM MgCl2, 10% glycerol, 0.5% CHAPS,
1 mM TCEP) and lysed through sonication.
Cells were cleared at 35,000 x g for 1 hour at
4°C. Supernatant was collected and incubated
with 1 ml bed volume of HaloLink resin
(Promega G1912) at 4°C overnight with gentle
rotation. The resin was sequentially washed
twice (10 min each) with wash buffer 1 (50-mM
HEPES, pH7.5, 500 mM NaCl, 1 mM MgCl2, 10%
glycerol, 0.5% CHAPS), wash buffer 2 (50 mM
HEPES, pH7.5, 1 M NaCl, 10% glycerol) followed by wash buffer 1.
To elute bound proteins, Halo resin was resuspended in lysis buffer and digested with
homemade GST-HRV-His protease overnight
at 4°C with gentle rotation. Resin was pelleted
and the HRV protease was removed from
the supernatant via Ni-NTA beads by affinity
chromatography (QIAGEN 30210). Flow through
was collected, concentrated, and further purified and buffer-exchanged via size exclusion

y g

1 March 2024

Purification of recombinant proteins
GBP1 to GBP4 and their mutants,
caspase-4C258A, and RFP-AtGBPL1

y

For 4Pi-SMS imaging, a customized microscope built using a vertical 4Pi cavity around
two opposing high-NA objective lenses as detailed elsewhere (28) was used to capture
high-resolution images of the GBP coat complex in IFN-g–activated HeLa cells. Cell samples were prepared on 25-mm diameter round
precision glass cover slips (Bioscience Tools,
San Diego, CA) that had been immersed in 1M
KOH and sonicated for 15 min in an ultrasonic
cleaner (2510 Branson, Richmond, VA). Sequential washes in Milli-Q water (EMD Millipore, Billerica, MA) and sterilization with 70%
ethanol was followed drying and polylysine
coating of coverslips. HeLa cells were grown
on coverslips for 24 to 48 hours before fixation in 4% paraformaldehyde (PFA). AntiGBP1 (1B1; Santa Cruz; 1:500 dilution) and
GBP2 (1E5; Origene; 1:200 dilution) antibodies
were labeled for 2 hours at room temperature
by goat anti-mouse Fab AF647 (Jackson
ImmunoResearch, PA) and goat anti-rabbit
IgG CF660C (Biotium, CA) at 1:200 dilution
as described (29).
1B1 antibody was raised against the fulllength GBP1 protein whereas anti-GBP2 targets an N-terminal 17-amino acid peptide
epitope. This enabled both antibodies to bind
their endogenous GBP targets because the
N-terminal LG is oriented toward the top of
the coat complex. Specificity was confirmed in
IFN-g–activated GBPD1q22.2, GBP1–/– or GBP2–/–
HeLa cells as negative controls: all went unlabeled at the single molecule level. At least
several hundred GBPs were detectable on
cytosolic bacteria using this technique, typically ~400 to 800 molecules. Although this
number represents <2 to 3% of the total enumerated by cryo-ET in cell-free systems, when
1B1 antibody staining was used on cell-free
GBP1-coated bacteria it gave similarly low
numbers. Hence, antibody accessibility (due
to the densely packed nature of the coat) together with larger secondary antibody detection
were limiting factors. Commercially available

nanobodies do not currently exist for GBP1 or
GBP2. Notably, however, this comparison discounted bacteria having fewer GBP coat proteins in cellulo compared with in vitro as both
gave similar results when labeled with the
same polyclonal antibody.
Sample mounting, image acquisition, and
data processing were mostly performed as previously described (29) except that the imaging
speed was 200 Hz with 642-nm laser intensity
of ~12.5 kW/cm2. Typically, 3000×100~200
frames were recorded. DME were used for
drift correction. All 4Pi-SMS images were rendered using Point Splatting mode (20 nm
particle size) with Vutara SRX 7.0.06 software
(Bruker, Germany).

g

Zhu et al., Science 383, eabm9903 (2024)

Stm SL1344 at MOI = 5~10 for 30 min and
further incubated with 100 mg of Gentamycin
for 50 min to eliminate extracellular bacteria.
Five- or six-color fluorescence images were
captured on Laser Scanning Confocal Microscope (SP8, Leica) with Z-stacks. A series of
lasers were used to excite fluorescence across
this broad wavelength and avoid bleed-through
as shown in fig. S4E. The corresponding laser
parameters (laser, % power, wavelength reception band) used were: pSapphire (Diode 405,
4, 424 to 444), pmEmerald (WLL 470, 20;
490 to 520), pmVenus (WLL 516, 6; 535 to
559), pmOrange (DPSS 561, 25, 566 to 593),
pmIFP24 (HeNe 633, 75, 680 to 730). Images
were analyzed by LAS X software (Leica) for
3D reconstruction.

p

(PDB ID 1GGX) calculated through the Voronoi
cell algorithm (http://proteinformatics.org/
voronoia), the total number of spatially constrained RFP-GBP1 molecules per voxel were
counted, along with coat kinetics. Euclidean
Voronoi cell algorithms enable the packing
densities of all atoms in a deposited protein
structure (PDB) to be estimated, accounting
for the volume inside an atom’s van der Waals
sphere, the solvent excluded volume of the
atom, and the position as well as approximate
radius of every cavity that can accommodate
a water molecule within the protein to yield
complete volumetric information (51). In addition, a compensatory ratio of 3D-SIM fluorescence versus EM volumes of GBP1-coated
bacteria was included since the former does
not resolve objects to the same precision as
EM. 3D-SIM volumes were on average 3.02
times larger (n = 17) than direct measurements by immuno-EM. Thus, 3D-SIM volumes
were divided by 3.0 to correct for fluorescence
enhancement. Long-term imaging used low light
OMX-SR conventional mode for up to 2.5 hours
continuous recording at 5-min intervals (fig. S3C).
For examining the GBP-COAT450-708 complex, we constructed fluorescent fusion proteins GBPs1-4, caspase-4, and GSDMD across
different spectral range to accommodate five
or possibly six proteins simultaneously along
with Stm. A color-coded matrix of 60 proteins
was generated by fusing each of the six coat
proteins to each of the following fluorescent
reporters: mAzurite- c1 (Addgene Plasmid
#54583), mT-Sapphire-c1 (Addgene Plasmid
#54545), pmTurqouise2-c1 (Addgene Plasmid #60560), pmEmerald-c1 (Addgene Plasmid #53975), pmVenus-c1 (Addgene Plasmid
#27794), pmOrange-c1 (Addgene Plasmid #54680),
pmKeima-Red-c1 (Addgene Plasmid #54546),
pmCardinal-c1 (Addgene Plasmid #54799), pmApplec1 (Addgene Plasmid #54631), and mIFP24-c1
(Addgene Plasmid #54820); Ex/Em range, 384/
450 to 684/708 nm. In brief, GBP1,2,3 and 4
were amplified by PCR with each primer from
IFNa-treated HeLa cell cDNA. Caspase-4 and
GSDMD were amplified by PCR with each
primer from caspase-4 cDNA (Sino Biologicals
#HG11158-M) and pET-SUMO-hGSDMD (Addgene,
Plasmid #111559), respectively. Amplicons were
cloned into reporter plasmids using SalI/
BamHI restriction enzyme digestion. All constructs were sequenced for validation. Multiarrayed testing found five combinations
of either pmIFP24-GBP1, pmOrange-GBP2,
pmVenus-GBP3, pmEmerald-GBP4, Sapphirecaspase 4 or -GSDMD, or pmOrange-GSDMD
could be used successfully with Stm.
GBP-COAT450-708 combinations were transfected into the GBPD1q22.2 septuple mutant
by TranslT-LT1 transfection reagent (Mirus Bio)
according to manufacturer’s instructions. After
24 hours, cells were activated with IFN-g for
16 to 18 hours and infected with SPI1-induced

RES EARCH | R E S E A R C H A R T I C L E

Reconstituted coat assays

Wild-type Stm and Stm mutants were freshly
streaked on LB plates. A single fresh colony of
bacteria was cultured in LB medium overnight
at 37°C. Before the coating assay, bacteria were
diluted 1/60 in fresh LB medium, cultured for
additional 2 hours and harvested via centrifugation (4000g, 5 min at 22°C). To remove
shed LPS which inhibits rRFP-hGBP1 targeting, bacteria was washed twice in coating
buffer [50 mM HEPES (pH 7.5), 150 mM NaCl,
1 mM MgCl2] and further diluted to an optical
density (OD600) of 0.1 in coating buffer. Im-

1 March 2024

Reconstituted killing assays

For bacterial killing assays, Stm were grown
to mid-log phase in LB medium, washed, and
incubated with 5 mM hGBP1 in coating buffer
(50 mM HEPES pH 7.4, 150 mM NaCl, 5 mM
MgCl2) with or without 2 mM GTP, 2 mM
GTPgS or GDP.AIF4– (2 mM GDP, 200 mM
AlCl3, 10 mM NaF). Stm was then pelleted
and resuspended in Buffer A [50 mM MES
pH 6.0, 100 mM potassium gluconate (KGluc)]
containing 5 mM rAPOL3 and incubated at 37°C
for 1 hour prior to plating on LB agar to enumerate colony forming units (CFU) in triplicate.
In vitro phase separation

Protein aliquots of rRFP-GBP1 and rRFP-GBPL1
(7) were thawed at room temperature, centrifuged at 14,000 × g for 5 min to remove any
aggregated protein. Droplet formation was induced by diluting protein to low salt buffers
(50 mM HEPES, pH7.5, 150 mM NaCl, 1 mM
MgCl2) by mixing with various volumes of no
salt buffer [20 mM HEPES (pH 7.5), 1 mM
TCEP] and analyzed in chambered coverglass (Grace Bio-labs) using a Lecia SP8 laser
scanning confocal at 63×/1.40 magnification.
Addition of Ficoll at 5, 10, 15, and 20% w/v had
no effect on rRFP-GBP1.
Electron microscopy
Negative-stain electron microscopy

To visualize rRFP-GBP1 on minicells, 100 mL
minicells were washed and dissolved in the
14 of 18

,

Zhu et al., Science 383, eabm9903 (2024)

Human GBP1, GBP2, GBP3, GBP4, GBP1 mutants (GBP1S52N, GBP1D184N, GBP1DD103,108LL,
GBP1R53A, and GBP1C589S) and caspase-4C258A
proteins were expressed as FLAG-tagged proteins in human embryonic kidney (HEK)-293E
cells and isolated from large-scale adherent
cell culture using Flag M2 beads as described
above. Recombinant human proteins were further purified to single peak using FPLC. Fluorescence anisotropy assays were conducted
at 37°C across different GBP concentrations
in binding buffer [50 mM Tris, 150 mM NaCl,
5 mM MgCl2, 0.3 mM GDP and AlFX (10 mM
NaF, 0.3 mM AlCl3), pH 7.0] and Flag- caspase4C258A (50 mM Tris, 150 mM NaCl, 1 mM DTT,
pH 7.0) for 15 min followed by addition of
Salmonella minnesota LPS–Alexa Fluor 488
to a final concentration 250 nM. After 1 hour
incubation with LPS, fluorescent anisotropy
values measured by SpectraMax i3x (Molecular Devices).

y

Thin layer chromatography (TLC) was used to
separate GTP, GDP, and GMP and performed
exactly as previously described (7–9). Briefly,
a-[32P]GTP (Perkin Elmer BLU006H250UC)
hydrolysis by purified recombinant proteins in
reaction buffer [20 mM HEPES (pH 7.0), 150 mM
NaCl, 5 mM KCl, 1 mM MgCl2, 100 mM GTP,
10 mCi a-(32P) GTP] was determined at 25°C
before quenching the reaction with 142 mM

LPS-binding assays

y g

Thin-layer chromatography

Overnight-cultured Stm was preincubated with
indicated inhibitors at 37°C with shaking for
2 hours. After centrifugation by 3500 rpm
room temperature for 20 min, bacteria were
washed three times by 10-fold volume of coating buffer [50 mM HEPES (pH 7.5), 150 mM
NaCl, 1 mM MgCl2, and 1 mM DTT] and finally
resuspended by 1 mL of coating buffer containing GTP (2 mM), GTPgS (2 mM), GMPPCP/
CppCp (10 mM), GDP.AIF4– (2 mM GDP, 200 mM
AlCl4, 10 mM NaF) or GMP.AIF3– (2 mM GMP,
200 mM AlCl3, 10 mM NaF]. Bacteria aliquoted
were incubated with or without 4 mM of recombinant GBP1 or its mutants at 30°C for
1 hour. Supernatants were carefully collected
by centrifugation (8,000 rpm, 22°C for 2 min)
and diluted 1:1000 to 1:5000. Released LPS
was measured by ToxinSensor Chromogenic
LAL Endotoxin Assay Kit (GenScript L00350)
according to manufacturer’s instructions.

y

His6-tagged GBP1 or GBP1 mutant proteins
were prepared as described above. FPLC was
performed on an AKTA system equipped with
a Superdex 200 10/300 GL size exclusion
column (GE Amersham). After preequilibrating the column with five volume equivalents
of buffer [20 mM Tris (7.5), 140 mM NaCl,
2 mM MgCl 2], 100 mg of each recombinant
GBP1 protein was examined. Column buffer
was then exchanged and equilibrated with
20 mM Tris (7.5), 140 mM NaCl, 2 mM MgCl2,
containing 200 mM GDP-AlF4– to test GBP1
assembly in the presence of the transition state
analog. Chromatographs were generated by absorbance measured at 280 nm over retention time.

LAL for lipid A detection in OM disruption assays

mediately before coating, rRFP-hGBP1 was
thawed at room temperature and centrifuged
at 12,000 g for 15 min to remove any insoluble
aggregates. For coating, 2 mM rRFP-hGBP1
and 2 mM GTP were added to each 0.1 OD of
bacteria, gently mixed and incubated for 1 hour
at 22°C before imaging.
To image rRFP-hGBP1-coated bacteria, coating reactions (20 mL) were transferred to 384well glass bottom plate (Cellvis; P384-1.5H-N).
The plate was centrifuged at 2000g for 1 min
to collect all liquid to the bottom of the well.
All images were acquired using a Leica SP8
laser-scanning confocal microscope with a
63×/1.40 oil immersion objective. Focal plane
was set to the bottom of the well. RFP was
excited at 561 nm and detected at 590 to
610 nm. The optical slices were acquired in
confocal mode (1 Airy unit) with an average
of six scans. Images were collected in a 512 ×
512 format. Image analysis was performed
with FIJI/ImageJ.
For coat complex assembly, samples were
doubly diluted from 2 mM rRFP-GBP1 until
loss of coating to establish a coating curve.
The observed behavior was plotted with bestfit interpolation and a sigmoidal curve emerged. Curve fitting together with regression
analysis found half maximal and Hill slope
values using Graph Pad Prism 9.1.1.

g

Size exclusion chromatography

EDTA after 7 hours. The resulting products were
separated by PEI cellulose (Sigma, Z122882)
with fluorescent indicator (UV 254) using
750 mM KH2PO4 (pH 3.5) as solvent and visualized by autoradiography.

p

chromatography (Superdex® 200 Increase; GE
Healthcare) equilibrated with storage buffer
[20 mM HEPES (pH 7.5), 150 mM NaCl, 1 mM
MgCl2, 1 mM TCEP]. Fractions were analyzed
by SDS-PAGE, pooled, concentrated and flash
frozen in liquid nitrogen before storing at
-80°C. Protein concentration of rRFP-tev-hGBP1
(abbreviated rRFP-hGBP1) was determined via
BCA and SDS-PAGE electrophoresis with BSA
standards running aside.
Recombinant human GBP1, GBP2, GBP3, and
GBP4 as well as the GBP1 mutants (hGBP1S52N,
hGBP1D184N, hGBP1DD103,108LL, hGBP1C589S, and
hGBP1R3A) were produced by cloning the respective genes into the mammalian expression
vector, pCMV-3Tag1A (Agilent 240195), with an
N-terminal FLAG tag. Human caspase-4 C258A
was cloned into pcDNA 3.1/Hygro (Thermo
V87020) for N-terminal FLAG attachment.
Plasmids were transfected via Mirus LT1 into
HEK-293 E cells. Cells were harvested after
14 hours and lysed for 2 hours in buffer 50 mM
Tris, 150 mM NaCl, 5 mM MgCl2, 1 mM DTT
containing 1% Triton X-100 and protease inhibitor. Supernatant containing the expressed
protein passed through anti-Flag M2 affinity
gel (Sigma A2220) and bound Flag protein
eluted using Flag peptide 150 mg/mL in buffer
50 mM Tris, 150 mM NaCl, 5 mM MgCl2 and
1 mM DTT. Further chromatographic purification was undertaken on an AKTA FPLC (GE
Amersham) instrument. All recombinant proteins made in human cells were subject to LAL
(ThermoFisher A39552) to confirm the absence of LPS contamination (<0.01 EU/mL,
lower detection limit).

RES EARCH | R E S E A R C H A R T I C L E

HEPES buffer (pH 7.4) to the OD600 equal to
0.1. The minicell suspensions were incubated
with rRFP-GBP1 with a final concentration
of 2 mM in the presence or absence of GTP at
22°C for 1 hour. The mixture was concentrated to 10mL for further TEM observation.
5 mL RFP-GBP1 coated and noncoated minicells were loaded onto glow-discharged carbon
film coated cooper EM grids (EMS, CF400Cu-50) for standing 1 min. The EM grid was
washed and stained with 2% uranyl formate
for 30 s. The grid was blotted by filter paper
(Whatman FILTER PAPERS 2; 1002-090) and
dried for 1 min. Grids were examined in
JEOL1400 plus electron microscope with acceleration voltage of 80 kV.
Cryo-ET
Cell seeding on EM grids

15 of 18

,

1 March 2024

Tomography data for the FIB-milled lamella
sample were collected using a Titan Krios G2
transmission electron microscope (Thermo
Fisher Scientific) equipped with a 300 kV
field-emission gun, a Volta phase plate (phase
shift located around 1/2 pi), a Quantum post
column energy filter (20-eV slit), and a Gatan
K2/K3 summit direct detection camera. We
first navigated the stage to the lamella and
collected a montage at medium magnification
(2250x); we used ImageJ to overlap the TEM
montage with the map taken by a second
CLEM. Three ice deposits as well as a doublemembrane vesicle on the lamella were used as
shared markers to overlay these two maps.
One of the successful targets with a GFP signal
was located beneath the largest ice deposit
and close to the lamella edge. The stage was
moved to the right of the target with fluorescent signal to collect tilt series. The images
were taken in a dose-fractionated mode at
near focus using SerialEM software. The resulting physical resolution is 0.45 nm/pixel. A
total dose of 80 e/Å2 distributed among 33 tilt
images was undertaken covering angles from
39° to –57° at tilting step of 3° and starting the
first tilt series at –9° in a continuous data

y

Cryo-CLEM demarcated grids were transferred into cryo-DualBeam microscope equipped
with cryostage and cryotransfer shuttle systems
(Thermo Fisher Scientific, Aquilos cryo-FIB
focused ion beam/scanning electron microscope). The lamella was prepared according to
Aquilos cryo-FIB protocols. The sample on the
O-ring side of the grid was sputter-coated with
platinum (1 kV, 30 mA, 10 pa, 15 s) to increase
sample conductivity during FIB milling. MAPS

Cryo-ET and image processing

y g

Milling lamella by FIB

y

EM grids seeded with EGFP-GBP1-expressing
HeLa CCL2 cells were prescreened using fluorescent light microscopy, clipped on a custom
manual plunger, blotted using Whatman #1
filter paper (GE, 1001-110), and frozen in liquid
ethane by the plunger. Frozen grids were
transferred onto a cryostage and clipped within
the O-ring and C-ring (ThermoFisher Scientific,
cryo-FIB autogrid and C-clip, respectively).
Cryo-CLEM (Leica Microsystems EM CryoCLEM microscope) was performed as previously described (7). Grids were transferred into
Leica-CLEM cartridge docked at a precooled
shuttle docking station and then transferred
onto a cryo-stage equipped with a pre-cooled
40x objective lens. A spiracle 4×4 grid region
was selected, and 16 subregions were imaged.
Montage EGFP-positive images were acquired
on the selected area using the Z-stack mode
within a stepwise of 0.35 mm. This montage
served as the navigating map in subsequent
milling assays by the focused ion beam method.

g

Zhu et al., Science 383, eabm9903 (2024)

Vitrification and cryo–correlative fluorescence
light microscopy (CLEM).

software was used to capture an EM grid
montage in the electron beam mode before
importing the fluorescent montage from cryoCLEM to generate a merged map to establish
potential targets for milling lamella. Eucentric
height of the targeting area was refined and
stage tilting angle determined at 16° for lamella.
At target positions on the grid organometallic
platinum coating was sprayed for 5 s by a gas
injection system (ThermoFisher Scientific, GIS).
Initially, rough milling was started from the
topside with 300 pA current and continued in
a small stepwise manner. Concurrently, SEM
images were taken to observe intracellular
elongated bacteria. Rough milling stopped
on the topside when elongated bacteria were
closely visible. Rough milling on the bottomside began with 300 pA current in a large
stepwise manner. When the lamella thickness arrived to 1.5 mm distance, the milling
current was changed to 100 pA until the
lamella thickness was ~0.8 mm. Thereafter
fine milling focused primarily on the bottomside at < 50 pA. An electron beam was used
to monitor the milling process at 5 keV and
13 pA. A final fine milling was done on both
sides of the lamella. Subsequently the grid
was sputter-coated with platinum (15 mA, 10 Pa,
10 s) to increase conductivity of milled lamella.
The grid was transferred from the FIB instrument using the transferring shuttle system
and stored (SubAngstrom, Pin type grid box)
in liquid nitrogen for further data collection
or immediately re-confirmed by CLEM for
correct positioning of GBP1-coated bacteria.

p

Pretreatment of EM grids. EM grids (Quantifoil R1/4 gold 200 mesh, Q250AR-14) were
placed onto 18 mm2 cover-glass that had been
washed and stored in 100% ethanol. They were
then placed into glass bottom dishes (Cellvis,
35 mm dish with 20 mm microwell cover
glass) and treated with 70% ethanol under
UV illumination for 10 min at 22°C. EM grids
were washed 6 times with sterile and degassed
water before treatment with 0.05 mg/mL collagen I (Gibco, A10644-01) for 60 min in a 37°C
incubator followed by washing in degassed/
distilled water. Collagen-coated grids were incubated in DMEM medium overnight at 37°C
incubator with 5% CO2 for further use.
The GBPD1q22.2 septuple mutant was cultured in 10 cm dish to the density of 4 × 106 cells
per dish. HeLa cells were transfected with
EGFP-GBP1 plasmid using Mirus LT1 kit for
overnight expression. Cells were subsequently
washed by PBS and treated with 0.05% TrypsinEDTA for detachment. Detached HeLa cells
were centrifugated and resuspended with
DPBS buffer before filtering via 40 mm nylon
mesh filter (Fisher Scientific, 22363547) to
remove larger debris. Filtered HeLa cell suspensions were transferred into 5 mL polystyrene round-bottom tube (Corning, 352235)
with cell-strainer cap for FACS sorting. GFPpositive cells were sorted into collection tubes,
centrifugated and resuspend to a final cell density of 2 × 104 cells/mL in DMEM supplemented
with 10% fetal bovine serum plus 100 mg/mL
Pen Strep (Gibco, 15140-122). EGFP-GBP1expressing cells were seeded onto pretreated
EM grids and further incubated overnight. For
the infection assay, StmpBAD-ftsZ induced by by
0.2% L-arabinose were cultured to a density of
OD600 1.0 and infected at MOI 200. EM grids
coated with GBPD1q22.2 HeLa cells and bacterial cells were further centrifuged at 800g for
5 min for spinning bacteria onto the HeLa cell
surface. EM grids were transferred into 37°C
incubator for additional 20 min before washing three times with DPBS and incubated with

complete DMEM plus 100 mg/mL gentamycin
for an additional 1 hour infection.
For preparation of minicells from StmDminD,
bacterial cultures were grown overnight at
37°C in LB medium in the presence of 100 mg/
mL ampicillin. 10 mL bacterial overnight cultures were added into 1 L fresh LB medium in
the presence of ampicillin and were grown to
late log phase. Batch cultures then underwent
3-step centrifugation at: (i) 6240g (Beckman
coulter, Avanti JXN-26, rotor, JLA-8.1000) for
10 min to remove parental rod-shaped bacilli;
(ii) 24,820g (rotor, JLA-16.250) for 10 min to
collect minicell fractions which were resuspended in HEPES buffer (pH 7.4) and filtered
by 0.45 mm PVDF membrane (Merck Millipore,
SLJVM33RS); (iii) 20,000g (rotor, JA-25.50)
for further 10 min. Minicell fractions were
adjusted to OD600 1.0 for rRFP-GBP1 in vitro
coating.
For preparation of minicells from
StmDwaaG::pBAD-ftsZ, fresh cultures were grown
in LB medium containing 100 mg/mL ampicillin and 0.2% L-arabinose for a continuous FtsZ induction at 37°C for 20 hours.
StmDwaaG::pBAD-ftsZ minicells were collected as
described above.

RES EARCH | R E S E A R C H A R T I C L E

A his-tag containing six histidine residues (His6)
with a GSG linker was fused to the N terminus
of GBP1. Purified His6-GBP1 diluted to 2 mM in
the coating buffer with 1 mM GTP was a
mixture with bacterial minicell suspension
as the reconstitution assay described. After
the full coatomers formed, the cell suspension
was centrifuged at 6200g and 4°for 5 min.
The pellets containing bacterial minicell and
GBP1 coat were dissolved in the blockingbinding buffer (20 mM Tris-HCl, 150 mM
NaCl, 0.1% Tween 20, 5% BSA, 1mM MgCl2)
for 10 min incubation at room temperature.
1.8 nm Ni-NTA-Nanogold beads (Nanoprobes)
diluted 1:10 were added in the blockingbinding buffer and incubated for 10 min. The
suspension underwent centrifugation at 6200g
for 5 min and was washed three times in
50 mM Tris-HCl, 100 mM Imidazole, 1 mM
MgCl2, 0.1% Tween 20, and 300 mM NaCl
(pH 7.4). The minicell pellets were finally
dissolved into the coating buffer and transferred onto glow-discharged carbon-coated copper grids for cryo-ET sample preparation.
Disulfide cross-linking assay

For testing the open conformer model of
GBP1, we undertook a Cysteine replacement
strategy for cross-linking. Six residues (Glu-389,
Ile-369, Glu-366, Ile-365, Arg-362, and Glu361) present in the GBP1 a7 helix of the middle
domain and face toward the a12 helix in GED
domain were chosen as prospective Cys-Cys
pairs. We measured the distance between the
Cb atoms of the above-selected residues and
16 of 18

,

1 March 2024

Ni-NTA-nanogold labeling

y

We enlisted AlphaFold2 to predict functional
homologs of the human GBP protein family.
Notably, the GBP5 protein (AF-A0A2K5PG15)
exhibited an elongated conformer resembling
some of the structural features revealed in the
subtomogram averaging of human GBP1. Recent findings suggested that GBP5 lacking its
C terminus can form dimers in the presence
of GDP.AIF3- (PDB ID 7E5A).
Based on this dimeric form of truncated
GBP5, we constructed a pseudomodel representing dimeric GBP1 in its activated state.
First, we aligned the N-terminal domain and MD
region from the human GBP1 monomer x-ray
crystal structure (6K1Z) using the MatchMaker
function in the Chimera package; it revealed a
147° rotation and a –14.7 Å shift along the axis
of the GTPase domain when compared to
one of the monomer subunits of the GBP5
dimer (PDB ID 7E5A). Then we aligned the
C-terminal region of GBP1 with the predicted
GBP5 monomer (AF-A0A2K5PG15) using Chimera. Lastly, pseudomodels of an open GBP1
conformer were built via AlphaFold2 and
docked into our low-resolution subtomogram
averaged GBP1 structures using Chimera. It

After reconstituting the coat complex, we collected untagged GBP1 coated minicells through
centrifugation at 14,000g for 5 min (Eppendorf
Centrifuge 5420, SN, 5420JQ303731) at room
temperature. The minicell pellets were washed
in the coating buffer [50 mM HEPES (pH7.5),
150 mM NaCl, 1 mM MgCl2] by gently pipetting
the solution 10 times. The GBP1-coated minicells were then collected again via centrifugation as above and subjected to the same
washing treatment. This was repeated one
more time for a total of 30 pipette-aided
washes and removal of the freed GBP1. The
coated minicells were then dissolved in the
coating buffer and ready for cryo-ET sample preparation. This method proved highly
efficient, reducing crowdedness from ~12,000
GBP1 molecules per minicell to an estimated
~100 to 120 molecules per minicell (>100-fold
depletion) based on direct counts of segmented tomograph series.

y g

Zhu et al., Science 383, eabm9903 (2024)

Computational modeling of human GBP1

Sequential wash method for reducing
GBP1 crowdedness

y

TomoSegMemTV and PySeg packages were
used for subtomogram analysis. For initial
membrane segmentation, 11 tomograms from
StmDminD dataset and 27 tomograms from
StmDwaaG::pBAD-ftsZ with a binning factor of
8 were selected for membrane segmentation
by using TomoSegMemTV (54). PySeg package
was used for tracing and picking the particles
located on the membrane based on Discrete
Morse theory (55). The initial orientation of
the particles perpendicular to the membrane
was determined by PySeg. The coordinates of
picked particles were multiplied by four to obtain the particles coordinates information in
the CTF corrected and weight back projection
reconstructed tomograms within a binning
factor of two. Subtomograms were extracted
and were initially aligned based on the feature of outer membrane (OM) by Relion package v3 (56).
After the OM was well aligned, general classification function in the PySeg was performed
to distinguish membrane and nonmembrane
features. Classes showing membrane features
were merged to have a further general classification by applying a cylinder mask on the
top of the membrane. The classification results

yielded elongated GBP1 monomer and dimer
models corresponding to native structures obtained from our monomer and dimer masks.

g

Subtomogram averaging

showing the extra density of GBP1 on top of
the membrane resulted in 21,587 particles obtained for further constrained refinement and
classification by the Relion package. During
the next step of Relion classification, 4 classes
resulting in 12,858 particles were averaged
by the constrained refinement with molecular masks that was used to remove overlapping density of GBP1 (fig. S8). During the
postprocessing, the final density map was
generated with Relion using a B-factor of
–1000 and a low-pass filter of 10 Å. Resolution estimation and Fourier shell correlation
for both dimeric and monomeric masks along
with their controls are shown in fig. S9.
For distance detection of the full coat complex of mRFP-GBP1, OM and GBP1 density
were manually segmented through IMOD
drawing contours. Mtk function was used
to detect the shortest distance between the
density of the GBP1 GTPase domain and the
density of the LPS outer leaflet. 11,889 GBP1
protomers on the DminD minicell having
wild type LPS; 12,487 GBP1 protomers on
the DwaaG minicell having O-antigen and
outer-core truncated LPS; 6107 GBP1 protomers on the DwaaG outer membrane vesicle
(OMV) were detected and analyzed by the
Graphpad Prism9 package. For distance detection of the washed coat complex of mRFPGBP1 on both minicells and OMVs, GBP1 density
in the 3D tomogram was drawn using IMOD
manually. Neural network–based segmentation on both washed minicell and washed
OMV was performed by EMAN2.23 and further refined by UCSF Chimera.

p

collection mode. The dataset for the minicell
was taken a dose symmetric scheme without
using phase plate. The tilt series for minicells
sample were taken at the physical resolution
of 0.14 nm/pixel starting from 0° and covered
angle ranges of 51° to –51° with 3° increments,
resulting in a total dose of 100 e/Å2.
Collected dose-fractioned data were subjected to motion correction for generating
drift-corrected image stack files. Stack files
were aligned using patch-tracking function
of IMOD. 3D tomograms were reconstructed
from aligned stack files by SIRT (Simultaneous Iterative Reconstruction Technique)
method using Tomo3D. For reconstruction
of binned tomograms, the aligned tilt series
was scaled to 3.6 nm as pixel size. IMOD was
used for visualizing the tomogram. Surface rendering of tomogram was done with EMAN2.23
and refined with UCSF chimera. Briefly, for
ribosomes, a template-matching strategy was
performed to determine all ribosome coordinates and orientations. A mammalian 80S ribosome structure (EMD-3418) determined by
Volta phase plate cryo-ET, and bacterial
ribosome cryo-EM structure (EMD-0076) (52)
were low-pass filtered to 20 Å and scaled to
3.6 nm/voxel to match to the tomograms.
EMAN2.3 was used to do segmentation on membrane, vesicles and GBP1 coat features. The
cryoEM map of type 3 secretion system (T3SS)
needle complex from Salmonella typhimurium
(EMD-11781) (53) was scaled to 3.6 nm/voxel
and fitted into the position of T3SS in the
tomograms.

RES EARCH | R E S E A R C H A R T I C L E

In silico protein sequence analysis

Zhu et al., Science 383, eabm9903 (2024)

1 March 2024

17 of 18

,

1. F. Randow, J. D. MacMicking, L. C. James, Cellular self-defense:
How cell-autonomous immunity protects against pathogens.
Science 340, 701–706 (2013). doi: 10.1126/science.1233028;
pmid: 23661752
2. J. D. Jones, R. E. Vance, J. L. Dangl, Intracellular innate
immune surveillance devices in plants and animals.
Science 354, aaf6395 (2016). doi: 10.1126/science.aaf6395;
pmid: 27934708
3. M. Shi, P. Zhang, S. M. Vora, H. Wu, Higher-order assemblies in
innate immune and inflammatory signaling: A general principle
in cell biology. Curr. Opin. Cell Biol. 63, 194–203 (2020).
doi: 10.1016/j.ceb.2020.03.002; pmid: 32272435

y

RE FE RENCES AND N OT ES

y g

Data were analyzed by GraphPad Prism 9.1.1.
and Excel v16.54 software. Unless otherwise
indicated, statistical significance was determined by t test (two-tailed) or one-way ANOVA
(Dunnett’s multiple comparison tests) with
Bonferroni post-hoc test.

y

Statistics

26. S. Huang, Q. Meng, A. Maminska, J. D. MacMicking, Cellautonomous immunity by IFN-induced GBPs in animals and
plants. Curr. Opin. Immunol. 60, 71–80 (2019). doi: 10.1016/
j.coi.2019.04.017; pmid: 31176142
27. K. Tretina, E. S. Park, J. D. Maminska, J. D. MacMicking,
Interferon-induced guanylate-binding proteins: Guardians of
host defense in health and disease. J. Exp. Med. 216, 482–500
(2019). doi: 10.1084/jem.20182031
28. F. Huang et al., Ultra-High Resolution 3D Imaging of Whole
Cells. Cell 166, 1028–1040 (2016). doi: 10.1016/
j.cell.2016.06.016; pmid: 27397506
29. D. Xu et al., PLSCR1 is a cell-autonomous defence factor
against SARS-CoV-2 infection. Nature 619, 819–827 (2023).
doi: 10.1038/s41586-023-06322-y; pmid: 37438530
30. W. D. Heo et al., PI(3,4,5)P3 and PI(4,5)P2 lipids target
proteins with polybasic clusters to the plasma membrane.
Science 314, 1458–1461 (2006). doi: 10.1126/science.1134389;
pmid: 17095657
31. C. J. Bradfield, “Sulfated DAMPs mobilize human GBPs for cellautonomous immunity against bacterial pathogens,” thesis,
Yale University, New Haven, CT (2016).
32. M. Wang, P. J. Casey, Protein prenylation: Unique fats make
their mark on biology. Nat. Rev. Mol. Cell Biol. 17, 110–122
(2016). doi: 10.1038/nrm.2015.11; pmid: 26790532
33. H. Nikaido, “Outer membrane” in Escherichia coli and
Salmonella: Cellular and Molecular Biology, vol. 1, F. C. Neidhardt,
R. Curtiss, J. L. Ingraham, Eds. (American Society for
Microbiology Press, ed. 2, 1996), pp. 29–47.
34. B. Hu, M. Lara-Tejero, Q. Kong, J. E. Galán, J. Liu, In Situ
Molecular Architecture of the Salmonella Type III Secretion
Machine. Cell 168, 1065–1074.e10 (2017). doi: 10.1016/
j.cell.2017.02.022; pmid: 28283062
35. H. Shi, B. P. Bratton, Z. Gitai, K. C. Huang, How to Build a
Bacterial Cell: MreB as the Foreman of E. coli Construction.
Cell 172, 1294–1305 (2018). doi: 10.1016/j.cell.2018.02.050;
pmid: 29522748
36. I. Hautefort et al., During infection of epithelial cells Salmonella
enterica serovar Typhimurium undergoes a time-dependent
transcriptional adaptation that results in simultaneous
expression of three type 3 secretion systems. Cell. Microbiol.
10, 958–984 (2008). doi: 10.1111/j.1462-5822.2007.01099.x;
pmid: 18031307
37. S. Okuda, D. J. Sherman, T. J. Silhavy, N. Ruiz, D. Kahne,
Lipopolysaccharide transport and assembly at the outer
membrane: The PEZ model. Nat. Rev. Microbiol. 14, 337–345
(2016). doi: 10.1038/nrmicro.2016.25; pmid: 27026255
38. Q. Kong et al., Effect of deletion of genes involved in
lipopolysaccharide core and O-antigen synthesis on virulence
and immunogenicity of Salmonella enterica serovar
typhimurium. Infect. Immun. 79, 4227–4239 (2011).
doi: 10.1128/IAI.05398-11; pmid: 21768282
39. J. Shi et al., Inflammatory caspases are innate immune
receptors for intracellular LPS. Nature 514, 187–192 (2014).
doi: 10.1038/nature13683; pmid: 25119034
40. Q. Kong et al., Palmitoylation state impacts induction
of innate and acquired immunity by the Salmonella
enterica serovar typhimurium msbB mutant. Infect.
Immun. 79, 5027–5038 (2011). doi: 10.1128/IAI.05524-11;
pmid: 21930761
41. J. A. Feulner et al., Identification of acyloxyacyl hydrolase, a
lipopolysaccharide-detoxifying enzyme, in the murine urinary
tract. Infect. Immun. 72, 3171–3178 (2004). doi: 10.1128/
IAI.72.6.3171-3178.2004; pmid: 15155618
42. E. G. Otten et al., Ubiquitylation of lipopolysaccharide by
RNF213 during bacterial infection. Nature 594, 111–116 (2021).
doi: 10.1038/s41586-021-03566-4; pmid: 34012115
43. X. Liu et al., Inflammasome-activated gasdermin D causes
pyroptosis by forming membrane pores. Nature 535, 153–158
(2016). doi: 10.1038/nature18629; pmid: 27383986
44. E. Kuru, S. Tekkam, E. Hall, Y. V. Brun, M. S. Van Nieuwenhze,
Synthesis of fluorescent D-amino acids and their use for
probing peptidoglycan synthesis and bacterial growth in situ.
Nat. Protoc. 10, 33–52 (2015). doi: 10.1038/nprot.2014.197;
pmid: 25474031
45. A. von Kügelgen et al., In Situ Structure of an Intact
Lipopolysaccharide-Bound Bacterial Surface Layer. Cell 180,
348–358.e15 (2020). doi: 10.1016/j.cell.2019.12.006;
pmid: 31883796
46. B. Barz, J. Loschwitz, B. Strodel, Large-scale, dynamin-like
motions of the human guanylate binding protein 1 revealed
by multi-resolution simulations. PLOS Comput. Biol. 15,
e1007193 (2019). doi: 10.1371/journal.pcbi.1007193;
pmid: 31589600

g

For lipid-binding motifs in human GBP1 compared with other related GTPases, evolutionary
tree history was inferred using the Maximum
Likelihood method and JTT matrix-based model. A bootstrap consensus tree inferred from
10,000 replicates was taken to represent the
evolutionary history of the taxa analyzed.
Branches corresponding to partitions reproduced in < 50% bootstrap replicates were collapsed. Initial tree(s) for the heuristic search
were obtained automatically by applying
Neighbor-Join and BioNJ algorithms to a
matrix of pairwise distances estimated using
the JTT model, and then selecting the topology with superior log likelihood value. A
discrete Gamma distribution was used to model evolutionary rate differences among sites
(5 categories (+G, parameter = 2.8088)). Evolutionary analyses were conducted in MEGA X.

4. H. Wu, Higher-order assemblies in a new paradigm of signal
transduction. Cell 153, 287–292 (2013). doi: 10.1016/
j.cell.2013.03.013; pmid: 23582320
5. L. Xiao, V. G. Magupalli, H. Wu, Cryo-EM structures
of the active NLRP3 inflammasome disc. Nature 613,
595–600 (2023). doi: 10.1038/s41586-022-05570-8;
pmid: 36442502
6. J. Wang et al., Reconstitution and structure of a plant NLR
resistosome conferring immunity. Science 364, eaav5870
(2019). doi: 10.1126/science.aav5870; pmid: 30948527
7. S. Huang, S. Zhu, P. Kumar, J. D. MacMicking, A phaseseparated nuclear GBPL circuit controls immunity in
plants. Nature 594, 424–429 (2021). doi: 10.1038/s41586-02103572-6; pmid: 34040255
8. J. H. Kim et al., Increasing the resilience of plant immunity to a
warming climate. Nature 607, 339–344 (2022). doi: 10.1038/
s41586-022-04902-y; pmid: 35768511
9. B.-H. Kim et al., A family of IFN-g-inducible 65-kD GTPases
protects against bacterial infection. Science 332, 717–721
(2011). doi: 10.1126/science.1201711; pmid: 21551061
10. A. R. Shenoy et al., GBP5 promotes NLRP3 inflammasome
assembly and immunity in mammals. Science 336, 481–485
(2012). doi: 10.1126/science.1217141; pmid: 22461501
11. S. D. Tyrkalska et al., Neutrophils mediate Salmonella
Typhimurium clearance through the GBP4 inflammasomedependent production of prostaglandins. Nat. Commun. 7,
12077 (2016). doi: 10.1038/ncomms12077; pmid: 27363812
12. M. P. Wandel et al., Guanylate-binding proteins convert
cytosolic bacteria into caspase-4 signaling platforms.
Nat. Immunol. 21, 880–891 (2020). doi: 10.1038/s41590-0200697-2; pmid: 32541830
13. R. G. Gaudet et al., A human apolipoprotein L with detergentlike activity kills intracellular pathogens. Science 373, eabf8113
(2021). doi: 10.1126/science.abf8113; pmid: 34437126
14. D. Fisch et al., Human GBP1 is a microbe-specific gatekeeper
of macrophage apoptosis and pyroptosis. EMBO J. 38,
e100926 (2019). doi: 10.15252/embj.2018100926;
pmid: 31268602
15. M. Kutsch et al., Direct binding of polymeric GBP1 to LPS
disrupts bacterial cell envelope functions. EMBO J. 39,
e104926 (2020). doi: 10.15252/embj.2020104926;
pmid: 32510692
16. J. C. Santos et al., Human GBP1 binds LPS to initiate assembly
of a caspase-4 activating platform on cytosolic bacteria.
Nat. Commun. 11, 3276 (2020). doi: 10.1038/s41467-02016889-z; pmid: 32581219
17. E. Kravets et al., Guanylate binding proteins directly attack
Toxoplasma gondii via supramolecular complexes. eLife 5,
e11479 (2016). doi: 10.7554/eLife.11479; pmid: 26814575
18. G. J. Praefcke, H. T. McMahon, The dynamin superfamily:
Universal membrane tubulation and fission molecules?
Nat. Rev. Mol. Cell Biol. 5, 133–147 (2004). doi: 10.1038/
nrm1313; pmid: 15040446
19. B. H. Kim, A. R. Shenoy, P. Kumar, C. J. Bradfield,
J. D. MacMicking, IFN-inducible GTPases in host cell defense.
Cell Host Microbe 12, 432–444 (2012). doi: 10.1016/
j.chom.2012.09.007; pmid: 23084913
20. B. Prakash, G. J. Praefcke, L. Renault, A. Wittinghofer,
C. Herrmann, Structure of human guanylate-binding protein
1 representing a unique class of GTP-binding proteins.
Nature 403, 567–571 (2000). doi: 10.1038/35000617;
pmid: 10676968
21. A. Ghosh, G. J. Praefcke, L. Renault, A. Wittinghofer,
C. Herrmann, How guanylate-binding proteins achieve
assembly-stimulated processive cleavage of GTP to GMP.
Nature 440, 101–104 (2006). doi: 10.1038/nature04510;
pmid: 16511497
22. C. Ji et al., Structural mechanism for guanylate-binding
proteins (GBPs) targeting by the Shigella E3 ligase IpaH9.8.
PLOS Pathog. 15, e1007876 (2019). doi: 10.1371/
journal.ppat.1007876; pmid: 31216343
23. W. Cui et al., Structural basis for GTP-induced dimerization and
antiviral function of guanylate-binding proteins. Proc. Natl.
Acad. Sci. U.S.A. 118, e2022269118 (2021). doi: 10.1073/
pnas.2022269118; pmid: 33876762
24. S. Shydlovskyi et al., Nucleotide-dependent farnesyl switch
orchestrates polymerization and membrane binding of human
guanylate-binding protein 1. Proc. Natl. Acad. Sci. U.S.A. 114,
E5559–E5568 (2017). doi: 10.1073/pnas.1620959114;
pmid: 28645896
25. B. H. Kim et al., Interferon-induced guanylate-binding proteins
in inflammasome activation and host defense. Nat. Immunol.
17, 481–489 (2016). doi: 10.1038/ni.3440; pmid: 27092805

p

their neighboring residues (Lys-520, Ser-523,
Glu-526, His-527, Gln-530, and Glu-533) located in the a12 in the GED domain and
picked the closest neighbor residues for crosslinking candidates (between 2 to 7 Angstroms
apart). After mutating these residues to Cysteines, we directly visualized their targeting
to cytosolic Stm in cellulo (Fig. S13A,B). This
confirmed no gross structural alterations that
would interfere with targeting. We chose the
middle R362C-E526C mutant with the shortest Cys-Cys interbond length (2 Å) to make
recombinant RFP-GBP1 for direct cell-free
coating assays. It was examined by in vitro
reconstitution assays in the presence or absence of DTT, a reducing agent. Specifically,
Cys-paired GBP1R362C-E526C tagged with mRFP
were purified and mixed with bacteria (OD =
0.1) in coating buffer (HEPES buffer with
150 mM NaCl, 1 mM MgCl2, pH 7.4). Reconstitution assays under reducing conditions
was performed by adding a final concentration of 1 mM DTT into the coating buffer.

RES EARCH | R E S E A R C H A R T I C L E

180, 519–530 (2012). doi: 10.1016/j.jsb.2012.09.006;
pmid: 23000701
57. S. Zhu, J. MacMicking, Native structure of a human GBP
defense complex for cell-autonomous immunity to infection,
v1, Zenodo (2021); doi: 10.1101/2021.08.26.457804
ACKN OWLED GMEN TS

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adm9903
Figs. S1 to S13
Tables S1 to S4
MDAR Reproducibility Checklist
Movies S1 to S6
Submitted 27 October 2021; resubmitted 12 September 2023
Accepted 17 January 2024
10.1126/science.abm9903

g

We thank R. Gaudet, E. Groisman, X. Liu, A. Martinez-Sanchez,
J. Nikolaus, G. Torrence, A. Tunaru, Z. Yu, and X. Zhao for
experimental advice and technical help; S. Wu, M. Llaguno,
J. Lin, Z. Yu, X. Zhao, and R. Yan for the help with cryo-ET data
collection; and J. Galan, R. Curtis III, S. Y. Wanda, B. Kazmierczak,
F. Isaacs, and H. Agaisse for bacterial strains. Funding: We
acknowledge financial support from the NIH National Institutes of
Allergy and Infectious Diseases (R01AI068041-13, R01AI108834-05)
to J.D.M; and the National Institute of General Medical
Sciences (R01 GM118486), National Institute of Diabetes and
Digestive and Kidney Diseases (P30 DK045735), and Wellcome
Trust (grant no. 203285/B/16/Z) to J.B. Author contributions:
S.Z. performed all cryo-EM and cryo-ET sample preparation,
data collection, and processing (3D reconstruction, segmentation,
and subtomogram averaging); C.J.B. discovered the GBP coat
complex on cytosolic Salmonella and GBP codependencies and
undertook GBP1 mutational analysis; A.M. performed all live
3D-SIM and 4Pi-SMS imaging plus in situ LPS release and
peptidoglycan and flagellin detection assays; E.S.P. and B.H.K.
developed the entire GBP-COAT450–708 platform, bacterial mutant and
reporter strains, CRISPR-Cas9 human cell lines and genetic
complementation, cytokine and cell death assays, in vitro LPS
release assays, and bacterial killing; P.K. and S.H. performed
recombinant protein preparation and analysis; M.K. undertook
3D-SIM imaging on GBP Cys-Cys mutants; and Y.Z. undertook 4PiSMS imaging with J.B. as a collaborator. J.D.M. conceptualized
the project and wrote the paper. All authors discussed the results

and commented on the manuscript. Competing interests:
J.B. declares financial interests in Bruker Corporation and
Hamamatsu Photonics. Y.Z. and J.B. have filed a patent application
on 4Pi-SMS (patent application no. 11209367). J.D.M is an
Investigator of the Howard Hughes Medical Institute. Data and
materials availability: Subtomogram averages of the native
GBP1 dimer and monomeric subunit of the dimer have been
deposited in the Worldwide Protein Data Bank (wwPDB) Deposition
and Annotation system (EMD-43091 and EMD-43153). Raw EM
images from tilt series and segmentation have been deposited in
EMPIAR (EMPIAR-11822). AlphaFold2 rendering of extended
GBP1 conformers have been deposited in Zenodo (accession code
10429400) (57). All other data are available in the main text or
the supplementary materials. License information: Copyright ©
2024 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse. This article is subject to
HHMI’s Open Access to Publications policy. HHMI lab heads have
previously granted a nonexclusive CC BY 4.0 license to the public
and a sublicensable license to HHMI in their research articles.
Pursuant to those licenses, the Author Accepted Manuscript (AAM)
of this article can be made freely available under a CC BY 4.0
license immediately upon publication.

p

47. J. Jumper et al., Highly accurate protein structure prediction
with AlphaFold. Nature 596, 583–589 (2021). doi: 10.1038/
s41586-021-03819-2; pmid: 34265844
48. S. W. Brubaker, K. S. Bonham, I. Zanoni, J. C. Kagan, Innate
immune pattern recognition: A cell biological perspective.
Annu. Rev. Immunol. 33, 257–290 (2015). doi: 10.1146/
annurev-immunol-032414-112240; pmid: 25581309
49. M. Du, Z. J. Chen, DNA-induced liquid phase condensation of
cGAS activates innate immune signaling. Science 361,
704–709 (2018). doi: 10.1126/science.aat1022;
pmid: 29976794
50. J. Mahamid et al., Visualizing the molecular sociology at the
HeLa cell nuclear periphery. Science 351, 969–972 (2016).
doi: 10.1126/science.aad8857; pmid: 26917770
51. R. Staritzbichler, N. Ristic, A. Goede, R. Preissner, P. W. Hildebrand,
Voronoia 4-ever. Nucleic Acids Res. 49 (W1), W685–W690
(2021). doi: 10.1093/nar/gkab466; pmid: 34107038
52. M. Graf et al., Visualization of translation termination
intermediates trapped by the Apidaecin 137 peptide during
RF3-mediated recycling of RF1. Nat. Commun. 9, 3053 (2018).
doi: 10.1038/s41467-018-05465-1; pmid: 30076302
53. S. Miletic et al., Substrate-engaged type III secretion system
structures reveal gating mechanism for unfolded protein
translocation. Nat. Commun. 12, 1546 (2021). doi: 10.1038/
s41467-021-21143-1; pmid: 33750771
54. A. Martinez-Sanchez, I. Garcia, S. Asano, V. Lucic,
J. J. Fernandez, Robust membrane detection based on tensor
voting for electron tomography. J. Struct. Biol. 186, 49–61
(2014). doi: 10.1016/j.jsb.2014.02.015; pmid: 24625523
55. A. Martinez-Sanchez et al., Template-free detection and
classification of membrane-bound complexes in cryo-electron
tomograms. Nat. Methods 17, 209–216 (2020). doi: 10.1038/
s41592-019-0675-5; pmid: 31907446
56. S. H. W. Scheres, RELION: Implementation of a Bayesian
approach to cryo-EM structure determination. J. Struct. Biol.

y
y g
y
,

Zhu et al., Science 383, eabm9903 (2024)

1 March 2024

18 of 18

RES EARCH

RESEARCH ARTICLE SUMMARY

◥

PLANT SCIENCE

Defective pollen tube tip growth induces
neo-polyploid infertility
Jens Westermann, Thanvi Srikant, Adrián Gonzalo, Hui San Tan, Kirsten Bomblies*

1 of 1

,

1 March 2024

y

Westermann et al., Science 383, 964 (2024)

y g

▪

y

CONCLUSION: Our work shows that
pollen tube tip growth is defective in newly formed A. arenosa
and A. thaliana polyploids, providing a previously unexplored component of neo-polyploid sterility
and a potential explanation for
Stebbins’s nonmeiotic sterility factor. Our transcriptome work provides insights into what problems
arise in neo-polyploid pollen tubes
and shows that established tetraploids did not simply return pollen tube gene expression to the
ancestral state during their evolution to overcome the neo-polyploid
challenge. Notably, we identify two
candidate genes and show genetically that derived tetraploid
alleles associate with improved
of pollen tubes. One grain has a pollen tube that burst immediately after
pollen tube tip growth. We also
germination. Several grains have just started germinating. Tubes are
RATIONALE: We used a “reverse geshow that it is the gametophyte,
stained with aniline blue, which marks callose, and the bright bands show
nomics of adaptation” approach
not the sporophyte, genotype that
callose plugs.
to identify what additional chalconditions these effects. The two
lenges to fertility might be imporgenes we identify are promising
perimentally induced neo-polyploids grow
tant in newly formed polyploids. We did this
candidates, motivating further studies to conslowly, with extensive abnormalities, includby focusing on a list of genes under selecfirm their mechanistic roles, find additional
ing stalling, branching, bulging, premature
tion that we had previously identified from
players, and, importantly, test whether they
bursting, disrupted calcium gradients, dea genome scan in a natural autotetraploid
can be transferred to other species to imfective actin filament formation, and ectopic
lineage of Arabidopsis arenosa. Among these
prove neo-polyploid fertility.
cell wall softening. In established tetraploids,
genes, we found two with strong signals of
which exist in A. arenosa but not A. thaliana,
positive selection that are known, from muInstitute of Molecular Plant Biology, Department of Biology,
ETH Zürich, Zürich, Switzerland.
the physiology and morphology of pollen
tant studies in Arabidopsis thaliana, to play
*Corresponding author. Email: kirsten.bomblies@biol.ethz.ch
tubes have returned to normal. The defects
a bona fide role in pollen tube tip growth.
Cite this article as J. Westermann et al., Science 383,
seen in pollen tubes of neo-tetraploids are
Given that pollen tube tip growth is esseneadh0755 (2024). DOI: 10.1126/science.adh0755
coupled with perturbations in gametophyttial for fertility across sexually reproducing
READ THE FULL ARTICLE AT
ic gene expression, including expression of
seed plants, we aimed to test whether it might
https://doi.org/10.1126/science.adh0755
several genes that could be involved in genindeed present a challenge for neo-polyploids

g

INTRODUCTION: Polyploidy arises from wholeand whether the novel alleles of the two cangenome duplication. It is implicated in evoludidate genes that led us to this hypothesis
tion and adaptation and thus is of fundamental
might play a role in solving it.
interest in evolutionary biology. AdaptaRESULTS: We show that pollen tube tip growth is
tion of polyploid lineages to their genomeseverely challenged in neo-polyploid A. arenosa
duplicated state provides an interesting case
and A. thaliana. The pollen tubes from exstudy for evolutionary cell biology: In newly formed polyploids, a diploidoptimized genome is suddenly in
a novel cellular context to which
it is not adapted, and this could
necessitate evolutionary retuning
of cellular functions. There is also
increasing interest in using polyploidy as a crop improvement tool,
because polyploids are often resistant to climate-relevant stresses—
most prominently, drought. However, for understanding their evolution or using them in agriculture,
the fact that most new polyploids
have very low fertility remains a key
stumbling block. Meiotic chromosome segregation is a well-known
challenge for new polyploids, but
we have recently shown that improving neo-polyploid meiotic stability
does not rescue fertility on its own.
This suggests, as G. L. Stebbins already proposed 75 years ago, that
Abnormal pollen tubes from a neo-tetraploid A. arenosa plant.
there is something else that afMicroscope image showing pollen grains and two defective pollen tubes
fects polyploid fertility—what that
from a neo-tetraploid A. arenosa plant growing in vitro. These defects show
might be has, to date, remained
that neo-tetraploids suffer a defect that disrupts the normal polar growth
mysterious.

p

erating the observed phenotypes; many, but
not all, of these return to the ancestral expression state in the evolved tetraploids. We
tested genetically whether genotypes of the
two strongly ploidy-differentiated genes we
identified, which encode a kinase AGC1.5
and a calcium-exporting adenosine triphosphatase ACA8, correlate with pollen tube performance in tetraploids. We find that diploid
pollen tubes homozygous for tetraploid alleles at both loci strongly outcompete pollen
tubes carrying diploid alleles in vivo when
used to fertilize tetraploid females. In segregating lines, homozygotes for tetraploid
alleles at AGC1.5 and ACA8 loci produce
pollen tubes with markedly improved growth
compared with plants homozygous for diploid alleles, both in terms of growth rate and
prevention of premature bursting. The two
loci thus additively associate with
improved growth of pollen tubes
after whole-genome duplication
and may represent good candidates for improving fertility of neopolyploids in other species.

RES EARCH

RESEARCH ARTICLE

◥

PLANT SCIENCE

Defective pollen tube tip growth induces
neo-polyploid infertility
Jens Westermann, Thanvi Srikant, Adrián Gonzalo, Hui San Tan, Kirsten Bomblies*
Genome duplication (generating polyploids) is an engine of novelty in eukaryotic evolution and a
promising crop improvement tool. Yet newly formed polyploids often have low fertility. Here we report
that a severe fertility-compromising defect in pollen tube tip growth arises in new polyploids of
Arabidopsis arenosa. Pollen tubes of newly polyploid A. arenosa grow slowly, have aberrant anatomy and
disrupted physiology, often burst prematurely, and have altered gene expression. These phenotypes
recover in evolved polyploids. We also show that gametophytic (pollen tube) genotypes of two tip-growth
genes under selection in natural tetraploid A. arenosa are strongly associated with pollen tube
performance in the tetraploid. Our work establishes pollen tube tip growth as an important fertility
challenge for neo-polyploid plants and provides insights into a naturally evolved multigenic solution.

1 of 10

,

We measured average in vitro PTGR for pollen
from A. arenosa diploids (“2X,” SNO) and
established autotetraploids from two genetically distinct populations (“est-4X,” TBG and
SBG). We also created neo-tetraploids in the
lab: “neo-4XC0” were generated by colchicineinduced whole-genome duplication of SNO, and
“neo-4XC1,” which were used for most of our
experiments, are next-generation neo-polyploids
created by inter crossing neo-4XC0 plants. The
latter are important because they have not
been directly exposed to colchicine. The pollen
tubes of neo-4XC0 and neo-4XC1 did not differ
significantly for any of the traits measured, nor
did those produced by 2X branches of chimeric
colchicine-treated plants differ from untreated
2X, verifying that the defects observed here are
not attributable to a toxic effect of colchicine.
PTGR was similar for haploid pollen tubes
from 2X plants and diploid pollen tubes from
est-4X plants but significantly lower in diploid
pollen tubes from neo-4XC0 plants (Fig. 1A;
P < 0.001 for both, Kruskal-Wallis with Dunn’s
multiple comparisons test). The same was true
in a second experiment using neo-4XC1 (see the
“Genetic analysis of ACA8 and AGC1.5 alleles”

y

1 March 2024

Results and discussion
Neo-tetraploids have slow pollen tube growth

y g

Westermann et al., Science 383, eadh0755 (2024)

y

*Corresponding author. Email: kirsten.bomblies@biol.ethz.ch

g

Institute of Molecular Plant Biology, Department of Biology,
ETH Zürich, Zürich, Switzerland.

Arabidopsis thaliana neo-tetraploids by reducing crossover numbers does not rescue fertility
on its own (11). This result raised the question
of what other factor (or factors) might contribute to the low fertility of neo-polyploids.
We set out to identify what else might affect
neo-polyploid fertility using a “reverse genetics
of adaptation” approach (12) by reexamining
lists of genes that we had previously identified
as targets of directional selection in a natural
autotetraploid lineage of A. arenosa (9, 13) that
arose once, about 30,000 generations ago (14).
Two genes among those most differentiated
between diploids and established tetraploids
encode proteins involved in pollen tube tip
growth (a type of polar cell growth found across
eukaryotes, which is essential for fertility in
seed plants). One encodes AGC kinase 1.5
(AGC1.5), which in A. thaliana is expressed
only in growing pollen tubes and controls Ca2+
homeostasis, actin dynamics, and vesicle transport (15, 16). The second encodes AUTOINHIBITED CA2+-ATPASE8 (ACA8). ACAs are
Ca2+ exporters (17), and ACA8, although expressed broadly, is among several ACAs important in pollen tube growth (18, 19). These
two genes were the only known pollen-related
genes with strong evidence of selection in tetraploid A. arenosa. Therefore, we hypothesized
that pollen tube tip growth might be a fertilitycompromising challenge for neo-polyploids and
that these genes might contribute to solving it.
To our knowledge, a failure in pollen tube
tip growth has not been specifically connected
to low fertility in neo-polyploids. There is, however, prior data that pollen tube growth rate
(PTGR) is slow in neo-polyploids relative to
diploid progenitors and that it recovers in
established polyploids (20). One possible cause
is that because genome duplication leads to
increased cell and pollen grain size (2, 3, 21),
cell wall deposition rates may not suffice for

p

G

enome duplication, giving rise to polyploids, can arise spontaneously in nature
through the fusion of unreduced gametes or can be induced, for example, by
application of chemical treatments in
the lab (1, 2). Polyploidy has important implications in evolution but also has potential as
a powerful crop improvement tool because
polyploids often have enhanced resilience to
climate-relevant abiotic stresses (1–3). In one
example, diploid and neo-tetraploid ryegrass
tested under current and future climate conditions for southern Britain showed that while
diploids currently perform better in terms of
biomass production, in simulated future conditions with more frequent drought, tetraploids
produced more biomass (4). However, newly
formed polyploids (“neo-polyploids”) often
have very low fertility; understanding why,
and what solutions can evolve, is important both
for understanding an evolutionarily important
phenomenon and for better exploiting polyploidy in agriculture.
Meiotic chromosome missegregation is a wellrecognized problem for neo-polyploids that can
compromise fertility because the multiple available homologs can associate into multivalents
that impair chromosome segregation (5, 6). We
showed previously that at least eight meiosis
genes are targets of directional selection in a
natural autotetraploid lineage of Arabidopsis
arenosa and, for at least three of them, that
naturally evolved tetraploid alleles help stabilize polyploid meiosis (7–9). But fitting with G. L.
Stebbins’s speculation from 75 years ago that
there is also an “unexplained nonmeiotic factor”
affecting neo-polyploid fertility (10), we recently
showed that reducing multivalent frequency in

larger pollen tubes of polyploids (22). Comparisons between diploid and established polyploid
species of the genera Betula and Handroanthus
show that polyploids have wider pollen tubes but
grow at the same rate as those of diploids, suggesting that evolution can compensate for size
changes (22). One direct comparison of diploids with both neo- and established polyploids
showed that siring success of pollen from neotetraploid Chamerion angustifolium is lower
than that of established autotetraploids (23), but
whether reduced fertility was caused by pollen
tube defects is unknown.
We hypothesize that slow PTGR in neopolyploids, which on its own would not necessarily cause low fertility, is a symptom of
broader physiological problems that do compromise fertility. Higher PTGR of evolved polyploids relative to neo-polyploids may thus arise
as a by-product of solving physiological problems, as opposed to higher PTGR per se being
targeted by selection. Consistent with this idea,
morphological abnormalities have been reported
in pollen tubes of several neo-polyploids (24–26).
Modeling studies show that coordination of cell
size, turgor, and cell wall properties is essential
for maintaining steady polar growth of pollen
tubes (27), so cellular defects could in principle
underlie both low PTGR and morphological
defects. Here, we show that pollen tube tip
growth is indeed defective in neo-polyploids
and then test whether the alleles of ACA8 and
AGC1.5 that were under selection in established
tetraploid A. arenosa might compensate for the
defects seen in neo-polyploids.

RES EARCH | R E S E A R C H A R T I C L E

1 March 2024

Transcriptomic differences in growing
pollen tubes

To further understand the nature of the pollen
tube defects in neo-tetraploids and the “rescue”
in evolved tetraploids, we sequenced transcriptomes of growing pollen tubes from 2X (n = 2),
neo-4XC1 (n = 3), and est-4X (TBG, n = 2).
2 of 10

,

Across plant species, an intracellular tip-focused
Ca2+ concentration gradient (highest [Ca2+] at
tip) is critical for processes such as vesicle
trafficking, pectin cross-linking, and signal
transduction during pollen tube growth (28–30).
We visualized intracellular [Ca2+] with Fluo4AM in actively growing pollen tubes in vitro
and quantified the [Ca2+] gradient (“m”) (see
materials and methods), where m < 0 indicates the expected “tip-focused” gradient. Pollen
tubes from neo-4XC1 plants frequently lacked a
tip-focused [Ca2+] gradient (mean m = 0.19;
22% of tubes m < 0; Fig. 2A). In contrast, tipfocused gradients were much more frequent
in pollen tubes from 2X (mean m = −1.20; 82%
with m < 0) and est-4X (mean m = −1.72; 76%

y

Pollen tubes of neo-tetraploids have
physiological defects

y g

Westermann et al., Science 383, eadh0755 (2024)

To test whether the pollen tube defects of neopolyploids we observed in vitro correlate with
fertility problems in vivo, we pollinated pistils
from est-4X and neo-4XC0 flowers with pollen
from both types in all four combinations and then
stained carpels with aniline blue after 24 hours
to visualize pollen tubes. At time of harvest, the
bulk of pollen tubes from neo-4XC0 covered a
smaller range of the female transmitting tract
than those of est-4X (Fig. 1H and fig. S6A).
Although we cannot discern whether these
tubes are bursting, or stalling for some other
reason, the result does suggest that a high rate
of pollen tube attrition also occurs in vivo for
pollen from neo-4XC0. Unfertilized ovules were
common in est-4X siliques pollinated by neo4XC0 plants (fig. S6B), and seed set per silique
was low (fig. S6, C and D). Some pollen tubes
from neo-4XC0 also showed evidence of not
targeting ovules normally (fig. S6B). Seed set
was not biased to apical ovules (fig. S6C),
suggesting that some pollen tubes from neo-4XC0
do grow long enough to fertilize ovules at the
base of the carpel, consistent with the few long
tubes we saw in vitro (Fig. 1G and fig. S4).

y

We observed prominent defects in morphology
of diploid pollen tubes from neo-4XC0 plants.
When growing in vitro, pollen tubes from neo4XC0 burst prematurely at a high frequency
compared to those of 2X or est-4X (P < 0.0001
in both cases, Kruskal-Wallis with Dunn’s multiple comparisons test, see table S4; Fig. 1, D to
F; see movie S1 for an example); the same was
true for neo-4XC1. PTGR and bursting rate per
genotype or cytotype correlated negatively,
suggesting that both may be symptoms of the
same underlying defects (fig. S3). In addition,
as opposed to the steady directional growth
observed in pollen tubes from 2X and est-4X,
pollen tubes from neo-4XC0 often showed bulging, branching, meandering (Fig. 1F), and reduced length (fig. S4). Nevertheless, a few pollen
tubes from neo-4XC0 grew quite long, suggesting
that defects arise stochastically (fig. S4).
Constant ratios of pollen tube width between the tip (wt, widest point 0 to 10 mm from
tip) and shank (ws, widest point 10 to 30 mm
from tip) are fundamental to normal polar

Neo-polyploids have low pollen fertility

with m < 0) relative to neo-4XC1 (P values <
0.027 and 0.040, respectively, c2 test with
Holm-Bonferroni correction; Fig. 2A, fig. S7,
and movie S2). 2X and est-4X did not differ
significantly for proportion of tubes with tipfocused gradients (P > 0.99 for both, c2 test
with Holm-Bonferroni correction). In a second
independent experiment, we saw similar trends,
but puzzlingly, a hybrid between est-4X and
neo-4XC0 (“hybrid-F1”) showed an even lower
percentage of tubes with a tip-focused gradient
(88%) than neo-4XC1 yet had intermediate
PTGR and bursting rates between those of
neo-4XC1 and est-4X (see the “Genetic analysis
of ACA8 and AGC1.5 alleles” section). The often
bright but even pattern seen in tubes from
neo-4XC1 plants hints that there might be a
problem with calcium export from the shank,
rather than import at the tip.
In A. thaliana pollen tubes, two features vital
for tip growth are subapical actin filaments
(“SA”) that transit tip-directed vesicle, and
an apical actin ring (“AR”) that maintains
directionality of turgor-driven growth (31).
Neo-4XC1 had a reduced frequency of tubes
with clearly visible SA (30%) compared with
2X and est-4X (78 and 73%, respectively; P <
0.0001 for both, c2 test adjusted with HolmBonferroni correction, table S4; Fig. 2B and
figs. S7 and S10B), whereas 2X and est-4X
did not significantly differ (P = 0.99, c2 test
with Holm-Bonferroni correction, table S4).
The frequency of AR in pollen tubes from
neo-4XC1 was not significantly different from
2X (P = 0.22, c2 test with Holm-Bonferroni
correction) but was significantly lower than
est-4X (P = 0.0012, c2 test with Holm-Bonferroni
correction; Fig. 2B and figs. S7 and S10B). The
maximum length of visible actin fibers was also
lower in pollen tubes from neo-4XC1 (fig. S10C).
Cell wall softening is important in tip growth
and is associated with the presence of acidic
pectin esters (32). For 2X and est-4X, most in
vitro–grown pollen tubes had the expected
single tip–localized acidic pectin signal (86
and 90.4%, respectively), but pollen tubes from
neo-4XC1 plants commonly had multiple sites
(Fig. 2C), and significantly fewer had just a
single signal at the tip (27.0%; P < 0.0001 for
both, c2 test with Holm-Bonferroni correction). The multiple signals in pollen tubes from
neo-4XC1 are associated with the bulging,
branching, and bending points (Fig. 2C), but
whether aberrant pectin acidification is a
cause or consequence of these morphological
problems is not clear.

g

Pollen tubes from neo-tetraploids have severe
morphological defects

growth and, when compromised, can lead to
growth defects (27). The wt/ws ratios for growing pollen tubes from 2X and est-4X fall in a
tight distribution around 1, as expected (s = 0.1;
Fig. 1G). For pollen tubes from neo-4XC0, the
mean also centered on 1, but the spread of
ratios was significantly higher (s = 0.4; Fig. 1G;
all P values < 0.001, F-test with Holm-Bonferroni
correction, table S4). The extent to which wt/ws
ratios deviated from 1 correlated negatively with
PTGR (Pearson correlation coefficient r = −0.54).
We also tested whether pollen tubes of neo4XC0 A. thaliana (neo-4XAt) show similar defects to those from neo-4X A. arenosa. Indeed,
in vitro–grown pollen tubes from neo-4XAt individuals of the Col-0 accession also showed
elevated bursting, bulging, branching, and
growth cessation relative to 2X (fig. S5).

p

section). Many pollen tubes from neo-4XC0 and
neo-4XC1 showed little or no growth, so to determine whether their calculated slow average PTGR
reflects global problems or instead arises simply
because a higher proportion of “zero or nearzero growth” tubes are included in calculating
the mean, we analyzed just the 20% fastestgrowing tubes from the second experiment
(which had sufficient sample sizes). Although
the differences are less extreme, the fastest
pollen tubes from neo-4XC1 plants still grew
significantly slower than those of either 2X or
est-4X plants (P < 0.0001 for both; KruskalWallis with Dunn’s multiple comparisons test;
fig. S1). Viability of ungerminated pollen estimated by fluorescein diacetate/propidium
iodide (FDA/PI) and Alexander staining was
high for all genotypes, albeit lower for neo-4XC0
(fig. S2, A to C). There was no significant difference in pollen germination rate in vitro (fig.
S2D). These results suggest that neo-4XC0 defects are mostly related to tube growth rather
than resulting from a catastrophic loss of pollen
grain viability.
Pollen grains of neo-4XC0 plants were significantly larger than those of either 2X or est-4X
(Fig. 1B), and pollen tubes from neo-4XC0 were
also wider on average than those from 2X or one
est-4X genotype (TBG; Fig. 1C). A second est-4X
genotype (SBG), however, produced pollen tubes
of equal width to those of neo-4XC0 (Fig. 1C) yet
had PTGR comparable to that of the other
est-4X (TBG versus SBG; Fig. 1A). Thus, even
if increased tube width might initially be a
problem for neo-tetraploids, reversing the size
increase is apparently not essential for recovering from it, consistent with results from
Betula and Handroanthus (22).

RES EARCH | R E S E A R C H A R T I C L E

b

2

a

400
300

16

80

14
12
10
8
6

40
a

a

c

10
0

Est-4X
(TBG)
Est-4X
(SBG)

2X

Neo-4X C0

T S T S T S T S

Est-4X
(TBG)
Est-4X
(SBG)

Neo-4X C0

0
2X

Est-4X
(SBG)

Est-4X
(TBG)

0

Neo-4X C0

0

2X

100

50

20

2

1

b

60

30

4

200

70

Est-4X
(TBG)
Est-4X
(SBG)

4

c

a

500

90

2X

c

600

100

18
Bursting rate (%)

Pollen grain size ( m 2 )

5

3

b

700
c

D
20

Tube width ( m)

a
6

PTGR ( m/min)

C
800

7

Neo-4X C0

B

A

E
p
g

Neo-4XC0

2X

F

Est-4X

Normal (2X)

Neo-4X C0 morphology

Est - 4X (SBG)

1

10
8
6

2
0
0

1

2

3

0

1

2

3

0

1

2

Tip to shank width ratio (w t /w s )

Fig. 1. Pollen defects in neo-tetraploids. (A) Pollen tube growth rate
(micrometers per minute). n ≥ 95 tubes per genotype, five plants per cytotype.
Different letters indicate significant differences (P < 0.05, Kruskal-Wallis, Dunn’s
multiple comparison test). (B) Pollen grain size. n ≥ 95 tubes per genotype,
five plants per cytotype. Different letters indicate significant differences (P < 0.05,
Kruskal-Wallis, Dunn’s multiple comparison test). (C) Box plots of maximal tube
width 0 to 10 mm from tip (T) and shank (S; 10 to 30 mm from tip; n ≥ 95 tubes per
genotype). Whiskers indicate minimum and maximum (Q1-1.5IQR), and black dots
are outliers. (D) Rate of bursting in vitro (n ≥ 532 grains per genotype). Different
letters indicate significant differences (P < 0.05, c2 test, Holm-Bonferroni
Westermann et al., Science 383, eadh0755 (2024)

1 March 2024

3

0

1

2

3

a a

80
60

b b

40
20
0

,

4

100

y

PTGR ( m/min)

12

H

Est - 4X x Est-4X
Neo- 4XC0 x Est-4X
Est-4X x Neo - 4XC0
Neo-4XC0 x Neo - 4XC0

Est-4X (TBG)

Neo-4X C0
0.1

y g

2X

14

Range of transmitting
tract covered (%)

y

G

correction). Black dots in (A), (B), and (D) indicate individual means. Error
bars in (A) and (B) show standard deviation. (E) Examples of in vitro–grown pollen
tubes. Scale bar, 200 mm. Arrowheads indicate “bursting” tubes. (F) In–vitro
grown pollen tubes from the diploid (far left panel) and neo-4XC0 (right
panels). Scale bar, 50 mm. (G) Ratio of greatest width at tube tip versus shank
(see methods), plotted against growth rate (micrometers per minute).
Black horizontal lines indicate mean PTGR. (H) Box plot showing range of
transmitting tract covered for pollen tubes from range measured in fig. S6A
[a versus b, significantly differing groups (Welch’s ANOVA, Dunnett’s T3 multiple
comparisons test)].
3 of 10

RES EARCH | R E S E A R C H A R T I C L E

A

2X

Neo-4XC1

Est-4X

[Ca2+]

Fluorescence

180

m = -1.79

m = 0.19

m = -1.20
130
80
30
-20 0

10

20

30

0

10

3 0 0

20

10

20

30

Distance from tip ( m)

B

C

F-actin

Esterified pectin

2X

2X

4 of 10

,

Neo-4XC1
Est-4X
1 March 2024

y

Westermann et al., Science 383, eadh0755 (2024)

of ectopic pectin acidification and frequent
bursting in these tubes (Fig. 2C). AGP2 encodes
an arabinogalactan protein important for cell
wall deposition during pollen tube growth [e.g.,
(35)]. Of the nine DEGs in the neo-4XC1 versus
2X contrast, three (RBR1, AGP2, and RALFL11)
were also DEGs between neo-4XC1 and est-4X.
In the diploid pollen tubes of est-4X, RBR1 and
AGP2 expression returned to levels seen in
haploid pollen tubes of diploids. In contrast,
RALFL11, a member of a peptide family that
maintains cell wall integrity during pollen tube
growth [e.g., (36, 37)], is expressed at an even
higher level in pollen tubes of est-4X than in
those of neo-4XC1 (Fig. 3E). Other notable
genes on the list of 27 pollen or Ca2+-related
DEGs include MIRO2, which encodes a mitochondrial guanosine triphosphatase associated
with calcium signaling (38); CDI, which encodes
a protein that affects male fertility by regulating
pollen tube growth (39); ACA7, which is closely
related to ACA8; and BON3, whose interaction

y g

Among 15,361 genes with sufficient read counts
across all samples, 790 genes were differentially
expressed in 2X versus neo-4XC1, or in est-4X
versus neo-4XC1, or both. The contrast est-4X
versus neo-4XC1 had the most differentially
expressed genes (DEGs; Fig. 3, A to D). Among
the 790 DEGs, we only found 27 whose A.
thaliana orthologs were previously directly or
indirectly associated with pollen tube growth
and/or Ca2+ homeostasis (Fig. 3E and table S1).
Nine of these 27 genes were DEGs comparing
2X versus neo-4XC1 (Fig. 3, B and E). Among
these are several interesting genes: PMEI5,
which is more highly expressed in neo-4XC1 than
in 2X, encodes a pectin methylesterase inhibitor
that regulates cell wall esterification (33). A
similar protein in maize, ZmPMEI1, is localized to the tip and sites of bending in pollen
tubes, and its external application induces cell
wall destabilization and pollen tube bursting
(34). Thus, overexpression of PMEI5 in pollen
tubes from neo-4XC1 fits with our observation

y

Fig. 2. Physiological abnormalities in pollen tubes from neo-polyploids. (A) (Top) Example images of
growing pollen tubes stained with Fluo4-AM to mark Ca2+. Scale bar, 20 mm. (Bottom) Average (mean)
Fluo4-AM fluorescence intensity along first 30 mm of tube calculated from 26 growing pollen tubes per
cytotype. Standard deviation shown in gray. m, mean inclination of signal (see methods). (B) Filamentous
actin cytoskeleton visualized with Phalloidin-AlexaFluor405. Scale bar, 10 mm. (C) Ruthenium red staining of
esterified pectins indicating expected tip-focused signal (white arrowhead) and ectopic sites (yellow
arrowheads). Scale bar, 50 mm. Additional representative images are shown in fig. S7. All assays had five
biological replicates per cytotype, with three technical replicates each.

The results presented thus far support the hypothesis that defects in tip growth of diploid pollen
tubes from neo-4X plants are associated with,
and likely caused by, physiological perturbations
including loss of the tip-focused [Ca2+]-gradient,
abnormal patterns of cell wall softening, and
cytoskeletal defects. The two tip growth–related
genes that experienced directional selection in
the tetraploid lineage, AGC1.5 and ACA8, are
both implicated in regulating these processes.
Both genes are strongly differentiated (i.e., have
sharply different allele frequencies) among all
2X and est-4X A. arenosa populations tested,
whereas other pollen tube–relevant genes show
weaker differentiation (e.g., the two in Fig.
3E) or no evidence of selection (9, 13); thus, we
chose to focus on AGC1.5 and ACA8 for genetic
characterization.
The diploid- and tetraploid-specific alleles of
AGC1.5 differ only by one amino acid, D148Y
(Asp148→Tyr), which, in a sample of nearly
300 A. arenosa individuals (43), is present at
90% frequency in natural (established) tetraploids and only 0.9% in diploids (fig. S9A).
Amino acid 148 lies at the junction between a
well-conserved kinase domain and an intrinsically disordered N-terminal region; its function is unknown (fig. S9, B and C). ACA8 has
two differentiated amino acids (I424V and
T493A), in the transmembrane and intracellular domains, respectively (fig. S9, A and
D). The derived variants are found in diploids
at about 7% frequency and in established
tetraploids at 69 and 72%, respectively (fig. S9A),
suggesting that for this gene, the tetraploid allele
was likely selected from standing variation already
present in diploids. The promoters do not show
evidence of differentiation in either locus, which
fits with their lack of differential expression.
To test whether these genes affect diploid
pollen tube performance in tetraploid plants,
we generated F2 populations segregating the
ancestral diploid (D) and derived tetraploid
(T) alleles of both AGC1.5 and ACA8. We
crossed neo-4XC0 plants (homozygous DDDD
genotype at both loci) with est-4X plants (TBG,
homozygous TTTT genotype at both loci) to
generate hybrid-F1 plants (DDTT at both loci).
We intercrossed hybrid-F1 plants to generate

g

Est-4X

Genetic analysis of ACA8 and AGC1.5 alleles

p

Neo-4X C1

with ACAs is important in pollen tube growth
(18). Among the 790 DEGs, there is a significant functional enrichment for genes encoding
proteins involved in glutathione metabolism
(n = 9) and components of the exocyst complex (n = 6) (Fig. 3, F to H, and fig. S8). Glutathione redox state affects pollen tube vigor
and size (40). The exocyst complex is important for exocytosis during pollen tube tip growth
(41, 42). Neither ACA8 nor AGC1.5 is among the
790 DEGs across our experiment, fitting with
the observation that their differentiated sites
are all in coding regions (see below).

RES EARCH | R E S E A R C H A R T I C L E

B

#DEGs

600

400

200

C

20

−log10(adjusted p-value)

Downregulated
Upregulated

−log10(adjusted p-value)

A

15

10

5

Downregulated
Non−DEG
Upregulated

D

0

neo-4XC1
vs 2X

Est-4X vs 2X

Est-4X vs
neo-4XC1

0
−

−

−30 −20 −10 0
10 20
log2 Fold Change (est-4X - neo4XC1)

E

F

Genotype
Cytotype

DEGs unique to
neo4XC1 vs 2X
contrast (N=115)

GGCT2.2

GSTL3

I

Transformed
read counts
15
12.5
10
7.5
5

y

G

Genotype Cytotype
2X
SNO(4X)
Neo-4XC1
SNO(2X)
TBG (4X) Est-4X

g

Transformed
read counts
Shared DEGs*
DEGs shared
between neo-4XC1 vs 2X
& Est-4Xvs neo-4XC1 contrasts

p

DEGs unique to
Est-4X vs neo4XC1
contrast (N=584)

Est-4X
vs
neo-4XC1

AT1G64980 / CDI
AT5G01700
AT2G31430 / PMEI5
AT3G12280 / RBR1
AT4G09740 / GH9B14
AT2G22470 / AGP2
AT5G06970 / PATROL1
AT4G02730 / WDR5b
AT3G20790
AT1G16300 / GAPCP-2
AT1G28220 / PUP3
AT2G18960 / AHA1
AT3G01280 / VDAC1
AT2G38720 / MAP65-5
AT5G35930
AT1G51450 / TRO
AT5G01810 / CIPK15
AT1G53210 / NCL
AT1G08860 / BON3
AT5G01820 / CIPK14
AT2G03150 / RSA1
AT3G63150 / MIRO2
AT5G51050 / APC2
AT4G30160 / VLN4
AT1G08450 / CRT3
AT2G19030 / RALF11
AT2G22950 / ACA7

Genes associated with
Genes associated with pollen,
Ca2+ homeostasis/signalling pollen-tube growth, gametogenesis

Shared DEGs*
(N=91)

neo4XC1
vs
2X

log2 Fold Change
10
5
log2Fold Change = NA
0
−5
−10

Identified as DEG

K

Genotype
Cytotype

neo4XC1
vs
2X

Est-4X
vs
neo-4XC1

neo4XC1
vs
2X

Est-4X
vs
neo-4XC1

AT4G31290 / GGCT2.2

GGT1

AT5G02780 / GSTL1
AT1G57720

GSTU12

AT5G42150

y g

AT5G45020

GSTL1

AT3G47680
AT5G02790 / GSTL3

AT5G45020
AT5G42150

L

Genotype
Cytotype
AT1G76850 / SEC5A
AT1G71820 / SEC6
AT5G50380 / EXO70F1
AT1G21170 / SEC5B
AT1G07000 / EXO70B2
AT5G52350 / EXO70A3

SEC5B
SEC6
EXO70F1

Fig. 3. Pollen tube transcriptome variation. (A) Number of differentially
expressed genes (DEGs) across contrasts. (B and C) Volcano plots showing log2
fold change and adjusted P value of 15,361 genes. Black lines indicate DEG
cutoffs. (D) Expression levels of DEGs in different contrasts. Colors represent
read counts. (E) Expression levels of 27 genes from (D) with known associations
with pollen, pollen tube development, and/or Ca2+ signaling (see table S1). Genes
with evidence of directional selection in tetraploid A. arenosa (9, 13) and highly
ploidy-differentiated (43) are written in purple. Corresponding log2 fold changes
across neo-4XC1 versus 2X and est-4X versus 2X contrasts are shown in the
Westermann et al., Science 383, eadh0755 (2024)

AT1G69920 / GSTU12

,

EXO70A3

AT3G47680

J

EXO70B2
SEC5A

AT4G39640 / GGT1

y

H

AT1G57720

Transformed
read counts
12.5
10
AT3G47680
7.5

1 March 2024

adjacent heatmap (F), where asterisks indicate significant DEGs. (G and
H) STRING network of protein–protein interactions of genes enriched in the
union of DEGs from transition 1 and transition 2 for glutathione metabolism
(G) and exocytosis (H). (I and J) Expression levels of genes responsible for
enrichment of glutathione metabolism and exocytosis, respectively, with the
A. thaliana orthologs of each gene indicated for every row. Adjacent heatmaps
(K and L) indicate log2 fold changes, with significant DEGs indicated with an
asterisk. Dark gray indicates DEGs with high log2 fold change due to one or
both samples having zero reads.
5 of 10

RES EARCH | R E S E A R C H A R T I C L E

B
a

20

PTGR ( m/min)

15
10

2

Est- 4X (TBG)

Hybrid-F1

Neo - 4X C1

T ACA8
T AGC1.5

Calcium gradient inclination

e

40

b

a

8
6
4
2
0
-2

y

-4
-6
-8

D
D

T
D

D
T

T ACA8
T AGC1.5

D T D T ACA8
D D T T AGC1.5

y g

Est- 4X (TBG)

-10

Hybrid -F1

2X

0

Neo -4X C1

10

Est- 4X (TBG)

Bursting rate (%)

c

30

D
D

10

70

f

Hybrid -F1

0

c

20 ab

D
T

1

90

d

T
D

1.5

D

50

bc

0.5

100

80

c
b

b

2X

AGC1.5 (TTTT)

ACA8 (TTTT)

c

2.5

2X
Neo -4X C1

b

0

60

a

c

3

5

C

a

3.5

Expectation
(Mendelian)

Frequency (%)

25

a

4

g

statistically significant (P > 0.99 for neo-4XC1
versus AGC1.5-D ACA8-D; P = 0.47 for est-4X
versus AGC1.5-T ACA8-T; Kruskal-Wallis with
Dunn’s multiple comparisons test). For bursting
rate, one comparison is statistically significant
(P = 0.06 for neo-4XC1 versus AGC1.5-D ACA8D; P = 0.04 for est-4X versus AGC1.5-T ACA8-T;
c2 test with Holm-Bonferroni correction).

1 March 2024

Diploid pollen tubes produced by AGC1.5-D
ACA8-D plants frequently lacked tip-focused
[Ca2+] gradients when grown in vitro (mean
m = 0.78; 14.0% with m < 0) and did not differ
significantly from neo-4XC1 (P > 0.99, c2 test
adjusted with Holm-Bonferroni correction;
Fig. 4D). In contrast, the diploid pollen tubes
produced by AGC1.5-T ACA8-T (mean m = −1.15;
6 of 10

,

Fig. 4. Genotypes of ACA8 and AGC1.5 strongly associate with ploidy-relevant differences in pollen
tube growth and physiology. (A) Frequency of F2 individuals homozygous for T alleles from a cross of TTDD
F1 plants compared with expected Mendelian frequency for an autotetraploid. Different letters indicate
significant differences (c2 test, P < 0.0001). (B) PTGR of control and F2 individuals homozygous for different
combinations of diploid (D) and tetraploid (T) alleles of ACA8 and AGC1.5 (n ≥ 110 pollen tubes per
genotype). Each dot represents mean from a single biological replicate. Error bars show standard deviation.
Different letters indicate significant differences (P < 0.05, Kruskal-Wallis, Dunn’s multiple comparisons test).
(C) Pollen tube bursting rate (n ≥ 825 pollen tubes per genotype). Genotypes and cytotypes as in (B). Error
bars show standard deviation. Different letters indicate significant differences (P < 0.05, c2 test, HolmBonferroni correction). (D) Box plots showing directionality of intracellular [Ca2+] gradient (n ≥ 26 pollen
tubes per genotype). Horizontal bars represent mean per genotype. Biological replicates per genotype:
2X = 6; neo-4XC1 = 4; hybrid-F1 = 6; est-4X = 6; ACA8 D AGC1.5 D = 4; ACA8 T AGC1.5 D = 5; ACA8 D AGC1.5
T = 6; and ACA8 T AGC1.5 T = 6. For each biological replicate, there were two technical replicates.

y

Westermann et al., Science 383, eadh0755 (2024)

A

p

tetraploid F2 populations segregating D and T
alleles of both genes. We saw a significant overrepresentation in the F2 of individuals homozygous for T alleles for both AGC1.5 and ACA8
(especially AGC1.5) relative to autotetraploid
Mendelian expectations (Fig. 4A), suggesting
that for each locus, among possible diploid
pollen grain genotypes produced by tetraploid
hybrid-F1 plants (DD, DT, TT), TT pollen tubes
had a substantial fertilization advantage over
DD or DT pollen tubes. The T-biased segregation shows that it is the pollen (gametophytic)
genotype at these loci, rather than the parental
(sporophytic) genotype, that is associated with
pollen tube performance and fertilization success in the tetraploids. This is consistent with
work that has shown that growing pollen
tubes contain almost exclusively gametophytegenerated mRNAs (44).
Pollen from F2 plants homozygous for T
versus D alleles of AGC1.5 and ACA8 did not
differ significantly in germination rate (fig. S10A),
but when pollen tubes were grown in vitro,
PTGR and bursting rate differed significantly
among genotypes (Fig. 4, B and C). Notably,
the pollen tubes from plants homozygous for
D alleles at both AGC1.5 and ACA8 (diploid
pollen tubes with genotype DD at both loci,
like neo-4X plants) had similarly slow PTGR
(3.09 mm/min) and high bursting rate (38%)
when grown in vitro as those from neo-4XC1
plants (0.89 mm/min PTGR: P > 0.99 KruskalWallis with Dunn’s multiple comparisons test;
37.5% bursting: P = 0.064 c2 test with HolmBonferroni correction). In contrast, the diploid
pollen tubes from tetraploids homozygous for
T alleles at both loci (pollen tubes TT at both
loci) grew nearly as fast as (2.54 mm/min) and
burst nearly as rarely as (14.5%) those of est-4X
(3.03 mm/min PTGR, 9.4% bursting; PTGR:
P = 0.47 Kruskal-Wallis test with Dunn’s multiple comparisons test; bursting: P = 0.042, c2
test with Holm-Bonferroni correction; Fig. 4, B
and C). Pollen tubes from homozygotes for T
alleles at only one of the two genes show
intermediate PTGR and bursting rate, suggesting that these loci have additive effects on both
traits (Fig. 4, B and C).
Homozygotes for T alleles at either AGC1.5
or ACA8 individually (homozygous D at the
other locus in each case) have nearly equal
PTGR (1.65 and 1.68 mm/min, respectively), in
both cases significantly higher than that of
D homozygotes (P = 0.033 and P < 0.0001
Kruskal-Wallis with Dunn’s multiple comparisons test, respectively). Bursting is less frequent
in pollen tubes homozygous for T alleles of
AGC1.5 than ACA8 (18.9% for AGC1.5-T ACA8-D
versus 29.5% for AGC1.5-D ACA8-T, P < 0.0001,
c2 test with Holm-Bonferroni correction; Fig. 4,
B and C). Small discrepancies remain in phenotypes of pollen tubes from AGC1.5-T ACA8-T
versus est-4X, or AGC1.5-D ACA8-D versus
neo-4XC1. For PTGR, these differences are not

RES EARCH | R E S E A R C H A R T I C L E

Generation of neo-tetraploids

To induce polyploidy, we treated apical meristems of 2X (SNO) seedlings grown in soil
7 of 10

,

A. arenosa seedlings of the accessions SNO
(2X), TBG (est-4X), and SBG (est-4X) [see (42)
for full information on each accession] were
grown on 50% soil and 50% sand under longday conditions [16 hours light, cool white, with
photosynthetic photon flux density (PPFD) ~
60 mmol m−2 s−1, 21°C; and 8 hours darkness,
12°C, 60% humidity each] for 6 weeks. Plants
with fully developed rosettes were vernalized
under short-day conditions (8 hours light, 6°C,
cool white, PPFD ~ 50 mmol m −2 s −1 ; and
16 hours darkness, 6°C, and 60% humidity) for
6 weeks, after which plants were returned to
long-day conditions (as above) to induce flowering within 2 to 3 weeks. All pollen-related
assays were performed on pollen from plants
within the first 2 to 4 weeks of their flowering
period from recently opened flowers to avoid
any defects that may be linked to plant or floral
senescence. A. thaliana plants of the Col-0
accession were grown in the same conditions
as A. arenosa, but without vernalization.

y

1 March 2024

Materials and methods
Plant cultivation

y g

Westermann et al., Science 383, eadh0755 (2024)

We show here that pollen tube tip growth is a
fertility-compromising challenge for neo-tetraploid
A. arenosa and A. thaliana. Neo-polyploid pollen
tubes show extensive physiological abnormalities, growth aberrations, and gene expression
perturbations. Pollen tube growth phenotypes
reported in other species [e.g., (24–26)] suggest that these phenomena are broadly relevant. Moreover, tip growth–related genes have
appeared as outliers in genome scans for selection in other polyploid plants [see (3) for a
review], suggesting that similar adaptations,
although not necessarily the same genes, may
be targeted in independent polyploidy events.
Our work also provides insights into what
might be the genetic basis of an evolved solution to the tip-growth challenge: In tetraploids,
pollen tubes homozygous for tetraploid alleles
of AGC1.5 and ACA8 that were under selection in the tetraploid lineage of A. arenosa outcompete and show more-normal pollen tube
growth and morphology than pollen tubes
carrying diploid alleles. This result supports

y

We generated transcriptome data for growing
pollen tubes from plants homozygous for D or
T alleles of ACA8 and AGC1.5 (fig. S11). When
comparing transcriptomes of pollen tubes
homozygous for T alleles for one or the other
gene (D homozygote at the other) with neo-4XC1
(D homozygous for both) and est-4X (T homozygous for both), we observed expression differ-

Conclusions

the hypothesis that the tetraploid alleles of
these genes likely contribute to polyploid pollen tube stabilization (neither gene is linked
to other ploidy-differentiated loci or other pollen tube–related genes). AGC1.5 and ACA8 genotypes strongly correlate not only with the same
morphological features we see differentiating
pollen tubes of neo-4X and est-4X plants but
also with gene expression differences. The overrepresentation of T alleles in the F2 population also shows that it is the gametophyte
genotype that matters for their performance,
consistent with prior work showing that pollen
tube growth is driven by gametophytic gene
expression (44). As we cannot formally rule
out that linked variation could contribute to
this trait by chance, we emphasize that functional studies will need to be done to confirm
their mechanistic role. This was not feasible
here, as transformation is currently prohibitive
in A. arenosa, but in the future, careful followup in a heterologous system should be done.
Although further work is needed to confirm
their role and understand how these genes
might control these effects, linking them to
polyploid fertility makes them useful candidates for engineering solutions to rescue fertility of artificially generated neo-polyploids
for plant breeding. Finally, an important aspect
of this work is that it demonstrates that the
“reverse adaptation genomics” approach we
used, in which the identification of genes under
selection in genome scans is used to generate
new hypotheses about undiscovered adaptions
(10), can work as a discovery tool.

g

Transcriptomes of AGC1.5 and ACA8
pollen tubes

ences for 13 of the 27 DEGs described in Fig. 3E.
In most cases, expression differences between
neo-4XC1 and AGC1.5 or ACA8 T homozygotes
mirrored the trend when comparing neo-4XC1
with est-4X (fig. S12), suggesting that T alleles
of AGC1.5 and ACA8 contribute at least partially (directly or indirectly) to the expression
levels characteristic of est-4X.
Several patterns are evident when comparing
expression levels among all genotypes tested:
For AGP2 and MIRO2, the diploid pollen tubes
of est-4X and ACA8 or AGC1.5 single T homozygotes show a return to expression levels characteristic of haploid pollen tubes of diploids.
For CDI, AT5G35930, and TRO, est-4X and
ACA8 or AGC1.5 single T homozygotes also
“correct” divergent neo-4XC1 expression but
overshoot the level seen in haploid pollen tubes
of 2X. For AT3G20790, RALFL11, and CIPK15,
est-4X and ACA8 or AGC1.5 single T homozygotes exacerbate a change that neo-4XC1 already
exhibited relative to 2X. For GH9B14, GAPCP-2,
and VLN4, a different expression level relative to
that of the haploid tubes from 2X occurs in
the diploid pollen tubes of est-4X (but not
neo-4XC1), and ACA8 or AGC1.5 T homozygotes have either the same as est-4X or, more
commonly, intermediate levels (fig. S12). Thus,
while some genes that are misexpressed in
diploid pollen tubes from neo-4XC1 plants return
in diploid pollen tubes of est-4X or T homozygotes for either AGC1.5 or ACA8 to expression
levels observed in haploid pollen tubes of 2X,
many do not. This finding suggests that evolution in the polyploid lineage has tinkered with
the novel situation created by genome duplication to create a new situation and did not
simply return gametophyte gene expression
to the haploid state.

p

78.3% with m < 0) were more like haploid pollen
tubes from 2X plants (m = −1.20; 82.8% with
m < 0) or diploid pollen tubes from est-4X
(mean m = −1.72; 77.1% with m < 0; frequencies: P > 0.99 in both cases, c2 test with HolmBonferroni correction).
Subapical actin filaments were evident in a
lower proportion of analyzed tubes from plants
homozygous for D alleles at both loci (68%,
AGC1.5-D ACA8-D) than those from 2X (94%,
P = 0.015), est-4X (89%, not significant, P =
0.067), or F2 plants homozygous for T alleles at
one or both loci (87%, P = 0.021; all three
c2 test with Holm-Bonferroni correction;
fig. S10B, table S4, and data S2). For the presence of the apical ring, results were more ambiguous: Pollen tubes from AGC1.5-D ACA8-D
plants differed significantly in AR presence
only from those of ACA8-D AGC1.5-T and ACA8-T
AGC1.5-D plants (P = 0.14 and P > 0.99, respectively, c2 test with Holm-Bonferroni correction). The maximum length of subapical
F-actin was low in pollen tubes from AGC1.5-D
ACA8-D or neo-4XC1, whereas the remaining
genotypes all had higher maximum F-actin
length (fig. S10C). For pectin acidification, the
frequency of single, tip-localized signals was
significantly lower in pollen tubes from AGC1.5-D
ACA8-D (39%) versus AGC1.5-T ACA8-T (65%,
P = 0.01, c2 test with Holm-Bonferroni correction; fig. S10D). Although the frequency of
normal pectin signals was higher for AGC1.5-D
ACA8-D (39%) than for neo-4XC1 (27%), the
difference was not statistically significant (P =
0.56, c2 test with Holm-Bonferroni correction).
AGC1.5-T ACA8-T (65%) was significantly lower
than est-4X (90%, P = 0.01, c2 test with HolmBonferroni correction; fig. S10D), suggesting
that at least one additional gene is involved.
The remaining differences in each trait value
between AGC1.5-D ACA8-D and neo-4XC1, or
AGC1.5-T ACA8-T and est-4X, suggest that one
or more additional genes, albeit with smaller
effects, are also involved in the stabilization of
pollen tube growth in est-4X. Nevertheless, it is
notable that AGC1.5-D ACA8-D versus AGC1.5-T
ACA8-T have nearly as wide a difference in in
vitro pollen tube performance (PTGR, bursting,
etc.) as we observe between neo-4X and est-4X,
suggesting that the alleles that came under selection in the tetraploids at these loci are likely
major players in the evolution of improved
pollen tube growth in polyploid A. arenosa.

RES EARCH | R E S E A R C H A R T I C L E

To assess pollen viability, anthers collected
from freshly opened flowers were kept for
30 min in a growth chamber (see above) at
100% humidity. Pollen grains were co-stained
in 0.01% FDA and 0.02% PI in Millipore water.
PI associates with cell wall pectins of living
(i.e., intact) cells, remaining external to the cell,
but it can enter through disordered membranes
upon cell death, leading to red cytoplasmic fluorescence. Thus, cells with cytoplasmic PI signal
were classified as dead cells. FDA is esterasedependent and causes green fluorescence of
living cells, and it was used as a counterstain
to confirm viability of PI-negative cells. Viability proportions were further verified with an
alternative staining method using commercial Alexander’s solution (Morphisto), whose

1 March 2024

For comparisons of numerical variables between two groups, the t test was used (in
Microsoft Excel), given that samples had a
normal distribution (checked with KolmogorovSmirnov test in Prism GraphPad). For comparisons between multiple groups, if the samples
did not have normal distribution (checked with
Kolmogorov-Smirnov test in Prism GraphPad),
the Kruskal-Wallis test was used followed by
Dunn’s test to adjust the P value for multiple
comparisons (in Prism GraphPad); whereas
Welch’s analysis of variance (ANOVA; in Prism
GraphPad) was used in the only case with
normal distribution, because the groups were
not homoscedastic (verified with F-test in Prism
GraphPad). For comparisons of categorical variables, the Chi-square test was used (in Microsoft
Excel), because the calculated expected values
exceeded 5 in every case. A summary of statistical tests not reported in the figures is included
in table S4, and all analyses are summarized in
the corresponding tabs of data S2.
Visualization and analysis of pollen tube
subcellular features

,

Westermann et al., Science 383, eadh0755 (2024)

Pollen viability analysis

Statistical analyses of pollen and pollen
tube features

y

To study pollen tube growth in vitro, freshly
opened flowers were collected in the morning
hours and kept under light at 100% humidity
for 30 min to allow pollen grains to hydrate.
Hydrated pollen was brushed onto a microscope
slide covered with Arabidopsis pollen germination medium [5 mM KCl, 1 mM MgSO4, 0.01%
H3BO3 (w/v), 5 mM CaCl2, and 10% sucrose
(w/v), dissolved in deionized (Millipore) water,
adjusted to pH 7.5 and solidified with 1% agarose, as previously described (46)]. Pollen germination was triggered by incubation at 28°C
for 45 min, and pollen was subsequently grown
at 21°C for 75 min in a growth chamber. Pollen
samples were then imaged with a Leica Thunder
Imager 3D Tissue microscope using differential
interference contrast (DIC). Germination rates
were calculated as the fraction of germinated

Receptive pistils of both evolved tetraploid
(TBG) and diploid (SNO) were crossed with
pollen from diploids (SNO), neo-tetraploids
(neo-4XC0), and established tetraploids (SBG).
The purpose of using SBG pollen was to avoid
erroneous results arising from the fact that
A. arenosa is self-incompatible and partial or
complete cross-incompatibilities have frequently
been observed for TBG intrapopulation crosses
(because TBG is quite inbred). Mature siliques
were harvested 4 weeks after pollination and
fixed in 25% acetic acid in ethanol. Cleared
pistils were then imaged and dissected to
enable reliable seed counts.

y g

In vitro pollen germination and growth assay

To visualize pollen tubes in vivo, receptive pistils
of mature flowers were pollinated with pollen
from the donor plant of choice. After 24 hours,
pistils were harvested and fixed in 25% glacial
acetic acid in ethanol (v/v) overnight at 21°C.
Then, pollinated pistils were subjected to an
ethanol series (75, 50, and 25%; incubation
under gentle shaking for 5 min each). Afterward,
siliques were washed twice for 5 min in sodium phosphate buffer (100 ml: 9.3 ml of 1 M
Na2HPO4, 6.8 ml of 1 M NaH2PO4, dissolved in
double-distilled water; pH 8.0). Ovule clearing
was performed by incubating siliques in 5%
chloral hydrate (in 24 ml: 32 g chloral hydrate,
16 ml double-distilled water, and 8 ml glycerol)
at 65°C for 5 to 10 min, and then washing
three times in sodium phosphate buffer. Pistils
were then incubated in 5 M NaOH at 65°C for
5 min, washed twice in sodium phosphate buffer,
and then immediately mounted in staining solution [0.1% aniline blue in sodium phosphate
buffer (w/v)] and imaged through epifluorescence microscopy (Leica Thunder Imager
3D Tissue).

Seed set assay

y

Differentiation of ACA8 and AGC1.5 between
A. arenosa diploid and tetraploid accessions
was verified with recent whole-genome resequencing data from nearly 300 individuals (42).
On the basis of ploidy-differentiated singlenucleotide polymorphisms (SNPs), two DNA
markers were designed using the dCAPS method
(45) (table S2). Hybrid-F1 plants were genotyped
for ACA8 and AGC1.5 to verify heterozygosity
at both loci and tetraploidy was confirmed by
flow cytometry (fig. S13). F1 individuals from
independent families were crossed to generate
F2 progeny in which both allelic variants segregate. Three hundred thirty F2 individuals were
genotyped to identify homozygous diploid and
tetraploid individuals and to determine allelic
segregation ratios.

Aniline blue staining of fixed pollen tubes in vivo

acid fuchsin and malachite components stain
the cytoplasm and cell wall, respectively. Acid
fuchsin–negative and malachite-positive pollen
grains were counted as dead cells.

g

Identification of allelic variants and generation
of homozygous plant lineages

grains divided by the total number of grains, and
bursting rates were calculated as the fraction
of grains and tubes that burst during the 75-min
experiment, divided by the sum of grains or
tubes examined. PTGR was determined by tracking tube growth for 4 min and dividing the
distance grown by time. Note that, as the length
values are limited by microscope resolution, they
are recorded in discrete pixel count values, resulting in some values being identical to a high
number of decimal points [data S1 (47) and
data S2] because discrete pixel values are divided by the same time value. Maximal width
measurements of the tube apex and shank regions were performed on 40x close-up images.
Maximal apex width was measured as the
widest point from 0 to 10 mm from the tip, and
shank width was measured as the widest point
in the region 10 to 30 mm from the apex.

p

with a solution of 0.05% colchicine and 0.05%
Silwet- L77 dissolved in water, at 15 to 21 days
after sowing (when the second pair of true
leaves started to emerge). Seedlings were
allowed to recover under long-day conditions
and were then transferred to single pots 7 days
after treatment. After induction of flowering
(see above,) ploidy was measured using flow
cytometry of petal tissue to avoid effects of
endoreduplication, which is common in leaves
(fig. S13). Petals from three different flowers
of each inflorescence were used. Only those
stems that consistently yielded tetraploid flow
cytometry profiles (fig. S13) were considered
neo-tetraploid (neo-4XC0) and used for experiments. Neo-tetraploid individuals (neo-4XC0)
were crossed (i) to each other to generate nextgeneration neo-tetraploids, which were also
confirmed by flow cytometry (hereafter “neo4XC1”) and (ii) with est-4X TBG individuals to
generate hybrid lineages (hereafter “hybrid-F1”)
in which all genes with distinct diploid and
tetraploid allelic variants should be heterozygous (DDTT).

For Ca2+ live-imaging, pollen was germinated
as described above and then stained by means
of a previously described staining protocol
using the intracellular Ca2+-dye Fluo-4 acetoxymethyl ester (Fluo-4/AM; ThermoFisher) (48).
Fluorescence intensity was captured from actively growing pollen tubes in the first 30 mm
from the tube tip using confocal microscopy
[excitation(ex)/emission(em) at 488/516 nm;
Zeiss LSM780]. Images were smoothed using
mean normalization per neighboring pixel to
reduce noise. A fluorescence profile was plotted
along the longitudinal axis of each tube using
ImageJ. Linear regression was performed, and
the slope m of the corresponding function was
determined as a proxy for presence (m < 0) or
8 of 10

RES EARCH | R E S E A R C H A R T I C L E

DEG-calling and enrichment analyses

Gene models of ACA8 (AL8G33030; scaffold_8:17902227..17911190; reverse strand) and
AGC1.5 (AL3G24370; scaffold_3:5074347..5077333;
reverse strand) were taken from the A. lyrata
reference genome v2.1. Genomic and local

1 March 2024

9 of 10

,

Protein structure predictions for diploid and
tetraploid variants of ACA8 and AGC1.5

1. D. H. Touchell, I. E. Palmer, T. G. Ranney, In vitro ploidy
manipulation for crop improvement. Front. Plant Sci. 11, 722
(2020). doi: 10.3389/fpls.2020.00722; pmid: 32582252
2. L. Comai, The advantages and disadvantages of being
polyploid. Nat. Rev. Genet. 6, 836–846 (2005). doi: 10.1038/
nrg1711; pmid: 16304599
3. K. Bomblies, When everything changes at once: Finding a new
normal after genome duplication. Proc. Biol. Sci. 287, 20202154
(2020). doi: 10.1098/rspb.2020.2154; pmid: 33203329
4. M. A. Lee, V. Howard-Andrews, M. Chester, Resistance of
multiple diploid and tetraploid perennial ryegrass (Lolium
perenne L.) varieties to three projected drought scenarios for
the UK in 2080. Agronomy (Basel) 9, 159 (2019). doi: 10.3390/
agronomy9030159
5. L. Grandont, E. Jenczewski, A. Lloyd, Meiosis and its deviations
in polyploid plants. Cytogenet. Genome Res. 140, 171–184
(2013). doi: 10.1159/000351730; pmid: 23817089
6. K. Bomblies, Learning to tango with four (or more): The
molecular basis of adaptation to polyploid meiosis. Plant
Reprod. 36, 107–124 (2023). doi: 10.1007/s00497-02200448-1; pmid: 36149479
7. C. Morgan, E. Knight, K. Bomblies, The meiotic cohesin subunit
REC8 contributes to multigenic adaptive evolution of
autopolyploid meiosis in Arabidopsis arenosa. PLOS Genet. 18,
e1010304 (2022). doi: 10.1371/journal.pgen.1010304;
pmid: 35830475
8. C. Morgan, H. Zhang, C. E. Henry, F. C. H. Franklin, K. Bomblies,
Derived alleles of two axis proteins affect meiotic traits in
autotetraploid Arabidopsis arenosa. Proc. Natl. Acad. Sci. U.S.A. 117,
8980–8988 (2020). doi: 10.1073/pnas.1919459117;
pmid: 32273390
9. L. Yant et al., Meiotic adaptation to genome duplication in
Arabidopsis arenosa. Curr. Biol. 23, 2151–2156 (2013).
doi: 10.1016/j.cub.2013.08.059; pmid: 24139735
10. G. L. StebbinsJr.., Types of polyploids; their classification and
significance. Adv. Genet. 1, 403–429 (1947). doi: 10.1016/
S0065-2660(08)60490-3; pmid: 20259289
11. A. Gonzalo, P. Parra-Nunez, A. L. Bachmann,
E. Sanchez-Moran, K. Bomblies, Partial cytological
diploidization of neoautotetraploid meiosis by induced crossover rate reduction. Proc. Natl. Acad. Sci. U.S.A. 120,
e2305002120 (2023). doi: 10.1073/pnas.2305002120;
pmid: 37549263
12. K. Bomblies, C. L. Peichel, Genetics of adaptation. Proc. Natl.
Acad. Sci. U.S.A. 119, e2122152119 (2022). doi: 10.1073/
pnas.2122152119; pmid: 35858399
13. J. D. Hollister et al., Genetic adaptation associated with
genome-doubling in autotetraploid Arabidopsis arenosa. PLOS
Genet. 8, e1003093 (2012). doi: 10.1371/journal.
pgen.1003093; pmid: 23284289

y

For each sample, transcript counts and metadata were imported into the R software using
the packages “tximport” and “tximportData”
for creating a DESeq object of 13 samples (for
processing with the “DESeq2” package). A
total of 15,361 genes were retained after filtering genes with low read counts, and these
were subjected to DESeq analyses, modeled by
the sample genotype as a single factor, and
using default parameters (nbinomWald test).
DEGs were subsequently identified by generating various contrasts (sample groups) and
identifying genes with an adjusted P ≤ 0.01
and |log2 fold change| > 1. For every A. lyrata
gene identified as a DEG, the corresponding
orthologs in A. thaliana (50) were provided as
input to STRING (www.string-db.org). Functional enrichment results from STRING were
plotted and visualized using R. Protein–protein
interactions of certain enriched groups were
exported from STRING and visualized using
Cytoscape.

REFERENCES AND NOTES

y g

Westermann et al., Science 383, eadh0755 (2024)

RNA sequencing (RNA-seq) libraries were prepared by Novogene UK Ltd (Directional mRNA
enrichment libraries) and subjected to NovaSeq
paired-end sequencing (2 × 150 base pairs) at
an average coverage of 32 million total reads
per library. Fastq files of samples were aligned
using bowtie2 to the Arabidopsis lyrata reference genome (50), prepared using the
rsem-prepare-reference function of the RSEM
software. Aligned bam files were sorted and indexed using samtools V1.9. Gene transcript
counts for each sample were estimated using
rsem-calculate-expression.

y

We collected 20 to 50 flowers from each plant
in 50-ml falcon tubes. Twenty milliliters of
20% pollen germination medium (5 mM KCl,
1 mM MgSO4, 0.01% boric acid, 5 mM CaCl2,
20% sucrose, pH 7.5) was added to each tube,
followed by vortexing at high speed for up to
4 min per sample. The tubes were centrifuged
for 10 min at 4°C at 3000g. Floating flower
tissue debris was removed using forceps, and
15 ml of the supernatant was discarded, leaving
5 ml of medium and a pellet of pollen in each
tube. Centrifugation was repeated for 2 min at
3000g at 4°C. Debris and supernatant were
removed again until 1 ml of medium was left
in the tube. The pollen pellet was resuspended
and added to petri plates containing 5 ml per
well of 10% pollen germination medium (5 mM
KCl, 1 mM MgSO4, 0.01% boric acid, 5 mM
CaCl2, 10% sucrose, pH 7.5). The pollen suspension was split into two or more wells if the
concentration of pollen was high (or if >25
flowers were collected per plant). The plates
were incubated for 4 hours in a plant growth
chamber (humidity 60%, illumination 3000 lux,
20°C). At two time points (before the incubation
and then 2 hours into the incubation), the plates
were briefly checked for pollen tube growth
using a light microscope. After incubation, the
pollen tube suspension in each well was filtered
using 20-mm cell strainers, repeated once using
the flow-through. The cell strainers carrying

Library prep, sequencing, and read mapping

positions of ploidy-differentiated SNPs, their
allele frequency in diploids (AF2X) and tetraploids (AF4X), are based on SNP profiles from
(42). For AGC1.5 protein structures of A.
arenosa diploid and tetraploid allelic variants
were inferred on the basis of SWISS-MODEL
search (https://swissmodel.expasy.org/) using
amino acid sequences of A. lyrata AGC1.5
(AL3G24370) and the A. arenosa 2X and 4X
allelic variants as query. This yielded entry ‘4
gv1.1.A’ - RAC-alpha serine/threonine-protein
kinase as best fit when building our protein
model (QmeanDisCo score = 0.62 ± 0.05). Protein active sites for fig. S9, B and C, are based on
prosite scan (https://prosite.expasy.org) using
AL3G24370 modified to represent the A. arenosa
2X and est-4X alleles. ACA8 topological and
transmembrane domains were inferred from
UniProt entry Q9LF79 (https://www.uniprot.
org/uniprotkb/Q9LF79/entry) from A. thaliana
ACA8. The top five protein templates per query
sequence were chosen on the basis of global
model quality estimate (GMQE) score before
modeling (table S3).

g

RNA extraction from pollen tubes

the pollen tubes were flash cooled on dry ice or
aluminum foil boats placed on liquid nitrogen.
The frozen cell strainers were subsequently
placed on new falcon tubes, and lysis buffer +
B-ME mixture (from the Spectrum Plant Total
RNA extraction kit–Protocol A) was directly
added to the strainer. The flow-through was
used for next steps of RNA extraction using
the Sigma Spectrum Kit protocol. RNA from
each sample was checked for quality using a
Nanodrop spectrophotometer (Thermo Fisher).
A maximum of 2000 ng of RNA was subjected
to DNase I treatment (Thermo Scientific) at
37°C for 30 min, followed by cleanup using the
Zymo RNA Clean & Concentrator kit. Eluted
RNA concentration was measured using Qubit
RNA BR reagents.

p

absence (m > 0) of a tip-focused Ca2+ gradient.
For actin labeling, fixed pollen tubes were
stained with Phalloidin-AlexaFluor405, an
F-actin–specific probe, according to (49), followed by confocal microscopy (ex/em: 405/
450 nm; Zeiss LSM780). A fluorescence profile
was plotted longitudinally for the first 30 mm
from the cell tip and transversely at three
randomly chosen positions in the subapical
region (defined as the 10- to 30-mm region
measured from the cell tip). Presence of the subapical actin ring was defined as those cells that
showed a maximum fluorescence intensity of
≥1.5× the median in the tube apex (0 to 10 mm
from tip). Presence of subapical F-actin was
defined as those cells that showed a maximum
fluorescence intensity of ≥1.5× the median in
one or more of the transverse profiles. Maximum
F-actin length was measured manually using
ImageJ. For esterified pectin staining, pollen
tubes were incubated in 0.01% ruthenium red
in germination medium (15) directly before
(<2 min) microscopic analysis (Leica Thunder
Imager 3D Tissue; Camera DFC9000 GT color).
The number and location (tip-focused versus
subapical or basal) of active growth sites (i.e.,
stained sites per tube) were manually determined. Specimens were counted as “normal”
when displaying a single, tip-focused signal,
as opposed to multiple signals or absence from
the tip frequently observed in neo-tetraploids.

RES EARCH | R E S E A R C H A R T I C L E

AC KNOWLED GME NTS

We thank A. Widmer and C. Hughes for helpful comments on the
manuscript, M. Dukic for SNP profiles, Y. Wijnand Revaz for help
with genotyping, and I. Zurkirchen for plant cultivation. Funding:
This work was supported by core funds from ETH Zürich and a
grant from the Swiss National Science Foundation (310030192671; to K.B.), a grant from the Swiss National Science
Foundation (217182; to T.S.), and funding from the European
Union’s Horizon 2020 research and innovation program under
Marie Skłodowska-Curie (MSC) grant agreement no. 101029732
(to A.G.). Author contributions: K.B., J.W., and T.S. designed the
study and wrote the paper. J.W., T.S., A.G., and H.S.T. performed
the experiments. J.W., T.S., and A.G. analyzed the data. All authors
helped check and edit the paper. Competing interests: The
authors declare that they have no competing interests. Data and
materials availability: All images upon which these analyses were
based and all primary data are freely and openly available through
the ETH Research Collection (47). RNA-seq data generated for this
study are available through the European Nucleotide Archive
(https://www.ebi.ac.uk/ena/browser/home) under accession
number PRJEB65902. Seeds available on request: kirsten.
bomblies@biol.ethz.ch. License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse. This research was funded in whole or
in part by the Swiss National Science Foundation (Postdoctoral
Fellowship 217182 and 310030-192671), a cOAlition S organization.
The author will make the Author Accepted Manuscript (AAM)
version available under a CC BY public copyright license.

y g

SUPPLEMENTARY MATERIALS

y

science.org/doi/10.1126/science.adh0755
Figs. S1 to S13
Tables S1 to S4
MDAR Reproducibility Checklist
Movies S1 and S2
Data S2

10 of 10

,

Submitted 8 February 2023; resubmitted 18 October 2023
Accepted 24 January 2024
10.1126/science.adh0755

y

1 March 2024

46. L. C. Boavida, S. McCormick, Temperature as a determinant
factor for increased and reproducible in vitro pollen
germination in Arabidopsis thaliana. Plant J. 52, 570–582
(2007). doi: 10.1111/j.1365-313X.2007.03248.x; pmid: 17764500
47. K. Bomblies, T. Srikant, J. Westermann, Dataset S1 for
“A severe polyploidy-induced pollen defect rescued
by evolved alleles of two tip-growth genes,” ETH
Research Collection (2023); https://doi.org/10.3929/
ethz-b-000630044.
48. H. Qu, W. Xing, F. Wu, Y. Wang, Rapid and inexpensive method of
loading fluorescent dye into pollen tubes and root hairs. PLOS ONE
11, e0152320 (2016). doi: 10.1371/journal.pone.0152320;
pmid: 27055240
49. X. Qu, Q. Wang, H. Wang, S. Huang, Visualization of actin
organization and quantification in fixed Arabidopsis pollen
grains and tubes. Bio Protoc. 10, e3509 (2020). doi: 10.21769/
BioProtoc.3509; pmid: 33654717
50. V. Rawat et al., Improving the annotation of Arabidopsis lyrata
using RNA-seq data. PLOS ONE 10, e0137391 (2015).
doi: 10.1371/journal.pone.0137391; pmid: 26382944

g

Westermann et al., Science 383, eadh0755 (2024)

31. A. Lovy-Wheeler, K. L. Wilsen, T. I. Baskin, P. K. Hepler, Enhanced
fixation reveals the apical cortical fringe of actin filaments as a
consistent feature of the pollen tube. Planta 221, 95–104 (2005).
doi: 10.1007/s00425-004-1423-2; pmid: 15747143
32. P. K. Hepler, C. M. Rounds, L. J. Winship, Control of
cell wall extensibility during pollen tube growth.
Mol. Plant 6, 998–1017 (2013). doi: 10.1093/mp/sst103;
pmid: 23770837
33. K. Müller et al., Overexpression of a pectin methylesterase
inhibitor in Arabidopsis thaliana leads to altered growth
morphology of the stem and defective organ separation. Plant
Signal. Behav. 8, e26464 (2013). doi: 10.4161/psb.26464;
pmid: 24675171
34. M. Woriedh et al., External application of gametophyte-specific
ZmPMEI1 induces pollen tube burst in maize. Plant Reprod. 26,
255–266 (2013). doi: 10.1007/s00497-013-0221-z;
pmid: 23824238
35. B. Levitin, D. Richter, I. Markovich, M. Zik, Arabinogalactan
proteins 6 and 11 are required for stamen and pollen function
in Arabidopsis. Plant J. 56, 351–363 (2008). doi: 10.1111/
j.1365-313X.2008.03607.x; pmid: 18644001
36. Z. Ge et al., Arabidopsis pollen tube integrity and sperm
release are regulated by RALF-mediated signaling. Science
358, 1596–1600 (2017). doi: 10.1126/science.aao3642;
pmid: 29242234
37. M. A. Mecchia et al., RALF4/19 peptides interact with LRX
proteins to control pollen tube growth in Arabidopsis. Science
358, 1600–1603 (2017). doi: 10.1126/science.aao5467;
pmid: 29242232
38. C. G. Sørmo, T. Brembu, P. Winge, A. M. Bones, Arabidopsis
thaliana MIRO1 and MIRO2 GTPases are unequally redundant in
pollen tube growth and fusion of polar nuclei during female
gametogenesis. PLOS ONE 6, e18530 (2011). doi: 10.1371/
journal.pone.0018530; pmid: 21494602
39. H. M. Li, H. Chen, Z. N. Yang, J. M. Gong, Cdi gene is required
for pollen germination and tube growth in Arabidopsis. FEBS
Lett. 586, 1027–1031 (2012). doi: 10.1016/j.febslet.2012.02.046;
pmid: 22569258
40. E. García-Quirós, J. D. Alché, B. Karpinska, C. H. Foyer,
Glutathione redox state plays a key role in flower development
and pollen vigour. J. Exp. Bot. 71, 730–741 (2020).
doi: 10.1093/jxb/erz376; pmid: 31557297
41. D. Bloch et al., Exocyst SEC3 and phosphoinositides define
sites of exocytosis in pollen tube initiation and growth. Plant
Physiol. 172, 980–1002 (2016). doi: 10.1104/pp.16.00690;
pmid: 27516531
42. R. A. Cole, L. Synek, V. Zarsky, J. E. Fowler, SEC8, a subunit of
the putative Arabidopsis exocyst complex, facilitates pollen
germination and competitive pollen tube growth. Plant Physiol.
138, 2005–2018 (2005). doi: 10.1104/pp.105.062273;
pmid: 16040664
43. P. Monnahan et al., Pervasive population genomic
consequences of genome duplication in Arabidopsis arenosa.
Nat. Ecol. Evol. 3, 457–468 (2019). doi: 10.1038/s41559-0190807-4; pmid: 30804518
44. B. Nelms, V. Walbot, Gametophyte genome activation occurs
at pollen mitosis I in maize. Science 375, 424–429 (2022).
doi: 10.1126/science.abl7392; pmid: 35084965
45. M. M. Neff, J. D. Neff, J. Chory, A. E. Pepper, dCAPS, a simple
technique for the genetic analysis of single nucleotide
polymorphisms: Experimental applications in Arabidopsis
thaliana genetics. Plant J. 14, 387–392 (1998). doi: 10.1046/
j.1365-313X.1998.00124.x; pmid: 9628033

p

14. B. Arnold, S. T. Kim, K. Bomblies, Single geographic origin
of a widespread autotetraploid Arabidopsis arenosa lineage
followed by interploidy admixture. Mol. Biol. Evol. 32,
1382–1395 (2015). doi: 10.1093/molbev/msv089
pmid: 25862142
15. E. Li et al., AGC1.5 kinase phosphorylates RopGEFs to control
pollen tube growth. Mol. Plant 11, 1198–1209 (2018).
doi: 10.1016/j.molp.2018.07.004; pmid: 30055264
16. Y. Zhang, J. He, S. McCormick, Two Arabidopsis AGC kinases
are critical for the polarized growth of pollen tubes. Plant J. 58,
474–484 (2009). doi: 10.1111/j.1365-313X.2009.03792.x;
pmid: 19144004
17. V. Demidchik, S. Shabala, S. Isayenkov, T. A. Cuin, I. Pottosin,
Calcium transport across plant membranes: Mechanisms and
functions. New Phytol. 220, 49–69 (2018). doi: 10.1111/
nph.15266; pmid: 29916203
18. Y. Li, J. Guo, Z. Yang, D. L. Yang, Plasma membrane-localized
calcium pumps and copines coordinately regulate pollen
germination and fertility in Arabidopsis. Int. J. Mol. Sci. 19, 1774
(2018). doi: 10.3390/ijms19061774; pmid: 29914054
19. M. Schiøtt et al., A plant plasma membrane Ca2+ pump is
required for normal pollen tube growth and fertilization. Proc.
Natl. Acad. Sci. U.S.A. 101, 9502–9507 (2004). doi: 10.1073/
pnas.0401542101; pmid: 15197266
20. J. B. Reese, J. H. Williams, How does genome size affect the
evolution of pollen tube growth rate, a haploid performance
trait? Am. J. Bot. 106, 1011–1020 (2019). doi: 10.1002/
ajb2.1326; pmid: 31294836
21. J. Doyle, J. Coate, Polyploidy, the nucleotype, and novelty: The
impact of genome doubling on the biology of the cell. Int. J.
Plant Sci. 180, 1–52 (2019). doi: 10.1086/700636
22. J. H. Williams, P. E. Oliveira, For things to stay the same, things
must change: Polyploidy and pollen tube growth rates. Ann.
Bot. 125, 925–935 (2020). doi: 10.1093/aob/mcaa007;
pmid: 31957784
23. S. J. Baldwin, B. C. Husband, Genome duplication and the
evolution of conspecific pollen precedence. Proc. Biol. Sci. 278,
2011–2017 (2011). doi: 10.1098/rspb.2010.2208;
pmid: 21123263
24. I. Susín, J. M. Álvarez, Fertility and pollen tube growth in
polyploid melons (Cucumis melo L.). Euphytica 93, 369–373
(1997). doi: 10.1023/A:1002904817971
25. Y. W. Rim, P. R. Beuselinck, Cytology of 2N pollen formation
and pollen morphology in diploid Lotus tenuis (Fabaceae).
Am. J. Bot. 83, 1057–1062 (1996). doi: 10.1002/
j.1537-2197.1996.tb12803.x
26. L. P. Lobanova, A. Y. Kolesova, Maize tetraploid pollen studies.
Maize Genet. Coop. News Lett. 84, 37–38 (2010).
27. P. Fayant et al., Finite element model of polar growth in pollen
tubes. Plant Cell 22, 2579–2593 (2010). doi: 10.1105/
tpc.110.075754; pmid: 20699395
28. H. Li, Y. Lin, R. M. Heath, M. X. Zhu, Z. Yang, Control of pollen
tube tip growth by a Rop GTPase–dependent pathway that
leads to tip-localized calcium influx. Plant Cell 11, 1731–1742
(1999). pmid: 10488239
29. P. K. Hepler, J. G. Kunkel, C. M. Rounds, L. J. Winship,
Calcium entry into pollen tubes. Trends Plant Sci. 17,
32–38 (2012). doi: 10.1016/j.tplants.2011.10.007;
pmid: 22104406
30. L. Steinhorst, J. Kudla, Calcium - a central regulator of pollen
germination and tube growth. Biochim. Biophys. Acta Mol. Cell
Res.1833, 1573–1581 (2013). doi: 10.1016/j.bbamcr.2012.10.009;
pmid: 23072967

RES EARCH

RESEARCH ARTICLE SUMMARY

◥

IMMUNOLOGY

The immunopathological landscape of human
pre-TCRa deficiency: From rare to common variants
Marie Materna et al.

INTRODUCTION: ab and gd T lymphocytes con-

mice, only a few TCRab cells are detected in
the lymph nodes (LNs) (5% of normal levels).
In these studies, the mice remained healthy
in pathogen-free conditions but were not challenged with pathogens. They did not develop
overt phenotypes, but, to our knowledge, no
data have been published for Ptcra−/− mice beyond the age of 2 months.

y

RESULTS: We studied 10 patients with rare
biallelic loss-of-function PTCRA variants. These
patients had small thymi, and their blood naive
ab T cell counts had been low since childhood,
but their memory ab T cell counts were normal.

g

RATIONALE: The pre-TCRa chain appears to be
essential for ab T cell development in mice,
but it is unknown whether this is also the case
in humans. Immunological and clinical studies
of humans with inherited pre-TCRa deficiency
would address this question.

p

stitute two of the three lineages of adaptive
immunity in jawed vertebrates. They are generated from progenitor stem cells that differentiate in the thymus. In mice, branching into
the ab and gd T cell lineages occurs during
early thymopoiesis, with rearrangements of
TRD [encoding d T cell receptor (TCRd)], TRG
(encoding TCRg), and TRB (encoding TCRb).
Productive TRD and TRG rearrangements
lead to maturation into gd T cells. Alternatively,
following productive TRB rearrangement, expression of the pre-TCRa chain (encoded by
PTCRA) leads to the surface expression of heterodimeric TCRb–pre-TCRa, which facilitates
the TCRb selection of thymocytes during TRAD
(TCRa) rearrangements, resulting in the development of ab T cells. In 4-week-old mice, preTCRa loss is associated with a >95% decrease
in ab T cell precursor counts. Although peripheral
T cells have not been extensively studied in these

y g

Their naive CD4−CD8− ab and gd T cell counts
were high. Moreover, six of these patients remained healthy at ages ranging from 2 to
65 years. The other four patients had infections, lymphoproliferation, or autoimmunity
beginning at ages ranging from 13 to 25 years.
This relatively mild clinical phenotype reflected
an age-dependent accumulation of normal counts
of diverse functional memory ab T cells. These
data raised questions about how ab T cells
develop in the absence of pre-TCRa. TRAD
rearrangements were biased in ab T cells from
pre-TCRa–deficient individuals. The TCRa repertoire suggested that these TCRa rearrangements occurred preferentially with a TCRd1
template. Similar to controls, most ab T cell
clones did not carry productive TRG rearrangements, suggesting that most of the patients’ ab
T cells were unlikely to have differentiated directly from gd+ thymocytes. Moreover, TCRd1
could not act as a surrogate for pre-TCRa in
pre-TCR formation with multiple TCRb. These
findings call for alternative hypotheses that
may account for ab T cell differentiation in the
absence of pre-TCRa and be consistent with
the associated rearrangement bias observed
at the TRAD locus. Finally, we also identified
two common PTCRA variants responsible for
partial pre-TCRa deficiency in homozygotes.
The hypomorphic p.Tyr76Cys PTCRA variant
was found to be homozygous in about 1 in
73,000 individuals from Africa. Moreover,
about 1 in 4000 individuals from the Middle
East and South Asia were homozygous for
the hypomorphic p.Asp51Ala variant. This
missense was located in the extracellular domain and affected an acidic residue, which
is important for the interaction between
pre-TCRa and TCRb. Homozygotes for the
p.Asp51Ala variant had high circulating naive
gd T cell counts and a significantly higher
incidence of autoimmunity when compared
with the general population.
CONCLUSION: Inherited complete pre-TCRa

Functional ab T cells and late-onset immunological conditions in humans with complete or partial
inherited pre-TCRa deficiency. Although complete pre-TCRa deficiency is very rare, partial pre-TCRa
deficiency is common in South Asia and the Middle East, affecting about 1 in 4000 individuals. DN, doublenegative. [Figure created with BioRender.com.]
Materna et al., Science 383, 966 (2024)

1 March 2024

All author names and affiliations are available in the full article online.
*Corresponding author: Vivien Béziat (vivien.beziat@inserm.fr)
Cite this article as M. Materna et al., Science 383, eadh4059
(2024). DOI: 10.1126/science.adh4059

READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.adh4059
1 of 1

,

▪

y

deficiency is rare in humans. It is less severe
than anticipated, as the patients have ab T cells
and can survive into adulthood, often without
clinical manifestations. Their TCRa repertoire
is biased, which suggests that noncanonical
thymic differentiation pathways can rescue
ab T cell development. Additionally, a partial
form of pre-TCRa deficiency was found to be
less rare than anticipated, affecting about
1 in 4000 individuals in South Asia and the
Middle East, where it is a monogenic etiology of
autoimmunity with incomplete penetrance.

RES EARCH

RESEARCH ARTICLE

◥

IMMUNOLOGY

The immunopathological landscape of human
pre-TCRa deficiency: From rare to common variants

1 of 18

,

These 10 patients came from seven unrelated
families and were of four different ethnicities (Fig. 1D, data S1, and supplementary
text). Six of 10 patients with predicted pre-TCRa

y

1 March 2024

Clinical features of patients with biallelic
pLOF PTCRA variants

y g

Materna et al., Science 383, eadh4059 (2024)

and, during a process known as b-selection, it
promotes a burst of proliferation and differentiation into CD4+CD8+ double-positive (DP)
thymocytes. The TRA loci on the DP thymocytes
then undergo successive waves of rearrangement (5–8), leading to the expression of TCRab
heterodimers on the cell surface, and downregulation of the pre-TCRa chain (8, 9). After
undergoing negative and positive selection,
TCRab+ thymocytes eventually differentiate
into CD4+ or CD8+ single-positive (SP) mature
T cells and migrate to the periphery (10, 11). In
4-week-old mice, pre-TCRa loss is associated
with a >95% decrease in DP thymocyte counts
(12). Although peripheral T cells have not been
extensively studied in these mice, only a few
TCRab cells are detected in lymph nodes (LNs)
(5% normal levels), with the cells displaying
normal TCR diversity (12, 13). In these studies,
the mice remained healthy in pathogen-free

y

a

b and gd T lymphocytes constitute
two of the three cellular lineages of
adaptive immunity in jawed vertebrates. In a process clarified in mice,
they are generated from progenitor
stem cells by differentiation in the thymus (1).
Double-negative (DN) thymocytes, which lack
both CD4 and CD8, are the most immature
cells. They differentiate into mature ab T cell
receptor (TCRab)– or TCRgd–expressing T cells.
Cells branch off into these two lineages during
early thymopoiesis, which occurs at the same
time as TRD, TRG, and TRB locus rearrangements (2–4). Productive TRD and TRG rearrangements then lead to TCRgd expression on the cell
surface, promoting maturation into gd T cells.
Alternatively, after productive TRB locus rearrangement, a TCRb chain may dimerize with
a pre-TCRa protein to generate a pre-TCR. This
heterodimer is expressed on the cell surface

PTCRA encodes two functional isoforms in
humans and mice (15). Isoform B is 106 amino
acids shorter than isoform A and lacks part of
the extracellular domain (Fig. 1, A and B). We
reanalyzed a public RNA sequencing (RNA-seq)
dataset corresponding to eight sorted thymocyte
subsets from healthy controls (Fig. 1C) (16) and
found that isoform A was the principal preTCRa isoform in all human thymocyte subsets
(supplementary text). Unless otherwise specified, we refer below to isoform A. We searched
for biallelic predicted loss-of-function (pLOF)
variants of the PTCRA isoform A, including
large deletions, frameshift insertions or deletions, premature stop codons, and variants
affecting essential splice sites or the start codon.
No biallelic pLOF variants meeting these criteria
have ever been reported in public databases
(17–19). In our in-house database containing
data for >25,000 patients, including four other
unrelated patients identified by newborn screening (P1, P2, P9, and P10), we identified 10 patients
from seven kindreds, all carrying biallelic pLOF
variants (Fig. 1D; fig. S1, A to C; and supplementary text). The seven pLOF variants in these
individuals were present in the homozygous
state in five kindreds and in the compound
heterozygous state in two kindreds. Five variants were private to the kindreds identified,
and two were reported in major public databases but only in the heterozygous state, with
a minor allele frequency (MAF) of <10−4 (17–19).
Two variants were predicted to affect a splice
site, two were small frameshift deletions, two
led to premature stop codons, and one was a
large deletion. The c.58G>C substitution was a
missense variant (p.Gly20Arg) but was considered to be pLOF because it was predicted to
impair splicing between exons 1 and 2. Apart
from the seven pLOF variants identified, only
15 biallelic coding variants—all missense and
not predicted to be LOF—were found in public
databases or in the HGID in-house database
(fig. S1A).

g

We describe humans with rare biallelic loss-of-function PTCRA variants impairing pre–a T cell receptor
(pre-TCRa) expression. Low circulating naive ab T cell counts at birth persisted over time, with normal
memory ab and high gd T cell counts. Their TCRa repertoire was biased, which suggests that
noncanonical thymic differentiation pathways can rescue ab T cell development. Only a minority of these
individuals were sick, with infection, lymphoproliferation, and/or autoimmunity. We also report that 1
in 4000 individuals from the Middle East and South Asia are homozygous for a common hypomorphic
PTCRA variant. They had normal circulating naive ab T cell counts but high gd T cell counts. Although
residual pre-TCRa expression drove the differentiation of more ab T cells, autoimmune conditions
were more frequent in these patients compared with the general population.

Identification of rare biallelic predicted
loss-of-function PTCRA variants in
seven kindreds

p

Marie Materna1,2, Ottavia M. Delmonte3†, Marita Bosticardo3†, Mana Momenilandi1,2†,
Peyton E. Conrey4†, Bénédicte Charmeteau-De Muylder5‡, Clotilde Bravetti6,7‡,
Rebecca Bellworthy8‡, Axel Cederholm9‡, Frederik Staels10‡, Christian A. Ganoza11‡,
Samuel Darko12‡, Samir Sayed4‡, Corentin Le Floc’h1,2‡, Masato Ogishi13‡, Darawan Rinchai13‡,
Andrea Guenoun14‡, Alexandre Bolze15‡, Taushif Khan14,16‡, Adrian Gervais1,2‡, Renate Krüger17,
Mirjam Völler17, Boaz Palterer3, Mahnaz Sadeghi-Shabestari18, Anne Langlois de Septenville6,
Chaim A. Schramm12, Sanjana Shah12, John J. Tello-Cajiao4,19, Francesca Pala3, Kayla Amini3,
Jose S. Campos4, Noemia Santana Lima12, Daniel Eriksson20, Romain Lévy1,2,21, Yoann Seeleuthner1,2,
Soma Jyonouchi4, Manar Ata14, Fatima Al Ali14, Caroline Deswarte1,2, Anaïs Pereira1,2,
Jérôme Mégret22, Tom Le Voyer1,2, Paul Bastard1,2,13,21, Laureline Berteloot23, Michaël Dussiot2,24,
Natasha Vladikine1,2, Paula P. Cardenas25, Emmanuelle Jouanguy1,2,13, Mashael Alqahtani26,
Amal Hasan27, Thangavel Alphonse Thanaraj28, Jérémie Rosain1,2, Fahd Al Qureshah13, Vito Sabato29,
Marie Alexandra Alyanakian30, Marianne Leruez-Ville31, Flore Rozenberg5,32, Elie Haddad33,
Jose R. Regueiro25, Maria L. Toribio34, Judith R. Kelsen35, Mansoor Salehi36,37, Shahram Nasiri38,
Mehdi Torabizadeh39, Hassan Rokni-Zadeh40, Majid Changi-Ashtiani41, Nasimeh Vatandoost37,42,
Hossein Moravej43, Seyed Mohammad Akrami44,45, Mohsen Mazloomrezaei45, Aurélie Cobat1,2,13,
Isabelle Meyts46,47, Etsushi Toyofuku48§, Madoka Nishimura49, Kunihiko Moriya50,
Tomoyuki Mizukami49, Kohsuke Imai50, Laurent Abel1,2,13, Bernard Malissen51,52, Fahd Al-Mulla28,
Fowzan Sami Alkuraya26,53, Nima Parvaneh54, Horst von Bernuth17,55,56,57, Christian Beetz11,
Frédéric Davi6,7, Daniel C. Douek12¶, Rémi Cheynier5¶, David Langlais8¶, Nils Landegren9,58¶,
Nico Marr14,59¶, Tomohiro Morio48#, Mohammad Shahrooei45,60#, Rik Schrijvers10#,
Sarah E. Henrickson4,61,62#, Hervé Luche52#, Luigi D. Notarangelo3#,
Jean-Laurent Casanova1,2,13,63,64#, Vivien Béziat1,2,13*#

conditions but were not challenged with pathogens. They did not develop overt phenotypes
but, to our knowledge, no data have been published for Ptcra−/− mice beyond the age of
2 months (12–14). The consequences of preTCRa deficiency in humans remain unknown.
We therefore searched for patients with biallelic
germline PTCRA variants likely to cause preTCRa deficiency.

RES EARCH | R E S E A R C H A R T I C L E

velopment or medullary thymic epithelial cells
(20–24).
Patient alleles cause mRNA decay or
premature translational termination

Patient variants are LOF, and two
variants from public databases
are severely hypomorphic

We assessed the ability of the pre-TCRa variants to stabilize TCRb and CD3 at the cell
surface in the TCRa-deficient JR3.11 Jurkat cell
line (25, 26). The transduction of these cells
with the WT isoform A of pre-TCRa restored
the expression of TCRb and CD3 at the cell
surface (Fig. 2, F to I, and fig. S3, C to F). As
expected, none of the cDNAs encoding variants
from the patients except the cDNA encoding
the p.Gly20Arg (c.58G>C) variant restored the

1

3

2 of 18

,

1 March 2024

y

Materna et al., Science 383, eadh4059 (2024)

y g

*Corresponding author. Email: vivien.beziat@inserm.fr
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
§Present address: Department of Rheumatology and Allergology, St. Marianna University School of Medicine, Kawasaki, Japan.
¶These authors contributed equally to this work.
#These authors contributed equally to this work.

y

Laboratory of Human Genetics of Infectious Diseases, Necker Branch, INSERM, Necker Hospital for Sick Children, Paris, France. 2Imagine Institute, University of Paris-Cité, Paris, France.
Laboratory of Clinical Immunology and Microbiology, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Bethesda, MD, USA. 4Division of Allergy Immunology,
Department of Pediatrics, Children’s Hospital of Philadelphia, Philadelphia, PA, USA. 5University of Paris, Institut Cochin, INSERM U1016, CNRS UMR8104, Paris, France. 6Department of Biological
Hematology, Hôpital Pitié-Salpêtrière, Assistance Publique–Hôpitaux de Paris (AP-HP) and Sorbonne Université, Paris, France. 7Sorbonne University, Paris Cancer Institute CURAMUS, INSERM
U1138, Paris, France. 8Deptartment of Human Genetics, Dahdaleh Institute of Genomic Medicine, McGill University, Montreal, QC, Canada. 9Science for Life Laboratory, Department of Medical
Biochemistry and Microbiology, Uppsala University, Uppsala, Sweden. 10Allergy and Clinical Immunology Research Group, Department of Microbiology, Immunology and Transplantation, KU
Leuven, Leuven, Belgium. 11Centogene GmbH, Rostock, Germany. 12Human Immunology Section, Vaccine Research Center, National Institute of Allergy and Infectious Diseases, National Institutes
of Health, Bethesda, MD, USA. 13St. Giles Laboratory of Human Genetics of Infectious Diseases, Rockefeller Branch, The Rockefeller University, New York, NY, USA. 14Department of Human
Immunology, Sidra Medicine, Doha, Qatar. 15Helix, San Mateo, CA, USA. 16The Jackson Laboratory, Farmington, CT, USA. 17Department of Pediatric Respiratory Medicine, Immunology and Critical
Care Medicine, Charité - Universitätsmedizin Berlin, corporate member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health (BIH), Berlin, Germany.
18
Immunology Research Center, TB and Lung Disease Research Center, Mardaniazar Children Hospital, Tabriz University of Medical Science, Tabriz, Iran. 19Department of Pathology, Children’s
Hospital of Philadelphia, Philadelphia, PA, USA. 20Department of Immunology, Genetics and Pathology, Uppsala University and University Hospital, Section of Clinical Genetics, Uppsala, Sweden.
21
Pediatric Immunology, Hematology and Rheumatology Unit, Necker Hospital for Sick Children, AP-HP, Paris, France. 22Cytometry Core Facility, SFR Necker, INSERM US24-CNRS UAR3633,
Paris, France. 23Department of Pediatric Radiology, University Hospital Necker-Enfants Malades, AP-HP, Paris, France. 24Laboratory of Molecular Mechanisms of Hematological Disorders and
Therapeutic Implications, INSERM UMR 1163, Paris, France. 25Department of Immunology, Ophthalmology and ENT, Complutense University School of Medicine and 12 de Octubre Health
Research Institute (imas12), Madrid, Spain. 26Department of Translational Genomics, Center for Genomic Medicine, King Faisal Specialist Hospital and Research Center, Riyadh, Saudi Arabia.
27
Department of Translational Research, Research Division, Dasman Diabetes Institute, Dasman, Kuwait City, Kuwait. 28Department of Genetics and Bioinformatics, Research Division, Dasman
Diabetes Institute, Dasman, Kuwait City, Kuwait. 29Department of Immunology, Allergology and Rheumatology, University of Antwerp, Antwerp University Hospital, Antwerp, Belgium.
30
Immunology Laboratory, Necker Hospital for Sick Children, Assistance Publique-Hôpitaux de Paris (AP-HP), Paris, France. 31Necker Hospital for Sick Children, AP-HP, Paris, France. 32Virology,
Cochin Hospital, AP-HP, APHP-CUP, Paris, France. 33Department of Pediatrics, Department of Microbiology, Immunology and Infectious Diseases, University of Montreal, CHU Sainte-Justine,
Montreal, QC, Canada. 34Immune System Development and Function Unit, Centro de Biología Molecular Severo Ochoa, Consejo Superior de Investigaciones Científicas (CSIC), Universidad
Autónoma de Madrid (UAM), Madrid, Spain. 35Division of Gastroenterology, Hepatology and Nutrition, Children’s Hospital of Philadelphia, Philadelphia, PA, USA. 36Cellular, Molecular and Genetics
Research Center, Isfahan University of Medical Sciences, Isfahan, Iran. 37Department of Genetics and Molecular Biology, Medical School, Isfahan University of Medical Sciences, Isfahan, Iran.
38
Department of Pediatric Neurology, Children’s Medical Center of Abuzar, Jundishapur University of Medical Sciences, Ahvaz, Iran. 39Golestan Hospital Clinical Research Development Unit,
Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran. 40Department of Medical Biotechnology, School of Medicine, Zanjan University of Medical Sciences (ZUMS), Zanjan, Iran. 41School
of Mathematics, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran. 42Pediatric Inherited Diseases Research Center, Research Institute for Primordial Prevention of NonCommunicable Disease, Isfahan University of Medical Sciences, Isfahan, Iran. 43Neonatal Research Center, Shiraz University of Medical Sciences, Shiraz, Iran. 44Medical Genetics Poursina St.,
Genetic Department, Medical Faculty, Tehran University of Medical Sciences, Tehran, Iran. 45Dr. Shahrooei Laboratory, Tehran, Iran. 46Laboratory for Inborn Errors of Immunity, Department of
Microbiology, Immunology and Transplantation, Department of Pediatrics, University Hospitals Leuven, KU Leuven, Leuven, Belgium. 47Department of Pediatrics, University Hospitals Leuven, KU
Leuven, Leuven, Belgium. 48Department of Pediatrics and Developmental Biology, Tokyo Medical and Dental University, Tokyo, Japan. 49Department of Pediatrics, NHO Kumamoto Medical
Center, Kumamoto, Japan. 50Department of Pediatrics, National Defense Medical College, Saitama, Japan. 51Immunology Center of Marseille-Luminy, Aix Marseille University, Inserm, CNRS,
Marseille, France. 52Immunophenomics Center (CIPHE), Aix Marseille Université, Inserm, CNRS, Marseille, France. 53Department of Anatomy and Cell Biology, College of Medicine, Alfaisal
University, Riyadh, Saudi Arabia. 54Division of Allergy and Clinical Immunology, Department of Pediatrics, Tehran University of Medical Sciences, Tehran, Iran. 55Berlin Institute of Health at
Charité – Universitätsmedizin Berlin, Berlin, Germany. 56Berlin-Brandenburg Center for Regenerative Therapies (BCRT), Charité - Universitätsmedizin Berlin, Berlin, Germany. 57Labor Berlin
GmbH, Department of Immunology, Berlin, Germany. 58Center for Molecular Medicine, Department of Medicine (Solna), Karolinska Institute, Stockholm, Sweden. 59College of Health and Life
Sciences, Hamad Bin Khalifa University, Doha, Qatar. 60Clinical and Diagnostic Immunology, Department of Microbiology, Immunology, and Transplantation, KU Leuven, Leuven, Belgium.
61
Institute for Immunology and Immune Health, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. 62Department of Microbiology, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA, USA. 63Department of Pediatrics, Necker Hospital for Sick Children, AP-HP, Paris, France. 64Howard Hughes Medical Institute, The Rockefeller
University, New York, NY, USA.

g

We then investigated the impact of the pLOF
variants on PTCRA mRNA and protein. We
were unable to test primary cells from the patients because pre-TCRa is expressed only in
the thymus. First, using an artificial construct
containing the genomic DNA (gDNA) sequence
of PTCRA from the 5′ untranslated region
(5′UTR) to the end of exon 2 (Fig. 2A), we
demonstrated that two of the seven pLOF
variants (c.58G>C and c.58+5G>A) severely
impaired pre-TCRa expression in vitro by mRNA
decay (Fig. 2, B to D; fig. S3A; and supplementary
text). Second, we transfected HEK293T cells with
C-terminally DDK-tagged complementary DNAs
(cDNAs) encoding the wild-type (WT) pre-TCRa,
one of the six coding pLOF variants identified in
the patients, or one of the 15 non-pLOF missense
variants identified in the homozygous state in
public databases or in our in-house cohort
(fig. S1A). Cell extracts were subjected to SDS–
polyacrylamide gel electrophoresis (SDS-PAGE)
followed by immunoblotting and immunodetection with a monoclonal antibody against

DDK- or the N terminus of pre-TCRa (Fig. 2E,
fig. S3B, and supplementary text). All variants
found in the homozygous state in public databases or in our in-house HGID cohort were
normally expressed in this system. By contrast,
cDNAs encoding pLOF variants yielded a truncated protein or no protein at all, except for the
p.Gly20Arg (c.58G>C) variant, which produced
normal amounts of protein in this cDNA overexpression system (Fig. 2E) but was subject to
mRNA decay in our artificial gene system (Fig. 2,
B to D). Thus, the pLOF variants identified in
the patients impair PTCRA expression by mRNA
decay or premature translation termination.

p

deficiency, including the four identified by neonatal screening, were clinically asymptomatic
at their most recent evaluation (at the ages
of 2, 2, 4, 7, 8, and 65 years). The other four
patients (13, 24, 31, and 66 years of age) displayed infection, lymphoproliferation, and/or
autoimmunity with an onset during their
teens or in adulthood (age at onset: 13, 13, 15,
and 25 years, respectively). One of these patients died from severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pneumonia
at the age of 24 years. P9 had a small thymus
on magnetic resonance imaging (MRI) at
the age of 2 years, whereas P5 and P6 had no
visible thymus at the ages of 13 and 8 years,
respectively (Fig. 1E). Three of the nine patients with pLOF PTCRA variants tested
were found to produce autoantibodies, several of which were associated with clinical
manifestations (fig. S2, A to E, and supplementary text). Antithyroid autoantibodies
and/or clinically overt thyroiditis were found
in three of the nine patients. P7, who suffered from recurrent herpes infections, had
autoantibodies against type I interferons
(IFNs) (fig. S2A). All known genetic etiologies
of these antibodies disrupt T cell tolerance as
a result of mutations affecting thymocyte de-

RES EARCH | R E S E A R C H A R T I C L E

A

B

C

p
g
y

D

E

y g
y
,

Fig. 1. Autosomal recessive pre-TCRa deficiency. (A to B) Schematic
representations of the isoform A (A) and isoform B (B) proteins encoded by
PTCRA. (C) Abundance of the indicated pre-TCRa isoforms in transcripts per
million (TPM) across thymocyte developmental stages (DN1, DN2, DN3, ISP, DP
early, DP late, SP8, and SP4). The proportion of total PTCRA transcripts
corresponding to isoform A in each thymocyte subset is indicated on the graph
(dashed gray line). (D) Pedigree of the seven unrelated families displaying
familial segregation of the mutant PTCRA alleles. The indicated mutant alleles,
Materna et al., Science 383, eadh4059 (2024)

1 March 2024

each with a specific color code, are labeled “M” in the pedigree. Individuals of
unknown genotype are labeled “E?”. Clinically asymptomatic individuals are annotated
with a vertical bar. (E) MRI on axial sections at the level of the aortic arch: T1-weighted
sequences after gadolinium injection (P5) and T2-weighted sequences (P6, P9, and
controls), for P5, P6, P9, and age- and sex-matched controls. In patients, the thymic
lodge, located between the sternum and the aortic arch (asterisk), appears empty
(P5 and P6) or small (P9). By contrast, the thymus is clearly visible in controls, even
after the onset of puberty (14-year-old girl).
3 of 18

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 2. Patient mutations and two mutants from gnomAD are LOF or
severely hypomorphic. (A) Schematic representation of the artificial gene
created to study the splicing between exons 1 and 2 of PTCRA. The two
mutations tested are depicted. TSS, transcription start site. (B to D) HEK293T
cells were transfected with an empty vector (EV) or with plasmids encoding the
artificial gene with the WT or mutant PTCRA sequence described in (A). (B) The
RNA was subjected to RT-qPCR for PTCRA with a probe spanning the splice
junction between exons 1 and 2. Data are displayed as 2−DCt values after
normalization relative to an endogenous control (DCt). The bar graphs show the
means ± SEMs of three technical replicates and are representative of three

Materna et al., Science 383, eadh4059 (2024)

1 March 2024

independent experiments. (C) Exon trapping. The bar graphs show the
proportion of canonical or noncanonical PTCRA transcripts in the transfected
HEK293T cells. (D) Total protein extracts were subjected to immunoblotting with
an antibody against the DDK tag or GAPDH. Data are representative of three
independent experiments. (E) HEK293T cells were transfected with an empty
plasmid or with a plasmid carrying a C-terminal DDK-tagged cDNA encoding the
WT or the indicated variants of PTCRA isoform A. Total protein extracts were
subjected to immunoblotting with an antibody against the DDK tag, pre-TCRa, or
vinculin. Data are representative of four independent experiments. (F to K)
TCRa-deficient Jurkat cells were transduced with an EV or with a plasmid

4 of 18

RES EARCH | R E S E A R C H A R T I C L E

encoding the WT isoform A, the WT isoform B, or the indicated variant of
pre-TCRa. The expression of TCRb [(F) and (G)], CD3e [(H) and (I)] or CD69
[(J) and (K)] at the cell surface was evaluated by flow cytometry on the
transduced cell lines. Data are representative of three independent experiments.
(F, H, and J) Histograms showing the mean fluorescence intensity (MFI) of

5 of 18

,

The asymptomatic pre-TCRa–deficient patients
(P1, P2, P6, P8, P9, and P10)—like the patient with
a mild clinical presentation (P5)—had normal or
near-normal distributions of leukocyte subsets
other than T cells and normal antibody responses
to antigens (data S1, fig. S5, and supplementary
text). By contrast, patients with clinical autoimmunity (P3, P4, and P7) were diagnosed
with common variable immunodeficiency (CVID)
and presented progressive cytopenia for multiple
cell types. Pre-TCRa deficiency affects thymocyte differentiation in mice. We consequently
investigated the blood T cell compartment
of the patients. Except for P10, all patients (P1,
P2, and P9) followed from birth displayed T
cell lymphopenia early in life (Fig. 3A). Their
total T cell counts remained stable over time,
reaching counts at the low end of the normal
range by the age of 3 years, when a physiological decline of CD3+ T cell counts is observed in
normal individuals. Relative to age-matched
controls, all patients other than P3 and P7 (aged

y

1 March 2024

Low CD3+ T cell counts in newborns with
complete pre-TCRa deficiency

y g

Materna et al., Science 383, eadh4059 (2024)

We investigated the impact of the p.Asp51Ala
variant on immunological phenotypes by analyzing the reported phenotypes of individuals
homozygous and heterozygous for this variant
among the South Asians included in the UK
Biobank. The frequencies of autoimmunity
(~20%) and hypothyroidism (~10%) codes
were similar in individuals heterozygous for
p.Asp51Ala and in controls (table S4). By contrast, three (75%) homozygous carriers had
autoimmunity-related codes, and one (25%)
had a hypothyroidism-related code. Homozygote
1 suffered from hypothyroidism and lichen
planus at the ages of 48 and 52 years, respectively. Homozygote 2 presented with thrombocytopenia and Henoch–Schönlein purpura
at the age of 50 years, and homozygote 3
suffered from rheumatoid arthritis at the age
of 50 years. No autoimmunity was reported in
homozygote 4, but he suffered from hypoxemic COVID-19 pneumonia at age 61. No lymphoproliferation was reported in any of the
four homozygotes. We also analyzed the phenotype of the homozygotes identified in other
cohorts (table S1). One of the two homozygotes
in the Qatar Biobank was asymptomatic at the
age of 46 years, and the other suffered from
hypothyroidism with autoantibodies against
thyroid peroxidase (TPO) at the age of 31 years.
Two homozygotes were identified in a Saudi
database, and clinical data were available for
only one—an otherwise healthy 38-year-old
man with vitiligo. His 40-year-old sister was

y

The p.Asp51Ala and p.Tyr76Cys variants identified in gnomAD affect residues interacting
with TCRb (fig. S4A) (27). The p.Asp51Ala variant affects a charged residue in the extracellular domain. In the mouse, knock-in mutations
of such residues impair the interaction between
pre-TCRa and TCRb, which leads to a decrease
in the count of DP thymocytes and an increase
in gd T cell counts (28, 29). The p.Asp51Ala and
p.Tyr76Cys variants may, therefore, impair dimerization between pre-TCRa and TCRb. In
gnomAD v2.1.1 and the Centogene Biodatabank,
the pTyr76Cys variant was most frequent in subSaharan Africans, with a MAF of ~0.0037 versus
0.0003 in the global population from gnomAD
V2.1.1 (fig. S4B and tables S1 to S3). Thus,
~0.001% of Africans would be expected to
have a partial deficiency of pre-TCRa (~1/
73,000 individuals). In various databases,
the p.Asp51Ala variant is more frequent in
individuals from South Asia and the Middle
East, whose MAF is ~0.01. By contrast, the
MAF for the global population from gnomAD
V2.1.1 is ~0.002 (fig. S4C and tables S1 to S3).
In these populations—which together account

Homozygosity for the Asp51Ala allele is a risk
factor for autoimmunity

shown, by Sanger sequencing, to be homozygous for the variant but was asymptomatic.
In an Iranian database of individuals recruited on the basis of neurological phenotypes and Sanger sequencing data for the
relatives of the proband, we identified three
homozygous carriers of the p.Asp51Ala variant
(P11, P12, and P13) (fig. S4D and supplementary text). None of these individuals presented
unusual susceptibility to infection. However,
two of the three children suffered from hypothyroidism. The thymic compartment of P11
(9 years old) contained tissue with abnormal
properties on MRI, which suggested that the
content of the thymus was abnormal (fig. S4E).
Thus, evidence of autoimmunity was obtained
for seven of the 11 (64%) homozygotes for whom
clinical information was available. Finally,
using the Centogene cohort, we identified
51 additional individuals homozygous for
p.Asp51Ala (tables S5 and S6 and supplementary text). In this cohort, the association
between autoimmunity and homozygosity
for the p.Asp51Ala variant was confirmed, with
an odds ratio (OR) of 5.02 relative to heterozygotes and WT subjects (95% CI, 1.750054 to
11.816898; adjusted P = 0.009965). Thus,
homozygosity for the p.Asp51Ala variant appears to be a significant risk factor for the
development of autoimmune disease in individuals of Middle Eastern and South Asian
origin.

g

Population genetics of the Asp51Ala and
Tyr76Cys variants

for almost 2 billion individuals—the p.Asp51Ala
allele can be regarded as “common” (MAF >
1%). Thus, 1/1000 to 1/10,000 Middle Eastern
and South Asian individuals would be predicted to have a partial form of recessive preTCRa deficiency. We analyzed the exomes of
two Iranian kindreds carrying the homozygous p.Asp51Ala variant and estimated that
the most recent common ancestor (MRCA) carrying the variant lived about 8000 years ago [95%
confidence interval (CI), 2511 to 29,430 years].
This finding suggests that there is no strong
depletion of individuals homozygous for the
p.Asp51Ala variant in these populations. Thus,
considering only the p.Asp51Ala and p.Tyr76Cys
alleles, ~1/180,000 individuals worldwide may
have a partial form of pre-TCRa deficiency. In
particular, the p.Asp51Ala variant is found in
the homozygous state in 1/1000 to 1/10,000 individuals in the populations of South Asia and
the Middle East.

p

cell-surface expression of TCRb and CD3. By
contrast, 13 of the 15 biallelic variants reported
in public or in-house databases restored the
expression of TCRb and CD3. The p.Asp51Ala
and p.Tyr76Cys variants induced only very low
levels of TCRb and CD3 expression (Fig. 2, F to I,
and fig. S3, C to F). Pre-TCR can signal autonomously when expressed at the cell surface. Its
successful expression is, therefore, associated
with cell-surface expression of the CD69 activation marker (25). Accordingly, all the pre-TCRa–
encoding constructs that restored the expression
of TCRb and CD3 at the cell surface also induced
weak CD69 expression (Fig. 2, J and K, and fig.
S3, G and H). Neither the pLOF variants from
the patients nor the p.Asp51Ala and p.Tyr76Cys
variants from the public Genome Aggregation
Database (gnomAD) V2.1.1 induced CD69 expression on JR3.11 Jurkat cells. Similar findings
were obtained when the deleterious variants
were tested for their impact on isoform B (Fig.
2, F to K, and fig. S3, C, D, and G). Thus, the
seven alleles from the patients are biochemically
LOF, and the patients are predicted to have an
autosomal recessive complete form of pre-TCRa
deficiency. Moreover, two missense variants
(p.Asp51Ala and p.Tyr76Cys) found in the homozygous state in the general population are highly
deleterious for pre-TCRa function.

cells transduced with the indicated PTCRA allele normalized against the
MFI for EV. (G, I, and K) Representative flow cytometry histogram plot for
the indicated PTCRA alleles. Cells transduced with the PTCRA alleles
(black line, unshaded area) are compared with the cells transduced with
the EV (shaded).

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 3. T cell immunophenotyping for patients with pre-TCRa deficiency.
(A and B) CD3+ T cell counts as a function of age. The control range is
represented by the gray area. (C) Thymic function was assessed in pre-TCRa–
deficient patients (red and blue dots) and healthy local controls (black dots;
n = 101). The concentration of sjTRECs in the blood (sjTRECs per 105 PBMCs) is
presented as a function of age. (D) TCRab+ T cell counts as a function of age for
naive and memory T cells. (E) TCRgd+ T cell counts as a function of age. The
control range is represented by the gray area. (F) TCRgd+ T cell counts as a
function of age for naive and memory T cells. (G) Frequency of TCRgd+ T cells
among total naive (CD3+CD45RA+CCR7+) and memory (defined as non-naive
Materna et al., Science 383, eadh4059 (2024)

1 March 2024

CD3+) T cells from patients (P3 to P6 and P8) and controls (n = 46).
(H) Frequency of TCRgd1+ and TCRgd2+ T cells among naive TCRgd+ T cells
from patients (P4, P8, and P9) and controls (n = 8). (I) Cell counts as a function
of age for CD4+CD8− T cells and CD4−CD8+ T cells. The control range is
represented by the gray area. (J) Naive ab T cell counts as a function of age for
CD4+CD8− T cells and CD4−CD8+ T cells. ab T cells are defined here as CD3
+
TCRgd− cells. (K) Frequency of CD4−CD8− cells in ab (defined here as CD3
+
TCRgd−) naive (right) and memory (left) T cells from patients (P3 to P6 and P8)
and controls. (L to N) Phenotyping of individuals homozygous for the p.Asp51Ala
mutation (D51A) and controls (n = 12 to 18). (L) Frequency of TCRgd+
6 of 18

RES EARCH | R E S E A R C H A R T I C L E

T cells among total naive and memory T cells. (M) TCRgd+ T cell
counts for naive and memory T cells. (N) Frequency of CD4−CD8 − cells
among ab (defined here as CD3+TCRgd−) naive (right) and memory (left)
T cells. (B, D, F, and J) P3 suffered from severe enteropathy and was

1 March 2024

We assessed the impact of the patients’ PTCRA
genotype on the early stages of T cell differentiation by isolating blood CD34+ cells from
7 of 18

,

Materna et al., Science 383, eadh4059 (2024)

Pre-TCRa deficiency impairs the generation of
TCRab+ but not TCRgd+ T cells in vitro

y

We then investigated the ab T cell compartment in more depth. The infants had low total
CD4−CD8+ and CD4+CD8− T cell counts (Fig. 3I),
which normalized between childhood and

Blood gd T cells typically have a CD4−CD8−/lo
phenotype, defining a T cell lineage that does
not pass through the CD4+CD8+ DP stage in
the thymus. In TCR-transgenic mice expressing
TCRab at the DN stage in the thymus, a small
abnormal population of TCRab cells with a
CD4−CD8− phenotype is observed in the periphery (4, 30–32). Fate mapping has shown
that these cells do not pass through the CD4
+
CD8+ DP stage. Instead, they are thought to
use the gd differentiation pathway, despite
their expression of a TCRab (33). We therefore
investigated whether a fraction of TCRab+ cells
in the periphery in pre-TCRa deficient patients
harbored the same phenotype. We found no
difference in the frequency of CD4−CD8− cells
among memory T cells from controls and patients (Fig. 3K). However, the frequency of
CD4 −CD8− DN cells among naive TCRab+
T cells from pre-TCRa–deficient patients was
higher (median = 4.2%; range, 2.5 to 8.9%)
compared with that in age-matched controls
(median = 0.6%; range, 0.2 to 1.5%) (Fig. 3K).

We tested the hypothesis that homozygosity
for the hypomorphic p.Asp51Ala variant affects
T cell differentiation. We determined the
sjTREC levels of three patients (P11 to P13).
These levels were low in the two youngest
patients (Fig. 3C). We also performed extensive immunophenotyping on these two
children, which showed their counts and proportions of myeloid, B, and NK cells to be
normal (fig. S7, A to D). The T cell counts of
these patients were within the normal range
for age-matched controls, as were the proportions of naive and memory T cell subsets, and
other TH subsets, Treg cells, iNKT cells, and
MAIT cells (fig. S7, E to I). Nevertheless, the
counts and proportions (among naive T cells)
of blood gd T cells were higher in the p.Asp51Ala
homozygotes compared with controls (Fig. 3, L
and M). In contrast to the findings for patients
with complete pre-TCRa deficiency (Fig. 3K),
the proportion of CD4−CD8− cells in the naive
ab T cell compartment was normal (Fig. 3N).
Compared with patients with complete preTCRa deficiency, p.Asp51Ala homozygotes
generally had a narrower but still distinctive
immunological phenotype, with higher proportions of gd T cells among naive T cells. This
was reminiscent of mice with mutations that
affect similarly charged residues of pre-TCRa
(28, 29).

y g

Patients with complete pre-TCRa deficiency
have normal memory ab T cell counts
and low mucosal-associated invariant
T cell counts

Patients with complete pre-TCRa deficiency
have a high proportion of CD4−CD8− DN ab
T cells among naive T cells

Low TREC levels and a high proportion
of gd T cells among the naive T cells
of p.Asp51Ala homozygotes

y

The mouse pre-TCRa is essential for ab T cell
development but is dispensable for gd T cell
development (12). We therefore studied the
impact of complete human pre-TCRa deficiency
on the two major T cell lineages. Patients had
lower blood counts of naive ab T cells compared
with age-matched controls but normal counts of
memory ab T cells (Fig. 3D). Total gd T cell
counts were high from early childhood (Fig. 3E).
In children and adults, both naive and memory
gd T cell counts remained normal to high (Fig.
3F). Accordingly, the proportion of gd T cells
among naive T cells was higher in patients
(median = 32.3; range, 3.7 to 62.3) compared
with controls (median = 0.6; range, 0.1 to 2.2)
(Fig. 3G). However, the proportion of gd T cells
among memory T cells was above the upper
limit of the control range only in P4, P5, and
P10. The proportions of d1+ and d2+ gd T cells
among naive T cells were normal in patients
with pre-TCRa deficiency (Fig. 3H). Thus, preTCRa deficiency has different impacts on the
thymic outputs of ab and gd T cells, impairing
the production of ab T cells and favoring the
production of gd T cells. Nevertheless, most circulating T cells (including naive T cells) were
TCRab+.

These CD4−CD8− DN cells did not have high
levels of HLA-DR or CD38 expression and were,
therefore, probably not chronically activated (fig.
S6I) (34). Moreover, only small proportions of
naive CD4−CD8−TCRab+ T cells from the preTCRa–deficient patients expressed MAIT cells
(CD161+TCRVa7.2+), iNKT cells (TCRVa24-Ja18+),
Treg cells (CD127−CD25+), or intraepithelial lymphocyte (IEL) markers (CLA, CD103, NKG2C,
and NKG2A), which suggests that most DN
ab T cells from the patients do not belong to
an unconventional ab T cell subset (fig. S6J).
Thus, pre-TCRa deficiency is associated with
an approximately eightfold increase in the proportion of CD4−CD8− T cells in the naive TCRab+
T cell compartment. As in mice (4), DN ab T cells
in humans may therefore develop through
an alternative T cell differentiation pathway.

g

Patients with complete pre-TCRa deficiency
have high naive gd and low naive ab
T cell numbers

adulthood (fig. S6A). However, as expected from
their low naive ab T cell counts, pre-TCRa–
deficient children and adults had low counts
of naive CD4+ and CD8+ T cells (Fig. 3J). The
low naive T cell counts of the patients were
accompanied by a higher proportion of both CD4
and CD8 effector memory T cells (fig. S6, B and
C). The proportion of regulatory T cells (Treg cells)
among CD4+ T cells was in the range of controls
for all patients (fig. S6D). Within the memory
CD4+ T cell compartment, the frequencies of
T helper (TH) subsets were within or near the
control range (fig. S6E). Accordingly, comparisons with controls revealed no major differences
in the production of TH1 (IFN-g), TH2 (IL-13),
and TH17 (IL-17A) cytokines by the patients’
memory CD4+ T cells after stimulation (fig. S6F).
In addition, pre-TCRa–deficient patients had lower
levels of CD161+TCRVa7.2+ mucosal-associated
invariant T (MAIT) cells compared with controls
and normal frequencies of invariant natural
killer T (iNKT) cells among T cells (fig. S6G).
The frequency of TCRVa7.2+ cells was low
among memory ab T cells but normal among
naive ab T cells (fig. S6H), which suggests that
the low frequency of MAIT cells was not because of impaired V(D)J rearrangement. Thus,
patients with complete pre-TCRa deficiency
have low total naive ab T cell counts, normal
ab T memory-cell counts from childhood onward,
and a low frequency of MAIT cells.

p

31 and 66 years, respectively, both displaying
progressive pancytopenia) had normal or nearnormal blood counts of total CD3+ T cells at
their most recent follow-up visit (Fig. 3, A and B).
Moreover, all patients under the age of 30
years had low proportions of single-joint TCR
excision circles (sjTRECs) among peripheral
blood mononuclear cells (PBMCs), suggesting
poor thymic output (Fig. 3C). These data suggested that pre-TCRa deficiency impairs T cell
development, resulting in low T cell counts in
infancy, facilitating detection by newborn
TREC level screening. However, the total T cell
counts of the patients gradually increased, eventually reaching the normal range for age-matched
controls (Fig. 3A). Moreover, these T cells proliferated normally upon mitogen stimulation in
vitro (data S1).

on rituximab treatment. P7 received chemotherapy for lymphoma. These
two patients are therefore depicted with triangles. C, controls; LOF, patients
homozygous for LOF variants; D51A, individuals homozygous for the
p.Asp51Ala mutation.

RES EARCH | R E S E A R C H A R T I C L E

p
g
y

Fig. 4. Impaired generation of TCRab+ T cells in pre-TCRa–deficient ATOs.
In vitro T cell differentiation from positively selected peripheral blood CD34+ cells
obtained from five healthy controls, three patients with the p.Asp51Ala variant
(D51A), and three patients with LOF PTCRA mutations (P1, P5, and P6) after
5 weeks of culture in the ATO system. (A) Flow cytometry plots showing the
expression of early and late T cell differentiation markers (CD7, CD5, CD1a,

The unexpectedly modest impact of pre-TCRa
deficiency on ab T cell development in vivo
raised the question of how these cells devel-

1 March 2024

8 of 18

,

Sequencing of the TRAD locus in ab
T cells reveals an enrichment in proximal
TCRd1 and a depletion of distal MAIT
cell rearrangements

oped in the absence of a major TCR component during the b-selection process. In patients
with complete pre-TCRa deficiency, the circulating ab and gd TCR repertoire diversities
were slightly low and normal, respectively,
(Fig. 5, A and B; fig. S8A; table S7; and supplementary text). We then investigated whether the patients displayed preferential usage
of productive V-J rearrangements at the TRAD
locus in gDNA from purified naive and memory
ab T cells. The most common productive V-J
recombination at the TRAD locus in the naive
and memory ab T cells of the patients was
TRDV01:TRDJ01 (i.e., TCRd1) (Fig. 5C and fig.
S8B). The percentages of productive and nonproductive TRD rearrangements among total
TRAD rearrangements were significantly higher
and lower, respectively, in the patients’ naive ab
T cells compared with those of the controls (Fig.
5D). Productive TRD rearrangements (involving
any TRDV) consequently accounted for ~70% of
the total TRD rearrangements detected in the ab
T cells of patients with complete pre-TCRa

y

Materna et al., Science 383, eadh4059 (2024)

TCRgd+ cells in the three pre-TCRa–deficient
patients studied, relative to controls (Fig. 4, A
and B). In particular, the ratio of TCRgd+ cells
to TCRab+ cells was markedly higher in the
pre-TCRa–deficient patients (~5) compared
with controls (~0.1) (Fig. 4C). Finally, ATOs
generated with CD34+ cells from the three
patients homozygous for the p.Asp51Ala variant had a phenotype intermediate between
those of the controls and pre-TCRa–deficient
patients (Fig. 4). Thus, complete pre-TCRa deficiency almost completely abolishes human
ab T cell differentiation in vitro, whereas partial deficiency due to p.Asp51Ala homozygosity
has a milder impact.

y g

three pre-TCRa–deficient patients and five
healthy controls and inducing their differentiation in vitro in an artificial thymic organoid
(ATO) system (Fig. 4) (35). After 5 weeks of
culture, control CD34+ cells remained highly
viable (~74%) (Fig. 4A). Efficient differentiation into CD4+CD8+ DP cells (mean, ~52% of
CD45+CD56− cells), TCRab+CD3+ SP cells (~23%),
and TCRgd+CD3+ cells (~1.8%) was observed. By
contrast, after 5 weeks of culture, viability was
much lower for the CD34+ cells isolated from
all three pre-TCRa–deficient patients (7 to 39%).
The differentiation of these cells into T cells was
impaired, with a block at the CD7+CD1a
+
CD4−CD8b− DN stage and an almost total
absence of CD4+CD8+ DP cells (mean, ~2% of
CD45+CD56− cells) and TCRab+CD3+ SP cells
(~0.5%). However, a significant fraction of the
cells were TCRgd+CD3+ (~3.5%). A determination of absolute counts per ATO of cells at
various stages of differentiation confirmed the
deficit of CD4+CD8b+ DP and TCRab+CD3+ cells
and the presence of a significant number of

CD4, CD8b, TCRab, TCRgd, and CD3) after gating on LIVE/DEAD–CD45+CD56–
cells. The data shown correspond to one control, one p.Asp51Ala patient, and
P1. (B) Plots of absolute counts per ATO for the various stages of T cell
differentiation, for the cells isolated from the ATOs. (C) Bar graphs showing
the ratio of absolute counts of TCRgd+ cells to absolute counts of TCRab+
cells per ATO.

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g

Materna et al., Science 383, eadh4059 (2024)

to TRDV01 (TRAV01 to TRAV23) were enriched
in the patients’ ab T cells (Fig. 5C and fig. S8B).
TCRd1 accounted for ~70% of total naive gd T
cells (Fig. 3H), so such a pattern would be expected for TCRa repertoires preferentially rearranged from a TCRd1 template, with TRAV23
becoming the most proximal TRAV gene after
successful TCRd1 (TRDV01:TRDJ01) rearrangement (fig. S8D). Thus, our TRAD repertoire an-

1 March 2024

alysis suggests that, in absence of pre-TCRa,
TCRa rearrangements preferentially occur from
a productive TCRd1 template.
TCRd1 is not a surrogate for pre-TCRa

Having excluded the possibility that most ab T
cells preferentially differentiate from gd+ thymocytes in the absence of pre-TCRa (Fig. 6, A
and B; tables S8 and S9; and supplementary
9 of 18

,

deficiency but only ~20% of those in healthy
controls (P < 0.0001) (Fig. 5E and fig. S8C). An
analysis of the TRAD locus from purified naive
and memory ab T cells showed a depletion of
the TRAV genes removed during TCRd1 rearrangement (TRAV24 to TRAV41) in the productive TRAD rearrangements in the ab T cells
of patients relative to controls (Fig. 5, C, F and
G, and fig. S8B). As a result, TRAV genes distal

TCRd1 (TRDV1:TRDJ1) rearrangement is indicated with a black circle. (D) Fraction
of TCRd rearrangements in total productive TRAD rearrangements from sorted
naive and memory ab T cells from controls (black; n = 4) and pre-TCRa–deficient
individuals (red; P1, P2, P4, P8, and P9). (E) Fraction of productive TCRd
rearrangements among total TCRd rearrangements in naive and memory ab T cells
from controls (black; n = 4) and pre-TCRa–deficient individuals (red; P1, P2, P4, P8,
and P9) (red). (F) Schematic representation of the TRAD locus before and after
TCRd1 rearrangement. (G) Percentage of productive TRA rearrangements involving
TRAV24-41 in sorted naive and memory ab T cells from controls (black; n = 4)
and pre-TCRa–deficient individuals (red; P1, P2, P4, P8, and P9). Unpaired t tests
were used for comparisons in (A), (B), (D), and (E).

y

Fig. 5. Biases in the TRAD rearrangement repertoire indicate that TCRa
chains are mostly generated by rearrangement of a TCRd1 template in
pre-TCRa–deficient individuals. (A) Shannon’s entropy for TCRg rearrangements in naive and memory gd T cells from controls (black; n = 4) and preTCRa–deficient individuals (red; P1, P2, P4, and P8). (B) Shannon’s entropy for
TCRa and TCRb rearrangements in naive and memory ab T cells from controls
(black; n = 4) and pre-TCRa–deficient individuals (red; P1, P2, P4, P8, and P9).
(C) Heatmap of paired gene rearrangements at the TRAD locus for naive ab
T cells from four controls compared with five pre-TCRa–deficient individuals
(P1, P2, P4, P8, and P9). The red color highlights V-J gene pairings overused in
patients and the blue color highlights V-J gene pairings overused in controls. The

RES EARCH | R E S E A R C H A R T I C L E

p

We found that the peripheral ab T cell counts
of pre-TCRa–deficient humans normalized with
age as a result of the physiological decrease in
T cell counts with age in healthy individuals
and an accumulation of memory ab T cells in the
patients. Four-week-old Ptcra−/− mice have been
reported to have 5% normal T cells in the LNs
(12). However, T cell dynamics in the thymus
and periphery have not been studied during
aging. We therefore sought to reassess and
extend mouse immunophenotyping longitudinally by studying the thymus, blood, spleen,
and LNs of 1-, 4-, 12-, and 24-week-old Ptcra−/−

1 March 2024

10 of 18

,

Longitudinal study of peripheral T cells
in 1- to 24-week-old Ptcra−/− mice

and control mice. The skewed thymocyte differentiation observed in pre-TCRa–deficient
mice remained stable in the aging thymus between the ages of 1 month and at least 6 months
(Fig. 7, A to C, and supplementary text). Contrasting with the previous report of 5% normal
T cells, we found that the CD4+ and CD8+ ab
T cell counts in the LNs of 4-week-old mice
corresponded to 23% and 14% of the normal
level, respectively (Fig. 7D) (12). CD4+ ab T cell
counts were 39% and 26% of the normal values
in 12- and 24-week-old Ptcra−/− mice, respectively, whereas CD8+ ab T cell counts were 14%
and 45% of the normal values in 12- and
24-week-old Ptcra−/− mice, respectively. Circulating CD4+ ab T cell counts in Ptcra−/− mice
were 2%, 11%, and 15% of the normal level,
whereas circulating CD8+ ab T cell counts were
1%, 5%, and 15% of the normal levels in 4-, 12-,
and 24-week-old mice, respectively (Fig. 7D).
Splenic CD4+ ab T cell counts in Ptcra−/− mice
were 17%, 18%, and 40% of the normal values,
whereas splenic CD8+ ab T cell counts were
3%, 5%, and 19% of the normal values in 4‐,
12‐, and 24‐week-old mice, respectively (Fig. 7D).

y

the rearrangement frequently found in pre-TCRa–
deficient patients and a TCRb chain previously
suggested to stabilize TCRd1 expression (36).
By contrast, pre-TCRa or TCRa stabilized CD3
expression at high levels on the cell surface
after cotransduction with any TCRb construct
(Fig. 6D). Thus, in this system, TCRd1 and TCRg
are unable to replace the pre-TCRa to stabilize
TCRb expression at the cell surface.

y g

Materna et al., Science 383, eadh4059 (2024)

and P8). The red color highlights V-J gene pairings overused in patients, and the
blue color highlights V-J gene pairings overused in controls. (D) TCRab-deficient
Jurkat cells were stably transduced with an empty plasmid or with a plasmid
encoding TCRa, pre-TCRa, TCRd1, or TCRg. Each of the resulting cell lines was then
cotransduced with another empty plasmid or with a plasmid encoding one of
eight selected TCRb chains. The expression of CD3 at the cell surface was evaluated
by flow cytometry. Representative flow cytometry histogram plot for three
independent experiments is shown. Representative flow cytometry data are
shown on the left. A recapitulative bar graph of the MFI for CD3 for each cell line
is shown on the right.

y

text), we hypothesized that TCRd might act as
a surrogate for pre-TCRa in the formation of a
pre-TCR complex with specific TCRb rearrangements and CD3. Consistent with this hypothesis, we found that the TCRb repertoire was
biased in patients with pre-TCRa deficiency,
with an enrichment in rearrangements involving the middle TRBV genes and any TRBJ
gene or involving the distal TRBV02-1 gene
and any TRBJ02 gene (Fig. 6C and fig. S8,
E and F). We transduced TCRab-deficient
Jurkat cells with TCRd1, pre-TCRa, TCRa, or
TCRg cDNA. These stable cell lines were cotransduced with an empty vector or one of
eight selected TCRb chains, and CD3 stabilization at the cell surface was assessed by flow
cytometry. TCRd1 and TCRa alone stabilized
low amounts of CD3 on the cell surface, whereas
neither pre-TCRa, TCRb, nor TCRg alone could
stabilize CD3 expression at detectable levels on
the cell surface. Relative to transduction with
single chains, we observed no enhancement of
CD3 stabilization after cotransduction with
TCRd1 or TCRg together with any of the tested
TCRb chains, including the TCRb chain with

g

Fig. 6. gd+ thymocytes do not preferentially differentiate into ab T cells
in the absence of pre-TCRa, and TCRd1 cannot act as a surrogate for
pre-TCRa. (A) Fraction of productive TCRg among total TCRg templates in
sorted naive and memory ab (left) or gd (right) T cells from controls (black;
n = 4) and pre-TCRa–deficient individuals (red; P1, P2, P4, and P9). Unpaired
t tests were used for all comparisons. (B) Fraction of expanded ab T cell clones
from two controls and two pre-TCRa–deficient individuals with a productive TRG
rearrangement at the gDNA level. Notably, ~90% of these clones were CD4+CD8−.
(C) Heatmap of paired gene rearrangements of the TRB locus for naive ab
T cells from controls compared with five pre-TCRa–deficient individuals (P1, P2, P4,

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g

Materna et al., Science 383, eadh4059 (2024)

mice were higher than those of control mice
at all time points, with a greater increase in
the spleen [10- to 20-fold between weeks 4 and
24 (W4 and W24)] and the LNs (7- to 16-fold
between W4 and W24) compared with in the
blood (2- to 10-fold increase between W4 and
W24) (Fig. 7D). As in humans, circulating ab
CD4−CD8− DN T cells were more abundant in

1 March 2024

Ptcra−/− compared with WT mice (Fig. 7D).
Splenic and LN ab CD4−CD8− DN T cell counts
did not significantly differ between Ptcra−/−
and mice, however. Thus, CD4+ and CD8+ ab
T cell counts increase in aging Ptcra−/− mice
mostly because of an expansion of the memory T cell compartment. These findings are
consistent with our pre-TCRa–deficient patient
11 of 18

,

Treg cell counts mirrored the counts of total
CD4+ ab T cells in the LNs and the blood (Fig.
7D). The increase in CD4+ and CD8+ ab T cell
counts in the periphery in aging Ptcra−/− mice
was driven by an accumulation of memory T
cells, except for blood CD8+ T cells, which remained mostly naive in 24-week-old mice (Fig.
7, E and F). The gd T cell counts of Ptcra−/−

(CD3+TCRgd−CD4−CD8+), DN (CD3+TCR−gd−CD4−CD8−), Treg (CD3+TCRgd−CD4
CD8−CD25+), and gd (CD3+TCRgd−CD4+CD8−). (E) Representative flow
cytometry plots of naive and memory cell staining for CD4 and CD8 ab T cells
from the spleen and LNs of 12-week-old WT (black) and Ptcra−/− mice (red). EM,
effector memory; CM, central memory; N, naive. (F) Frequency of naive cells
among CD4-SP and CD8-SP ab T cells, and of gd T cells in the indicated tissue
of Ptcra−/− mice (red) and WT mice (black) from 0 to 24 weeks of age. (B, D,
and F) The data shown are the means and standard deviations of four to six
animals at each age.
+

y

Fig. 7. Longitudinal studies of Ptcra−/− mice. (A) Schematic representation of
thymocyte differentiation stages in mice. (B) Cell counts per thymus in Ptcra−/−
mice and WT mice aged 0 to 24 weeks for the various thymocyte developmental
stages. mut, mutant. (C) Intracellular (IC) and membrane (Mb) expression of
TCRb for the indicated thymocyte subsets from Ptcra−/− mice (bottom) and
WT mice (top). Cytometry data are representative of six Ptcra−/− mice, and six
WT mice are shown. (D) Counts of cells per microliter of blood or per spleen or
per LN for Ptcra−/− mice (red) and WT mice (black) aged 0 to 24 weeks for
the indicated T cell subsets, including CD4-SP (CD3+TCRgd−CD4+CD8−), CD8-SP

RES EARCH | R E S E A R C H A R T I C L E

data, which highlights the importance of studying aging animals when exploring thymic
phenotypes.
Discussion

Founder effect analysis for the p.Leu99Hisfs*68
and p.Asp51Ala variants

We used the UK Biobank plink-formatted
population-level exome original quality functional
12 of 18

,

Analysis of the UK Biobank data

y

The occurrence of homozygosity for the p.
Leu99Hisfs*68 and p.Asp51Ala variants in different kindreds suggested a founder effect. An
analysis of the whole-exome sequencing data
revealed that P4 (kindred C) and P5 (kindred D),
both of whom are homozygous for p.Leu99Hisfs68*, have a homozygous haplotype around
PTCRA and encompassing a 1.38 Mb region
corresponding to 73 single-nucleotide variants
(SNVs) in common. The ESTIAGE method estimated the age of the MRCA of the two patients
at 60 generations (95% CI, 19 to 258 generations) (47). Assuming a generation time of
27 years (48), the MRCA of these patients with
the p.Leu99Hisfs68* mutation would have lived
about 1600 (513 to 6966) years ago. Similarly, an
analysis of the whole-exome sequencing data of
P11 (kindred H) and P13 (kindred I), both homozygous for p.Asp51Ala, showed that they had a
homozygous haplotype around PTCRA encompassing a 580 kb region corresponding to
39 SNPs in common. The ESTIAGE method
estimated the age of the MRCA of these two
patients at 301 generations (95% CI, 93 to 1090
generations) (47). Assuming a generation time
of 27 years (48), the MRCA of these patients
with the p.Asp51Ala mutation would have lived
about 8000 (2511 to 29,430) years ago.

y g

1 March 2024

P1, P9, and P10 underwent whole-exome sequencing as part of clinical care after a presumptive positive result for neonatal SCID
screening. As a sibling of P1, P2 underwent
neonatal SCID screening (presumptive positive
result) and confirmatory testing for the familial
mutations identified in P1. Commercial wholegenome sequencing (Macrogen) was performed
for P3, P4, P5, and P7, due to their clinical history
of immunodeficiency or autoimmunity. P6 and
P8 were identified by regular Sanger sequencing
for familial segregation analysis. The frequency
of the p.Asp51Ala [Chr6(GRCh37):g.42890858A>C)]
variant was evaluated in different cohorts including the South Asians from the UK biobank (44),
800 Qataris from the Qatar Genome Program
(QGP) and Qatar Biobank (QBB) longitudinal
study (45), Kuwaitis (46), Iranians (neurological
phenotypes only), and Saudis (patients with
suspected Mendelian diseases and their parents).

y

Written informed consent was obtained from
legally authorized representatives in accordance
with the Declaration of Helsinki. The study was
approved by the ethics committee of INSERM
(RCB 2010-A00634-35 et 2008-A01078-47), the
UZ/KU Leuven ethical committee for research
(reference number s58466), the Children’s
Hospital of Philadelphia Institutional Review
Board, and Tokyo Medical and Dental Univer-

Whole-exome sequencing, whole-genome
sequencing, and Sanger sequencing

g

Materna et al., Science 383, eadh4059 (2024)

Materials and methods
Informed consent

sity (G2000-103). For studies of in vitro T cell
differentiation and high-throughput sequencing of TCR repertoires, patients gave consent
to participate in protocol 18-I-0128 approved by
the NIH IRB and registered at www.clinicaltrials.gov as protocol NCT03610802.

p

The mouse and human PTCRA genes were
discovered in 1994 and 1995, respectively, and
the first pre-TCRa–deficient mice were described
in 1995 (12, 37, 38). We report 10 humans from
seven kindreds and three distant ancestries
with autosomal recessive complete pre-TCRa
deficiency. Given the severity of ab T cell deficiency in pre-TCRa–deficient mice (12), it seemed
likely that pre-TCRa–deficient humans would
suffer from life-threatening infections in infancy.
To our surprise, this was not the case, and six
of our 10 patients, aged 2 to 65 years, have
remained healthy. The remaining four patients
have exhibited severe infection, lymphoproliferation, or autoimmunity beginning between
the ages of 13 and 25 years. This relatively mild
clinical phenotype is likely the result of an agedependent accumulation of normal numbers
of diverse, functional memory ab T cells. With
hindsight, these findings do not conflict with
the reported role of pre-TCRa in mice. The
impact of the Ptcra−/− genotype on thymocytes
and peripheral T cells has not been studied
in aging mice and young mice, which survived in pathogen-free conditions and have
not been challenged with pathogens. As in
humans, we observed a progressive increase
in mouse blood ab T cell counts with age,
driven by memory cell accumulation. These
findings highlight the need for caution when
extrapolating phenotypes from mutant mice,
which are often studied at a young age and in
a narrow range of experimental conditions,
to humans (39, 40). They also suggest that it
can be productive to revisit mouse phenotypes
on the basis of human studies. In both mice
and humans, ab T cells can develop in the absence of pre-TCRa.
These findings raise questions about how
diverse naive ab T cells develop in the absence
of pre-TCRa. Our first hypothesis was that early
productive proximal TRA rearrangements may
permit ab T cell development (41). However, we
observed a depletion of productive TRAD rearrangements involving proximal TRAV genes
(TRAV24 to TRAV41) and an enrichment in
rearrangements involving distal TRAV genes
(TRAV1 to TRAV23) in the patients’ naive ab
T cells. Moreover, an abnormal enrichment in
the productive TCRd1 (TRDV01:01-TRDJ01:01)
rearrangement was observed in ab T cells from
patients. Because TRAV segments preferentially
recombine with symmetric TRAJ segments (proximal V with proximal J, distal V with distal J)
(42, 43), the TCRa repertoire observed in the
absence of pre-TCRa—with a depletion of rearrangements involving proximal TRAV and
an enrichment in rearrangements involving distal
TRAV—suggests that these TCRa rearrangements

occurred preferentially with a TCRd1 template
(fig. S8D). We therefore tested the hypothesis
that TCRd permits ab T cell development.
However, we found that TCRd1 was unable to
act as a surrogate for pre-TCRa in the formation of a pre-TCR. Moreover, similar to controls,
most CD4+ SP ab T cell clones from the patients
did not carry a productive TRG rearrangement,
which suggests that most of the patients’ T cells
were unlikely to have differentiated directly
from gd+ thymocytes. These findings call for
alternative hypotheses that may account for ab
T cell differentiation in the absence of pre-TCRa,
which are consistent with the associated rearrangement bias observed at the TRAD locus.
We also identified two alleles affecting residues interacting with TCRb responsible for
partial pre-TCRa deficiency in homozygotes.
Homozygosity for the p.Tyr76Cys variant affects
about 1/76,000 individuals in sub-Saharan Africa.
Homozygosity for p.Asp51Ala is more common,
affecting between 1/1000 and 1/10,000 individuals in South Asian and Middle Eastern countries.
The p.Asp51 residue was previously identified
as potentially crucial for pre-TCRa function
owing to its conservation and negative charge
(28). We show that this amino acid plays a
crucial role in pre-TCRa dimerization with
TCRb. Humans homozygous for p.Asp51Ala
have a partial form of pre-TCRa deficiency.
They have normal cell counts of blood ab T cell
subsets but high counts of naive gd T cells and
low sjTREC levels. This phenotype is reminiscent of that of transgenic mice with substitutions in the extracellular domain of pre-TCRa,
which have a phenotype intermediate between
those of WT and Ptcra−/− mice (29). The clinical
phenotype of individuals with partial pre-TCRa
deficiency is milder than that of individuals
with complete pre-TCRa deficiency. The patients
are asymptomatic or display isolated autoimmune manifestations. Homozygosity for hypomorphic mutations of PTCRA should be
considered in patients with isolated autoimmunity, particularly the p.Asp51Ala substitution in individuals of South Asian or Middle
Eastern origin. Collectively, our findings demonstrate that human pre-TCRa is largely redundant for ab T cell development, but its complete
or partial deficiency can result in late-onset
clinical manifestations (including autoimmunity
in particular) with incomplete penetrance.

RES EARCH | R E S E A R C H A R T I C L E

Sanger sequencing and TA cloning

The C-terminally Myc/DDK-tagged pCMV6
empty vector and the human PTCRA expression
vectors were purchased from Origene (NM_138296;
no. RC215794). Constructs carrying mutant alleles
were generated by direct mutagenesis with the
CloneAmp Hifi premix and polymerase (no.
639298, Takara). The resulting PCR products
were digested with DpnI (no. R0176L, New
England Biolabs) for 1 hour at 37°C, amplified
in competent E. coli cells (no. C3019H, New
England Biolabs), and purified with a Maxiprep
kit (no. 12663, Qiagen). Isoform B of pre-TCRa
and the mutant allele lacking exons 1 to 3 were
obtained by opening the Isoform A WT vector
by PCR with primers flanking the region to be
deleted and then using the Quick Blunting Kit
(no. E1201L, New England Biolabs) and the
Quick Ligation Kit (no. M2200S, New England
Biolabs) as recommended by the manufacturer.
Lentiviral plasmids carrying the various PTCRA
variants were generated by inserting the cDNA
from the pCMV6 plasmids into an empty pTripSFFV-DNGFR vector (modified pTRIP-SFFVmtagBFP-2A; addgene, plasmid no. 102585).
This was achieved by digesting the empty pTripSFFV-DNGFR vector with XhoI and BamHI
for 1 hour at 37°C. The cDNA of interest was
amplified by PCR and inserted into the vector
by homologous recombination with the In-Fusion
HD Cloning Kit according to the manufacturer’s
instructions (no. 638911, Takara). Lentiviral
plasmids pTrip-SFFV-DNGFR encoding TCRa,
TCRd, TCRg, or TCRgd and pTrip-SFFV-GFP
encoding various forms of TCRb were synthesized by TwistBioscience, after onboarding our
empty vector plasmid. HEK293T cells were
transiently transfected in the presence of the
X-tremeGENE 9 DNA Transfection Reagent
13 of 18

,

1 March 2024

Plasmids and transient transfection

y

Genomic DNA was obtained from whole blood
from the patients. The PTCRA mutations identified by WES were checked by amplifying the
corresponding gDNA regions with a recombinant
Taq polymerase (Thermo Fisher Scientific). Polymerase chain reaction (PCR) products were
purified by centrifugation through Sephadex
G-50 Superfine resin (Merck) before and after
the sequencing reaction, which was performed
with the BigDye Terminator Cycle Sequencing
Kit (Applied Biosystems) and the primers previously used to amplify the region of interest.
Purified sequencing products were analyzed
with an ABI Prism 3500 apparatus (Applied
Biosystems) and aligned with the genomic
sequence of PTCRA (Ensembl) with Serial
Cloner 2.6 software. We checked that the compound mutations found in P7 and P8 really were
in two different alleles by cloning the PCR

HEK293T cells were cultured in Dulbecco’s
modified eagle medium (DMEM) (no. 61965059,
Gibco) supplemented with 10% fetal bovine
serum (FBS) (Sigma-Aldrich). TCRa-deficient
JR3.11 Jurkat cells and TCRab-deficient J76
Jurkat cells were cultured in RPMI (no. 61870044,
Gibco) supplemented with 10% FBS (25). All
cell lines were cultured at 37°C under an
atmosphere containing 5% CO2. For transfection, HEK293T cells were plated at a density
of 8×105 cells per well in six-well plates. PBMCs
were isolated from whole-blood samples
by Ficoll-Hypaque centrifugation (Amersham-Pharmacia-Biotech).

y g

Materna et al., Science 383, eadh4059 (2024)

Data for biological triplicates of RNA-seq performed on sorted primary human thymic
T cell subsets (Thy1/DN1, Thy2/DN2, Thy3/
DN3, ISP4/ISP, DP early, DP late, and singlepositive SP8 and SP4) were downloaded with the
SRA toolkit and the fastq-dump v2.9.6 tool (BioProject dataset accession number: PRJNA741323)
(16). The sequence reads were aligned with the
human hg38 reference genome assembly with
HISAT2 v2.2.1, using the -k 1 function (49).
The principal pre-TCRa isoforms were detected with a combination of cufflinks v2.2.1
de novo transcript assembly (50) and the manual
curation of transcript databases: Ensembl hg38
v96 and NCBI Refseq genes v110. Expression was
estimated for five pre-TCRa isoforms, along with
the full gene list in Ensembl v96, with kallisto
v0.46.1, for each sample (51). Estimated normalized isoform expression, expressed in transcripts per million (TPM), was used to compare
expression levels across thymocyte developmental stages and to calculate the abundance
of isoform A relative to the other isoforms. The
aligned reads were converted to BAM format
with samtools v1.14 and the triplicates were
combined with the merge function and loaded
onto the Integrated Genome Viewer for figure
preparation (52, 53).

Cell culture

y

Exome and genome analyses performed at
Centogene up to 21 April 2023 were analyzed
(Centogene started exome and genome sequencing in 2014 and 2016, respectively). Only variants
with a base coverage ≥10, read frequency ≥30,
and variant quality ≥220 (only for exome sequencing) were retained for further analyses.
We then selected all related genetic information for the variants NM_138296.3:c.152A>C
[PTCRA p.(Asp51Ala)] and NM_138296.3:
c.227A>G [PTCRA p.(Tyr76Cys)]. We extracted
information concerning the patient’s year of
birth, sex, country of origin, family relationship,
reported genetic test results of ES/GS-tested
individuals, and HPO-encoded clinical information from our database when available. All the
available deidentified data were aggregated at
individual level.
We calculated p.(Asp51Ala) allele frequencies
and their binomial CIs by country and geographic region. For the phenotypic analysis, we
stratified the cohort by PTCRA p.(Asp51Ala)
genotype [homozygotes (HOM), heterozygotes
(HET), and WT] and counted the number of
occurrences per individual of any HPO term
from a predefined list of 24 autoimmunityrelated terms (table S5). We analyzed the difference in the proportions of individuals with
a matching autoimmune-related phenotype (having at least one of 24 predefined HPO terms)
by PTCRA p.(Asp51Ala) genotype. We tested
the hypothesis of an association between the
individual’s PTCRA p.(Asp51Ala) genotype and

RNA-seq analysis of sorted human
thymocyte subsets

amplicons from the gDNA of these patients
with the TOPO TA cloning kit (Thermo Fisher
Scientific) and using them for the one-shot
transformation of TOP10 chemically competent
Escherichia coli cells (Thermo Fisher Scientific).
PCR with the M13 primers supplied with the
TA cloning kit was performed on individual
colonies before sequencing.

g

Analysis of the Centogene Biodatabank

the presence of an autoimmune-related phenotype by enrichment analysis. Briefly, we calculated the OR of having a positive match to
the autoimmunity-related HPO terms from the
predefined list, comparing all PTCRA p.(Asp51Ala)
genotype groups in Fisher’s exact test. Statistical
analyses and figures were produced with RStudio
(version 2023.03.1 Build 446, Posit Software,
rstudio.com), using R Statistical Software
(version 4.3.0, R Core Team 2023, R-project.org)
and the tidyverse package (version 2.0.0, Posit
Software, tidyverse.org).

p

equivalent exome files for n = 454,713 individuals
(field 23155, with genotypes set to missing when
read depth was <7 for single-nucleotide variations). For the UK Biobank cohort, the participants were aged 50 to 87 years as of 2021, and
55% were female. We further restricted our analysis to participants with South Asian ancestry,
using field 21000. We considered a participant
to be of South Asian ancestry if they had one of
the following data codes for field 21000: 3001
(Indian), 3002 (Pakistani), 3 (Asian or Asian
British), 3003 (Bangladeshi), or 3004 (any other
Asian background). ICD10 codes and associated
dates were collected from inpatient data (category 2000), cancer registries (category 100092)
and first occurrences (category 1712), defined as
the earliest occurrences of ICDs in the general
practice, inpatient and death data, at three-digit
resolution. The ICD10 codes used for autoimmune diseases were as follows: E03 (hypothyroidism), E10 (type 1 diabetes), M35 (Sjogren’s
disease), G73.7 (Addison’s disease), K90.0 (celiac
disease), M33 (dermatomyositis), M34 (systemic
sclerosis), E05.0 (Graves’ disease), G35 (multiple
sclerosis), M06 (rheumatoid arthritis), G70 (myasthenia gravis), D51.0 (pernicious anemia), M32
(systemic lupus erythematosus), D69.6 (autoimmune cytopenia), L80 (vitiligo), and L40
(psoriasis).

RES EARCH | R E S E A R C H A R T I C L E

(no. 6365787001, Roche), in accordance with
the manufacturer’s instructions.
Choice of TCRa, TCRb, TCRg, and TCRd1
cDNA sequences

Intracellular cytokine staining

Intracellular cytokine staining (ICS) was performed as previously described (57). Briefly,
PBMCs were thawed in cRPMI (1X RPMI
14 of 18

,

Transduced JR3.11 Jurkat cells and J76 Jurkat
cells were stained with antibodies against CD271
(no. 557196, BD, 1:500), CD3 (no. 555333, BD,
2:50), C1b TCR (no. 565776, BD, 2:50), or CD69
(no. 310912, Biolegend, 1:400) and incubated
with the Aqua Live/Dead Cell Stain Kit (Thermo
Fisher Scientific) for 1 hour at room temperature
before analysis with a Gallios (Beckman Coulter)
flow cytometer. All data were analyzed with
FlowJo 10.5.3 software.

y

1 March 2024

Flow cytometry analysis of JR3.11 Jurkat cells
and J76 Jurkat cells

y g

The exon-trapping primers were designed to
retain the translation frame of the Myc/DDK
tag from the pCMV6 plasmid (Fig. 2A). The
end of exon 2 was, therefore, fused to a Myc/
DDK tag for detection of the artificial fusion protein. Total protein was extracted from HEK293T
cells after 24 hours of transfection with the WT,
c.58G>C or c.58+5G>A exon-trapping vectors. We
assessed the expression of PTCRA variants by
extracting total protein from HEK293T cells
48 hours after transfection with the various
pCMV6 plasmids encoding the PTCRA variants.
Total protein extracts were obtained by incubating cells with lysis buffer (50 mM Tris, pH 7.4,
150 mM NaCl, 2 mM EDTA, and 0.5% Triton
X-100). A mixture of protease and phosphatase
inhibitors was added to the buffers: aprotinin
(10 mg/ml; Sigma-Aldrich), PMSF (1 mM; SigmaAldrich), leupeptin (10 mg/ml; Sigma-Aldrich),
protease inhibitor cocktail (Sigma-Aldrich). After
30 min of lysis at 4°C, the cells were centrifuged
for 10 min at 16,000g, and the supernatant was
collected for immunoblotting. For each variant,
we separated 20 mg of total protein by SDSPAGE and immunoblotting was performed
with antibodies against the DDK Tag (1:3000,
HRP-coupled, M2, no. A8592, Sigma-Aldrich),

The lentiviruses used for the transduction of
TCRa-deficient JR3.11 Jurkat cells and TCRabdeficient J76 Jurkat cells were produced by
transfecting HEK293T cells with pCMV-VSVG (0.2 mg) (56), pHXB2 env (0.2 mg; NIH-AIDS
Reagent Program; no. 1069), psPAX2 (1 mg; gift
from D. Trono; Addgene plasmid no. 12260)
and a vector containing the sequence for transduction. The vectors containing the sequences
for transduction were pTrip-SFFV-DNGFR
(empty vector), pTrip-SFFV-GFP (empty vector),
pTrip-SFFV-DNGFR-PTCRA-WT, the other pTripSFFV-DNGFR vectors containing the PTCRA
variants studied, pTrip-SFFV-DNGFR-TCRa,
pTrip-SFFV-DNGFR-TCRd, pTrip-SFFV-DNGFRTCRg, pTrip-SFFV-DNGFR-TCRgd, and the
pTrip-SFFV-GFP-TCRb vectors. HEK293T cells
were transfected in six-well plates and the medium was replaced after 6 hours of incubation.
The virus-containing supernatant was collected
and passed through a 0.2-mm filter 24 hours
after the medium was changed. Protamine sulfate (8 mg/ml) was added to the virus-containing
supernatant, which was then added to Jurkat
cells (immediately after seeding), which were
spinoculated for 2 hours at 1200g and 25°C.
The cells were then cultured for 48 hours at
37°C under an atmosphere containing 5% CO2,
without shaking. Transduction efficiency was
then checked by flow cytometry with the green
fluorescent protein (GFP) tag or an anti-CD271
antibody (no. 557196, BD, 1:500). Transduced
cells were sorted with a magnetic MACS Column
and the CD271 MicroBead Kit (no. 130-099023, Miltenyi Biotec), as recommended by
the manufacturer.

y

Cell lysis and immunoblotting

We cloned PTCRA gDNA from a control, as described in Fig. 2A. The inserted PTCRA gDNA
sequence, extending from the 5′UTR to the end
of exon 2, is represented in blue (exons) and
gray (intron 1). Briefly, PTCRA gDNA containing exon 1 and the first 909 nucleotides of
intron 1 was amplified with CloneAmp Hifi
premix (Takara), the forward primer 5′- GAGATCTGCCGCCGCGTAGAAGGCAGTCTTGTGGGTGC-3′, and the reverse primer 5′AAGGAACTCAGTTCCTCCAGGACTCAACCTCCAGA-3′. Similarly, PTCRA gDNA containing
the last 724 nucleotides of intron 1 and exon 2
was amplified with the forward primer 5′GGAACTGAGTTCCTTGAGAGCAGGGACAATGACTTAC-3′ and the reverse primer 5′-CTCGAGCGGCCGCGTACGCGTTGACAGATGCATGGGCTGTGTAC-3′. The Infusion Cloning kit
(Clontech) was used to insert both PCR products between the ASIS1 and Mlu1 cloning sites
of the pCMV6 entry vector (Origene) by homologous recombination. The c.58+G>C or c.58
+5G>A mutation was generated by mutagenesis. We extracted mRNA from HEK293T cells
after 24 hours of transfection with the WT,
c.58G>C or c.58+5G>A exon-trapping vectors.
Materna et al., Science 383, eadh4059 (2024)

Total RNA was extracted from the indicated
cells with the RNeasy Extraction Kit (Qiagen).
RNA was reverse-transcribed with the SuperScript II reverse transcriptase (Thermo Fisher
Scientific) and oligo-dT primers (Thermo Fisher
Scientific). We then performed qPCR with the
Applied Biosystems Assays-on-Demand probes/
primers specific for PTCRA-FAM (Hs00300125_m1)
on 100 ng cDNA. The data were normalized relative
to the expression (DCt) of GUS (13-glucuronidaseVIC, 4326320E) and are expressed as 2–DCt values.

Lentivirus production and transduction

g

Artificial gene and exon trapping for the
c.58G>C and c.58+5G>A alleles

mRNA purification and reverse transcription
quantitative PCR (RT-qPCR)

PTCRA (1:3000, PA5-95578, Invitrogen), glyceraldehyde-3-phosphate dehydrogenase (GAPDH)
(1:5000, FL335, no. sc47724 HRP, Santa Cruz
Biotechnology) or vinculin (1:5000, EPR8185,
no. ab129002, Abcam). Staining was detected
with the Clarity Western ECL substrate (Biorad,
no. 1705061) or SuperSignal West Femto (ThermoScientific, no. 34096) with ChemiDoc MP (Biorad).

p

The TCRa sequence was obtained from addgene
plasmid no. 128544 (after removal of the intron).
The choice of TCRb was based on a comparison
of the TRB repertoire of naive ab cells from
patients and controls. The TRBV2*01|TRBJ21*01, TRBV7-8*01|TRBJ2-1*01, TRBV5-4*01|
TRBJ1-6*01, and TRBV7-8*01|TRBJ1-6*01 rearrangements were selected because they were
found to be overused in patients’ cells. Conversely, the TRBV19*01|TRBJ1-5*01 and TRBV291*01|TRBJ1-1*01 rearrangements were selected
because they were less frequently detected in the
patients’ cells than in control cells. The TRBV123*01|TRBJ1-2*01 and TRBV18*01|TRBJ1-2*01
TCRb chains were selected because they are the
chains expressed in the Jurkat and DN-D41 cell
lines, respectively (54). These two cell lines were
used as controls. It has been suggested that DND41 expresses the TCRd1/TCRb heterodimer at
the cell surface (36). The full-length TCRd1, TCRg,
and TCRb sequences were assembled with
stitchR software (55). The TCRg sequence used
was the example data provided by stitchR. The
CDR3 of the TCRd1 and TCRb sequences were
randomly picked from the most frequent ab or
gd T cell clones of the patients with the V-J of
interest according to our TCR bulk sequencing
data. The full sequences of all the TCRs used in
this study are provided in the supplementary
materials.

The cDNA for the PTCRA transcript was amplified with a recombinant Taq polymerase
(Thermo Fisher Scientific), a forward primer
5′-TAGAAGGCAGTCTTGTGGGTGC-3′ binding
to the 5′UTR of PTCRA, and a reverse primer
5′-CATTTGCTGCCAGATCCTCTT-3′ binding to
the in-frame C-terminal Myc/DDK tag. The
PCR products were then cloned with the TOPO
TA cloning kit (Thermo Fisher Scientific) and
used for the one-shot transformation of TOP10
chemically competent E. coli cells (Thermo
Fisher Scientific). Splice variants of PTCRA from
individual colonies were amplified with the M13
primers supplied with the TA cloning kit before
sequencing and alignment with the PTCRA
cDNA (NM_138296.2), with SnapGene software
used to identify alternative splicing variants. We
screened 82 colonies for the WT PTCRA exontrapping vector, 65 for the c.58G>C PTCRA exontrapping vector, and 83 for the c.58+5G>A
PTCRA exon-trapping vector.

RES EARCH | R E S E A R C H A R T I C L E

1640 supplemented with 10% heat-inactivated
FBS, 1% L-glutamine and 1% Pen-Strep) and
centrifuged before being resuspended at a
concentration of 2×106 cells/ml in cRPMI. Cell
pellets were resuspended in surface antibody
cocktail (table S10) and incubated for 20 min.
Cells were then permeabilized with permeabilization reagent (Invitrogen) and incubated for
20 min followed by a wash with PERM buffer.
Cell suspensions were centrifuged, and the
pellets stained with intracellular antibody cocktail (table S10) for 60 min. Finally, cells were washed
with PERM buffer before being resuspended in
1.6% paraformaldehyde (PFA) to fix. The fixed
cells were stored overnight at 4°C and analyzed
on Aurora Spectral flow cytometer (Cytek).
Mass cytometry (CyTOF)

The blocking activity of anti–IFN-a, anti–IFN-b,
and anti–IFN-w autoantibodies was assessed
in a luciferase reporter assay, as described
elsewhere (58).
Protein array for assessing autoantibodies

Protein arrays (HuProt v4.0 from CDI laboratories) for assessing autoantibodies were performed as previously described (24).
Single-cell RNA-seq (5′ transcriptomics,
ab and gd TCR)

Antibody profiling by phage immunoprecipitation sequencing (PhIP-Seq) was performed
on plasma samples from patients and controls
as previously described (60).
In vitro T cell differentiation in the ATO system

In vitro T cell differentiation was studied by
coculturing peripheral blood CD34+ cells with
15 of 18

,

VirScan—phage immunoprecipitation sequencing

y

We processed 10x single-cell transcriptome
libraries with Cellranger (v6.1.1) and Seurat
(v4.0.4). The TCRab and TCRgd libraries were
demultiplexed and cell barcodes were assigned
with Minnn (v10.1). TCR libraries were annotated with MiXCR (v3.0.13) and then separated
by subject. The numbers of ab or gd TCRs for
the patients and controls were calculated by
counting the numbers of cells expressing both
TRA and TRB V-J genes and both TRG and
TRD V-J genes. The final counts corresponded
to the intersection of cells expressing combinations of TRA, TRB, TRG or TRD genes, accounting for 10,365, 21,755, 2233, and 1440 cells,
respectively, for the controls (n = 11) and 5963,
11,416, 2095, and 529 cells, respectively, for the
patients (n = 5). The diversity of a, b, g, and
d TCRs was estimated by calculating Shannon’s
entropy (H) index. Entropy was calculated by
summing the frequencies of each clone (CDR3
amino acid sequence) and multiplying by the
base 2 logarithm of the same frequency over all
cells expressing TRA, TRB, TRG or TRD V-J
genes. Higher H-index values indicate a more
diverse distribution of CDR3 clones (59).

y g

1 March 2024

Evaluation of TCR entropy and TCR chain
combinations in single-cell RNA-seq data

y

CD3+ T cells were sorted by flow cytometry from
the PBMCs of P1, P2 (two independent samples analyzed, collected at the ages of 12 and
18 months), P3, P4, and healthy controls matched
for age. Cryopreserved PBMCs in R10 medium
(RPMI 1640, 10% FBS, 2 mM L-glutamine,
100 U/ml of penicillin, and 100 mg/ml of streptomycin) were thawed and immediately centrifuged to obtain a cell pellet. The cells were
then incubated for 15 min at 37°C in an incubator containing 5% CO2 in the presence of
Benzonase (Millipore Sigma, cat. 70664) diluted
1:1000 in R10 medium. The cells were then
washed once in R10 and once in fluorescenceactivated cell sorting (FACS) buffer (2% FBS in
PBS). For staining, cells were resuspended in
50 ml of a mixture of LIVE/DEAD Fixable Blue
Dead Cell Stain (cat. L34962) diluted 1:200 in
PBS and anti-CCR7 APC-Cy7 antibody (Biolegend,
cat. 353212) and incubated for 10 min at 37°C in
an incubator containing 5% CO2. Cells were
then labeled with 1 ml of oligonucleotide-linked
hashing antibody (Totalseq-C, Biolegend) and
stained by incubation with 50 ml of antibody
cocktail diluted in Brilliant Stain Buffer (BD
Biosciences, cat. 566349) for 20 min at room temperature. The antibody cocktail contained the following antibodies: anti-CD14 BV510 (Biolegend,
cat. 301842), anti-CD19 BV510 (Biolegend, cat.
302242), anti-CD56 BV510 (Biolegend, cat. 362534),
anti-CD8 BV785 (Biolegend, cat. 301046), antiCD4 PECy7 (Biolegend, cat. 300512), anti-CD95
AlexaFluor 700 (Biolegend, cat. 305648), antiPD-1 BV750 (Biolegend, cat. 329966), anti-CD69
FITC (Biolegend, cat. 310904), anti-CD40L BV421
(Biolegend, cat. 310824), anti-CD3 BUV805 (BD
Biosciences, cat. 741999), anti-CD45RA PECF594
(BD Biosciences, cat. 562298), anti-CD25 BUV661
(BD Biosciences, cat. 741685), anti-CXCR3 BV711
(BD Biosciences, cat. 563156), anti-HLA-DR PECy5.5
(Invitrogen, cat. MHLDR18), and anti-CXCR5
APC (Invitrogen, cat. 17-9185-42). The cells were
washed twice with FACS buffer and resuspended
in R10 for sorting. We sorted 12,000 T cells
(CD3+CD14−CD19−CD56−Live/Dead−) from each
sample with a BD FACSymphony S6 Cell Sorter instrument (BD Biosciences) running BD

g

Materna et al., Science 383, eadh4059 (2024)

Luciferase reporter assays for autoantibodies
against IFNs

FACSDiva Software version 9.5.1 (BD Biosciences). Sorted cells were pooled four by four,
and each pool was loaded in a different lane of
the 10x Genomics Chromium Chip for sequencing. For the sequencing of single-cell V(D)J repertoires for sorted T cells, the cell suspension
was loaded on the 10x Genomics Chromium
Instrument according to the manufacturer’s
protocol for the Next GEM Single-Cell 5′ Kit
v1.1 (10x Genomics PN-1000165) to generate gel
bead-in-emulsions and for GEM-RT and the
amplification of total cDNA. After purification
with SPRIselect beads (Beckman Coulter), specific TCR targets were amplified from the cDNA
with the PTCR1 primer (5′-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTC-3′) and constant region primers: TRB (5′TGCTTCTGATGGCTCAAACACAGCGACCT-3′), TRA
(5′-TCTCAGCTGGTACACGGCAGGGTCAGGGT3′), TRG (5′-GAAGGAAGAAAAATAGTGGGCTTGGGGGAAAC-3′), or TRD (5′-CACCAGACAAGCGACATTTGTTCC-3′) with a barcode and
the P7 sequence added to the constant region
primers. The Illumina-ready libraries were sequenced by paired-end MiSeq with 2×300 base
pair reads to obtain VDJ sequences.

p

CyTOF was performed with various strategies.
One involved the use of whole blood in the
Maxpar Direct Immune Profiling Assay, 30
Markers (Standard Biotools, ref: 201334), according to the manufacturer’s instructions. All the
samples for whole-blood staining were processed
within 24 hours of collection. P10, P12, and P13
were phenotyped by the same protocol but with
a customized antibody panel (table S11). We
investigated the T cell subsets, including IEL
markers, with another CyTOF staining panel
for cryopreserved samples (IEL panel, table S12).
PBMCs were thawed and 4×106 cells were
immediately stained according to the Standard
Biotools protocol. The antibodies against TCR
Vd1 and TCR Vd2 were added after 10 min of
staining with the other antibodies to prevent
interference with the binding of the TCRgd
antibody. For both whole blood and IEL panels,
cells were frozen at −80°C after iridium staining and stored at the same temperature until
acquisition on a Helios machine (Standard
Biotools). In addition to whole-blood immunophenotyping, we also performed immunophenotyping on cryopreserved PBMCs for some
patients (table S13). Single-cell suspensions were
centrifuged to obtain a cell pellet, which was
then incubated with 20 mM lanthanum-139
(Trace Sciences)–loaded maleimido-mono-amineDOTA (Macrocyclics) in PBS for 10 min at room
temperature for live-dead discrimination (LD).
Cells were washed in staining buffer and resuspended in surface antibody cocktail, incubated
for 30 min at room temperature, washed twice in
staining buffer, fixed, permeabilized with the
FoxP3 staining buffer set (eBioscience), and
subjected to intracellular staining for 60 min
at room temperature. Cells were washed twice
and then fixed by overnight incubation in 1.6%
PFA (Electron Microscopy Sciences) solution
supplemented with 125 nM iridium at 4°C.
Before data acquisition on a CyTOF Helios
flow cytometer (Standard Biotools), cells were
washed twice in PBS and once in dH2O. Custom conjugation to isotope-loaded polymers
was performed with the MAXPAR kit (Stan-

dard Biotools). The data were analyzed with
OMIQ software.

RES EARCH | R E S E A R C H A R T I C L E

TREC levels

Statistics

Analyses were performed with GraphPad Prism
V10.1.1 software. Two-tailed Mann–Whitney
tests or unpaired t tests were used for single
comparisons of independent groups. In the
corresponding figures, n.s. indicates not significant, ****P < 0.0001, ***P < 0.001, **P <
0.01, and *P < 0.05.
REFERENCES AND NOTES

1. M. A. Yui, E. V. Rothenberg, Developmental gene networks: A
triathlon on the course to T cell identity. Nat. Rev. Immunol. 14,
529–545 (2014). doi: 10.1038/nri3702; pmid: 25060579
2. J. P. Rast et al., a, b, g, and d T cell antigen receptor genes
arose early in vertebrate phylogeny. Immunity 6, 1–11 (1997).
doi: 10.1016/S1074-7613(00)80237-X; pmid: 9052832

16 of 18

,

sjTRECs were quantified by nested qPCR, with
the primers and standard curve plasmid described by Dion et al. (64). The qPCR protocol
was adapted as previously described (65)
using ~500 ng of purified gDNA for each
quantification.

y

1 March 2024

T cell clones were obtained bIting naive T cells
(CD3+CD45RA+CCR7+TCRab+TCRgd−) with a
BD FACSAria III SORP cell sorter (Becton
Dickinson, San Jose, CA) and DIVA 9.1 software.
Cells were sorted, one cell per well, in 96-well
plates containing 50 ml of ImmunoCult-XF
T Cell Expansion Medium (StemCell Technologies, REF no. 10981) supplemented with IL-2
(1 ng/ml) and ImmunoCult Human CD3/
CD28/CD2 T cell Activator (StemCell Technologies, REF no. 10990,1:40) per well. Every
2 days, fresh medium with IL-2 (1 ng/ml) was
added to the cells. Clones were visible under a
microscope 1 week after sorting. Clones were
reactivated every 3 weeks with ImmunoCult
Human CD3/CD28/CD2 T cell Activator (StemCell
Technologies, no. 10990,1:80). DNA was extracted
from clones with the DNeasy Blood & Tissue
Kit (no. 69504 ; Qiagen). The TRG gene repertoire was investigated by next-generation sequencing (NGS). For library preparation, PCR was
performed on 100 ng of genomic DNA with a
published protocol (63), but with adaptation
of the primers for a NGS version of the assay
(table S16). Dual barcoding of the primers
made the simultaneous multiplexing of samples possible. After library purification, sequencing was performed on an Illumina MiSeq
platform. Sequencing data analysis, including
demultiplexing, quality control and clonotype
assignment, was performed with the Vidjil
pipeline (https://www.vidjil.org). IMGT V-QUEST
(https://www.imgt.org/IMGT_vquest/analysis)
was used for further TRGV and TRGJ annotation
and for CDR3 characterization.

y g

Materna et al., Science 383, eadh4059 (2024)

PBMCs were stained with antibodies against
CD3 (no. 565491, BD, 1:50), CD45RA (no. 130-092247, Miltenyi Biotec, 2:50), CCR7 (no. 130-120600, Miltenyi Biotec, 2:50), TCRgd (no. 331218,
Biolegend, 1:50), or TCRab (no. 555548, BD,
2:50) and incubated with the Aqua Live/Dead
Cell Stain Kit (Thermo Fisher Scientific) for
30 min at room temperatINaive and memory
ab and gd T cells were sorted with a FACSAria
cell sorter (Becton Dickinson, San Jose, CA) on
the basis of CD45RA and CCR7 expression.
DNA extraction was performed with the DNeasy
Blood & Tissue Kit (no. 69504; Qiagen). The
rearranged TRAD, TRB, and TRG genomic loci
were sequenced by Adaptive Biotechnologies
(Seattle, WA) as a commercial service. The
data were then analyzed with ImmunoSeq
online tools (Adaptive Biotechnologies) and
custom R scripts. The frequencies of productive and nonproductive TRD, TRG, TRB, and
TRA rearrangements were analyzed for both
unique and total TRD, TRG, TRB, or TRA sequences obtained from the sorted ab and gd
T cell subsets. The frequency distributions
for individual clonotypes (including TRBV-toTRBJ pairing and TRAV-to-TRAJ pairing) were
analyzed within unique sequences. Diversity
indices were calculated and heat-map representations of the frequencies of individual
TRAV/TRDV to TRAJ/TRD gene pairs and
TRBV-to-TRBJ gene pairs were produced with
R software version 4.2.0 (2022-04-22 ucrt) and

HTS of the human TRG locus from the gDNA of
clonally expanded T cells

y

Mice were bred under specific pathogen-free
conditions in CIPHE animal facilities (agreement
number: B1301407) and handled in accordance
with institutional committee and European
guidelines for animal care. C57BL/6 mice
were purchased from Janvier Laboratories.
Ptcratm1(icre)Hjf KO mice have been described
elsewhere (62) and were rederived from the
INFRAFRONTIER/EMMA archive (EM:08347).
Multiparametric immunophenotyping was performed at the CIPHE-PHENOMIN (INSERM,
US012) flow cytometry facility. Peripheral blood
(PB) was collected by submandibular puncture
into Microvette 500 K3 EDTA tubes (Sarstedt).
Hematological analysis was performed on a
Procyte Dx (IDDEX) machine, in accordance
with the manufacturer’s recommendations.
Peripheral blood leukocytes were analyzed with
a Lyse No Wash protocol and 1X FACS Lysing
Solution (BD Biosciences). Leukocytes from the
spleen and thymus were extracted according to
the protocol of the International Mouse Phenotyping Consortium (IMPC_IMM_002). Red blood
cells were not lysed for thymic leukocyte preparations. LN T cells were isolated from pooled
inguinal, brachial, axillary, and submandibular
LNs. Briefly, organs were disrupted with the
OctoGentleMACS system (Miltenyi Biotec),
using 600 Mandl units of collagenase D (Roche
Life Science) and 30 mg of DNAse I (Sigma), for

High-throughput sequencing (HTS) of the
human TCR repertoire from the gDNA
of sorted I and memory T cells

the R packages Tidyverse (1.3.2) and Immunarch (0.6.9).

g

Mouse experiments

20 min at room temperature. The cell suspension was filtered, and the cells were counted.
Red blood cells were lysed by incubation for
1 min at room temperature with ammoniumchloride-potassium (ACK) lysis solution (eBioscience). Before staining, the cells were incubated
for 10 min on ice with an anti-CD16/32 (2.4G2)
antibody to block Fc receptors. In all experiments,
4′,6-diamidino-2-phenylindole (DAPI) (Invitrogen)
staining was used to exclude dead cells from
the analysis. Multiparameter FACS acquisition
was performed on a Fortessa LSRII SORP or
Canto 10C system (BD Biosciences). The analysis
was performed with FACSDiva 9.01 (BD Biosciences) software. Doublets were systematically
excluded on the basis of side scatter (SSC) and
forward scatter (FSC) parameters. The antibodies used for immunophenotyping are listed
in table S15. The thymocyte subsets were defined
as ETP (CD4−CD8a−CD3e−CD44+CD25−ckit+),
TN2 (CD4−CD8a−CD3e−CD44+CD25+ckit+), TN3
(CD4−CD8a−CD3e−CD44−CD25+gd−), TN4 (CD4−
CD8a−CD3e−CD44−CD25−gd−), ISP (CD4−CD8a+
CD3e−CD44−CD25−gd−), iDP (CD4+CD8a+CD3e−
CD44−CD25−gd−), mDP (CD4+CD8a+CD3e+
CD44−CD25−gd−), SP4 (CD4+CD8a−CD3e+CD44−
gd−), SP8 (CD4−CD8a+CD3e+CD44−gd−), gd TN3
(CD3e+CD25+gd+), or gd (CD3e+CD25−gd+).

p

a DLL4-expressing stromal cell line (MS5-hDLL4)
in the ATO system, as previously described (61),
but with minor modifications. Briefly, CD34+
peripheral blood cells from five normal donors,
three patients with partial pre-TCRa deficiency
(P11 to P13), and three patients with complete
pre-TCRa deficiency (P1, P5, and P6) were
positively selected with the CD34 MicroBead
UltraPure kit (Miltenyi Biotec) on an AutoMACS
Pro Separator. We mixed 1000 to 1500 CD34+
cells with 150,000 MS5-hDLL4 cells per ATO.
Each ATO (5 ml) was then plated in a 0.4 mM
Millicell Transwell insert and placed in one
well of a six-well plate in 1 ml complete RB27
medium supplemented with rhIL-7 (5 ng/ml),
rhFlt3-L (5 ng/ml), and 30 mM L-ascorbic acid
2-phosphate sesquimagnesium salt hydrate.
Each insert contained a maximum of two ATOs.
For the first 3 weeks of culture, the medium was
also supplemented with 10 ng/ml of rhSCF.
After 5 weeks in culture, MACS buffer (PBS
supplemented with 0.5% BSA and 2 mM EDTA)
was added to each well and ATOs were dissociated by manual pipetting. The cells were then
collected into a pellet by centrifugation, resuspended in FACS buffer (2% FBS in PBS),
counted, and stained with the antibody cocktail
described in table S14. Events were acquired on
a BD LSR II Fortessa flow cytometer (BD Biosciences, San Jose, CA) and analyzed with
FlowJo software version 10.6.1 (FlowJo, LLC,
Ashland, OR).

RES EARCH | R E S E A R C H A R T I C L E

y g
y
,

17 of 18

y

1 March 2024

44. C. Bycroft et al., The UK Biobank resource with deep
phenotyping and genomic data. Nature 562, 203–209 (2018).
doi: 10.1038/s41586-018-0579-z; pmid: 30305743
45. T. Khan et al., Human leukocyte antigen class II gene diversity
tunes antibody repertoires to common pathogens. Front.
Immunol. 13, 856497 (2022). doi: 10.3389/
fimmu.2022.856497; pmid: 36003377
46. P. Hebbar et al., Genome-wide landscape establishes novel
association signals for metabolic traits in the Arab
population. Hum. Genet. 140, 505–528 (2021). doi: 10.1007/
s00439-020-02222-7; pmid: 32902719
47. E. Genin, A. Tullio-Pelet, F. Begeot, S. Lyonnet, L. Abel,
Estimating the age of rare disease mutations: The example of
Triple-A syndrome. J. Med. Genet. 41, 445–449 (2004).
doi: 10.1136/jmg.2003.017962; pmid: 15173230
48. R. J. Wang, S. I. Al-Saffar, J. Rogers, M. W. Hahn, Human
generation times across the past 250,000 years. Sci. Adv. 9,
eabm7047 (2023). doi: 10.1126/sciadv.abm7047;
pmid: 36608127
49. D. Kim, J. M. Paggi, C. Park, C. Bennett, S. L. Salzberg,
Graph-based genome alignment and genotyping with HISAT2
and HISAT-genotype. Nat. Biotechnol. 37, 907–915 (2019).
doi: 10.1038/s41587-019-0201-4; pmid: 31375807
50. C. Trapnell et al., Differential gene and transcript expression
analysis of RNA-seq experiments with TopHat and Cufflinks.
Nat. Protoc. 7, 562–578 (2012). doi: 10.1038/nprot.2012.016;
pmid: 22383036
51. N. L. Bray, H. Pimentel, P. Melsted, L. Pachter, Near-optimal
probabilistic RNA-seq quantification. Nat. Biotechnol. 34,
525–527 (2016). doi: 10.1038/nbt.3519; pmid: 27043002
52. P. Danecek et al., Twelve years of SAMtools and BCFtools.
Gigascience 10, giab008 (2021). doi: 10.1093/gigascience/
giab008; pmid: 33590861
53. H. Thorvaldsdóttir, J. T. Robinson, J. P. Mesirov, Integrative
Genomics Viewer (IGV): High-performance genomics data
visualization and exploration. Brief. Bioinform. 14, 178–192
(2013). doi: 10.1093/bib/bbs017; pmid: 22517427
54. K.-T. Tan et al., Profiling the B/T cell receptor repertoire of
lymphocyte derived cell lines. BMC Cancer 18, 940 (2018).
doi: 10.1186/s12885-018-4840-5; pmid: 30285677
55. J. M. Heather et al., Stitchr: Stitching coding TCR
nucleotide sequences from V/J/CDR3 information. Nucleic
Acids Res. 50, e68 (2022). doi: 10.1093/nar/gkac190;
pmid: 35325179
56. S. A. Stewart et al., Lentivirus-delivered stable gene silencing
by RNAi in primary cells. RNA 9, 493–501 (2003). doi: 10.1261/
rna.2192803; pmid: 12649500
57. P. E. Conrey et al., IgA deficiency destabilizes homeostasis
toward intestinal microbes and increases systemic immune
dysregulation. Sci. Immunol. 8, eade2335 (2023). doi: 10.1126/
sciimmunol.ade2335; pmid: 37235682
58. P. Bastard et al., Autoantibodies neutralizing type I IFNs are
present in ~4% of uninfected individuals over 70 years old and
account for ~20% of COVID-19 deaths. Sci. Immunol. 6,
eabl4340 (2021). doi: 10.1126/sciimmunol.abl4340;
pmid: 34413139
59. C. E. Shannon, The mathematical theory of communication.
1963. MD Comput. 14, 306–317 (1997). pmid: 9230594
60. V. Béziat et al., Humans with inherited T cell CD28 deficiency
are susceptible to skin papillomaviruses but are otherwise
healthy. Cell 184, 3812–3828.e30 (2021). doi: 10.1016/
j.cell.2021.06.004; pmid: 34214472
61. C. S. Seet et al., Generation of mature T cells from human
hematopoietic stem and progenitor cells in artificial thymic
organoids. Nat. Methods 14, 521–530 (2017). doi: 10.1038/
nmeth.4237; pmid: 28369043
62. H. Luche et al., In vivo fate mapping identifies pre-TCRa
expression as an intra- and extrathymic, but not prethymic,
marker of T lymphopoiesis. J. Exp. Med. 210, 699–714 (2013).
doi: 10.1084/jem.20122609; pmid: 23509324
63. M. Armand et al., A New and Simple TRG Multiplex PCR
Assay for Assessment of T-cell Clonality: A Comparative
Study from the EuroClonality Consortium. HemaSphere 3, e255
(2019). doi: 10.1097/HS9.0000000000000255;
pmid: 31723840
64. M.-L. Dion, R.-P. Sékaly, R. Cheynier, in Immunological
Tolerance: Methods and Protocols, P. J. Fairchild, Ed., vol. 380
of Methods in Molecular Biology (Humana Press, 2007),
pp. 197–213. doi: 10.1007/978-1-59745-395-0_12
65. H. M. Roux et al., Genetically determined thymic function
affects strength and duration of immune response in COVID
patients with pneumonia. Sci. Adv. 9, eadh7969 (2023).
doi: 10.1126/sciadv.adh7969; pmid: 37738336

g

Materna et al., Science 383, eadh4059 (2024)

24. T. Le Voyer et al., Autoantibodies against type I IFNs in humans
with alternative NF-kB pathway deficiency. Nature 623,
803–813 (2023). doi: 10.1038/s41586-023-06717-x;
pmid: 37938781
25. A. R. Ramiro et al., Differential developmental regulation and
functional effects on pre-TCR surface expression of human
pTaa and pTab spliced isoforms. J. Immunol. 167, 5106–5114
(2001). doi: 10.4049/jimmunol.167.9.5106; pmid: 11673521
26. J. Arnaud et al., The interchain disulfide bond between TCR
alpha beta heterodimers on human T cells is not required for
TCR-CD3 membrane expression and signal transduction.
Int. Immunol. 9, 615–626 (1997). doi: 10.1093/intimm/
9.4.615; pmid: 9138023
27. S. S. Pang et al., The structural basis for autonomous
dimerization of the pre-T-cell antigen receptor. Nature 467,
844–848 (2010). doi: 10.1038/nature09448; pmid: 20944746
28. S. Yamasaki et al., Mechanistic basis of pre–T cell receptor–
mediated autonomous signaling critical for thymocyte
development. Nat. Immunol. 7, 67–75 (2006). doi: 10.1038/
ni1290; pmid: 16327787
29. E. Ishikawa, Y. Miyake, H. Hara, T. Saito, S. Yamasaki,
Germ-line elimination of electric charge on pre–T-cell receptor
(TCR) impairs autonomous signaling for b-selection and TCR
repertoire formation. Proc. Natl. Acad. Sci. U.S.A. 107,
19979–19984 (2010). doi: 10.1073/pnas.1011228107;
pmid: 21030681
30. J. Buer, I. Aifantis, J. P. DiSanto, H. J. Fehling, H. von Boehmer,
T-cell development in the absence of the pre-T-cell receptor.
Immunol. Lett. 57, 5–8 (1997). doi: 10.1016/S0165-2478(97)
00078-3; pmid: 9232417
31. K. Terrence, C. P. Pavlovich, E. O. Matechak, B. J. Fowlkes,
Premature expression of T cell receptor (TCR)ab suppresses
TCR gd gene rearrangement but permits development of gd
lineage T cells. J. Exp. Med. 192, 537–548 (2000).
doi: 10.1084/jem.192.4.537; pmid: 10952723
32. L. Bruno, H. J. Fehling, H. von Boehmer, The ab T cell receptor
can replace the gd receptor in the development of gd lineage
cells. Immunity 5, 343–352 (1996). doi: 10.1016/S1074-7613
(00)80260-5; pmid: 8885867
33. I. Aifantis et al., The Ed enhancer controls the generation of
CD4−CD8− abTCR-expressing T cells that can give rise to
different lineages of ab T cells. J. Exp. Med. 203, 1543–1550
(2006). doi: 10.1084/jem.20051711; pmid: 16754716
34. Z. M. Ndhlovu et al., Magnitude and Kinetics of CD8+ T Cell
Activation during Hyperacute HIV Infection Impact Viral Set
Point. Immunity 43, 591–604 (2015). doi: 10.1016/j.
immuni.2015.08.012; pmid: 26362266
35. M. Bosticardo et al., Artificial thymic organoids represent a
reliable tool to study T-cell differentiation in patients with
severe T-cell lymphopenia. Blood Adv. 4, 2611–2616 (2020).
doi: 10.1182/bloodadvances.2020001730; pmid: 32556283
36. F. Hochstenbach, M. B. Brenner, T-cell receptor d-chain can
substitute for a to form a bd heterodimer. Nature 340,
562–565 (1989). doi: 10.1038/340562a0; pmid: 2528071
37. P. Del Porto, L. Bruno, M. G. Mattei, H. von Boehmer,
C. Saint-Ruf, Cloning and comparative analysis of the human
pre-T-cell receptor alpha-chain gene. Proc. Natl. Acad. Sci. U.S.A.
92, 12105–12109 (1995). doi: 10.1073/pnas.92.26.12105; pmid:
8618853
38. C. Saint-Ruf et al., Analysis and expression of a cloned pre-T
cell receptor gene. Science 266, 1208–1212 (1994).
doi: 10.1126/science.7973703; pmid: 7973703
39. P. Gros, J.-L. Casanova, Reconciling Mouse and Human
Immunology at the Altar of Genetics. Annu. Rev. Immunol. 41,
39–71 (2023). doi: 10.1146/annurev-immunol-101721-065201;
pmid: 36525691
40. K. Medetgul-Ernar, M. M. Davis, Standing on the shoulders of
mice. Immunity 55, 1343–1353 (2022). doi: 10.1016/
j.immuni.2022.07.008; pmid: 35947979
41. S. J. C. Mancini et al., TCRA gene rearrangement in immature
thymocytes in absence of CD3, pre-TCR, and TCR signaling.
J. Immunol. 167, 4485–4493 (2001). doi: 10.4049/
jimmunol.167.8.4485; pmid: 11591775
42. C. Aude-Garcia, M. Gallagher, P. N. Marche, E. Jouvin-Marche,
Preferential ADV-AJ association during recombination in the
mouse T-cell receptor alpha/delta locus. Immunogenetics 52,
224–230 (2001). doi: 10.1007/s002510000266;
pmid: 11220624
43. N. Pasqual et al., Quantitative and qualitative changes in V-J a
rearrangements during mouse thymocytes differentiation:
Implication for a limited T cell receptor a chain repertoire. J.
Exp. Med. 196, 1163–1173 (2002). doi: 10.1084/jem.20021074;
pmid: 12417627

p

3. M. S. Krangel, Mechanics of T cell receptor gene
rearrangement. Curr. Opin. Immunol. 21, 133–139 (2009).
doi: 10.1016/j.coi.2009.03.009; pmid: 19362456
4. T. Kreslavsky, M. Gleimer, H. von Boehmer, ab versus gd
lineage choice at the first TCR-controlled checkpoint. Curr.
Opin. Immunol. 22, 185–192 (2010). doi: 10.1016/
j.coi.2009.12.006; pmid: 20074925
5. M. Malissen et al., A T cell clone expresses two T cell receptor
a genes but uses one ab heterodimer for allorecognition and
self MHC-restricted antigen recognition. Cell 55, 49–59 (1988).
doi: 10.1016/0092-8674(88)90008-6; pmid: 3262424
6. J. L. Casanova, P. Romero, C. Widmann, P. Kourilsky,
J. L. Maryanski, T cell receptor genes in a series of class I
major histocompatibility complex-restricted cytotoxic
T lymphocyte clones specific for a Plasmodium berghei
nonapeptide: Implications for T cell allelic exclusion and
antigen-specific repertoire. J. Exp. Med. 174, 1371–1383 (1991).
doi: 10.1084/jem.174.6.1371; pmid: 1836010
7. P. Borgulya, H. Kishi, Y. Uematsu, H. von Boehmer, Exclusion
and inclusion of a and b T cell receptor alleles. Cell 69,
529–537 (1992). doi: 10.1016/0092-8674(92)90453-J;
pmid: 1316241
8. B. L. Brady, N. C. Steinel, C. H. Bassing, Antigen receptor allelic
exclusion: An update and reappraisal. J. Immunol. 185,
3801–3808 (2010). doi: 10.4049/jimmunol.1001158;
pmid: 20858891
9. H. von Boehmer, F. Melchers, Checkpoints in lymphocyte
development and autoimmune disease. Nat. Immunol. 11,
14–20 (2010). doi: 10.1038/ni.1794; pmid: 20016505
10. Y. Xing, K. A. Hogquist, T-cell tolerance: Central and peripheral.
Cold Spring Harb. Perspect. Biol. 4, a006957 (2012).
doi: 10.1101/cshperspect.a006957; pmid: 22661634
11. L. Klein, B. Kyewski, P. M. Allen, K. A. Hogquist, Positive and
negative selection of the T cell repertoire: What thymocytes
see (and don’t see). Nat. Rev. Immunol. 14, 377–391 (2014).
doi: 10.1038/nri3667; pmid: 24830344
12. H. J. Fehling, A. Krotkova, C. Saint-Ruf, H. von Boehmer,
Crucial role of the pre-T-cell receptor a gene in development of
ab but not gd T cells. Nature 375, 795–798 (1995).
doi: 10.1038/375795a0; pmid: 7596413
13. S. Mancini et al., TCR a-chain repertoire in pTa-deficient mice
is diverse and developmentally regulated: Implications for preTCR functions and TCRA gene rearrangement. J. Immunol. 163,
6053–6059 (1999). doi: 10.4049/jimmunol.163.11.6053;
pmid: 10570293
14. J. Buer, I. Aifantis, J. P. DiSanto, H. J. Fehling, H. von Boehmer,
Role of different T cell receptors in the development of pre–T
cells. J. Exp. Med. 185, 1541–1548 (1997). doi: 10.1084/
jem.185.9.1541; pmid: 9151891
15. C. Saint-Ruf, O. Lechner, J. Feinberg, H. von Boehmer, Genomic
structure of the human pre-T cell receptor a chain and
expression of two mRNA isoforms. Eur. J. Immunol. 28,
3824–3831 (1998). doi: 10.1002/(SICI)1521-4141(199811)
28:11<3824::AID-IMMU3824>3.0.CO;2-9; pmid: 9842925
16. V. Sun et al., The Metabolic Landscape of Thymic T Cell
Development In Vivo and In Vitro. Front. Immunol. 12, 716661
(2021). doi: 10.3389/fimmu.2021.716661; pmid: 34394122
17. Z. Ren et al., ATAV: A comprehensive platform for populationscale genomic analyses. BMC Bioinformatics 22, 149 (2021).
doi: 10.1186/s12859-021-04071-1; pmid: 33757430
18. E. M. Scott et al., Characterization of Greater Middle Eastern
genetic variation for enhanced disease gene discovery. Nat.
Genet. 48, 1071–1076 (2016). doi: 10.1038/ng.3592;
pmid: 27428751
19. K. J. Karczewski et al., The mutational constraint spectrum
quantified from variation in 141,456 humans. Nature 581,
434–443 (2020). doi: 10.1038/s41586-020-2308-7;
pmid: 32461654
20. M. S. Anderson et al., Projection of an immunological self
shadow within the thymus by the aire protein. Science 298,
1395–1401 (2002). doi: 10.1126/science.1075958;
pmid: 12376594
21. J. E. Walter et al., Broad-spectrum antibodies against selfantigens and cytokines in RAG deficiency. J. Clin. Invest. 125,
4135–4148 (2015). doi: 10.1172/JCI80477; pmid: 26457731
22. J. M. Rosenberg et al., Neutralizing Anti-Cytokine
Autoantibodies Against Interferon-a in Immunodysregulation
Polyendocrinopathy Enteropathy X-Linked. Front. Immunol. 9,
544 (2018). doi: 10.3389/fimmu.2018.00544; pmid: 29651287
23. P. Bastard et al., Preexisting autoantibodies to type I IFNs
underlie critical COVID-19 pneumonia in patients with APS-1.
J. Exp. Med. 218, e20210554 (2021). doi: 10.1084/
jem.20210554; pmid: 33890986

RES EARCH | R E S E A R C H A R T I C L E

66. M. Materna et al., Data from: The immunopathological
landscape of human pre-TCRa deficiency: From rare to
common variants, Dryad (2024); https://doi.org/10.5061/
dryad.9zw3r22m8.
67. M. Materna et al., Data from: The immunopathological landscape
of human pre-TCRa deficiency: From rare to common variants,
Adaptive Biotechnologies (2024); https://doi.org/10.21417/
MM2024S.
ACKN OW LEDG MEN TS

M.At., F.A.A., C.D., A.P., J.M., T.L.V., P.B., L.B., M.D., N.Vl., P.P.C.,
E.J., M.Al., A.H., T.A.T., M.A.A., M.L.-V., F.R., E.H., J.R.R., M.L.T., J.R.K.,
H.R.-Z., M.C.-A., S.M.A., M.Maz., A.Co., I.M., L.A., B.M., F.A.-M., F.S.A.,
C.Be., F.D., D.C.D., R.C., D.L., N.L., N.M., T.Mo., M.Sh., R.S., S.E.H., H.L.,
L.D.N., J.-L.C., and V.B. Competing interests: I.M. received consultancy
fees from Boehringer-Ingelheim and a research grant from CSL Behring,
outside this work and paid to KU Leuven. S.H. declares that she was
on the ad hoc advisory board for Horizon Therapeutics, without relation
to this work. The other authors have no competing interests to declare.
Data and materials availability: The materials and reagents used are
commercially available and nonproprietary. All raw and processed data
and biological materials, including immortalized cell lines from patients,
are available from the corresponding author through a material transfer
agreement with INSERM. The RNA-seq data for sorted primary human
thymic T cell subsets have already been published in the BioProject
repository under the accession number PRJNA741323 (16). Single-cell
RNA-seq data are available in the MIAME compliant gene expression
omnibus database (GEO: GSE243927). Raw data for the immunoblots
and qPCR are available from Dryad (66). The entire TCR sequencing
dataset is accessible through the Adaptive Biotechnologies website (67).
License information: Copyright © 2024 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse. This
research was funded in part by the French National Research Agency
(ANR) (ANR-10-IAHU-01, ANR-10-LABX-62-IBEID, and ANR-21-CE150034) and the Horizon Europe Framework Programme (HORIZON)
(01057100; UNDINE), cOAlition S organizations. This article is also
subject to HHMI’s Open Access to Publications policy. HHMI lab heads
have previously granted a nonexclusive CC BY 4.0 license to the
public and a sublicensable license to HHMI in their research articles.
Pursuant to those licenses, the Author Accepted Manuscript (AAM) of
this article can be made freely available under a CC BY 4.0 license
immediately upon publication.

p
g

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adh4059
Supplementary Text
Figs. S1 to S8
Tables S1 to S16
References (68–86)
MDAR Reproducibility Checklist
Data S1

y

We thank the patients and their families for participating in the
study. We thank the members of the HGID laboratory for providing
excellent comments on the paper. We also wish to thank L. Hadjem
from the CIPHE facility and the members of the CYPS mass
cytometry core facility team (Pitié Salpêtrière Hospital) for
providing outstanding technical assistance. We thank A. Liston,
S. Humblet-Baron, and M. Willemsen (Laboratory of Adaptive
Immunology, KU Leuven). We thank the National Facility for
Autoimmunity and Serology Profiling at SciLifeLab for their
excellent technical support with the protein microarray studies. We
thank Qatar Genome and the Qatar Biobank (QBB) management
and staff for allowing us to access and analyze QBB/QGP samples
and data and the Integrated Genomics Services team of Sidra
Medicine for generating and processing WGS data for QBB study
participants. We also thank S. Elledge (Brigham and Women’s
Hospital and Harvard Medical School) for providing the VirScan
phage library used in this study. This research was performed with
the UK Biobank resource under application no. 40436. Funding:
This study was supported in part by a grant from the St. Giles
Foundation; the Rockefeller University; Institut National de la Santé
et de la Recherche Médicale (INSERM); Paris Cité University;
the National Center for Research Resources; the National Center
for Advancing Sciences of the National Institutes of Health
(UL1TR001866); the French National Research Agency (ANR)
under the “Investments for the Future” program (ANR-10-IAHU01); the Integrative Biology of Emerging Infectious Diseases
Laboratory of Excellence (ANR-10-LABX-62-IBEID); ANR CARMIL2
(ANR-21-CE15-0034); the ANR-RHU program (ANR-21-RHUS08-COVIFERON); the HORIZON-HLTH-2021-DISEASE-04 program
under grant agreement 01057100 (UNDINE); the French
Foundation for Medical Research (EQU201903007798); ITMO
Cancer of Aviesan and INCa within the framework of the 2021–
2030 Cancer Control Strategy (funds administered by Institut
National de la Santé et de la Recherche Médicale); the Square
Foundation; W. E. Ford, General Atlantic’s Chairman and Chief
Executive Officer; G. Caillaux, General Atlantic’s Co-President,

Managing Director and Head of business in EMEA, and the General
Atlantic Foundation; the Qatar National Research Fund (PPM11220-150017); Sidra Medicine (SDR400048); the SCOR Corporate
Foundation for Science; Institut National de la Santé et de la
Recherche Médicale; and the University of Paris. Open Access
funding was provided by Rockefeller University. D.L. was supported
by a Fonds de Recherche du Québec – Santé Chercheur-Boursier
Junior 1 award. T.L.V. and P.B. were supported by the BettencourtSchueller Foundation and the MD-PhD program of the Imagine
Institute. A.H., T.A.T., and F.A.-M. are supported by institutional
funding from the Kuwait Foundation for the Advancement of
Sciences. M.Mo. is supported by the ANRS. P.B. was supported by
the FRM (EA20170638020). The work by A.Ce., D.E., and N.L. was
funded by grants from the Swedish Research Council (2021-03118)
and the Göran Gustafsson Foundation (2141 and 2227) to N.L.
L.D.N. is supported by the Division of Intramural Research, National
Institute of Allergy and Infectious Diseases, NIH, Bethesda, MD,
USA (grant ZIA AI001222). F.S. (11B5520N, fellow), V.S.
(1804523N), I.M. (1805821N), and R.S. (1805518N, 1805523N,
senior clinical investigator fellow) are supported by the Fonds
Wetenschappelijk Onderzoek - Vlaanderen National Fund for
Scientific Research (FWO). E.H. holds the Bank of Montreal Chair
for Pediatric Immunology. F.A.Q. was supported by the Ibn Rushd
Fellowship Award – King Abdullah University of Science and
Technology and King Abdulaziz City for Science and Technology.
D.E. was supported by Clas Groschinsky Memorial Foundation
(M21116). This study was supported by the VIB Grand Challenge
program (translational science initiative on PID, GC01-C01 for
I.M. and R.S.). I.M. and R.S. are members of the European
Reference Network for Rare Immunodeficiency, Autoinflammatory
and Autoimmune Diseases (project ID no. 739543). I.M. is funded
by the FWO Vlaanderen G0B5120N and by C16/18/007 KU
Leuven and by the Jeffrey Modell Foundation and is a senior
clinical investigator at FWO Vlaanderen. S.E.H. is supported by a
K08AI135091 grant, the Burroughs Wellcome Fund, and the CHOP
Research Institute. Author contributions: Conceptualization:
S.E.H., H.L., L.D.N., J.-L.C., and V.B. Supervision: S.E.H., L.D.N., J.-L.C.,
and V.B. Writing – original draft: V.B. Funding acquisition: S.E.H.,
L.D.N., J.-L.C., and V.B. Resources: M.Mo., R.K., M.V., M.S.-S., S.J.,
L.B., F.A.Q., V.S., M.A.A., M.L.-V., F.R., J.R.K., M.Sa., S.N., M.To.,
N.Va., H.M., E.T., M.N., K.M., T.Mi., K.I., N.P., H.V.B., M.Sh., and R.S.
Methodology, data curation, visualization, software, validation,
formal analysis, investigation, and writing – review & editing: M.Mat.,
O.M.D., M.B., P.E.C., B.C.-D.M., C.Br., R.B., A.Ce., F.S., C.A.G.,
S.D., S.Sa., C.L.F., M.O., D.R., A.Gu., A.B., T.K., A.Ge., B.P., A.L.D.S.,
C.A.S., S.Sh., J.J.T.-T., F.P., K.A., J.S.C., N.S.L., D.E., R.L., Y.S.,

Submitted 1 March 2023; accepted 26 January 2024
10.1126/science.adh4059

y g
y
,

Materna et al., Science 383, eadh4059 (2024)

1 March 2024

18 of 18

RES EARCH

RESEARCH ARTICLE SUMMARY

◥

NEUROSCIENCE

SynGAP regulates synaptic plasticity and cognition
independently of its catalytic activity
Yoichi Araki†, Kacey E. Rajkovich†, Elizabeth E. Gerber†, Timothy R. Gamache†, Richard C. Johnson,
Thanh Hai N. Tran, Bian Liu, Qianwen Zhu, Ingie Hong, Alfredo Kirkwood, Richard Huganir*

INTRODUCTION: Experience-dependent changes

disability, characterized by intellectual disability, autistic-like features, and epilepsy.

▪

The list of author affiliations is available in the full article online.
*Corresponding author. Email: rhuganir@jhmi.edu
†These authors contributed equally to this work.
Cite this article as Y. Araki et al., Science 383, eadk1291 (2024).
DOI: 10.1126/science.adk1291

y g

SynGAP in neuronal cultures and replaced
it with wild-type and GAP mutant SynGAP
and found that mutation of the GAP domain
did not affect its ability to rescue LTP in neuronal cultures in vitro. We confirmed this in
vivo using mice containing inactivating GAP
mutations. These mice show normal viability,
LTP, and behaviors that are deficient in the
heterozygote Syngap1-knockout mice. We investigated how the structural properties of

SynGAP’s GAP activity is not required for synaptic plasticity and several cognitive behaviors. These data do not suggest that GAP
activity is unimportant, and further work with
these mice is needed to understand the role
of SynGAP GAP activity in brain function.
Finally, these results are relevant for developing treatments for SYNGAP1-related intellectual disability. Our findings suggest that
treatments that regulate Ras activity or its
downstream signaling will not be sufficient
as a therapy and that rescuing SYNGAP1 haploinsufficiency by increasing the expression of
the normal allele will be a more effective therapeutic approach.

y

RESULTS: We knocked down endogenous

CONCLUSION: These results indicate that

g

RATIONALE: SynGAP is one of the most abundant proteins at excitatory synapses, suggesting that it may play a structural role in the
PSD in addition to its role in regulating Ras
activity. SynGAP was recently found to have
unique structural properties and to undergo
liquid-liquid phase separation (LLPS) with
PSD95. Dispersion of SynGAP from the synapse during LTP induction would be predicted
to free up PSD95-binding sites, allowing other
PSD95-binding proteins to dynamically change
the composition of the synapse. To differentiate the role of GAP activity from its structural properties, we examined the function of
SynGAP with mutations that inactivate GAP
activity in vitro in neuronal cultures and in
vivo using knock-in mice containing inactivating GAP mutations.

p

in the strength of synaptic connections in
the brain are essential for neuronal development and for brain processes such as learning
and memory. Long-term potentiation (LTP) of
synapses is a key form of synaptic plasticity
that is widely recognized as a cellular model
for the study of memory. Many forms of synaptic plasticity, including LTP, are mediated
by long-lasting changes in the level of AMPA
receptors (AMPARs), the major neurotransmitter receptors at excitatory synapses. Excitatory synapses contain a complex structure
called the postsynaptic density (PSD), which
includes hundreds of proteins that orchestrate synaptic structure and function and dynamic changes during synaptic plasticity. One
of these is SynGAP, a RasGAP that binds to
the major synaptic scaffolding protein PSD95
and is highly abundant in the PSD in excitatory synapses. SynGAP is essential for normal
brain development and for LTP. During LTP
induction, SynGAP is phosphorylated, decreasing its affinity for PSD95, resulting in its dispersion from the synapse. This disinhibits Ras
activity and activates its downstream signaling processes, which were thought to be critical for synaptic potentiation. Heterozygote
Syngap1-knockout mice have deficits in synaptic plasticity, learning, and memory and exhibit seizures. De novo damaging SYNGAP1
mutations in humans result in haploinsufficiency and cause SYNGAP1-related intellectual

SynGAP could regulate AMPAR recruitment to
synapses and mediate synaptic potentiation.
Recent studies have shown that Transmembrane AMPAR Regulatory Proteins (TARPs),
essential components of the AMPAR protein
complex, also undergo LLPS with PSD95. A
simple hypothesis was that SynGAP directly
competes with the TARP-AMPAR complex,
and when SynGAP is dispersed from the synapse, tthis complex could replace it and be
recruited to the synapse. We tested whether
SynGAP competed with TARPs in forming
LLPS with PSD95 in vitro using purified proteins, heterologous cells, and neurons. We
found that SynGAP directly competed with
TARPs in forming LLPS with PSD95. This
competition with TARPs was not dependent
on GAP activity but required regions in the
C-terminal domain of SynGAP responsible for
LLPS with PSD95.

READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.adk1291

y

ILLUSTRATION: N. CARY/SCIENCE BASED ON BILL BLAKESLEY

,

AMPA receptor

TARP

PSD95 PSD95
SynGAP

PSD95

PSD95
SynGAP

PSD95

P

Basal

PSD95

SynGAP

PSD95

SynGAP
P

LTP induction

PSD95

PSD95

P

PSD95

SynGAP

PSD95 PSD95

SynGAP
P

LTP expression

Model of SynGAP regulation of synaptic plasticity. SynGAP regulates synapses by competing with AMPAR-TARP complexes to form LLPS condensates with
PSD95. During LTP induction, phosphorylation of SynGAP promotes the dispersal of SynGAP from the synapse and is replaced with AMPAR-TARP complexes, resulting
in the potentiation of synaptic transmission.
Araki et al., Science 383, 963 (2024)

1 March 2024

1 of 1

RES EARCH

RESEARCH ARTICLE

◥

NEUROSCIENCE

SynGAP regulates synaptic plasticity and cognition
independently of its catalytic activity
Yoichi Araki†, Kacey E. Rajkovich†, Elizabeth E. Gerber†, Timothy R. Gamache†, Richard C. Johnson,
Thanh Hai N. Tran, Bian Liu, Qianwen Zhu, Ingie Hong, Alfredo Kirkwood, Richard Huganir*

1 of 15

,

1 March 2024

y

Araki et al., Science 383, eadk1291 (2024)

y g

*Corresponding author. Email: rhuganir@jhmi.edu
†These authors contributed equally to this work.

y

Department of Neuroscience, Kavli Neuroscience Discovery
Institute, Johns Hopkins University School of Medicine,
Baltimore, MD 21205, USA.

activating protein (GAP) that negatively regulates small G protein signaling important for
activity-dependent changes in synaptic strength
(11–13). SynGAP is an abundant synaptic protein that is surpassed in copy number in the
PSD by only the PSD-95 family of proteins and
calcium/calmodulin-dependent protein kinase II
(CaMKII) (14). Previously, we and others have
shown that SynGAP undergoes a rapid change
in localization after neuronal activity (7, 8). At
baseline, PSD-enriched SynGAP regulates synaptic plasticity by inhibiting several G protein
signaling cascades involved in LTP, including
the Ras-Raf-MEK-ERK pathway, the activation
of which is required for the insertion of AMPARs
into the PSD (15). After an LTP-inducing stimulus, SynGAP is phosphorylated by CaMKII in
an NMDA receptor–dependent manner and is
rapidly dispersed from the PSD (7). SynGAP
dispersion leads to increases in dendritic spine
volume and synaptic AMPAR number (7). This
dispersion relieves the negative regulation of
synaptic Ras signaling and facilitates the induction and maintenance of changes underlying
activity-dependent synapse strengthening (7).
Its abundance in the PSD suggests that SynGAP
may occupy a substantial number of the finite
PDZ-binding slots under basal conditions, which
in turn may limit the number of “slots” for
AMPAR/TARP complexes (9). Indeed, reduced
SynGAP expression in heterozygous knockout
(KO) mice has been reported to be associated
with increased concentrations of TARPs and
AMPARs within the PSDs of forebrain neurons
in vivo (9, 10). However, Syngap1 heterozygous
mice also display enhanced activity of SynGAPregulated downstream signaling pathways
throughout development (16), making it difficult to determine whether the anticorrelation
between SynGAP and TARP protein amounts

To test whether SynGAP regulates PSD composition in a GAP-independent manner, we
used a knockdown-replacement strategy in
which we knocked down SynGAP with short
hairpin RNA (shRNA) and replaced it by transfecting wild-type (WT) or mutant SynGAP constructs in rat hippocampal neurons in vitro.
We have previously used this approach to
study SynGAP function during chemically induced LTP (cLTP) (7). cLTP causes SynGAP
dispersion from the synapse, recruitment of
AMPARs to synapses, and spine enlargement.
In our previous study, we found that SynGAP
KD increased synaptic Ras activity, enlarged
spines, and increased amounts of synaptic
AMPARs in the basal state, which occluded
further increases in spine size and receptor
content upon cLTP induction (7). Expression
of the WT SynGAP-a1 isoform rescued this
phenotype. However, SynGAP harboring
serine-to-alanine mutations at CaMKII phosphorylation sites critical for SynGAP dispersion
from synapses (SynGAP 2SA; S1108A; S1138A)
rescued the basal spine size and receptor content but failed to rescue cLTP because of deficits in SynGAP dispersion (7) (also see fig. S1
and Fig. 1). Here, we used this approach to
examine the role of GAP activity in cLTP. We
knocked down endogenous SynGAP by transfecting 19 to 21 days in vitro (DIV) rat hippocampal neurons with shRNA against SynGAP
(shRNA-SynGAP) and simultaneously expressing either an shRNA-resistant form of full-length
Azurite-tagged WT or mutant SynGAP-a1 constructs, along with Super-ecliptic pHluorin
(SEP)–tagged AMPAR GluA1 subunit (SEPGluA1) and a mCherry cytosolic cell fill to observe SynGAP, AMPAR, and spine size changes
during cLTP (Fig. 1). As observed previously,
when SynGAP KD was rescued with shRNAresistant WT SynGAP, cLTP stimulation resulted in the rapid dispersion of SynGAP from
dendritic spines and a concomitant increase
of both synaptic SEP-GluA1 signal and spine
size (Fig. 1, B to D) (7). These cLTP-dependent
changes were blocked when we transfected
neurons with Azurite-SynGAP harboring the

g

L

ong-term potentiation (LTP) is a major
form of synaptic plasticity in the brain that
is thought to underlie learning, memory,
and other higher-order brain processes
(1–3). LTP has been a central focus in
neuroscience for decades, and the biochemical
signaling cascades underlying it have been
investigated in great depth. Synaptic potentiation during LTP is mediated by increases in
synaptic AMPA receptors (AMPARs), the major excitatory neurotransmitter receptors in
the brain (1–3). However, it remains unclear
how LTP induction leads to the stable trapping of AMPARs at the synapse to establish
and maintain increased synaptic strength. One
leading hypothesis involves the diffusional trapping of plasma membrane–inserted AMPARs
by binding to proteinaceous binding “slots” in
the postsynaptic density (PSD) (4–6). According to the “slot” hypothesis, AMPARs associate
with the PSD through the binding of their auxiliary subunit transmembrane AMPAR regulating proteins (TARPs) to PDZ-domain-containing
scaffolding molecules in the PSD, including
PSD-95 and other members of the membraneassociated guanylate kinase (MAGUK) family
of proteins. As the PSD undergoes changes in
organization and composition after the induction of synaptic plasticity, these PDZ domains
can be dynamically occupied by AMPAR/TARP
complexes and other transmembrane and nontransmembrane molecules (6–10).
One such nontransmembrane molecule
is SynGAP, a synaptically localized GTPase-

SynGAP GAP activity is not required for
synaptic AMPAR recruitment in vitro

p

SynGAP is an abundant synaptic GTPase-activating protein (GAP) critical for synaptic plasticity,
learning, memory, and cognition. Mutations in SYNGAP1 in humans result in intellectual disability,
autistic-like behaviors, and epilepsy. Heterozygous Syngap1-knockout mice display deficits in
synaptic plasticity, learning, and memory and exhibit seizures. It is unclear whether SynGAP imparts
structural properties at synapses independently of its GAP activity. Here, we report that inactivating
mutations within the GAP domain do not inhibit synaptic plasticity or cause behavioral deficits.
Instead, SynGAP modulates synaptic strength by physically competing with the AMPA-receptor-TARP
excitatory receptor complex in the formation of molecular condensates with synaptic scaffolding
proteins. These results have major implications for developing therapeutic treatments for
SYNGAP1-related neurodevelopmental disorders.

in the PSD is caused by PDZ slot binding competition, changes in synaptic GAP activity and
downstream signaling, or both.
Here, we tested the role of SynGAP at PSD
both in vitro and in vivo using catalytically inactive SynGAP expression constructs and knockin (KI) mice containing inactivating mutations
within the SynGAP GAP domain. We provide
evidence for a distinct structural role for SynGAP
in the PSD that is independent of its role as a
regulator of G protein signaling and show that
SynGAP dispersion during LTP induction increases the number of PDZ-binding slots available for AMPAR and TARP complexes.

RES EARCH | R E S E A R C H A R T I C L E

*
G
+ AP
2S *
A

A

AP

G

2S

W

T

Normalized spine SynGAP

*
G
+ AP
2S *
A

A

AP

G

2S

T

W

*
G
+ AP
2S *
A

AP

A

G

2S

W
T

Normalized spine GluA1

Normalized spine volume

We next sought to determine whether the
structural contribution of SynGAP to AMPAR
trafficking that we observed in vitro could be
observed in vivo. To separate the role of G protein signaling and the structural properties of
2 of 15

,

SynGAP-GAP mutant mice have normal LTP

y

ner that is dissociable from the mechanisms
underlying spine enlargement.

y g

GAP* was enriched at synapses and underwent
dispersion from synapses after cLTP stimulation like WT SynGAP (Fig. 1, B to D). Dispersion of SynGAP-GAP* was sufficient to rescue
the cLTP-dependent enhancement of synaptic SEP-GluA1 signal, but not spine enlargement (Fig. 1, B to D). These data indicate that
SynGAP regulates AMPAR synaptic accumulation during cLTP in a GAP-independent man-

1 March 2024

y

Araki et al., Science 383, eadk1291 (2024)

g

CaMKII phosphorylation site mutations. We
then performed similar experiments using a
SynGAP construct harboring two point mutations in the GAP domain at residues known
to be critical for its GAP activity (SynGAP-GAP*;
F484A, R485L) (17, 18). These mutations eliminated SynGAP GAP activity in a RAS activation
pull-down assay in transfected human embryonic kidney (HEK293T) cells (fig. S2). SynGAP-

p

Chemical LTP

Baseline

Fig. 1. SynGAP GAP activity
GluA1
A
WT
2SA
GAP*
GAP*2SA
is not required for synaptic
AMPAR recruitment
in vitro. (A) Representative
Merged
live fluorescent confocal
images of a secondary dendrite from a rat hippocampal
SEP-GluA1
neuron transfected with
mCherry (cytosolic cell fill),
SEP-GluA1, and AzuritemCherry
tagged WT (WT) or mutant
SynGAP before (Baseline) or
Azuriteafter chemical LTP (cLTP).
SynGAP
Mutants included phosphodeficient SynGAP (2SA),
GAP-inactive SynGAP
(GAP*), and a combination
Merged
mutant with both (GAP*+2SA).
Endogenous SynGAP was
knocked down by shRNA and
SEP-GluA1
replaced by exogenous
shRNA-resistant AzuriteSynGAP. Arrowheads indicate
mCherry
representative synaptic spine
heads with SynGAP dispersion and SEP-GluA1 insertion.
AzuriteWhite arrowheads indicate
SynGAP
dendritic spines that enlarge
and exhibit SEP-GluA1 insertion
Baseline
Chem LTP
and SynGAP dispersion in
response to cLTP. Yellow
SynGAP
SEP-GluA1
B
C mCherry
D
arrowheads indicate dendritic
spines displaying SEP-GluA1
****
***
5
2.0 ****
5
***
**
insertion and SynGAP
*
4
4
1.5
dispersion without spine
3
3
enlargement. Blue arrow1.0
heads indicate spines with no
2
2
response during cLTP. Scale
0.5
1
1
bar, 5 mm. (B) Quantification
of SEP-GluA1 expression
0
0.0
0
before and after cLTP in
neurons transfected with WT
or various mutant constructs.
1.880 ± 0.181 A.U., cLTP 2.323 ± 0.205 A.U.). (D) Quantification of synaptic
Normalized total synaptic spine GluA1 contents by dendritic intensity is shown
SynGAP expression before and after cLTP induction in neurons transfected with
(WT: n = 6, Basal 1.000 ± 0.086 A.U., cLTP 2.265 ± 0.303 A.U.; 2SA: n = 6,
WT or various mutant constructs. Normalized total synaptic spine SynGAP
Basal 0.974 ± 0.055 A.U., cLTP 1.271 ± 0.088 A.U.; GAP*: n = 7, Basal 1.349 ±
0.145 A.U., cLTP 2.232 ± 0.211 A.U.; GAP*+2SA: n = 7, Basal 1.374 ± 0.189 A.U., contents by dendritic intensity is shown (WT: n = 6, Basal 1.000 ± 0.018 A.U.,
cLTP 0.435 ± 0.074 A.U.; 2SA: n = 6, Basal 1.153 ± 0.048 A.U., cLTP 0.944 ±
cLTP 1.631 ± 0.141 A.U.). (C) Quantification of the average change in spine
volume during cLTP in neurons expressing WT or various mutant constructs, as 0.073 A.U.; GAP*: n = 7, Basal 0.987 ± 0.058 A.U., cLTP 0.401 ± 0.059 A.U.;
GAP*+2SA: n = 7, Basal 1.091 ± 0.034 A.U., cLTP 0.981 ± 0.073 A.U.). For (A) to (D),
measured by mCherry cell fill. Normalized total synaptic mCherry by dendritic
intensity is shown (WT: n = 6, Basal 1.000 ± 0.141 A.U., cLTP 2.238 ± 0.182 A.U.; two-way ANOVA with repeated measures for chemical LTP treatment and multiple
comparisons with Šídák’s test were used. *P < 0.05, **P < 0.01, ***P < 0.001,
2SA: n = 6, Basal 1.058 ± 0.105 A.U., cLTP 1.326 ± 0.128 A.U.; GAP*: n = 7,
Basal 1.961 ± 0.160 A.U., cLTP 2.392 ± 0.284 A.U.; GAP*+2SA: n = 7, Basal
****P < 0.0001; n.s., not significant.

RES EARCH | R E S E A R C H A R T I C L E

A

GAP-deficient Syngap1 mutant mice
Syngap1 exon 9 FR → AL in GAP domain

R

H

E

L

I

F

R

E

N

T

L

A

...CGG GAA CAC CTC ATA TTC CGA GAG AAC ACG CTC GCC...
...GCC CTT GTG GAG TAT AAG GCT CTC TTG TGC GAG CGG...

gRNA Target
CRISPR/Cas9
94-nucleotide oligo primer

R

H

E

L

I

A

L

E

N

T

L

A

...CGG GAA CAC TTA ATC GCT CTC GAG AAT ACG CTA GCC...
...GCC CTT GTG AAT TAG CGA GAG CTC TTA TGC GAT CGG...
XhoI
NheI

Syngap1 GAP domain:

pERK

GAPDH

GAPDH

ERK

ERK

0.5

1.0
0.5
0.0

G

*

1.5
1.0
0.5
0.0

+/+

*

1.5
1.0
0.5
0.0

GAP*/GAP*

+/GAP*

Survival until P10
χ test
p < 0.05
2

H

χ2 test
n.s.

I
# of Mice

# of Mice

15
10
5

n.s.

50

0

1

2

3

+/-

-/-

G
G AP
AP */
*

AP

n.s.
n.s.

100
50

,

***

0

(%) Probability of Survival

100

Observed
Expected

y

(%) Probability of Survival

+/

G

+
+/

+/

-/-

-

+/
+

*

40
30
20
10
0

y g

0

to what has been seen previously in SynGAP
heterozygous KO (Syngap1+/−) mice (18). Whereas homozygote SynGAP KO (Syngap1−/−) mice
dies perinatally within 2 ro 3 days (16, 19),
homozygote GAP* KI mice (Syngap1GAP*/GAP*)
survives well beyond postnatal day 7 (P7) (Fig. 2,
H and I) into adulthood, are fertile, and can be
bred in homozygosity. Thus, whereas SynGAP
is required for viability, its GAP activity is not.

1 March 2024

2.0

y

0.0

2.0

g

1.0

F

1.5

[pERK]/[ERK]

D

1.5

[pERK]/[ERK]

C

+/+

Araki et al., Science 383, eadk1291 (2024)

+
+/

pERK

[SynGAP]/[GAPDH]

SynGAP

[SynGAP]/[GAPDH]

SynGAP

Weeks

SynGAP, we generated KI mice with the same
inactivating mutations in the GAP domain used
in vitro to eliminate GAP activity (Fig. 2A). Heterozygote mice harboring this GAP* KI mutation
(Syngap1+/GAP*) had normal SynGAP protein expression in the brain but showed increased
expression of phosphorylated extracellular signal–
regulated kinase (pERK) (Fig. 2, B to G), consistent with decreased GAP activity similar

G
G AP
AP */
*

AP
*
G

+

+/

+/

E
+/

+

AP
*
G
+/

+/

+

B

G
G AP
AP */
*

CGGGAACACTTAATCGCTCTCGAGAATACGCTAG
A L

p

Fig. 2. SynGAP-GAP KI mice exhibit normal SynGAP
protein expression but have elevated Ras-ERK
signaling in the brain. (A) Generation of Syngap1+/GAP* mice by
CRISPR-Cas9. gRNA was designed to make the
double-strand break near the target site, and the
GAP activity-deficient mutant was introduced (FR→AL, “GAP*”)
by homology-directed repair using a 94-nucleotide GAP-mutant
oligo donor. (B to D) Representative immunoblots and
quantification of SynGAP and GAPDH protein from whole brains of
Syngap1+/GAP*, Syngap1GAP*/GAP* mice and WT littermates
(Syngap1+/+). Syngap1+/+ (n = 2, mean ± SEM; 1.000 ± 0.046 A.U.)
versus Syngap1+/GAP* (n = 2, mean ± SEM; 1.076 ± 0.077 A.U.);
Syngap1+/+ (n = 2, mean ± SEM; 1.030 ± 0.069 A.U.)
versus Syngap1GAP*/GAP* (n = 2, mean ± SEM; 1.020 ± 0.021 A.U.).
*P < 0.05, Mann-Whitney test. (E to G) Representative
immunoblots and quantification of phospho-ERK and total
ERK protein from whole brains of Syngap1+/GAP*, Syngap1GAP*/GAP*
mice and WT littermates (Syngap1+/+). Syngap1+/+
(n = 4, mean ± SEM; 1.05 ± 0.043 A.U.) versus
Syngap1+/GAP*(n = 4, mean ± SEM; 1.288 ± 0.017 A.U.);
Syngap1+/+ (n = 4, mean ± SEM; 1.000 ± 0.029 A.U.) versus
Syngap1GAP*/GAP* (n = 4, mean ± SEM; 1.437 ± 0.092 A.U.).
*P < 0.05, Mann-Whitney test. (H) Survival of Syngap1+/−,
Syngap1−/− mice and WT littermates (Syngap1+/+) resultant
from Syngap1+/− × Syngap1+/− breeding until age P10.
Top panel: Observed number of mice (Syngap1+/+ = 8,
Syngap1+/− = 12, Syngap1−/− = 0) versus Expected number
of mice (Syngap1+/+ = 5, Syngap1+/− = 10, Syngap1−/− = 5);
*P < 0.05, chi-square test. No Syngap1−/− mice survived
until P10. Bottom panel: Survival plot. Log-rank (Mantel-Cox)
test was used; Syngap1+/+ and Syngap1+/− (P = 0.36, n.s.);
Syngap1+/+ and Syngap1−/− (***P = 0.0009). (I) Survival
of Syngap1+/GAP*, Syngap1GAP*/GAP* mice and WT littermates
cSyngap1+/+) resultant from Syngap1+/GAP* × Syngap1+/GAP*
breeding until P10. Top Panel: Observed number of mice
(Syngap1+/+ = 18, Syngap1+/GAP* = 27, Syngap1GAP*/GAP* = 13)
versus Expected number of mice (Syngap1+/+ = 14.5, Syngap1+/GAP* = 29,
Syngap1GAP*/GAP* = 14.5; chi-square test was used, n.s. (P = 0.57).
Bottom panel: Survival plot. Log-rank (Mantel-Cox) test
was used for Syngap1+/+ and Syngap1+/GAP* (P = 0.41, n.s.)
and Syngap1+/+ and Syngap1GAP*/GAP* (P = 0.11, n.s.).

0
0

1

2

3

Weeks
+/+

+/GAP*

GAP*/
GAP*

Previous studies have shown that spine size
and miniature excitatory postsynaptic current
(mEPSC) are both increased in the Syngap1
heterozygote KO (20). However, consistent with
our in vitro data above showing that after
SynGAP knockdown (KD), the GAP* mutant
did not rescue the increased synaptic spine size
but did rescue the normal AMPAR content in the
baseline conditions (Fig. 1), the synaptic spine
3 of 15

RES EARCH | R E S E A R C H A R T I C L E

% LTP

TBS

SynGAP-GAP mutant mice have normal activity,
working memory, and associative fear memory

Previous work has shown that SynGAP is required for normal locomotor activity, learning,

SynGAP+/+

180

B

160
140

SynGAP+/-

120

80
-20

0

20

40

60

10 ms
1 mV

180

140
100
+/+
+/-

+/+

TBS

D

180

% LTP

160

+/GAP*

140
120

80
-20

GAP*/GAP*
0

20

40

60

Time (min)

10 ms
1 mV

180

SynGAP competes with TARP-g8 for
binding to PSD-95

140
100

Araki et al., Science 383, eadk1291 (2024)

1 March 2024

4 of 15

,

Fig. 3. SynGAP-GAP KI mice have normal LTP. (A) Averaged population field CA1 recordings of TBS-LTP
time course obtained from brain slices of Syngap1+/− mice and Syngap1+/+ littermate controls. All data points
are normalized to the averaged baseline fEPSP slope. Inset: Example averaged fEPSP traces from Syngap1+/+ and
Syngap1+/− slices recorded during baseline (black) and 40 to 60 min after TBS-LTP induction (red).
(B) Quantification of averaged TBS-LTP in Syngap1+/− and Syngap1+/+ littermates. Individual data points are
superimposed. TBS-LTP is calculated by the ratio of the mean fEPSP slope measured 40 to 60 min after
TBS-LTP induction (yellow-shaded region) divided by the averaged fEPSP baseline slope within each recorded
sample (Syngap1+/+: n = 13, 150.9 ± 7.51% SEM; Syngap1+/−: n = 13, 123.0 ± 5.416% SEM). Mann-Whitney
rank sum test was used. (C) Averaged population field CA1 recordings of TBS-LTP time course obtained
from brain slices of Syngap1+/GAP* and Syngap1GAP*/GAP* mice as well as their Syngap1+/+ littermate controls.
Inset: Example averaged fEPSP traces from Syngap1+/+, Syngap1+/GAP*, and Syngap1GAP*/GAP* slices recorded
during baseline (black) and 40 to 60 min after TBS-LTP induction (red). (D) Quantification of averaged
TBS-LTP in Syngap1+/+, Syngap1+/GAP*, and Syngap1GAP*/GAP* littermates. Individual data points are
superimposed. (Syngap1+/+: n = 22, 145.5 ± 4.74% SEM; Syngap1+/GAP*: n = 19, 151.9 ± 6.57% SEM;
Syngap1GAP*/GAP*: n = 16, 154.8 ± 9.34% SEM). Nonparametric one-way ANOVA and Kruskal-Wallis multiplecomparisons test were used. Error bars and shading represent SEM. *P < 0.05; n.s., not significant.

y

+/+
+/GAP*
GAP*/GAP*

Because GAP catalytic activity of SynGAP is
not required for normal LTP and memory, we
investigated whether SynGAP’s structural role
in the PSD is essential for neuroplasticity. Both
SynGAP (24) and TARP-g8 (25) (hereafter referred to as g8) undergo liquid-liquid phase separation (LLPS) with MAGUK family proteins
in cell-free systems. LLPS is a known mechanism of the formation of molecular condensates
that are composed of dynamic protein clusters
that exchange constituents with the adjacent
pool of freely diffusing proteins (26). In vitro,
many synaptic proteins are known to undergo
LLPS, which can facilitate the clustering of membrane proteins (27, 28). Previous studies have
shown that PSD-95 can form molecular condensates with both SynGAP (24) and TARPs
(27, 28). To explore whether this property of
SynGAP is important for its role in LTP, we
first investigated the possibility that SynGAP
competes with the g8-terminal cytosolic region
(198 amino acids, hereafter referred to as g8CT)
for binding to PSD-95 to regulate the composition of synaptic PDS-95 molecular clusters

y g

100

% LTP (40-60 min)

220

n.s.
n.s. n.s.

y

Time (min)
C

*

g

100

220

and memory (20–23). However, whether the
GAP activity of SynGAP is required for these
behaviors remains to be fully elucidated. We
performed a series of behavioral experiments
in 2- to 4-month-old mice. In open-field testing, Syngap1+/− mice showed hyperactivity
compared with WT littermates (Fig. 4A), consistent with prior work (22). By contrast, both
Syngap1+/GAP*and Syngap1 GAP*/GAP* mice
showed normal activity that was indistinguishable from that of their WT littermates
(Fig. 4B). We next compared working memory
using the Y-maze spontaneous alternation
task. Consistent with previous studies (20, 23),
Syngap1+/− mice had reduced spontaneous alternations compared with WT littermates (Fig.
4C). However, the percentages of alternations
for Syngap1+/GAP* and Syngap1GAP*/GAP* mice
were not different from those of their WT
littermates (Fig. 4D). To explore whether SynGAP
GAP activity is required for associative learning, we then performed auditory cued and contextual fear conditioning. Consistent with prior
studies (21, 22), Syngap1+/− mice had impaired
learning of a shock-associated auditory cue, a
conditioned stimulus (CS), as assessed by measuring the amount of time spent freezing in
response to the presentation of the CS after
conditioning (Fig. 4E). Syngap1+/GAP* mice
showed no impairment in fear conditioning
and showed increases in freezing that were
no different from those of WT mice (Fig. 4F).
Taken together, these data show that whereas
Syngap1+/− mice exhibit hyperactivity and deficits in both working memory and fear learning,
these impairments are not found in heterozygous and homozygous GAP* KI mice.

p

A

viously observed LTP deficits with Syngap1 haploinsufficiency (Fig. 3, A and B) (19). By contrast,
slices prepared from both Syngap1+/GAP* mice
unexpectedly exhibited normal TBS-LTP expression compared with recordings obtained from
brain slices of WT littermates (Fig. 3, C and D).
Moreover, we found that Syngap1GAP*/GAP* mice
had normal LTP (Fig. 3, C and D). These data
suggest that the structural presence of SynGAP
at synapses is sufficient for normal LTP to occur and demonstrate that the GAP activity of
SynGAP is dispensable for the expression of
hippocampal LTP.

% LTP (40-60 min)

size in the CA1 hippocampal region of homozygote GAP* KI mice (Syngap1GAP*/GAP*) was
enlarged but the mEPSC amplitude remained
unchanged (fig. S3). These data support the idea
that GAP activity may be important for changes
in spine size but does not affect AMPA receptor
content in synapses in vivo.
To test whether SynGAP’s GAP activity is
required for synaptic plasticity in vivo, we
performed extracellular field recordings to
measure LTP in CA1 of the hippocampus induced by repeated theta-burst stimulation
(TBS) of the Schaeffer collateral pathway in
Syngap1+/−, Syngap1+/GAP*, and Syngap1GAP*/GAP*
mice, along with their respective WT littermates (Fig. 3). We measured a 54% reduction of
TBS-LTP expression in Syngap1+/− brain slices
compared with WT littermates, replicating pre-

RES EARCH | R E S E A R C H A R T I C L E

Distance (cm)

A

SYNGAP+/+ B
SYNGAP+/-

6000

Distance (cm)

Fig. 4. Syngap1+/GAP* KI mice have normal activity,
working memory, and associative fear memory.
(A) Distance traveled by Syngap1+/− mice (n = 15) and
Syngap1+/+ WT (WT) littermates (n = 18) during a
2-hour open-field test in 5-min intervals. Two-way
ANOVA with repeated measures for time only
and Šídák’s multiple-comparisons test were used.
(B) Distance traveled by Syngap1+/GAP* mice (n = 16),
Syngap1GAP*/GAP* mice (n = 14), and Syngap1+/+
WT littermates (n = 17) during a 2-hour open-field
test in 5-min intervals. Two-way ANOVA with
repeated measures for time only and Šídák’s
multiple-comparisons test were used. (C) Percentage
of spontaneous alternating arm visits (% alternation)
by Syngap1+/− mice (n = 48, 56.00 ± 1.29%
alternation) and Syngap1+/+ littermates (n = 37,
68.30 ± 1.58% alternation) during a 5-min Y-maze
exploration test. The red dotted line represents the
50% successful alternation rate expected due
to chance. Two-tailed Student’s t test was used.
(D) Percentage of spontaneous alternating arm visits
(% alternation) by Syngap1+/GAP* mice (n = 35,
65.63 ± 1.83% alternation), Syngap1GAP*/GAP* mice
(n = 18, 67.14 ± 2.37% alternation), and Syngap1+/+
littermates (n = 35, 66.74 ± 1.49% alternation)
during a 5-min Y-maze exploration test. The red dotted
line represents the 50% successful alternation rate
expected due to chance. One-way ANOVA and
Tukey’s test were used. (E) Average percentage of
time spent freezing per minute (% freezing) with
and without the conditioned stimulus (auditory cue,
CS) by Syngap1+/− mice (n = 14 29.12 ± 4.44%
freezing) and Syngap1+/+ littermates (n = 19, 27.50 ±
2.37% freezing). Two-way ANOVA with repeated
measures for CS only and Šídák’s multiple-comparisons
test were used. (F) Average percentage of time spent
freezing per minute (% freezing) with and without
presentation of the conditioned stimulus (auditory cue,
CS) by Syngap1+/GAP* mice (n = 15, 23.18 ± 2.84%
freezing), Syngap1GAP*/GAP* mice (n = 19, 20.11 ±
2.129% freezing) and Syngap1+/+ littermates (n = 18,
27.15 ± 2.203% freezing). Two-way ANOVA with
repeated measures for CS only and Šídák’s multiplecomparisons test were used. Error bars represent
SEM. *P < 0.05, **P < 0.01, ***P < 0.001, ****P <
0.0001; n.s., not significant..

****
4000
2000

6000
4000
2000

0

0

40
80
Time (min)

C

100

0

120

+/-

+/+

0

40
80
Time (min)

D

ns

100

120

GAP*/GAP*

+/GAP*

+/+

****

+/+
+/GAP*
GAP*/GAP*

ns

(%) Alternation

60
40
20

80
60
40

g

(%) Alternation

p

80

20
y

0

****

n.s.

F

80

100

60
40

**** **** ***

80
60
y g

% Freezing

100

% Freezing

E

0

40
20

20

y

during the expression of LTP. We transfected
full-length PSD-95-mCherry and GFP-g8CT
in COS cells in the absence and presence of
SynGAP. Consistent with data from cell-free
experiments (25), coexpression of PSD-95mCherry and GFP-g8CT in the absence of SynGAP
resulted in molecular clusters containing both
GFP-g8CT and PSD-95-mCherry (Fig. 5A). We
then cotransfected GFP-g8CT and PSD-95mCherry with increasing concentrations of
SynGAP and examined the PSD-95 clusters for
the presence of g8 and SynGAP. At low SynGAP
Araki et al., Science 383, eadk1291 (2024)

– + – +

expression, SynGAP co-clustered with g8CT and
PSD-95, but with increasing concentrations of
SynGAP, the presence of g8CT in the clusters was
eliminated (Fig. 5A), indicating that SynGAP can
compete with g8 for binding to PSD95 clusters. We tested whether g8CT/PSD-95 and PSD95/SynGAP clusters dynamically exchange their
contents with cytosolic pool using a fluorescence
recovery after photobleaching (FRAP) assay
(fig. S4), which confirmed that these contents
recovered within a few minutes after bleaching. These results suggest these clusters are

1 March 2024

0
CS

,

0
CS

– +

– + – +

not immobile aggregates of the two proteins
but rather are likely condensed liquid-phase
droplets similar to those seen in vitro.
We then characterized the structural requirements of SynGAP for competition with
g8. Coexpression of WT SynGAP eliminated
g8CT from PSD-95 clusters (Fig. 5A). Mutation
of the GAP domain had no effect on the ability of SynGAP to compete with g8CT (Fig. 5B).
By contrast, a series of mutations that regulate
cluster formation substantially affected the
ability of SynGAP to compete with g8CT (Fig.
5 of 15

RES EARCH | R E S E A R C H A R T I C L E

Dose-dependency of SynGAP competition to γ8-PSD95 puncta

+
+

PSD95
SynGAP

+

+

+

+

+

+

+

+

+

+

+

+

x0

x0.25

x0.5

x1

x2

x4

Merge

γ-8 CT

***

75
50

***

25

*** ***
x4

PS 8
8+ D95
PS
D
95
x0
.2
5
x0
.5

SynGAP

x2

0

PSD95

+Azurite-SynGAP

Structual element of SynGAP competition to γ8-PSD95 puncta

PSD95

+
+

SynGAP

+
+

+
+
WT

+
+
LDKD

+
+
ΔPDZ

+
+
LDKD+
ΔPDZ

+
+
ΔC143

+
+
ΔC580

+
+

+
+

GAP*

GAP*+
LDKD+
ΔPDZ

Merge

***

***
***

*** ***
75
50
25

* *
n.s.

0

SynGAP

y

P 8
8+ SD9
PS 5
D
95
W
LD T
KD
LD
KD PD
+Δ Z
PD
C Z
14
C 3
58
0
G
AP
LD
KD GA *
+Δ P*
PD +
Z

PSD95

100

g

γ-8 CT

p

γ-8 CT

% PSD95 puncta with γ8

B

n.s.

100

x1

γ-8 CT

% PSD95 puncta with γ8

A

+Azurite-SynGAP

Phase-in-phase separation of TARP-g8-PSD95
and SynGAP-PSD95 condensates within
phase-separated droplets of purified proteins

Next, we used purified proteins to further explore how g8 and SynGAP compete for binding
to PSD95. We first confirmed that g8CT-PSD95
and PSD95-SynGAPCC-PBM underwent LLPS
and formed liquid condensates (droplets) in
our assay system when the two pairs of purified proteins were mixed (24, 25) (Fig. 6A, top
and middle panels). Next, we explored the phase
separation of these three purified proteins
when combined. The proteins did not homogeneously mix within droplets and formed
separate phase-in-phase condensates within
individual liquid droplets; SynGAPCC-PBM-PSD95
clustered in the center, whereas g8CT-PSD95

1 March 2024

formed an adjacent protein condensation, a
ring-like structure around the periphery (Fig.
6A, bottom panels). These distinct localizations were found in nearly all droplets (fig.
S8A). When the localization of each protein
was plotted on the differential interference
contrast (DIC) image, faint boundaries detected
by refractive index changes were observed
around the inner droplet of SynGAPCC-PBMPSD95 (Fig. 6B). This observation indicates
that separate condensates with distinct lightscattering properties were forming within each
droplet, with SynGAP-PSD95 condensates forming inside the g8-PSD95 condensate. Plotting the
localization of g8CT-PSD95 by drawing a line
across the droplet center when only g8CT and
PSD-95 were mixed showed that g8CT-PSD95 had
a uniform distribution across the droplet (Fig.
6C, left panel). However, when SynGAPCC-PBM
was added, g8CT was more peripherally localized
6 of 15

,

Araki et al., Science 383, eadk1291 (2024)

sal role of SynGAP to compete with the TARPPSD95 interaction.

y

5B). Mutations in SynGAP’s PDZ ligand domain (DPDZ mutant) or in its coil-coil domain
(LDKD mutant), which we have previously
shown are important for cluster formation
and LLPS (24), decreased its ability to displace
g8CT (Fig. 5B; for structure of deletion constructs please see fig. S5). Combining these
two mutations (DPDZ/LDKD) almost completely eliminated SynGAP’s ability to compete
with g8. Deletion of the entire C-terminal coilcoil domain and disordered region (DC580), but
not the GAP domain, was required to eliminate SynGAP’s ability to displace g8CT. Similar experiments using full-length g8 showed
similar results (fig. S6). These results indicate
that SynGAP’s C-terminal structure is essential
for its ability to compete with g8 for cluster formation with PSD-95. The same experiment using
TARP-g2 (g2), another major TARP, yielded
similar results (fig. S7), suggesting a univer-

cells transfected with GFP-g8CT (“g8”), PSD95-mCherry, and different AzuriteSynGAP mutants (WT, LDKD, DPDZ, LDKD+DPDZ, DC143, DC580, GAP*, and
GAP*+LDKD+DPDZ). Scale bar, 5 mm. Right panel: percentage of PSD95 puncta
with g8 with different amounts of Azurite-SynGAP. One-way ANOVA and
Tukey’s multiple-comparisons test were used. Error bars represent SD. *P <
0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; n.s., not significant compared
with g8+PSD95+Azurite-SynGAP WT unless otherwise specified.

y g

Fig. 5. SynGAP-PSD95 and TARP-g8-PSD95 compete in vitro. (A) Confocal
microscopy of COS cells transfected with GFP-g8CT (“g8”), PSD95-mCherry,
and different amounts of Azurite-SynGAP (0.25×, 0.5×, 1×, 2×, and 4×). Scale
bar, 5 mm. Right panel: percentage of PSD95 puncta with g8 with different
amounts of Azurite-SynGAP. One-way ANOVA and Tukey’s multiple-comparisons
test were used. Error bars represent SD. *P < 0.05, **P < 0.01, ***P < 0.001;
n.s., not significant. compared with g8+PSD95. (B) Confocal microscopy of COS

RES EARCH | R E S E A R C H A R T I C L E

A

Merge
+DIC

DIC

Merge

γ8

PSD95

SynGAP

Zoom-Merge

γ8
PSD95

PSD95
SynGAP
γ8
PSD95
SynGAP

B

DIC+
SynGAP-PSD95 phase (Boundary)

γ8
SynGAP

SynGAP

PSD95
SynGAP

γ8

C

PSD95
γ8 SynGAP γ8

γ8 PSD95

PSD95

Merge

PSD95
SynGAP

γ8

γ8 (Top, Outer)

x-y

γ8

PSD95
SynGAP

x-y

PSD95
SynGAP
(Bottom, Inner)

Top

x-z

Optical slice

Bottom

1 March 2024

7 of 15

,

Araki et al., Science 383, eadk1291 (2024)

SynGAPCC-PBM-PSD95 droplet was not only
located inside but also had a propensity to
sink closer to the bottom, likely becase of its
higher density. Conversely, g8CT-PSD95 droplets were generally distributed peripherally and
positioned in the plane above SynGAPCC-PBMPSD95 (Fig. 6D). Using time-lapse imaging,
we investigated whether the droplets exhibited

y

with >50% higher g8CT concentration in the
periphery, whereas SynGAPCC-PBM was >95%
located centrally with PSD95 (Fig. 6C, right
panel), suggesting that the g8CT-PSD95 and
SynGAPCC-PBM-PSD95 droplets form different
layers of protein condensates.
Using confocal microscopy to scan both
the x–z and x–y planes, we found that the

y g

Fig. 6. SynGAP-PSD95 and TARP-g8-PSD95 show mutually exclusive phase-in-phase separation in
droplets. (A) Images of purified protein sedimentation assay by confocal microscopy. Purified proteins
included TARP-g8CT (“g8”) tagged with iFlour568 (green), PSD95 tagged with iFlour633 (red), and SynGAPCC-PBM
(last 156 amino acids of SynGAP: coiled-coil domain + PDZ ligand: “SynGAP”) tagged with iFlour488 (blue).
Left panels: merged fluorescence images with DIC images. Top: g8-PSD95 droplets. Middle: SynGAP-PSD95
droplets. Bottom: g8-PSD95-SynGAP droplets. High-power views are also shown (right panels). Scale bar,
3 mm. (B) Phase-in-phase separation of SynGAP-PSD95 droplets inside the g8-PSD95 droplets. Left panels:
blue arrows or circles delineate the inner rings of phase-in-phase separation. Right panels: merge of DIC
images with g8-PSD95-SynGAP droplets. Yellow arrows indicate regions of separation between SynGAP and
PSD95 phase. Scale bar, 3 mm. (C) Comparison between g8-PSD95 droplets and g8-PSD95-SynGAP droplets.
A line scan of protein condensations (yellow line) is shown to the right of each image. Scale bar, 3 mm.
(D) Optical sectioning microscopy of g8-PSD95-SynGAP protein droplets. Top panels: x–y view. Bottom
panels: x–z view. Optical slices (blue boxes) used to generate top (x–y) panels are shown. Scale bar, 5 mm.
Right panel: schematic of g8-PSD95-SynGAP droplets.

y

D

We next tested whether SynGAP GAP-activity is
required for synaptic recruitment of g8 during
cLTP using the same SynGAP KD/replacement
approach used above. After rat hippocampal
neurons were transfected with GFP-g8, we
observed a cLTP-dependent enhancement of
synaptic GFP-g8 fluorescence comparable to
that observed with SEP-GluA1. Synaptic recruitment of g8 required phosphorylation of
SynGAP but did not require SynGAP GAP activity (Fig. 7, A and B), revealing that SynGAP
regulates synaptic accumulation of g8 during
cLTP in a GAP-independent manner.
We then investigated whether SynGAP mutations that alter SynGAP condensate formation with PSD-95 could affect the expression
of cLTP. In these experiments, we knocked
down SynGAP and replaced it with either WT
SynGAP or SynGAP mutants that regulate
LLPS. In these experiments, we induced cLTP
at two glycine concentrations (10 and 200 mM)
to test the sensitivity of SynGAP dispersion
and cLTP induction to the strength of the induction stimulus. At 10 mM glycine, WT SynGAP
was not dispersed from spines and cLTP was
not expressed, as assayed by increases in spine
size or the recruitment of g8. By contrast, glycine at 200 mM resulted in clear WT SynGAP
dispersal and cLTP induction [Fig. 8 and (24)].
Replacement with the LDKD mutant also rescued cLTP using 200 mM glycine, whereas replacement with the PDZ mutant only partially
rescued cLTP, highlighting the importance of
the PDZ ligand sequence of SynGAP for occupying PSD-95 PDZ domains in the basal
state. At 10 mM glycine, the LDKD mutant was
dispersed and cLTP was expressed, in contrast to WT (Fig. 8). This outcome suggests
that the LDKD mutation lowers the threshold
for SynGAP dispersal from synapses, enhancing the sensitivity of cLTP induction by glycine
and consequently facilitating the recruitment
of g8 during cLTP. Neither the DPDZ mutant

g

γ8
PSD95
SynGAP

γ8

Mutations in SynGAP that affect condensate
formation and LLPS with PSD-95 regulate
recruitment of TARP-g8 to synapses
during cLTP

p

PSD95
SynGAP

one of the fundamental properties of phaseseparated bodies, the tendency to coalesce (fig.
S8B). We observed that within 1 to 2 min after
initial physical contact between droplets, the
outer layer of g8CT-PSD95 first fused, followed
by the inner layer of SynGAPCC-PBM-PSD95.
This observation suggests that both the outer
g8-PSD95 phase and the SynGAP-PSD95 phase
retain their droplet-like properties.
The results strongly support the idea that
SynGAP competes with g8 for PSD95 binding,
resulting in the formation of the different
layers of protein droplets rapidly and spontaneously. Critically, these phases appear incompatible with one another, exhibiting distinct
droplet properties.

RES EARCH | R E S E A R C H A R T I C L E

A

TARP-γ8

WT

2SA

GAP*

GAP*2SA

Baseline

Merged

TARPγ8

mCherry
AzuriteSynGAP

Chemical LTP

Merged

p

TARPγ8

mCherry

g

AzuriteSynGAP
Baseline
Chem LTP

***

***

1.5
1.0
0.5

*

G
+ AP
2S *
A

A

AP
G

2S

W
T

0.0

8 of 15

,

cLTP 1.162 ± 0.071 A.U.; GAP*: n = 7, Basal 1.438 ± 0.143 A.U.,
cLTP 2.666 ± 0.177 A.U.; GAP*+2SA: n = 7, Basal 1.675 ± 0.288 A.U.,
cLTP 1.867 ± 0.283 A.U.). (C) Quantification of the average change in spine
volume during cLTP in neurons expressing WT or various mutant
constructs as measured by mCherry cell fill. Normalized total synaptic
mCherry by dendritic intensity is shown (WT: n = 6, Basal 1.000 ±
0.088 A.U., cLTP 2.516 ± 0.234 A.U.; 2SA: n = 6, Basal 0.978 ± 0.118 A.U.,
cLTP 1.301 ± 0.132 A.U.; GAP*: n = 7, Basal 1.945 ± 0.158 A.U.,
cLTP 2.644 ± 0.333 A.U.; GAP*+2SA: n = 7, Basal 1.875 ± 0.085 A.U.,
cLTP 2.038 ± 0.180 A.U.). (D) Quantification of synaptic SynGAP
expression before and after cLTP induction in neurons transfected with
WT or various mutant constructs. Normalized total synaptic spine
SynGAP content by dendritic intensity is shown (WT: n = 6, Basal 1.000 ±
0.098 A.U., cLTP 0.435 ± 0.066 A.U.; 2SA: n = 6, Basal 1.077 ± 0.065 A.U.,
cLTP 0.997 ± 0.095 A.U.; GAP*: n = 7, Basal 1.032 ± 0.075 A.U., cLTP 0.497 ±
0.081 A.U.; GAP*+2SA: n = 7, Basal 1.019 ± 0.098 A.U., cLTP 0.912 ± 0.061 A.U.).
Two-way ANOVA with repeated measures for chemical LTP treatment and
multiple-comparisons with Šídák’s test were used. *P < 0.05, **P < 0.01,
***P < 0.001, ****P < 0.0001; n.s., not significant.

y

1 March 2024

*

0

Fig. 7. SynGAP GAP activity is not required for synaptic TARP-g8
recruitment in vitro. (A) Representative live fluorescent confocal images
of a secondary dendrite from a rat hippocampal neuron transfected with
GFP-TARP-g8, mCherry (cytosolic cell fill) and Azurite-tagged WT or mutant
SynGAP before (Baseline) or after chemical LTP (cLTP). Mutants include
phospho-deficient SynGAP (2SA), GAP-inactive SynGAP (GAP*), and a
combination mutant with both (GAP*+2SA). Endogenous SynGAP was
knocked down by shRNA and replaced by exogenous shRNA-resistant
Azurite-SynGAP. Arrowheads indicate representative synaptic spine heads
with SynGAP dispersion and g8 insertion. White arrowheads indicate
dendritic spines that enlarge and exhibit g8 insertion and SynGAP dispersion
in response to chemical LTP. Yellow arrowheads indicate dendritic spines
displaying g8 insertion and SynGAP dispersion without enlargement
(no structural plasticity). Blue arrowheads indicate spines with no response
during cLTP. Scale bar, 5 mm. (B) Quantification of synaptic GFP-g8
expression before and after cLTP induction in neurons expression WT or
various mutant constructs. Normalized total synaptic spine g8 contents
by dendritic intensity are shown (WT: n = 6, Basal 1.000 ± 0.090 A.U.,
cLTP 2.394 ± 0.185 A.U.; 2SA: n = 6, Basal 0.992 ± 0.042 A.U.,
Araki et al., Science 383, eadk1291 (2024)

Normalized spine SynGAP

1

G
+ AP
2S *
A

G
+ AP
2S *
A

*
AP
G

A
2S

2

SynGAP
2.0

y g

W
T

0

3

AP

1

*

A

2

***

4

G

3

5

2S

***

Normalized spine volume

Normalized spine γ8

***

4

D

mCherry

W
T

C

TARPγ8
5

y

B

RES EARCH | R E S E A R C H A R T I C L E

A

WT

LDKD

ΔPDZ

LDKDΔPDZ

Merge
Baseline

GFP-γ8
mCherry
AzuriteSynGAP

Chemical LTP (10 μΜ)

Merge
GFP-γ8

AzuriteSynGAP

Chemical LTP (200 μΜ)

Merge
GFP-γ8

y

mCherry
AzuriteSynGAP

Fig. 8. SynGAP phase-separation and PDZ-ligand binding capacity regulate
TARP-g8 trafficking during chemical LTP. (A) Representative live fluorescent
confocal images of a secondary dendrite from a rat hippocampal neuron
transfected with GFP-g8, mCherry (cytosolic cell fill), and Azurite-tagged WT
or mutant SynGAP before (Baseline) or after either weak cLTP (10 mM; Glycine)
or strong cLTP (200 mM; Glycine). Mutants include LDKD, DPDZ, or both.
Endogenous SynGAP was knocked down by shRNA and replaced with exogenous
shRNA-resistant Azurite-SynGAP. Green circles indicate spine heads with the
basal condition. Yellow circles and arrows indicate dendritic spines that enlarge
and exhibit g8 insertion, spine enlargements, and SynGAP dispersion in response
Araki et al., Science 383, eadk1291 (2024)

1 March 2024

2

***

***

ns

1
0
Z
ΔP

D

Z
KD

ΔP
D

KD

Normalized spine SynGAP

***

3

LD

ΔP

4

W
T

Z
D

Z
D

KD

***

LD

Baseline
Chem LTP (10 μΜ)
Chem LTP (200 μΜ)

LD

ΔP

0
ΔP

Z
D

Z
D

KD

1

LD

ΔP

KD
LD

W
T

2

***

ns

5

,

0

ns

***
3

SynGAP

y

1

*

4

KD

2

***

LD

*

***

T

***

***

3

5

W

4

ns

D

mCherry

y g

***

Normalized spine volume

C

TARPγ8
5

Normalized spine γ8

g

B

p

mCherry

to chemical LTP. Blue arrows indicate dendritic spines displaying g8 insertion
and large spine even in the basal state. Scale bar, 5 mm. (B) Quantification of
synaptic GFP-g8 expression in neurons transfected with WT or various mutant
constructs before and after cLTP. Normalized total synaptic spine g8 contents
by dendritic intensity are shown [WT: n = 5, Basal 1.218 ± 0.098 A.U. cLTP
(10 mM) 1.556 ± 0.157 A.U, cLTP (200 mM) 3.237 ± 0.099 A.U.; LDKD: n = 6,
Basal 1.356 ± 0.097 A.U. cLTP (10 mM) 3.164 ± 0.285 A.U, cLTP (200 mM)
3.540 ± 0.410 A.U. DPDZ: n = 6, Basal 1.853 ± 0.221 A.U., cLTP (10 mM)
1.996 ± 0.243 A.U, cLTP (200 mM) 2.748 ± 0.137 A.U.; LDKD+DPDZ: n = 5,
Basal 2.474 ± 0.353 A.U., cLTP (10 mM) 3.075 ± 0.300 A.U, cLTP (200 mM)
9 of 15

RES EARCH | R E S E A R C H A R T I C L E

3.312 ± 0.149 A.U.]. (C) Quantification of the average change in spine volume as
measured by mCherry cell fill in neurons transfected with WT or various mutant
constructs before and after cLTP. Normalized total synaptic mCherry by dendritic
intensity are shown [WT: n = 5, Basal 1.139 ± 0.070 A.U. cLTP (10 mM) 1.538 ±
0.160 A.U, cLTP (200 mM) 2.974 ± 0.109 A.U.; LDKD: n = 6, Basal 1.220 ±
0.092 A.U. cLTP (10 mM) 2.868 ± 0.142 A.U, cLTP (200 mM) 3.586 ± 0.249 A.U.
DPDZ: n = 6, Basal 1.816 ± 0.270 A.U., cLTP (10 mM) 1.954 ± 0.267 A.U,
cLTP (200 mM) 3.098 ± 0.242 A.U.; LDKD+DPDZ: n = 5, Basal 2.771 ± 0.225 A.U.,
cLTP (10 mM) 2.698 ± 0.152 A.U, cLTP (200 mM) 3.022 ± 0.156 A.U.].
(D) Quantification of synaptic SynGAP expression in neurons transfected with

Discussion

y
,

10 of 15

y g

1 March 2024

y

Araki et al., Science 383, eadk1291 (2024)

over the lifetimes of the animals tested. Here,
we have disentangled SynGAP signaling function from its structural properties both in vitro
and in vivo by generating mice with inactivating GAP mutations. The heterozygous
Syngap1+/GAP* mice have reduced SynGAP
GAP activity comparable to the heterozygous
KO mice but have normal total SynGAP protein expression and displayed normal LTP and
no apparent deficits in several behaviors despite diminished GAP activity. Moreover, the
homozygous Syngap1GAP*/GAP* mice are viable
and have normal LTP and behavior, indicating that LTP and viability are independent of
the GAP activity. These results indicate that
SynGAP binding in the PSD is required for
normal plasticity and cognition by regulating the number of PSD slots available for
binding TARP-AMPAR complexes and in turn
directly regulating synaptic strength.
Our data suggest that the GAP-dependent
signaling functions of SynGAP are important
for spine size changes during LTP. Further work
using these mice and other approaches is
needed to understand the role of SynGAP GAP
activity in brain function. SYNGAP1-related
intellectual disability has been classified as a
RASopathy resulting from loss of function
of the SYNGAP1 gene (32). Several therapeutic strategies to ameliorate aberrant biochemical signaling downstream of Ras as a result of
SYNGAP1 haploinsufficiency have been tested
(33, 34). However, the efficacy of this treatment approach remains inconclusive. Our
new data suggest that pharmacologically correcting dysregulated downstream GAP signaling of SynGAP may not be sufficient to
rescue disease phenotypes because these
strategies do not address the reduced PSD
slot occupancy by SynGAP haploinsufficiency.
In searching surveys of various ages and ethnicities [a total of 687,000 entries encompassing GnomAD (35), TOPMED (36), 8.3KJPN (37),
and ALFA], we found five human SYNGAP1
single nucleotide variant carriers with GAPdisabling mutations (17) (rs1224277120 C>T:
R485C, rs1248933822 G>A: R485H) that were
not associated with any neurological or mental diagnosis. This result suggests that heterozygous SynGAP mutations that impair GAP
signaling might be neither lethal nor sufficient
to lead to an neurodevelopmental disorder

g

The synaptic RasGAP SynGAP is essential for
synaptic plasticity and learning and memory,
and mutations in SYNGAP1 cause intellectual
disability, autistic-like behaviors, and epilepsy
in humans (16, 19, 29, 30). Recent studies have
shown that the dispersion of SynGAP from
synapses is required for the induction of LTP
(7). We have previously demonstrated that
SynGAP synaptic dispersion during cLTP
relieves the basal inhibition of synaptic Ras,
an important step that allows derepression of
ERK activity and AMPAR insertion (7). Whether SynGAP serves additional critical functions
for AMPAR recruitment beyond its GAP activity has been an open question. Moreover, a
comprehensive mechanistic understanding of
how AMPARs are up-regulated and maintained
at the synapse during LTP has remained elusive. The slot hypothesis of LTP suggests that
AMPAR-TARP complexes could bind to a finite number of available “slots” on scaffolding
proteins at the PSD (4–6, 9, 10). Here, we provide evidence for the pivotal role of SynGAP in
determining slot availability for the AMPARTARP complex independently of its GAP activity (see schematic model in fig. S10).
We have previously shown that phosphorylation of SynGAP by CaMKII is required for the
activity-dependent dispersion of SynGAP from
the PSD during synaptic plasticity (7). Here,
we reveal that AMPARs can be recruited to the
PSD after SynGAP dispersion in a manner that
is independent of the GAP activity of SynGAP.
Additionally, we found that SynGAP binding
to PSD-95 competes with TARPs for binding
to PSD-95 through the coiled-coil mediated
multivalent interaction that forms PSD-LLPS
(LLPS) and PDZ-ligand mediated protein binding (PDZ), and this antagonistic relationship is
regulated by CaMKII phosphorylation sites on
SynGAP (fig. S10). These data indicate that

CaMKII can act to molecularly tune PSD-95
binding partners at the PSD. Our data suggest that CaMKII may differentially regulate
the affinities and condensation properties of
SynGAP and TARPs for PSD-95 to promote
the recruitment of AMPARs. The tuning of
condensation properties is an attractive model
with which to describe the rapid and dynamic
changes to PSD composition and receptor density that occur during synaptic plasticity. Finally,
we showed that the elimination of SynGAP
GAP activity does not disrupt LTP in CA1 of the
hippocampus, and several behavioral phenotypes are normal in Syngap1 GAP mutant KI
mice, suggesting that the structural function
of SynGAP is a critical feature of its ability to
regulate synaptic plasticity and to promote
normal cognition.
These data strongly suggest that SynGAP is
a dominant driver of TARP enrichment in reconstituted condensates, because its CaMKIIdependent dispersion from synapses results in
the recruitment of TARP-g8 to synapses. It is
known that CaMKII phosphorylates TARP
C-terminal domains during synaptic plasticity (6). The phosphorylation of the TARP-g2
C terminus by CaMKII enhances its binding
affinity for PSD-95 and AMPAR activity at
synapses (31). Thus, it seems plausible that
CaMKII phosphorylation of TARP C termini
contributes to the compositional switch that
we observed. However, we found that mutation of two key CaMKII sites on SynGAP eliminated the CaMKII-dependent dispersion of
SynGAP and the recruitment of TARP-g8. In
addition, a recent report suggested that TARPg8 phosphorylation disrupts phase separation
with PSD-95, resulting in decreased clustering
(25). Additional experiments are required to
explore the potential contribution of TARP
phosphorylation in condensate composition
switching.
Previous studies have reported that the reduced SynGAP expression in heterozygous KO
mice is associated with increased concentrations of TARPs and AMPARs within the PSDs
of forebrain neurons in vivo, suggesting a
potential competition between SynGAP and
TARPs for slots (9, 10). However, experimental
results using Syngap1 heterozygous KO mice
are confounded by the effects of persistent upregulation of synaptic small GTPase activity

p

nor the LDKD mutant exhibited substantial
effects on the synaptic targeting of SynGAP
(fig. S9). These results show that the LDKD
mutation that modulates SynGAP’s ability to
compete with g8 for condensate formation with
PSD-95 can regulate the threshold for recruitment of g8 during cLTP induction, demonstrating that SynGAP’s ability to undergo LLPS is
critical for LTP expression.

WT or various mutant constructs before and after cLTP. Normalized total
synaptic spine SynGAP contents by dendritic intensity are shown [WT: n = 5,
Basal 4.530 ± 0.296 A.U. cLTP (10 mM) 4.194 ± 0.449 A.U, cLTP (200 mM)
1.852 ± 0.326 A.U.; LDKD: n = 6, Basal 3.909 ± 0.284 A.U. cLTP (10 mM) 1.569 ±
0.216 A.U, cLTP (200 mM) 1.254 ± 0.075 A.U. DPDZ: n = 6, Basal 3.646 ±
0.389 A.U., cLTP (10 mM) 3.348 ± 0.497 A.U, cLTP (200 mM) 1.392 ± 0.080 A.U.;
LDKD+DPDZ: n = 5, Basal 1.422 ± 0.120 A.U., cLTP(10 mM) 1.128 ± 0.080 A.U,
cLTP (200 mM) 1.098 ± 0.033 A.U.]. Two-way ANOVA with repeated measures
for chemical LTP treatment and multiple-comparisons with Šídák’s test were
used. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; n.s., not significant.

RES EARCH | R E S E A R C H A R T I C L E

diagnosis, unlike typical SYNGAP1 loss-offunction mutations (38, 39) and is consistent
with our results here. Our data indicate that
future therapeutic strategies for the treatment of SYNGAP1-related intellectual disability should focus on increasing the amount
of total SynGAP protein generated from the
spared allele. These strategies will be complicated because SynGAP is expressed as a
heterogeneous collection of structural isoforms
that serve distinct functions in neuronal development and synaptic plasticity (40). Future studies will be needed to determine the
structural and functional requirements for
a complete rescue of SYNGAP1 haploinsufficiency phenotypes, and this will help to guide
the development of treatments for SYNGAP1related intellectual disability and other severe
neurodevelopmental disorders.

1 March 2024

11 of 15

,

Total SynGAP was detected by the antibody
obtained from abcam (catalog #ab3344; 1:1000
dilution). Phospho-ERK and total-ERK were
detected by the antibody obtained from Cell

Live COS cells and cultured hippocampal neurons were imaged using a Cell Observer spinning disk confocal microscope (Carl Zeiss) or
an LSM 880 laser scanning confocal microscope (Carl Zeiss). For live imaging of COS
cells, cells were plated on collagen-coated
18-mm glass coverslips 24 to 36 hours before
the start of the experiment. COS cells were
transiently transfected with cDNA constructs
encoding proteins of interest using Lipofectamine
2000 (Invitrogen) 16 to 24 hours before the
start of the experiment. Coverslips containing

y

Araki et al., Science 383, eadk1291 (2024)

Antibodies

Confocal live cell imaging

y g

Brain tissue was excised from 5-month-old
mice. Tissue was lysed in 10 volumes of lysis
buffer (50 mM Tris, pH 8.0, 100 mM NaCl,
1 mM EDTA, 1 mM EGTA, 1% Triton X-100,
0.2% SDS, 0.5% sodium deoxycholate, with
cOmplete Protease inhibitor EDTA-free mix;
Roche/Sigma) and PhosSTOP phosphatase inhibitor mixture (Sigma, catalog #4906847001)
with a Dounce A homogenizer. Protein concentrations were measured with the Pierce
BCA assay kit (Pierce, catalog #23225). SDS
sample buffer (5×, 250 mM Tris, pH 6.8,
20% v/v glycerol, 10% w/v SDS, 12% v/v
b-mercaptoethanol, and 0.05% w/v bromophenol blue) was added to each sample (10 mg

Small GTPase (Ras) activity was measured
using a small GTPase(Ras)–GTP pull-down
assay. DNA constructs expressing a small G
protein (Ras) and SynGAP constructs (WT,
GAP*) were co-transfected into HEK293T
cells for 48 to 72 hours. Active Ras levels
were then assayed using a Ras activation
assay kit (EMD Millipore, catalog #17-218).
In brief, cells were lysed in Mg2+ lysis and
wash buffer (25 mM HEPES, pH 7.5, 150 mM
NaCl, 1% Igepal CA-630, 10 mM MgCl2, 1 mM
EDTA, and 10% glycerol), and active GTPbound small G-proteins were pulled down
using beads covalently bound to effector domains (Raf-1 RBD agarose). After washing
beads, active GTP-bound small G proteins
were recovered through the addition of 2×
SDS sample buffer followed by SDS-PAGE
and subsequent immunoblotting for the Ras
(anti-Ras, clone RAS10, EMD Millipore, catalog
#05-516, 1:1000 dilution).

COS cells or HEK293T cells originally obtained
from ATCC (COS cells: catalog #CRL-1651;
HEK293T cells: catalog #CRL-3216) were thawed
from liquid nitrogen and maintained in a medium consisting of Dulbecco’s modified Eagle’s
medium (DMEM) supplemented with 10%
fetal bovine serum (FBS) (Hyclone), and 1%
penicillin-streptomycin (10,000 U/ml) (Thermo
Fisher Scientific). Cells were maintained at
37°C in an incubator with 5% CO2 and passaged <20 total times. For the g8-PSD95-SynGAP
competition assay, the cells were fixed by 4%
paraformaldehyde and phosphate-buffered
saline (PBS) for 20 min at room temperature
after 16 to 24 hours of transfection. After three
washes with PBS, coverslips were mounted
in PermaFluor mounting medium (Thermo
Fisher Scientific). Cultured primary hippocampal neurons were prepared as described
previously (7) with some modifications. Briefly,
hippocampi were dissected from embryonic
day 18 (E18) rat pups (both female and male;
strain name: Sprague Dawley, envigo/inotiv,
catalog #002) before dissociation by papain
treatment and mechanical trituration. Cells
were plated on coverslips treated with 25 mM
poly-L-Lysine in Neurobasal medium (Gibco)
supplemented with 5% horse serum (Hyclone),
1% penicillin-streptomycin (10,000 U/ml) (Thermo Fisher Scientific), 2 mM GlutaMAX Supplement (Thermo Fisher Scientific), and 2% B27
supplement (Gibco) (NM5). One day after
plating, the medium was replaced with medium lacking horse serum (NM0). Neurons
were maintained in an incubator at 37°C with
5% CO2 for up to 21 days, and medium (NM0)
was changed once per week over the duration
of the culture.

y

Western blotting (SDS–polyacrylamide
gel electrophoresis)

Ras activation assay

Cell culture

g

All constructs used in this study were generated using standard restriction cloning
protocols. GFP-TARP g8 was generated by
synthesizing a double-stranded gene fragment
(gBlock, IDT) with an engineered 5′ SalI site
and a 3′ NotI site, followed by restriction subcloning into the CMV-driven GFP-C3 expression vector. For experiments involving purified
proteins, coding sequences of proteins of
interest(SynGAP CC-PBM [156 amino acids
(24)], PSD95 full length (24), and g8 C-terminal
tail [198 amino acids (25)] were cloned into
pGEX-6p bacterial expression vectors to generate protein products with an N-terminal
GST tag with an engineered PreScission Protease recognition site for tag removal after
isolation. Expression constructs for GFP-g8
C-terminal tail (198 amino acids) (25) were
cloned into pEGFP expression vectors to generate protein products with an N-terminal
GFP tag. Other constructs used in the chemical LTP assay and COS cell competition assay were described previously (7, 40, 41) or
fig. S5.

Signaling Technology (catalog #9102 and #9101)
(1:1000 dilution). GAPDH was detected by the
antibody obtained from Cell Signaling Technology (catalog #2118) (1:1000 dilution). Ras
antibody was described in Ras activation assay
section. IRDye 680RD or 800CW donkey antimouse IgG and goat anti-Rabbit IgG secondary antibodies (LI-COR Biosciences) were
used to detect primary antibodies of the appropriate species.

p

Materials and methods
Molecular biology and cloning

of total protein) for a final 1× concentration.
Samples were sonicated with a probe sonicator before heating at 90°C for 5 min to
facilitate complete protein denaturation. Samples were then loaded into precast Bolt 4 to
12% gradient Bis-Tris 1.0-mm gels (Invitrogen,
Thermo Fisher Scientific, catalog #NW04120BOX)
soaked in Bolt MOPS (Invitrogen, Thermo
Fisher, catalog #B000102) or MES SDS running buffers (Invitrogen, Thermo Fisher Scientific, catalog #B0002) depending on the
molecular weights of the proteins of interest.
Proteins were separated on the basis of molecular weight by SDS–polyacrylamide gel
electrophoresis (SDS-PAGE). Proteins were
then transferred to a 0.2-mm nitrocellulose
membrane (Amersham Protran NC). Membranes were blocked with Intercept Trisbuffered saline blocking buffer (LI-COR
Biosciences) for 30 min; incubated with
primary antibodies targeted to the proteins
of interest in a solution of Tris-buffered saline
with 0.1% Tween-20 (TBST) with 1.5% w/v
bovine serum albumin (BSA) and 0.1% sodium
azide overnight at 4°C with gentle rocking;
washed with TBST three times for 5 min each
before incubation with species-specific fluorescent secondary antibodies (LI-COR Biosciences) directed against bound primary
antibodies for 1 hour at room temperature
in the dark; washed with TBST three times
for 10 min each, followed by Tris-buffered
saline 1 time for 5 min; and then imaged using
the Odyssey CLx Imaging system. Protein bands
were quantified using Image Studio software.

RES EARCH | R E S E A R C H A R T I C L E

cLTP stimulation

All animals (rats and mice) used in this study
were housed in the Johns Hopkins University
School of Medicine (JHU SOM) animal facility
according to Institutional Animal Care and
Use Committee guidelines. SynGAP+/− KO mice
were generated using a BAC transgene at the
Transgenic Core Facility of JHU SOM (19). Mice
were backcrossed onto the C57BL/6 background.
SynGAP+/GAP* and SynGAPGAP*/GAP* mice were
engineered using CRISPR/Cas9 genome editing, resulting in two amino acid changes
(F484A and R485L) on a mixed C57BL/6J
background. Cas9 was targeted to the genomic
sequence of interest using the following guide
RNA (gRNA) sequence: 5′-GCGTGTTCTCTCGGAATATG-3′. The following 94-bp edited
oligo donor template containing nine point
mutations (five silent and four to generate
FR→AL mutations) was introduced: 5′-ATCAGT
CTCATATACTCTTCTATGGCTTTAGTGGCTAGCGTATTCTCGAG AGCGATTAAGTGTTCCCGCTCCATGAATCGGTCTACCTCTGACA-3′. The protospacer adjacent motif (PAM)
was destroyed, and two restriction sites (XhoI
and NheI) were engineered downstream of
the FR→AL mutations for founder screening.
Animals

SynGAP heterozygous KO mice that have been
previously described (19), SynGAP GAP mutants (+/GAP*, GAP*/GAP*), and WT (+/+)
12 of 15

,

1 March 2024

Generation of Syngap1-GAP-mutant KI mice

y

Live imaging and quantification of LTP were
performed as described previously (7). Hippocampal neurons from E18 rats were seeded on
25-mm poly-L-lysine–coated coverslips. The
cells were plated in Neurobasal medium (Gibco)
containing 50 U/ml penicillin, 50 mg/ml streptomycin, and 2 mM GlutaMax supplemented
with 2% B27 (Gibco) and 5% horse serum (Hyclone). Starting at 6 DIV, cells were maintained in glia-conditioned NM1 [Neurobasal
media with 2 mM GlutaMax, 1% FBS, 2%
B27, 1× FDU (5 mM uridine, Sigma catalog
#F0503), and 5 mM 5-fluro-2 -deoxyuridine
(Sigma catalog #U3003)]. Cells were transfected at 17 to 19 DIV with Lipofectamine
2000 (Invitrogen) in accordance with the manufacturer’s instructions. After 2 days, coverslips
were placed on a custom perfusion chamber
with basal ECS [143 mM NaCl, 5 mM KCl,
10 mM Hepes, pH 7.42, 10 mM glucose, 2 mM
CaCl2, 1 mM MgCl2, 0.5 mM tetrodotoxin (TTX),
1 mM strychnine, and 20 mM bicuculline], and
time-lapse images were acquired with either
an LSM880 (Carl Zeiss) or a spinning disk
confocal microscope controlled by Axiovision
software (Carl Zeiss). After 5 to 10 min of baseline recording, cells were perfused with 10 ml
of glycine without magnesium ECS (143 mM
NaCl, 5 mM KCl, 10 mM HEPES, pH 7.42, 10 mM
glucose, 2 mM CaCl2, 0 mM MgCl2, 0.5 mM TTX,
1 mM strychnine, 20 mM bicuculline, and 200 mM
glycine) for 10 min, followed by 10 ml of basal
ECS. To stabilize the imaging focal plane for
long-term experiments, we used Definite focus
(Zeiss). For quantification, we selected pyram-

y g

Araki et al., Science 383, eadk1291 (2024)

Purified proteins were labeled by iFluor-488/
568/633 succinimidyl ester (AAT Bioquest, catalog #1023/1049/1030) as previously described
(28). Fluorescence-labeled protein was mixed
with corresponding unlabeled protein at a ratio
of 1:20. For the imaging assay, g8 C-terminal
tail (50 mM) and PSD95 full-length (10 mM), and
SynGAP CC-PBM (10 mM) were mixed in cleavage buffer plus 5% polyethylene glycol in a
total volume of 10 ml. Each condition mixture
was injected into a flowmetry chamber composed of a #1.5H coverslip (Carl Zeiss, catalog
#474030-9000) and a slide glass separated by
two layers of double-sided tape as a spacer).
A Zeiss LSM880 confocal microscope [20×
Plan-Apochromat M27, air, numerical aperture (NA) 0.8, or 63× Plan-Apochromat M27,
oil-immersion, NA 1.4, Carl Zeiss] was used
for DIC and fluorescent imaging at room temperature. ImageJ software was used for analyzing images.

y

GST-tagged SynGAP CC-PBM (156 amino
acids) (24), PSD95 full length (24), and the g8
C-terminal tail (198 amino acids) (25) were
expressed in Escherichia coli BL21 in Luria
broth medium at 37°C for ~3 hours. Protein
expression was induced by 0.1 mM IPTG (final
concentration) at OD600 around 1.0 at 30°C.
Proteins were purified using glutathione agarose affinity column, and GST tags were cleaved
by PreScission Protease (Cytiva, catalog #270843-01) with a cleavage buffer containing
50 mM Hepes, pH 7.4, 50 mM NaCl, 1 mM
EDTA, and 1 mM dithiothreitol. Purified proteins eluted from the affinity column were
then collected and their concentrations determined by Nanodrop. g8 proteins were further

Protein fluorescence labeling and an
imaging-based assay of phase separation

idal neurons on the basis of morphology that
consisted of a clear primary dendrite and quantified all spines on the 30- to 40-mm stretch of
the secondary dendrite beginning just after the
branch from the primary dendrite. For identifying spine regions, we used the mCherry
channel to select the spine region that was
well separated from the dendritic shaft. These
ROIs in the mCherry channel were transferred
to the green/blue channel to quantify either
SEP-GluA1, GFP-g8, or Azurite-SynGAP content in synaptic spines. Total signals of each
channel were normalized by the signal density of adjacent dendritic regions to control
for different expression levels of each protein
in different cells. Total relative spine volume
[in arbitrary units (A.U.)] was calculated as
follows: average red signal at ROI (synaptic
spine) – average red density at background
region) * (area of ROI)/(average red signal at
ROI (adjacent dendritic region) – average red
density at background region). Total SEPGluA1, GFP-g8, or Azurite-SynGAP content at
synaptic spine was calculated as follows: average signal at ROI (synaptic spines) – average
density at background region) * (area of ROI)/
(average signal at ROI (adjacent dendritic
region) – (average density at background region).
All values are represented as the relative ratio
to the value of WT rescue before stimulation
(as 1.0).

g

Protein purification

concentrated using an Amicon Ultra centrifugal filter with MWCO 3K (Merck Millipore,
catalog #UFC500324).

p

cells were preincubated for 20 min in basal
extracellular solution (ECS) (150 mM NaCl,
3 mM KCl, 2 mM CaCl2, 10 mM HEPES, pH 7.4,
10 mM D-(+)-glucose, and 1 mM MgCl2) at
37°C with 0% CO2 before being transferred
to a custom-made imaging chamber filled
with basal ECS. Experiments were performed
at 37°C. For the FRAP assay, single z plane
images were taken while focused on the basal
coverglass-adjacent plasma membrane with
the same 63× objective and 2× digital zoom.
Images were acquired every 10 s using 488-,
561-, and 633-nm lasers. Five baseline images were acquired before photobleaching at a
single circular region of interest (ROI) drawn
by the experimenter (~2.5 mm square shape).
Images were analyzed using ImageJ. All FRAP
ROI mean intensity values were normalized to
both the intensity before bleach and bleach
depth to extract the recovery fraction. Cultured hippocampal neurons were transiently
transfected with the desired cDNA constructs
using Lipofectamine 2000 at 17 to 19 DIV 36
to 48 hours preceding the start of the experiment. The molecular replacement of SYNGAP1
in primary cultured neuron was described previously (7). Briefly, SynGAP pLKO KD vector
with the shRNA sequence (SYNGAP1-#5: CCT
GGA TGA AGA CTC CAT TAT, which corresponds to nucleotides 597 to 617 of rat SynGAP:
numbering according to NP_851606) was cotransfected with the Azurite-tagged shRNAresistant SynGAP (with TCT CGA GGA TTC
TAT CAT; silent mutations are indicated by
bold font at same positions of rat SynGAP)
with either WT or various mutants as rescue
constructs together with mCherry and SEPGluA1 or GFP-g8 before starting cLTP experiments. About 70% of endogenous SynGAP
was knocked down by this treatment (7), which
is well over the 50% loss that is universally
observed in SRID patient model mice (42).
Coverslips containing cultured neurons were
subjected to the same preincubation protocol
applied to COS cells before being loaded into
the imaging chamber.

RES EARCH | R E S E A R C H A R T I C L E

littermates were maintained on a mixed background of C57/B6J and 129/SvEv background
strains. Animals were allowed ad libitum access to food and water and reared on a typical
12-hour light-dark cycle. All animal experiments used both male and female mice (aged
3 to 7 months) and were conducted in accordance with the guidelines implemented by
the Institutional Animal Care and Use Committee at JHU under the protocol numbers
MO20M372, MO23M52, and RA23M46.
Acute slice preparation

Each test mouse was placed in a photobeamequipped (16 × 16 configuration with equal
spacing of 2.54 cm) plastic chamber (45 ×
45 cm) and allowed to explore free from interference for 120 min. Ambulatory movements
were tracked and analyzed using the Photobeam Activity System – Open Field (San Diego
Instruments).
Y-maze spontaneous alternation task

After a 30-min acclimatization period, mice
were placed in the center of a Y-maze in which
the three arms were oriented 120 degrees from
one another. Mice were allowed to explore the
apparatus for 5 min. Arm entries were recorded
when both of the mouse’s rear paws passed
over the boundary line between the apparatus’s center region and arm region. An arm
entry was recorded as an alternation when
the mouse fully entered an arm that it had not
visited most recently (e.g., arm A → arm B →
arm C = alternation; arm A → arm B → arm
A = not alternation). The percent alternation
was calculated as the number of alternating
arm entries divided by the total number of arm
entries. The Y-maze apparatus was thoroughly cleaned between trials. Arm entries were
13 of 15

,

1 March 2024

Open field task

y

AAV: AAV2/9-hSyn1-DIO-eGFP-CW3SL (45)
(titer: 1.79 × 1013 GC/ml) and AAV1.CamKII0.4.
Cre.SV40 (https://www.addgene.org/105558/)
(titer: 1.6 × 109 GC/ml) were injected to bilateral dorsal CA1 (coordinates: AP 1.9 mm;
ML 1.6 mm DV 1.2 mm) for 0.5 ml of AAV mixture per each position into C57B6J WT and
SynGAP GAP*/GAP* mice (2 months old) to
sparsely label the pyramidal neurons. After
9 days of injection, mice were transcardially
perfused with PBS followed by fixation solution
(4% paraformaldehyde in PBS) under isoflurane anesthesia immediately before decapitation. The brain was removed from the skull

Mice aged 2 to 4 months were subjected to
behavioral tests, including the open-field test,
to assess locomotion, the Y-maze spontaneous
alternation task to assess working memory
performance, and contextual plus cued fear
conditioning to assess associative memory. All
groups were approximately evenly divided
(45 to 55%) between males and females. All
animals were housed in the JHU SOM Miller
Research Building (MRB) animal facility. All
behavioral assays were performed in the JHU
SOM Animal Behavioral Core.

y g

Araki et al., Science 383, eadk1291 (2024)

Sparse GFP labeling of CA1 hippocampal
neurons and measure of synaptic spine volume

Behavior

y

Slices were placed in a submersion recording
chamber with recirculating aerated ACSF at
30°C. Synaptic field excitatory postsynaptic potentials (fEPSPs) were evoked in response to
electrical stimulation of the Schaeffer collateral inputs using a bipolar theta glass Ag/
AgCl electrode (3 MΩ) containing ACSF. Inputoutput curves were used to determine the
half-maximal fEPSP amplitude, which was
the stimulation intensity used to measure the
fEPSP slope over a stable 20-min baseline
period in response to a single 0.2-ms stimulation pulse delivered every 30 s. (A stable
baseline period of 10 min with baseline fEPSP
slope not drifting by >10% was the minimum
required inclusion criteria for LTP recordings.) To induce LTP, four episodes of thetaburst stimulation (TBS) were triggered at
0.1 Hz. Each TBS episode consisted of 10 stimulus trains administered at 5 Hz, with one train
consisting of 4 pulses at 100 Hz. After TBS,
the fEPSP slope was measured for 60 min by
delivering single electrical pulses every 30 s.
The magnitude of LTP was quantified by

Whole-cell recordings were performed as
previously described (43, 44). Briefly, paired
littermates of WT and SynGAPGAP*/GAP* mice
(P23 to P24) were anesthetized, and 300-mmthick transverse hippocampal slices were prepared in dissection buffer containing 210 mM
sucrose, 7 mM glucose, 26.2 mM NaHCO3,
2.5 mM KCl, 1 mM NaH2PO4 , and 7 mM
MgSO4. Slices were recovered in a submersion
chamber filled with oxygenated ACSF (119 mM
NaCl, 26.2 mM NaHCO 3 , 11 mM glucose,
2.5 mM KCl, 1 mM NaH2PO4, 2.5 mM CaCl2,
and 1.3 mM MgSO4) at 36°C for 30 min before
recordings. During all recordings, slices were
perfused in ACSF2 in the presence of 100 mM
picrotoxin at the flow rate of ~3 ml/min at
room temperature.
The micropipettes (2.5 to 3.5 mOhm) were
made of borosilicate glass (World Precision
Instruments) with a Sutter micropipette puller
(P-97) and filled with internal solution (115 mM
Cs-MeSO 3 , 0.4 mM EGTA, 5 mM TEACl, 2.8 mM NaCl, 20 mM Hepes, 3 mM Mg-ATP,
0.5 mM Na2-GTP, 10 mM Na phosphocreatine,
and 5 mM QX-314). Hippocampal CA1 excitatory neurons were held at –80 mV in a wholecell mode. Next, 1 mM TTX and 50 mM APV was
applied into ACSF2. Signals were measured
5 min after whole-cell mode established with
MultiClamp 700B amplifier and digitized
using a Digidata 1440A digitizer (Molecular
Devices). All data were sampled and digitized
at 10 kHz with Clampex 11.2 software, filtered
at 1 kHz and analyzed with Clampfit 10.7. Cells
with <100 miniature events were discarded.
Rise time was calculated at 20 to 80% of
mEPSC rise phase and decay time was calculated at 10 to 90%.

g

Extracellular LTP recordings

Extracellular mEPSC recordings

and incubated in 4% paraformaldehyde overnight for postfixation. Coronal brain slices
(50 mm thickness) were prepared using a
vibratome (Leica VT1200S). Brain slices were
mounted in PermaFluor mounting medium
(Thermo Fisher Scientific) on glass slides. Images were acquired with either LSM880 (Carl
Zeiss) (5× objective lens for entire hippocampus, 5-mm z-section for entire 50-mm thickness)
or spinning disk confocal microscopes controlled by Axiovision software (Carl Zeiss) (10×
objective lens for CA1 region; 2-mm z-section
for total 30-mm thickness, 63× objective lens for
synaptic spine imaging; 0.5-mm z-section for
total 30-mm thickness). Total relative spine
volume (arbitrary unit) was calculated as follows: average green signal at ROI (synaptic
spine) – average green density at background
region) * (area of ROI)/(average green signal
at ROI adjacent dendritic region) – (average
green density at background region).

p

SynGAP+/GAP*, SynGAPGAP*/GAP*, and SynGAP+/−
mice (3 to 7 months of age), along with their
respective WT (Syngap1+/+) littermates, were
transcardially perfused with ice-cold aerated
dissection buffer (212.7 mM sucrose, 5 mM
KCl, 1.25 mM Na2PO4, 10 mM glucose, 26 mM
NaHCO3, 0.5 mM CaCl2, 10 mM MgCl2) under
isoflurane anesthesia immediately before decapitation. The brain was rapidly removed
from the skull, and the hippocampi were extracted in a continuously oxygenated dissection
buffer. Acute transverse hippocampal slices
(400 mm thickness) were prepared using a vibratome (Leica VT1200S) and briefly washed of
the sucrose-based dissection buffer in aerated
artificial cerebrospinal fluid (ACSF) composed
of the following (in mM): 119 NaCl, 5 KCl,
1.25 Na2PO4, 26 NaHCO3, 10 glucose, 2.5 CaCl2,
and 1.5 MgCl2. Slices were then placed in a chamber containing ACSF at 30°C for 30 min and
then transferred to room temperature for at
least 60 min until being used for electrophysiological recordings. The experimenter was
blinded to the animals’ genotypes until all experiments and analyses were completed.

normalizing the fEPSP slope to the average
baseline response, then calculating the average fEPSP slope between 40 and 60 min after
TBS. Statistical comparisons were made exclusively between WT and mutant littermates
with a Student’s t test, Mann-Whitney test, or
one-way ANOVA.

RES EARCH | R E S E A R C H A R T I C L E

recorded manually, and the experimenter was
blind to the experimental conditions. The normality of the data was assessed using the
Kolmogorov-Smirnov test, and groups were
statistically compared using a one-way
ANOVA followed by Tukey’s post hoc multiplecomparisons test.
Contextual and cued fear conditioning

,

14 of 15

y

1 March 2024

y g

Araki et al., Science 383, eadk1291 (2024)

y

All statistical analyses were performed using
GraphPad Prism software. Graphs were prepared using GraphPad Prism 8.0/9.0/10.0 Software or Microsoft Excel 16.36. In the figures,
all error bars and shadows represent the SEM
unless otherwise stated. We followed International Committee of Medical Journal Editors
(ICMJE) guidelines for units of measurement
[International Systems of Units (SI)]. The normality of the data was assessed using the
Shapiro-Wilk test unless stated otherwise.
Details regarding sample numbers (mouse
number, neurons, and synaptic spines that
were imaged) and experimental replicates
were described in each figure legend. We did
not exclude any outliers. We estimated sample

1. H. W. Kessels, R. Malinow, Synaptic AMPA receptor plasticity
and behavior. Neuron 61, 340–350 (2009). doi: 10.1016/
j.neuron.2009.01.015; pmid: 19217372
2. R. L. Huganir, R. A. Nicoll, AMPARs and synaptic plasticity: The
last 25 years. Neuron 80, 704–717 (2013). doi: 10.1016/
j.neuron.2013.10.025; pmid: 24183021
3. R. A. Nicoll, A brief history of long-term potentiation.
Neuron 93, 281–290 (2017). doi: 10.1016/j.neuron.2016.12.015;
pmid: 28103477
4. S. Shi, Y. Hayashi, J. A. Esteban, R. Malinow, Subunit-specific
rules governing AMPA receptor trafficking to synapses in
hippocampal pyramidal neurons. Cell 105, 331–343 (2001).
doi: 10.1016/S0092-8674(01)00321-X; pmid: 11348590
5. J. Lisman, S. Raghavachari, A unified model of the presynaptic
and postsynaptic changes during LTP at CA1 synapses.
Sci. STKE 2006, re11 (2006). doi: 10.1126/stke.3562006re11;
pmid: 17033044
6. P. Opazo, M. Sainlos, D. Choquet, Regulation of AMPA receptor
surface diffusion by PSD-95 slots. Curr. Opin. Neurobiol. 22,
453–460 (2012). doi: 10.1016/j.conb.2011.10.010;
pmid: 22051694
7. Y. Araki, M. Zeng, M. Zhang, R. L. Huganir, Rapid dispersion of
SynGAP from synaptic spines triggers AMPA receptor insertion
and spine enlargement during LTP. Neuron 85, 173–189
(2015). doi: 10.1016/j.neuron.2014.12.023; pmid: 25569349
8. Y. Yang, J. H. Tao-Cheng, T. S. Reese, A. Dosemeci, SynGAP
moves out of the core of the postsynaptic density upon
depolarization. Neuroscience 192, 132–139 (2011).
doi: 10.1016/j.neuroscience.2011.06.061; pmid: 21736925
9. W. G. Walkup IV et al., A model for regulation by SynGAP-a1 of
binding of synaptic proteins to PDZ-domain ‘Slots’ in the
postsynaptic density. eLife 5, e16813 (2016). doi: 10.7554/
eLife.16813; pmid: 27623146
10. T. L. Mastro et al., A sex difference in the response of the
rodent postsynaptic density to synGAP haploinsufficiency.
eLife 9, e52656 (2020). doi: 10.7554/eLife.52656;
pmid: 31939740
11. J. H. Kim, D. Liao, L. F. Lau, R. L. Huganir, SynGAP: A synaptic
RasGAP that associates with the PSD-95/SAP90 protein
family. Neuron 20, 683–691 (1998). doi: 10.1016/S0896-6273
(00)81008-9; pmid: 9581761
12. H. J. Chen, M. Rojas-Soto, A. Oguni, M. B. Kennedy, A synaptic
Ras-GTPase activating protein (p135 SynGAP) inhibited by
CaM kinase II. Neuron 20, 895–904 (1998). doi: 10.1016/
S0896-6273(00)80471-7; pmid: 9620694

g

Statistics

RE FERENCES AND NOTES

13. T. R. Gamache, Y. Araki, R. L. Huganir, Twenty years of SynGAP
research: From synapses to cognition. J. Neurosci. 40,
1596–1605 (2020). doi: 10.1523/JNEUROSCI.0420-19.2020;
pmid: 32075947
14. M. Sheng, E. Kim, The postsynaptic organization of synapses.
Cold Spring Harb. Perspect. Biol. 3, a005678 (2011).
doi: 10.1101/cshperspect.a005678; pmid: 22046028
15. J. J. Zhu, Y. Qin, M. Zhao, L. Van Aelst, R. Malinow, Ras and
Rap control AMPA receptor trafficking during synaptic
plasticity. Cell 110, 443–455 (2002). doi: 10.1016/S0092-8674
(02)00897-8; pmid: 12202034
16. N. H. Komiyama et al., SynGAP regulates ERK/MAPK signaling,
synaptic plasticity, and learning in the complex with
postsynaptic density 95 and NMDA receptor. J. Neurosci. 22,
9721–9732 (2002). doi: 10.1523/JNEUROSCI.22-2209721.2002; pmid: 12427827
17. K. Scheffzek et al., The Ras-RasGAP complex: Structural basis
for GTPase activation and its loss in oncogenic Ras mutants.
Science 277, 333–339 (1997). doi: 10.1126/
science.277.5324.333; pmid: 9219684
18. G. Rumbaugh, J. P. Adams, J. H. Kim, R. L. Huganir, SynGAP
regulates synaptic strength and mitogen-activated protein
kinases in cultured neurons. Proc. Natl. Acad. Sci. U.S.A. 103,
4344–4351 (2006). doi: 10.1073/pnas.0600084103;
pmid: 16537406
19. J. H. Kim, H. K. Lee, K. Takamiya, R. L. Huganir, The role
of synaptic GTPase-activating protein in neuronal
development and synaptic plasticity. J. Neurosci. 23,
1119–1124 (2003). doi: 10.1523/JNEUROSCI.23-0401119.2003; pmid: 12598599
20. J. P. Clement et al., Pathogenic SYNGAP1 mutations impair
cognitive development by disrupting maturation of dendritic
spine synapses. Cell 151, 709–723 (2012). doi: 10.1016/
j.cell.2012.08.045; pmid: 23141534
21. X. Guo et al., Reduced expression of the NMDA receptorinteracting protein SynGAP causes behavioral abnormalities
that model symptoms of Schizophrenia.
Neuropsychopharmacology 34, 1659–1672 (2009).
doi: 10.1038/npp.2008.223; pmid: 19145222
22. R. Nakajima et al., Comprehensive behavioral analysis of
heterozygous Syngap1 knockout mice. Neuropsychopharmacol.
Rep. 39, 223–237 (2019). doi: 10.1002/npr2.12073;
pmid: 31323176
23. E. D. Ozkan et al., Reduced cognition in Syngap1 mutants is
caused by isolated damage within developing forebrain
excitatory neurons. Neuron 82, 1317–1333 (2014).
doi: 10.1016/j.neuron.2014.05.015; pmid: 24945774
24. M. Zeng et al., Phase transition in postsynaptic densities
underlies formation of synaptic complexes and synaptic
plasticity. Cell 166, 1163–1175.e12 (2016). doi: 10.1016/
j.cell.2016.07.008; pmid: 27565345
25. M. Zeng et al., Phase separation-mediated TARP/MAGUK
complex condensation and AMPA receptor synaptic
transmission. Neuron 104, 529–543.e6 (2019). doi: 10.1016/
j.neuron.2019.08.001; pmid: 31492534
26. B. Wang et al., Liquid-liquid phase separation in human health
and diseases. Signal Transduct. Target. Ther. 6, 290 (2021).
doi: 10.1038/s41392-021-00678-1; pmid: 34334791
27. S. Banjade, M. K. Rosen, Phase transitions of multivalent
proteins can promote clustering of membrane receptors.
eLife 3, e04123 (2014). doi: 10.7554/eLife.04123;
pmid: 25321392
28. M. Zeng et al., Reconstituted postsynaptic density as a
molecular platform for understanding synapse formation and
plasticity. Cell 174, 1172–1187.e16 (2018). doi: 10.1016/
j.cell.2018.06.047; pmid: 30078712
29. F. F. Hamdan et al., Mutations in SYNGAP1 in autosomal
nonsyndromic mental retardation. N. Engl. J. Med. 360,
599–605 (2009). doi: 10.1056/NEJMoa0805392;
pmid: 19196676
30. M. H. Berryer et al., Mutations in SYNGAP1 cause intellectual
disability, autism, and a specific form of epilepsy by inducing
haploinsufficiency. Hum. Mutat. 34, 385–394 (2013).
doi: 10.1002/humu.22248; pmid: 23161826
31. A. Sumioka, D. Yan, S. Tomita, TARP phosphorylation regulates
synaptic AMPA receptors through lipid bilayers. Neuron 66,
755–767 (2010). doi: 10.1016/j.neuron.2010.04.035;
pmid: 20547132
32. W. E. Tidyman, K. A. Rauen, Expansion of the RASopathies.
Curr. Genet. Med. Rep. 4, 57–64 (2016). doi: 10.1007/s40142016-0100-7; pmid: 27942422
33. E. H. Cook, J. T. Masaki, S. J. Guter, F. Najjar, Lovastatin
treatment of a patient with a de novo SYNGAP1 protein

p

Contextual and cued fear conditioning was
performed as previously described (21). Animals were handled (picked up and held by the
experimenter for 30 s) daily for 14 days before
training. Before this test, mice were kept in a
soundproof room separate from the testing
room for 30 min. To assess fear‐related learning and memory, mice were placed singly in
an acrylic chamber of PVC plastic walls (33 ×
25 × 28 cm) with a stainless‐steel grid floor
(0.2 cm diameter, spaced 0.5 cm apart). Between mice for conditioning and context testing, the walls and grids of the chamber were
wiped with 70% ethanol. In the cued test, the
walls and floor were cleaned with 1% acetic
acid. Animals were placed in the conditioning
chamber (Video Freeze; Med Associates) and
allowed to explore for 120 s, after which a tone
(90 dB) was presented for 30 s, which coterminated with a footshock (1 s, 0.5 mA). This
tone-shock pairing occurred three times per
session, with an intertrial interval of 90 s.
After the third and final shock, the animals
remained in the training chamber for 90 s.
Twenty-four hours later, the animals were returned to the chamber. Contextual memory
was assessed by the percentage of time spent
freezing for 600 s (no shock presentation).
Twenty-four hours later, the dimensions, as
well as the visual, tactile, and olfactory cues
(cleaning with 1% acetic acid) of the conditioning chamber, were altered to create a
novel context for the mice. Cued fear learning
was assessed in this novel context by measuring the average percent freezing per minute for 300 s without and 300 s with the
presentation of the 90 dB auditory cue (CS).

sizes on the basis of previous experiments in
the laboratory, which are comparable to the
other studies in the field. We randomly allocated culture dishes and mice in every set of
experimental conditions. Each data points are
shown in figures. All behavioral experiments
were conducted blind to the genotype and/or
treatment. Only after all data were collected
was a correspondence table opened, linking
the provisional experimental group names with
the formal names where the genotype and/or
treatment details were known. Western blot data
were analyzed using a one-way ANOVA followed
by Tukey’s post hoc multiple-comparisons test
unless stated otherwise. For all behavior testing, the normality of the data was assessed
using the Kolmogorov-Smirnov test. In the
Y-maze spontaneous alternation task, groups
were compared using a one-way ANOVA followed by Tukey’s post hoc multiple-comparisons
test. For contextual and cued fear conditioning, mutant mice were compared with their
WT littermates during the context trial using
a one-way repeated-measures ANOVA. With
and without cue was compared within each
genotype using a one-way ANOVA. *P < 0.05,
**P < 0.01, ***P < 0.001, ****P < 0.0001.

RES EARCH | R E S E A R C H A R T I C L E

34.

35.

36.

37.

38.

39.

41. X. Chen et al., Arc weakens synapses by dispersing AMPA
receptors from postsynaptic density via modulating PSD phase
separation. Cell Res. 32, 914–930 (2022). doi: 10.1038/
s41422-022-00697-9; pmid: 35856091
42. Y. Araki et al., Mouse models of SYNGAP1-related intellectual
disability. Proc. Natl. Acad. Sci. U.S.A. 120, e2308891120
(2023). doi: 10.1073/pnas.2308891120; pmid: 37669379
43. H. L. Tan, S. L. Chiu, Q. Zhu, R. L. Huganir, GRIP1 regulates
synaptic plasticity and learning and memory. Proc. Natl. Acad.
Sci. U.S.A. 117, 25085–25091 (2020). doi: 10.1073/
pnas.2014827117; pmid: 32948689
44. A. R. Graves et al., Visualizing synaptic plasticity in vivo by
large-scale imaging of endogenous AMPA receptors. eLife 10,
e66809 (2021). doi: 10.7554/eLife.66809; pmid: 34658338
45. I. Hong et al., Calcium-permeable AMPA receptors govern PV
neuron feature selectivity. bioRxiv, 2023.2007.2020.549908
(2023). doi: 10.1101/2023.07.20.549908
ACKN OWLED GMEN TS

We thank all members of the Huganir lab for discussion and support
throughout this work, especially S. Ju, R. Oba, S. Myung, S. H. Kwon,
and S. Glavaris for reagent preparation and critical reading of the
manuscript. We also thank the Bridge the Gap SYNGAP Education and
Research Foundation, the SynGAP Research Fund, and all the patients
of SYNGAP1-related intellectual disability and their families for their
outreach and advocacy. Funding: This work was supported by the
National Institutes of Health (grants R01MH112151 and R01NS036715
to R.H. and grant T32MH015330 to K.R.) and the SynGAP Research

Fund (R.H.). Author contributions: Conceptualization: Y.A., T.G.,
R.H.; Formal analysis: Y.A., K.R., E.G., T.G., R.J., T.H.N.T., Q.Z.; Funding
acquisition: R.H.; Investigation: Y.A., K.R., E.G., T.G., R.J., T.H.N.T.,
Q.Z.; Methodology: Y.A., B.L., I.H.; Project administration: R.H.;
Supervision: R.H., A.K.; Visualization: Y.A., K.R., E.G., T.G.; Writing –
original draft: Y.A., K.R., E.G., T.G., R.H.; Writing – review and editing:
Y.A., E.G., I.H., R.H. Competing interests: R.H. is a scientific
cofounder and scientific advisory board (SAB) member of Neumora
Therapeutics and SAB member of MAZE Therapeutics. Y.A, R.C.J,
I.H., and R.L.H are listed as inventors on pending patent claims
(US 2021/0180062, PCT/US2023/065370, and PCT/US2023/
065364) filed by Johns Hopkins University covering therapeutic
strategies for SYNGAP1-related intellectual disability. The remaining
authors declare no competing interests. Data and materials
availability: All data are available in the main text or the
supplementary materials. License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk1291
Figs. S1 to S10
MDAR Reproducibility Checklist
Submitted 4 August 2023; accepted 28 December 2023
10.1126/science.adk1291

p

40.

truncating variant. J. Child Adolesc. Psychopharmacol. 29,
321–322 (2019). doi: 10.1089/cap.2018.0159; pmid: 30901256
G. Kluger, C. von Stülpnagel-Steinbeis, S. Arnold,
K. Eschermann, T. Hartlieb, Positive short-term effect of
low-dose rosuvastatin in a patient with SYNGAP1-associated
epilepsy. Neuropediatrics 50, 266–267 (2019). doi: 10.1055/
s-0039-1681066; pmid: 30875700
K. J. Karczewski et al., The mutational constraint spectrum
quantified from variation in 141,456 humans. Nature 581, 434–443
(2020). doi: 10.1038/s41586-020-2308-7; pmid: 32461654
D. Taliun et al., Sequencing of 53,831 diverse genomes from
the NHLBI TOPMed Program. Nature 590, 290–299 (2021).
doi: 10.1038/s41586-021-03205-y; pmid: 33568819
S. Tadaka et al., jMorp updates in 2020: Large enhancement of
multi-omics data resources on the general Japanese
population. Nucleic Acids Res. 49, D536–D544 (2021).
doi: 10.1093/nar/gkaa1034; pmid: 33179747
D. R. M. Vlaskamp et al., SYNGAP1 encephalopathy: A
distinctive generalized developmental and epileptic
encephalopathy. Neurology 92, e96–e107 (2019). doi: 10.1212/
WNL.0000000000006729; pmid: 30541864
C. Mignot et al., Genetic and neurodevelopmental spectrum of
SYNGAP1-associated intellectual disability and epilepsy.
J. Med. Genet. 53, 511–522 (2016). doi: 10.1136/jmedgenet2015-103451; pmid: 26989088
Y. Araki et al., SynGAP isoforms differentially regulate synaptic
plasticity and dendritic development. eLife 9, e56273 (2020).
doi: 10.7554/eLife.56273; pmid: 32579114

g
y
y g
y
,

Araki et al., Science 383, eadk1291 (2024)

1 March 2024

15 of 15

RES EARCH

RESEARCH ARTICLE

◥

NEUROSCIENCE

Retrograde endocannabinoid signaling at inhibitory
synapses in vivo
Barna Dudok1,2*†, Linlin Z. Fan3†, Jordan S. Farrell2,4,5, Shreya Malhotra2,
Jesslyn Homidan2, Doo Kyung Kim3, Celestine Wenardy3, Charu Ramakrishnan6,
Yulong Li7, Karl Deisseroth3,8,9, Ivan Soltesz2

Departments of Neurology and Neuroscience, Baylor College
of Medicine, Houston, TX 77030, USA. 2Department of
Neurosurgery, Stanford University, Stanford, CA 94305, USA.
3
Department of Bioengineering, Stanford University,
Stanford, CA 94305, USA. 4F.M. Kirby Neurobiology Center
and Rosamund Stone Zander Translational Neuroscience
Center, Boston Children’s Hospital, Boston, MA 02115, USA.
5
Department of Neurology, Boston Children’s Hospital,
Harvard Medical School, Boston, MA 02115, USA. 6Cracking
the Neural Code (CNC) Program, Stanford University,
Stanford, CA 94305, USA. 7State Key Laboratory of Membrane
Biology, School of Life Sciences, Peking University, Beijing
100871, China. 8Department of Psychiatry and Behavioral
Sciences, Stanford University, Stanford, CA 94305, USA.
9
Howard Hughes Medical Institute, Stanford, CA 94305, USA.
*Corresponding author. Email: barna.dudok@bcm.edu
†These authors contributed equally to this work.

y
,

1 March 2024

y g

Fig. 1. Rapid eCB signaling in the hippocampus
in vivo. (A) GRABeCB2.0 and jRGECO1a were
expressed in CA1 neurons. Head-fixed mice
ran on a linear treadmill during multiphoton
imaging. (B) Event-aligned average single-cell
calcium and eCB responses during calcium
transients. Plot shows mean responses
(line) ± SEM (shaded area) for n = 4 sessions
from n = 4 vehicle-treated mice, 607 ± 241
ROIs per session, and 5.2 ± 1.1 peaks per ROI.
Labels show decay time constants of exponential
fits. (C) Analysis of place cells. Average tuning
curves (solid black line) were calculated
for each session by aligning location-averaged
place cell traces on their preferred location.
(D) Average spatial tuning curves (±SEM)
are shown centered on the preferred
location of place cells (red indicates calcium)
together with the tuning curves of eCB signals
from the same cells (blue) or after shuffling cells
within sessions (gray). One-sided, one-sample
t test with alternative hypothesis m > 0: P = 5.67 × 10–5,
n = 4 male mice; shuffle: P = 0.88. Plots show
average tuning curves (line) ± SEM (shaded area),
n = 4 sessions from n = 4 drug-naïve mice and
161 ± 35 place cell ROIs per session.

Dudok et al., Science 383, 967–970 (2024)

y

triggers eCB synthesis and the retrograde activation of cannabinoid type-1 receptors (CB1s),
which in turn suppresses GABA release. In the
CA1 region of the hippocampus, the highest
CB1 expression is found on axons of perisomatically projecting GABAergic basket cells that
also express cholecystokinin (CCKBCs) (6–8).
Conversely, the other major basket cell type,
parvalbumin-expressing basket cells (PVBCs),
do not express CB1s. Correspondingly, DSI is

1

g

S

trong depolarization of neurons can induce a transient suppression of their inhibitory synaptic inputs in acute brain
slices (1, 2). Such retrograde, activitydependent suppression of GABAergic
synapses, referred to as depolarization-induced
suppression of inhibition (DSI), is mediated
by endocannabinoid (eCB) signaling (3–5).
In vitro studies have shown that robust postsynaptic calcium (post-Ca) increase during DSI

p

Endocannabinoid (eCB)–mediated suppression of inhibitory synapses has been hypothesized, but
this has not yet been demonstrated to occur in vivo because of the difficulty in tracking eCB dynamics
and synaptic plasticity during behavior. In mice navigating a linear track, we observed locationspecific eCB signaling in hippocampal CA1 place cells, and this was detected both in the postsynaptic
membrane and the presynaptic inhibitory axons. All-optical in vivo investigation of synaptic responses
revealed that postsynaptic depolarization was followed by a suppression of inhibitory synaptic
potentials. Furthermore, interneuron-specific cannabinoid receptor deletion altered place cell tuning.
Therefore, rapid, postsynaptic, activity-dependent eCB signaling modulates inhibitory synapses on a timescale
of seconds during behavior.

maximally potent at CCKBC inputs to pyramidal cells and is capable of completely muting
these synapses (9, 10).
DSI has been hypothesized to also occur
in vivo, but the specific neuronal activity patterns that give rise to DSI remain unknown
(11). When mammals navigate their environment, individual hippocampal pyramidal cells
discharge at specific place fields (11, 12), and
several observations are consistent with the
possibility that place cell firing in behaving
animals may engage a DSI-like phenomenon.
In vitro, externally imposed place cell–like activity can drive DSI (13), and disinhibition of
the postsynaptic cells by DSI can facilitate excitatory synapse plasticity (14, 15). In vivo, place
cell formation is supported by reduced inhibition (16, 17). A potential role of DSI in hippocampal place field properties has been proposed

1 of 4

RES EARCH | R E S E A R C H A R T I C L E

(18); however, the steps that would underlie
a retrograde, eCB-mediated, DSI-like plasticity
in vivo have remained speculative, and the hypothesis that DSI contributes to place cell disinhibition has remained untested. Here, we
used optical methods in mice navigating a
linear track to test (i) whether place cell activity in behaving animals is sufficient to trigger eCB synthesis in the postsynaptic cell,
(ii) whether eCB signals affect presynaptic
CB1s on GABAergic terminals in vivo, and (iii)
whether DSI-like plasticity can modulate place
cell activity patterns.
Location-specific eCB signaling by place cells

,

2 of 4

y

1 March 2024

y g

Dudok et al., Science 383, 967–970 (2024)

y

Fig. 2. Spatially tuned presynaptic
eCB signals in the hippocampus
in vivo. (A) Labeling strategy
for in vivo imaging. Interneuronal
GRABeCB2.0 and pan-neuronal
jRGECO1a expression were combined.
Bottom panels show the segmentation approach. Neuron cell bodies
were segmented in the jRGECO1a
channel (post-Ca). The ROIs
were enlarged by binary dilation
for measuring signals in the
neighboring axons in the GRABeCB2.0
channel (pre-eCB). (B) Average
spatial tuning curves (±SEM) are shown
centered on the preferred location
of place cells (red indicates calcium)
together with the tuning curves of
eCB signals from the corresponding
pre-eCB ROIs (blue) or after shuffling
ROIs within sessions (gray), n = 18
sessions from n = 5 mice and
193 ± 130 ROIs per session.
(C) Quantification of signal intensity
at the preferred location. Boxes
indicate median ± interquartile range;
whiskers: nonoutlier range; markers:
recording sessions. pre-eCB: P = 0.002, n = 5 mice (n = 3 males and n = 2 females); shuffle: P = 0.69. (D) Spatial tuning curves are shown after injecting mice
with JZL-184 to inhibit the enzymatic breakdown of the eCB 2-AG by MGL or after vehicle injection. (E) Quantification of location-specific pre-eCB signals, P = 0.0004,
Mann-Whitney test, n = 14 vehicle sessions from n = 5 mice and n = 6 JZL sessions from n = 3 mice.

g

mobilization in place cells during exploration
is specific to the cell’s preferred location. By
contrast, non-place cells had lower calcium
and accompanying eCB transient amplitudes
compared with place cells in the same field of
view (fig. S1G).
Although the molecular mechanisms of retrograde eCB transport are not precisely understood, there is general agreement that DSI
requires the postsynaptically generated eCBs
to engage presynaptic CB1s on interneuronal
terminals impinging on the activated neuron
(4, 27). Thus, we specifically allowed the expression of GRABeCB2.0 only in interneurons using
Dlx5/6-Cre transgenic mice (28) to enable presynaptic eCB measurements (fig. S2A). The
distribution of GRABeCB2.0, a chimera of CB1
and a green fluorescent protein variant, resembled membrane-enriched CB1 targeting
(29) in interneuron axon terminals, with no
detectable postsynaptic expression in principal cells and relatively low expression in
interneuron somata (fig. S2, B to F).
For simultaneously imaging somatic calcium and axonal eCB transients, we combined
interneuronal GRABeCB2.0 and pan-neuronal,
red-shifted calcium sensor expression (Fig. 2A
and fig. S2E). We generated somatic, putatively post-Ca ROI sets as above and measured
nearby axonal, putatively presynaptic eCB
(pre-eCB) signals after enlarging the somatic

p

The genetically encoded G protein–coupled receptor activation based eCB reporter GRABeCB2.0
enables the recording of eCB dynamics with
high spatial resolution in vivo (19, 20). eCB mobilization during DSI depends on post-Ca influx
(21). To characterize eCB signaling related to
calcium transients, we expressed GRABeCB2.0
and the red-shifted calcium sensor jRGECO1a
(22) in CA1 neurons. We performed two-photon
dual calcium and eCB imaging in the pyramidal
layer while mice ran several laps on a linear
treadmill track with tactile cues (Fig. 1A) (23).
We segmented regions of interest (ROIs) corresponding to neuronal somata (most of which
in the pyramidal layer are expected to belong
to pyramidal cells) (24) and measured calcium

and eCB signals in the same ROIs. We analyzed calcium transients by finding peaks on
traces of fluorescence change over baseline
(DF/F) (Fig. 1B). Transient eCB signals were
detected concomitant with calcium peaks (Fig.
1B), with a peak delayed by 1.04 ± 0.16 s relative to calcium and an average decay time
constant of 3.53 ± 0.75 s. To investigate which
eCB ligand contributes to the transients, we
performed the latter analysis on datasets that
we previously recorded in the presence of ligandspecific inhibitors of eCB synthesis or metabolism (20). Calcium peak–coupled eCB transients
were suppressed by inhibiting the synthesis of
2-arachidonoylglycerol (2-AG), the eCB species
involved in CA1 DSI in vitro (25). Furthermore,
eCB transient durations were extended after
we treated mice with JZL 184 to inhibit monoacylglycerol lipase (MGL) and thus 2-AG degradation (25, 26) (fig. S1, A to C). Conversely,
manipulations altering the synthesis or degradation of the other major eCB species, anandamide (AEA), had no effect on the in vivo
eCB transients (fig. S1, D to F).
Next, to investigate eCB dynamics specifically in place cells, we identified place cells
by calculating location-specific average calcium
signals (Fig. 1C). Average eCB signals were
elevated around the same track locations where
calcium was high in the same individual place
cells (Fig. 1D). These results indicate that eCB

RES EARCH | R E S E A R C H A R T I C L E

ROIs (Fig. 2A). Similar to eCB signals measured in place cell somata (Fig. 1D), pre-eCB
signals in interneuronal axons surrounding
place cells were elevated at the same track
locations where post-Ca was high (Fig. 2, B
and C). These results indicate that place cell

activations during behavior are accompanied by eCB signaling at perisomatic inhibitory axons.
Similarly to DSI in vitro (30) and calcium
transient–related post-eCB signals in vivo (fig.
S1B), location-specific pre-eCB signals around

place cells were magnified by pharmacological inhibition of 2-AG degradation (Fig. 2, D
and E), consistent with a prominent role of
2-AG in inhibitory axon eCB signaling while
not ruling out the partial involvement of
other eCBs such as AEA.

p
g
y
y g
y
,

Fig. 3. Inhibitory synaptic plasticity in behaving mice. (A) Labeling strategy for
the all-optical assay of CCKBC synaptic function in vivo. (B) Top: example unfiltered
fluorescence traces from four CA1 neurons [(a) to (d)]. Bottom: spike raster (n =
30 neurons from n = 5 mice). Cyan bars indicate CCKBC photostimulation onset
(488 nm, 10 ms duration, 9.5 to 20 mW/mm2, 0.5 Hz). (C) Mean subthreshold
postsynaptic waveforms after presynaptic CCKBC photostimulation (n = 30 neurons
from n = 5 mice). (D) Unfiltered example traces of plateau-driven complex spikes
(CS, red arrows) preceding photostimulation events. (E) Additional example traces from
the same cells as in (D) without complex spikes occurring within 1 s before the
stimulation. (F) Stimulus-triggered average (mean ± SEM) oeIPSP (black: with CS;
orange: without CS). (G) Quantification of neuronal depolarization before stimulation
Dudok et al., Science 383, 967–970 (2024)

1 March 2024

and oeIPSP amplitudes (negative values) during trials with or without preceding
complex spikes (depolarization: P = 0.0076, paired t test, n = 15 cells from n = 4
mice; oeIPSP amplitude: P = 0.0045). (H) Histograms of place field sizes of
individual place cells in control mice and after cell-type-specific CB1 KO in GABAergic
neurons (GABA-CB1-KO). n = 420 ± 254 place cells from n = 5 control and n = 3
GABA-CB1-KO mice. (I) Quantification of place cell place field size and spatial
information. n = 13 sessions from n = 2 male and n = 2 female control mice; n =
19 sessions from n = 3 male GABA-CB1-KO mice. Markers and box plots show
individual sessions (boxes: median ± interquartile range, whiskers: nonoutlier range).
Place field size: P = 0.032, c2(1) = 4.59; spatial information: P = 0.004, c2(1) =
8.5, linear mixed effects models and likelihood ratio test.
3 of 4

RES EARCH | R E S E A R C H A R T I C L E

Postsynaptic activity–dependent modulation
of inhibitory postsynaptic potentials

8.
9.
10.
11.
12.
13.

1 March 2024

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk3863
Materials and Methods
Figs. S1 to S3
References (42–45)
MDAR Reproducibility Checklist
Submitted 18 August 2023; accepted 24 January 2024
10.1126/science.adk3863

4 of 4

,

6.
7.

I. Llano, N. Leresche, A. Marty, Neuron 6, 565–574 (1991).
T. A. Pitler, B. E. Alger, J. Neurosci. 12, 4122–4132 (1992).
A. C. Kreitzer, W. G. Regehr, J. Neurosci. 21, RC174 (2001).
R. I. Wilson, R. A. Nicoll, Nature 410, 588–592 (2001).
T. Maejima, K. Hashimoto, T. Yoshida, A. Aiba, M. Kano, Neuron
31, 463–475 (2001).
I. Katona et al., J. Neurosci. 19, 4544–4558 (1999).
K. Tsou, S. Brown, M. C. Sañudo-Peña, K. Mackie, J. M. Walker,
Neuroscience 83, 393–411 (1998).
G. Marsicano, B. Lutz, Eur. J. Neurosci. 11, 4213–4225 (1999).
C. Földy, A. Neu, M. V. Jones, I. Soltesz, J. Neurosci. 26,
1465–1469 (2006).
S.-H. Lee et al., J. Neurosci. 35, 10039–10057 (2015).
J. O’Keefe, J. Dostrovsky, Brain Res. 34, 171–175 (1971).
K. C. Bittner et al., Nat. Neurosci. 18, 1133–1142 (2015).
F. Dubruc, D. Dupret, O. Caillard, J. Neurophysiol. 110,
1930–1944 (2013).

y

1.
2.
3.
4.
5.

We thank A. Ortiz, C. Porter, S. Linder, and K. Patron for technical and
administrative support. Funding: This work was supported by the
National Institute of Neurological Disorders and Stroke (NINDS) of the
National Institutes of Health (NIH). The content is solely the
responsibility of the authors and does not necessarily represent the
official views of the NIH. This work was supported by the NIH (grants
R01NS99457, R01NS131728, and R01NS133381 to I.S.; grant
R00NS117795 to B.D.; grant K99MH132871 to L.Z.F.; and grant
K99NS126725 to J.S.F.); the Knight Initiative for Brain Resilience
(grant KCG-116 to I.S.); a McNair scholarship from the McNair Medical
Institute at The Robert and Janice McNair Foundation to B.D.; a
Helen Hay Whitney fellowship to L.Z.F.; a Burroughs Wellcome Fund
Career Award at the Scientific Interface to L.Z.F.; a Stanford University
Bio-X Undergraduate Summer Research Program grant to C.W.;
and the National Institute of Mental Health, National Institute on Drug
Abuse, National Science Foundation, Gatsby Foundation, Fresenius
Foundation, AE Foundation, Tarlton Foundation, and NOMIS
Foundation to K.D. Author contributions: Conceptualization: B.D.,
L.Z.F., K.D., I.S.; Formal analysis: B.D., L.Z.F.; Funding acquisition:
B.D., K.D., I.S.; Investigation: B.D., L.Z.F., J.S.F., S.M., J.H., D.K., C.W.,
C.R.; Methodology: B.D., L.Z.F., J.S.F.; Resources: Y.L.; Supervision:
K.D., I.S.; Visualization: B.D., L.Z.F.; Writing – original draft: B.D., L.Z.F.,
I.S.; Writing – review & editing: all authors. Competing interests: I.S.
declares unrelated consultant activity for Actio Biosciences, CODA
Biotherapeutics, MapLight Therapeutics, Praxis Precision Medicines,
and Ray Therapeutics. K.D. declares unrelated consultant activity
for MapLight Therapeutics and Stellaromics. The remaining authors
declare no competing interests. Data and materials availability:
All data, code, and materials are available from the authors upon
reasonable request. License information: Copyright © 2024 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse

y g

Dudok et al., Science 383, 967–970 (2024)

RE FERENCES AND NOTES

AC KNOWLED GME NTS

y

The above results provide evidence for postsynaptic neuronal activity–dependent modulation of CCKBC synapses in vivo. A suppression
of inhibition could disinhibit place cells during
place field traversal, contributing to locationspecific place cell activity (16, 36). To determine whether preventing inhibitory synaptic
eCB signaling may lead to altered place fields,
we knocked out CB1 selectively in forebrain
GABAergic neurons (GABA-CB1-KO, lacking
CB1 from perisomatic and dendritic interneurons) (28) (fig. S3B) and recorded place cell calcium signals during a spatial navigation task

g

Interneuron cannabinoid receptors modulate
place cell activity patterns

14. G. Carlson, Y. Wang, B. E. Alger, Nat. Neurosci. 5, 723–724 (2002).
15. V. Chevaleyre, P. E. Castillo, Neuron 43, 871–881 (2004).
16. M. Valero, A. Navas-Olive, L. M. de la Prida, G. Buzsáki, Cell Rep.
40, 111232 (2022).
17. S. V. Rolotti et al., Neuron 110, 783–794.e6 (2022).
18. T. F. Freund, I. Katona, D. Piomelli, Physiol. Rev. 83, 1017–1066
(2003).
19. A. Dong et al., Nat. Biotechnol. 40, 787–798 (2022).
20. J. S. Farrell et al., Neuron 109, 2398–2403.e4 (2021).
21. Y. Hashimotodani, T. Ohno-Shosaku, M. Kano, Curr. Opin.
Neurobiol. 17, 360–365 (2007).
22. H. Dana et al., eLife 5, e12727 (2016).
23. N. B. Danielson et al., Neuron 91, 652–665 (2016).
24. P. Kaifosh, J. D. Zaremba, N. B. Danielson, A. Losonczy,
Front. Neuroinform. 8, 80 (2014).
25. Y. Hashimotodani, T. Ohno-Shosaku, T. Maejima, K. Fukami,
M. Kano, Neuropharmacology 54, 58–67 (2008).
26. J. Z. Long, D. K. Nomura, B. F. Cravatt, Chem. Biol. 16,
744–753 (2009).
27. E. Albarran et al., Nat. Neurosci. 26, 997–1007 (2023).
28. K. Monory et al., Neuron 51, 455–466 (2006).
29. B. Dudok et al., Nat. Neurosci. 18, 75–86 (2015).
30. B. Pan et al., J. Pharmacol. Exp. Ther. 331, 591–597 (2009).
31. L. L. Glickfeld, M. Scanziani, Nat. Neurosci. 9, 807–815 (2006).
32. B. Dudok et al., Neuron 109, 997–1012.e9 (2021).
33. L. Z. Fan et al., Cell 186, 543–559.e19 (2023).
34. H. Tian et al., Nat. Methods 20, 1082–1094 (2023).
35. J. Epsztein, M. Brecht, A. K. Lee, Neuron 70, 109–120 (2011).
36. S. Royer et al., Nat. Neurosci. 15, 769–775 (2012).
37. Ö. Albayram, S. Passlick, A. Bilkei-Gorzo, A. Zimmer,
C. Steinhäuser, Pflugers Arch. 468, 727–737 (2016).
38. I. Del Pino et al., Nat. Neurosci. 20, 784–792 (2017).
39. C. Varga, P. Golshani, I. Soltesz, Proc. Natl. Acad. Sci. U.S.A.
109, E2726–E2734 (2012).
40. M. Bartos et al., Proc. Natl. Acad. Sci. U.S.A. 99, 13222–13227
(2002).
41. A. D. Milstein et al., eLife 10, e73046 (2021).

p

Retrograde eCB signaling through CB1 inhibits
CCKBC to pyramidal cell synapses in vitro
(9, 31). On the basis of our results showing
eCB transients time-locked to calcium transients, we expected to observe an activitydependent modulation of CCKBC synapses.
We used a CCKBC-specific (Sncg-FlpO) mouse
line to test this hypothesis (32) and developed
an all-optical method to study synaptic transmission between CCKBCs and postsynaptic
neurons. These animals express the FlpO
recombinase enzyme specifically in gammasynuclein (Sncg)–expressing cells. Sncg is expressed selectively in CCKBCs; therefore, FlpO
will be expressed specifically in this cell population in Sncg-FlpO mice. We expressed FlpOdependent excitatory opsin (sombC1C2TG) (33)
in CCKBCs and a soma-localized genetically
encoded voltage indicator (GEVI, somQuasAr6a)
(34) in sparsely labeled CA1 neurons in SncgFlpO mice (Fig. 3A and fig. S3A). We imaged
GEVI in awake mice head-fixed on a spherical
treadmill while activating CCKBCs with photostimulation (Fig. 3B). Brief CCKBC activation
elicited time-locked CA1 neuronal hyperpolarization, consistent with optogenetically
evoked inhibitory postsynaptic potentials
(oeIPSP; Fig. 3C).
Plateau-driven complex spikes in CA1 pyramidal cells are particularly important for
synaptic plasticity (12, 33, 35). We identified
plateau-driven complex spikes with voltage imaging and then grouped the photostimulationinduced responses based on the presence or
absence of complex spikes during the 1 s before
the stimulus (Fig. 3, D and E). Whereas oeIPSPs
were detectable in the absence of a preceding
complex spike (Fig. 3E), the same postsynaptic
cells showed reduced oeIPSPs after complex
spikes (Fig. 3, D, F, and G). As expected, the
average postsynaptic depolarization before the
CCKBCs stimulus was higher in the presence
of complex spikes (Fig. 3G). Together, these
results demonstrate a transient suppression of
CCKBC inhibition after complex spikes, consistent with a DSI-like mechanism.

as mice foraged for a water reward. Both control (Dlx-Cre) and GABA-CB1-KO mice exhibited
spatially tuned calcium signals, suggesting
that CB1 expression by GABAergic neurons is
not required for place field formation per se
(fig. S3, C and D). However, we observed a
widening of place fields in GABA-CB1-KO mice
relative to mice with intact CB1 expression
(Fig. 3, H and I). Analyzing the properties of
individual place cells revealed that in the absence of interneuron CB1 expression, place cells
were active over a larger fraction of the belt and
altogether encoded less spatial information
(Fig. 3I and fig. S3, C to J). In GABA-CB1-KO
mice, place cells fired less reliably lap-to-lap,
and had fewer calcium transients near the
preferred location (fig. S3, H and I). As a
population, place cells in the GABA-CB1-KO
encoded mouse location less accurately compared with control despite the similar ratio of
place cells (fig. S3, E and J). The observed changes
in place cell activity patterns are consistent with
the reported impaired spatial learning performance of GABA-CB1-KO mice (37) and mice
with perturbed CCKBC development (38).
In this study, we report (i) rapid eCB signals
time-locked to calcium transients in hippocampal neurons including place cells, both in
the postsynaptic membrane and the presynaptic inhibitory axons; (ii) modulation of CCKBC
synapses correlated to past postsynaptic activity; and (iii) diminished place cell place field
properties in the absence of eCB signaling at
inhibitory synapses. Our results demonstrate
that an eCB-mediated, DSI-like plasticity is
capable of rapid modulation of inhibition in
vivo on the behaviorally relevant timescale
of seconds. Because of the selective expression of CB1 at synapses of CCK-expressing but
not PV-expressing interneurons, DSI may enable recently activated place cells to maintain
elevated excitability without suppressing the
ability of PVBC synapses to synchronize the PC
population activity dynamics to theta and
gamma oscillations (39, 40). Such a selective,
lasting suppression of inhibition involving CB1
signaling may also contribute to maintaining
an eligibility trace for non-Hebbian activity–
dependent plasticity (41).

RES EARCH

THERMODYNAMICS

Variance sum rule for entropy production
I. Di Terlizzi1,2†, M. Gironella3,4†, D. Herraez-Aguilar5, T. Betz6,7, F. Monroy8,9, M. Baiesi2,10, F. Ritort3,11*
Entropy production is the hallmark of nonequilibrium physics, quantifying irreversibility, dissipation,
and the efficiency of energy transduction processes. Despite many efforts, its measurement at
the nanoscale remains challenging. We introduce a variance sum rule (VSR) for displacement and force
variances that permits us to measure the entropy production rate s in nonequilibrium steady states.
We first illustrate it for directly measurable forces, such as an active Brownian particle in an optical
trap. We then apply the VSR to flickering experiments in human red blood cells. We find that s is
spatially heterogeneous with a finite correlation length, and its average value agrees with calorimetry
measurements. The VSR paves the way to derive s using force spectroscopy and time-resolved
imaging in living and active matter.

ð1Þ

where the left-hand side includes the variances of the displacements Dxt ¼ xt  x0 , and
t
of time-cumulative forces [ SF ðt Þ ¼ ∫ 0 dsFs ].
The total variance V T ðt Þ ¼ V Dx ðt Þ þ m2 V SF ðt Þ
equals the free diffusion term 2Dt plus a nonequilibrium contribution S ðt Þ denoted as excess variance
t

Sðt Þ ¼ 2m∫ 0 ds½CxF ðsÞ  CFx ðsÞ

ð2Þ

that measures the breakdown of time-reversal
symmetry, with CAB ðsÞ ¼ As B0  A s B 0 the
correlation function in the NESS. In equilib-

1 March 2024

s¼

v2
1
m
þ @ 2 V Dx jt¼0 þ V F
2
m 4m t

ð4Þ

To illustrate the VSR, we consider two examples of a NESS where Ft equals the force in
the measurement device, Ft ¼ FtM, and FtI ¼ 0.
Methods

Experiments with colloidal particles (Figs. 1 and
2) were done in a miniaturized version of an
optical tweezers instrument described in (22).
Human red blood cells (RBCs) were obtained
by finger pricking of a healthy donor for the
RBC experiments. The phosphate-buffered saline (PBS) solution contains 130 mM NaCl,
20 mM K/Na phosphate buffer, 10 mM glucose,
and 1 mg/ bovine serum albumin per milliliter
of solution. For optical tweezer (OT)-stretching
experiments, 4 ml of blood was diluted in 1 ml
of PBS. RBCs were treated and biotinylated for
OT sensing as described in (11). For optical microscopy (OM) measurements, the RBC pellet
obtained after centrifugation (5000g for 10 min
at 4°C) was resuspended (1:15) in PBS solution (23). Contact areas in OT experiments were
estimated using a multiscale feature extractor
based on a Gaussian pyramid representation
of the raw image followed by a Laplacian reconstruction. For OT sensing, we used estimates
from (11).
Bead dragged through water

The first system is an optically trapped colloidal particle dragged through water (friction coefficient g ¼ 1=m) at speed v. The bead’s
dynamics can be analytically solved, and
the VSR (Eq. 1) verified (materials and methods S3). Equation 4 follows with S ¼ 0 and
s ¼ gv2, as expected. Figure 1C shows the experimental validation of the VSR (Eq. 1). The
right inset shows measurements of s  gv2
for several repetitions of the experiment and
1 of 6

,

V Dx ðt Þ þ m2 V SF ðt Þ ¼ 2Dt þ S ðt Þ




y

Di Terlizzi et al., Science 383, 971–976 (2024)




where v ¼ x
 is the particle’s average velocity
and s is expressed in power units (e.g., kB T =s).
By using Eq. 1 along with Eq. 3, we derive
the formula for the rate of entropy production in terms of the static variance of the force
V F ¼ F 2  F 2 and the convexity of the meansquared displacement V Dx at time 0

y g

*Corresponding author. Email: ritort@ub.edu
†These authors contributed equally to this work.

We introduce a variance sum rule (VSR) to derive s in experiments where a measurement
probe is in contact with a system in a NESS
(Fig. 1A). Dynamics are described
by a Langevin
pffiffiffiffiffiffi
equation, xðt Þ ¼ mFt þ 2Dht , with probe mobility m, diffusivity D, and a Gaussian white noise
term, ht. The total force acting on the probe
Ft ≡ Ft ðxt Þ equals the sum of the force exerted
by the measurement device, FtM , plus a probesystem interaction, FtI , Ft ¼ FtM þ FtI (arrows
in Fig. 1A). In most experimental settings, FtI
remains inaccessible, so Ft and s cannot be
directly measured. Our approach focuses on
how observablesQt on average spread in time, as
quantified by their variance V Q ðt Þ ¼ Qt2  Qt 2
with ð…Þ the dynamical average in the NESS.
The VSR is an equality for integrated quantities in an arbitrary time interval (0, t), which
imposes a tight constraint on the fluctuations
in a stochastic diffusive system over the experimental timescales. By integrating the Langevin
equation over the interval (0,t) and by taking
the variance of both sides, a time-preserved
identity can be obtained (materials and methods S1). The VSR for position and force fluctuations reads

ð3Þ

y

Max Planck Institute for the Physics of Complex Systems,
Nöthnitzer Straße 38, 01187 Dresden, Germany.
Dipartimento di Fisica e Astronomia, Università di Padova,
Via Marzolo 8, 35131 Padova, Italy. 3Small Biosystems Lab,
Condensed Matter Physics Department, Universitat de
Barcelona, C/ Marti i Franques 1, 08028 Barcelona, Spain.
4
Department of Medical Biochemistry and Cell Biology,
Institute of Biomedicine, The Sahlgrenska Academy,
University of Gothenburg, 40530 Gothenburg, Sweden.
5
Facultad de Ciencias Experimentales, Universidad Francisco
de Vitoria, Ctra. Pozuelo-Majadahonda Km 1,800, 28223
Pozuelo de Alarcón, Madrid, Spain. 6Third Institute of
Physics, Georg August Universität Göttingen, Göttingen,
Germany. 7Cluster of Excellence “Multiscale Bioimaging: from
Molecular Machines to Networks of Excitable Cells” (MBExC),
University of Göttingen, Göttingen, Germany. 8Departamento de
Química Física, Facultad de Química, Universidad Complutense,
28040 Madrid, Spain. 9Translational Biophysics, Instituto de
Investigación Sanitaria Hospital Doce de Octubre (IMAS12),
Av. Andalucía, 28041 Madrid, Spain. 10INFN, Sezione di Padova,
Via Marzolo 8, 35131 Padova, Italy. 11Institut de Nanociència i
Nanotecnologia (IN2UB), Universitat de Barcelona, 08028
Barcelona, Spain.
2

Variance sum rule

v2
1
þ @ 2 Sj
m 4m t t¼0

g

1

methods that estimate s more precisely are
needed to determine dissipative processes in
the nanoscale.

s¼

p

N

onequilibrium steady states (NESS) pervade nature, from climate dynamics (1)
to living cells and active matter (2). A
fundamental quantity is the entropy
production rate s at which energy is
dissipated to the environment, which is positive by the second law of thermodynamics
(3, 4). Entropy production measurements
remain challenging despite their relevance,
especially in microscopic systems with stochastic and spatially varying fluctuations and
limited access to microscopic variables (5, 6).
The entropy production rate s determines
the efficiency of energy transduction in classical and quantum systems (7, 8), the energetic costs and irreversible behavior of living
cells (9–12). It is an elusive quantity when
forces and currents are experimentally inaccessible. Bounds can be obtained from time
irreversibility (13, 14), the thermodynamic
uncertainty relation (15, 16), and coarse-graining (17–21). Most of these results provide lower
bounds that refine the second law of thermodynamics, s ≥ 0. However, the bounds are
often loose without upper limits and therefore
uninformative about the actual s. Alternative

rium, Sðt Þ ¼ 0 because of time-reversal symmetry. Figure 1B illustrates the VSR for a
generic NESS.
From the VSR, one can derive an equation
relating s to the variances of fluctuating variables. By taking the time derivative twice of
Eq. 2 and evaluating it at t ¼ 0, one obtains a
formula for s that depends on the convexity
of the excess variance S ðt Þ at t ¼ 0 (materials
and methods S2),

RES EARCH | R E S E A R C H A R T I C L E

using Eq. 4, finding s  gv2 ¼ 5 T 7 kB T =s.
Notice that S ¼ 0 implies that the two rightmost terms in Eq. 4 are of equal magnitude
but opposite sign, compensating each other,
mV F ¼  2m1 @t2 V Dx jt¼0 ¼ kB T =tr > 0 with tr ¼
g=k ¼ 0:35 ms the bead’s relaxation time (k ¼
70 pN=mm being the trap stiffness). The value
mV F ∼ 3  103 kB T =s is almost three orders of
magnitude larger than s  gv2 (T7 kB T =s).
The results S ¼ 0 and s ¼ gv2 are not restricted to a harmonic well but hold for an arbitrary time-dependent potential U ðx  vt Þ.
This gives a reversed thermodynamic uncertainty relation (16) for the work exerted on
the bead by the optical trap, Wt ¼ vSF ðt Þ ¼
t
v∫0 dsFs, and an upper bound for s (materials
and methods S4),
2

s
2Wt
≤
k B T t V W ðt Þ

ð5Þ

4  104 nm/pN  s, speed v ¼
10 mm=s, gv2 ¼ 610 kB T=s). The
lower inset plots s  gv2 ¼
m
1 2
4m @t V Dx jt¼0 þ 2 V F from
Eq. 4 for the experimental
realizations; the horizontal red
line shows the average over
all experiments [5 T 7 kB T=s]
with one standard deviation
(red band). The black
dashed line is the theoretical
prediction s ¼ gv2 . The upper
inset shows the experimental
test of the inequality (Eq. 5). Dashed vertical lines show the bead’s relaxation time tr .

p
g

In Fig. 1C (left inset), we experimentally test
Eq. 5. The upper bound becomes tight for
t ≫ tr , the difference between two terms in
Eq. 5 vanishing as tr =t, as expected from the
steady-state fluctuation theorem for Gaussian
work distributions (4).

Fig. 1. Variance sum rule
(VSR): Sketches and
experiments with a dragged
particle. (A) Experimental setup
for a NESS measured with
optical tweezers. (B) Illustration
of the VSR showing the different terms in Eq. 1. (C) Experimental test of the VSR for
an optically trapped bead
dragged through water at room
temperature (bead radius
R ¼ 1:5 mm, mobility m ¼

The stochastic switching trap

ð6Þ

with w ¼ wþ þ w , a ¼ wr =w , and wr ¼
1=tr ¼ k=g (the bead’s relaxation rate for a
resting trap). In Fig. 2C, we test the VSR and
Eq. 6 for three NESS conditions. The inset
shows the two terms contributing to the total
variance V T . For large times, S converges to a
Di Terlizzi et al., Science 383, 971–976 (2024)

w
w þ wr

ð7Þ

Figure 2D shows values of s measured in SST
experiments with Dl ¼ 280 nm using Eq. 4.
Their average sexp ¼ 4:6 T 4  103 kB T =s
agrees with the theoretical prediction (Eq. 7),
sth ∼ 5:3  103 kB T =s . Figure 2E compares
sexp with sth (Eq. 7) (black dashed line) for
varying Dl. Experiment and theory agree over
three decades of s.
Reduced VSR

Until now, we have considered the case of a
single degree of freedom where the total force
acting on the bead equals the measured force,
Ft ¼ FtM and FtI ¼ 0. For the case of multiple
degrees of freedom where positions and total
forces can be measured, Eqs. 1 and 3, can be
generalized (materials and methods S1 and
S2). Quite often, however, a measurement probe
(atomic force microscope tip, microbead, etc.)
is in contact with a system in a NESS, such as a
biological cell with metabolic activity (Fig. 1A).
In this case, FtI ≠ 0 is experimentally inaccessible, and Ft ¼ FtM þ FtI cannot be measured, making the VSR (Eq. 1) inapplicable.
Moreover, in many cases, only a spatial degree

1 March 2024

V Dx ðt Þ þ m2 k2 V Sx ðt Þ ¼ 2Dt þ S~ðt Þ

ð8Þ

,

að1  ewr t Þ  a2 ð1  ewt Þ
1  a2

sth ¼ ðkDlÞ2 qð1  qÞm

of freedom xt is monitored, e.g., in particletracking experiments (24, 25) or in detecting
cellular fluctuations (26, 27). To apply the VSR
in these situations, it is necessary to model the
NESS by making assumptions about the interaction FtI and the underlying degrees of
freedom. Specifically, for a linear-response measuring device (FtM ¼ kxt ), a reduced VSR for a
single degree of freedom can be derived and
expressed in terms of variances related to the
position xt only. In these conditions, the displacement variance,V Dx, along with the variance
t
of Sx ðt Þ ¼ ∫0 ds xs , V Sx ðt Þ , satisfy (materials
and methods S6),

y



finite value, and V T merges with the equilibrium line 2Dt (black dashed line) when plotted
in log-log scale. Equations 3 and 6 yield the theoretical prediction (v ¼ 0)

y g

S ðt Þ ¼ 4ðDlÞ2 qð1  qÞ

y

The second system we consider is the stochastic switching trap (SST) (22), where an active
force is applied to an optically trapped bead
by randomly switching the trap position lt between two values (lþ ; l ) separated by Dl ¼
lþ  l (Fig. 2A). Jumps occur at exponentially
distributed times with switching rates wþ ; w
at each position. The ratio w =wþ ¼ q=ð1  qÞ
defines the probability q of the trap to be at
position lþ . Figure 2B shows the measured
bead’s position xt and force Ft ¼ kðlt  xt Þ
for three cases with q ¼ 1=2 and varying Dl.
The bead follows the movement of the trap
(top), quickly relaxing to its new equilibrium
trap position at every jump (force spikes,
bottom). Figure 2C shows the total variance,
V T ðt Þ ¼ V Dx ðt Þ þ m2 V SF ðt Þ. V T deviates from
2Dt (dashed line) between 104 and 1 s, showing that S≠0 is comparable to V T (notice the
log-log scale). The SST model is analytically
solvable (materials and methods S5), giving
expressions for V Dx ðt Þ; V SF ðt Þ, and Sðt Þ. For
the latter, we find

Equation 8 is a general result which, however, does not permit one to derive a formula
for s like Eq. 3. Notice that S~ differs from S in
Eq. 1 and does not vanish in equilibrium. S~
can be expressed in terms of the generic interacting force FtI ; see eq. S38 in materials
and methods. To derive s using Eq. 8, we
use a solvable model for the experiment and
a procedure consisting of the following steps:
(i) Analytically derive expressions for the excess variances, Sðt Þ and S~ðt Þ for the model;
(ii) calculate sth from Sðt Þ using Eq. 3; (iii)
fit the reduced VSR (Eq. 8) to the experimental data using S~ðt Þ from the model to extract
the model parameters; (iv) insert the parameters in the analytical expression for sth
2 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g

Fig. 2. VSR and entropy production rate for experiments with a stochastic switching trap. (A) Schematics of the experiment. (B) Traces of position and
force for three Dl values [see legend in (C)]. (C) VSR (Eq. 1) and total variance V T : Symbols are experimental data, and lines represent the theory with known parameters
without fitting. The inset shows the different terms in the VSR. (D) Measurements of s for wþ ¼ w ¼ 10 s1 and Dl ¼ 280 nm; we show different experimental
realizations (squares), their average sexp and the theoretical value sth (Eq. 7). (E) s (red symbols) averaged over experimental realizations (orange circles) for
Dl = 18, 70, and 280 nm; black line is the analytical prediction (Eq. 7).

y
y g

Di Terlizzi et al., Science 383, 971–976 (2024)

of amplitude D, fta ¼ 0 , fta fsa ¼ D2 ejtsj=ta ,
with ta the active correlation time (Fig. 3A,
inset). The dynamics are described by the stochastic equation
xt ¼ kmxt þ



pffiffiffiffiffiffi
2Dht þ mfta

ð9Þ

with k the trap stiffness, m the particle mobility, and D ¼ kB T m the diffusion constant.
To test the reduced-VSR approach (Eq. 8) for
deriving s, we exploit the mapping of the ABP

1 March 2024

(Eq. 9) to the SST model discussed previously
(Fig. 2A). The mappingpfollows
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiby
ffi identifying parameters D ¼ kDl qð1  qÞ, ta ¼ 1=w,
wr ¼ km from which Eq. 7 follows [for q ¼ 1=2,
see also (28)]. We have used Eq. 8 to analyze
the data already used in the previous approach for the SST experiments
2) wi th

(Fig.

~ðt Þ ¼ 2D2 m2 ta t  ta 1  et=ta ( co m par e
S
eq. S37) where D and ta are fitting parameters.
Results are shown in Fig. 3A and residuals
in fig. S2A. Their values and s agree with the
3 of 6

,

to derive s. The approach remains applicable
to a vast category of NESS whenever the interacting force FI between the probe and NESS
is linear. This is a typical situation in mesoscopic systems where fluctuations are small
in the linear response regime. A model for the
experimental system that includes the degrees
of freedom contributing most to s is required.
For instance, consider an active Brownian
particle (ABP) in an optical trap subject to a
random time-correlated active force FtI ≡ fta

y

Fig. 3. Application of the reduced VSR to experiments (SST and RBCs) to extract the entropy production rate. (A) Test of Eq. 8 for the SST experimental
~ T ðtÞ ¼ V Dx ðtÞ þ m2 k2 V Sx ðtÞ, fitted to Eq. 8 for different
data, equivalent to the active ABP, in a harmonic trap (Eq. 9 and inset). Symbols are experimental values for V
~ T ðtÞ for the two-layer active model
~
Dl (lines). Blue and red circles are the two contributions to V T ðtÞ for Dl ¼ 18 nm. (B and C) Fits of the reduced VSR to V
(materials and methods S6). (B) Healthy RBCs in OT-stretching experiments at three trap stiffness values (Fig. 4A). To help visualization, the three different
~ value; (C) Healthy (active) and passive RBCs in OT-sensing experiments (Fig. 4B).
~ T ðtÞ values have been scaled with respect to a single D
V

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
stretching (A) varying the trap stiffness kb from high values (5  102 pN/nm,
rightmost points) to low values (7  104 pN/nm, leftmost points) for healthy
RBCs. (E) s measurements for OT sensing for healthy (red symbols) and
passive (blue symbols) RBCs. (F) Colored s map for OM measurements along
the equatorial cell contour, as in (C), for a healthy RBC (circles) and a passive
RBC (diamonds). The radial distance represents s in arbitrary units. The
orange curve is the s-smoothed profile. (G) Scatter plot of s versus V x for the
RBCs of (F), showing that they are partially anticorrelated. Orange circles are
s values averaged over windows of 50 nm2 in V x . (H) Spatial correlation
functions for s and position x are measured along the cell contour. (I) Values
of sRBC compared to calorimetry estimates. For OT stretching, the dark (light)
red bar corresponds to the lowest (highest) trap stiffness.

y g

Fig. 4. Application of the reduced VSR to RBCs. (A) OT-stretching
experiments. Video image of stretched RBC and schematics of contact area
estimation (left); (right) three selected bead position traces at a high (blue),
medium (orange), and low (red) trap stiffness. (B) OT-sensing experiments.
Experimental setup from (11) (left) and tracking bead position traces for a
healthy (red) and passive (blue) RBC (right). (C) Ultrafast OM measurements:
Healthy RBC (upper images) and position traces (right) for three selected
pixels (50 nm by 50 nm) along the cell contour with high (red), medium (yellow),
and low (green) variance V x ; passive RBC (lower images) and cell contour
traces for three selected pixels (blue, right). The right images also show a color
variance map along the cell contour. The color bar denotes variance levels (red,
highest; blue, lowest). (D) s and position variance Vx measurements for OT

y

Red blood cells

Finally, we apply the reduced-VSR to the challenging case of human RBCs (29). RBCs metabolize glucose into adenosine 5´-triphosphate
(ATP) via the glycolytic pathway, producing the
cell membrane’s active flickering with a consequent entropy creation (11, 23, 30, 31). The
RBC membrane is dynamically attached to the
spectrin cortex through multiprotein complexes, which actively bind and unbind in the
phosphorylation step of the glycolytic pathway
(32). We have carried out experimental RBC
Di Terlizzi et al., Science 383, 971–976 (2024)

measurements using three techniques (Fig. 4).
Two of them use OTs in different setups: (i)
mechanical stretching of RBCs using beads
nonspecifically attached to the membrane with
different optical trap stiffness (OT stretching,
Fig. 4A); (ii) mechanical sensing of a biotinylated RBC membrane using streptavidin functionalized beads using data from (11) (OT
sensing, Fig. 4B). The third technique measures
cell contour fluctuations by membrane flickering segmentation tracking of free-standing
RBCs using ultrafast OM (23, 33) (Fig. 4C). As
a first observation, a single-layer active model
(Eq. 9) with its S~ðt Þ in Eq. 8 does not describe
the experimental data. Instead, we consider
a two-layer model with one hidden position
variable for the active membrane–cortex in-

1 March 2024

teraction that is linearly coupled to the membrane outer layer x (probe) (materials and
methods S7). Similar active models have been
proposed in the study of hair-cell bundle dynamics (14, 34, 35). The two-layer active model
leads to a reduced VSR of the form (Eq. 8) that
fits the experimental data; the fitting procedure is described in materials and methods
S8 and S9. Some fits of the reduced VSR are
shown in Fig. 3, B and C, and residuals of the
fits are shown in fig. S2, B to F.
Figure 4, D and E, show s values obtained
from OT-stretching data in the range of trap
stiffnesses kb ¼ 5  102  7  104 pN/nm
and OT-sensing data with kb ∼ 2  105 pN/nm
for healthy and ATP-depleted (passivated) RBCs.
For OT stretching, s increases as kb decreases
4 of 6

,

expected ones (table S1 and fig. S3A.). Therefore, the reduced VSR (Eq. 8) permits us to infer
NESS parameters and s from xt measurements only.

RES EARCH | R E S E A R C H A R T I C L E

Discussion

M. S. Singh, M. E. O’Neill, Rev. Mod. Phys. 94, 015001 (2022).
C. Bechinger et al., Rev. Mod. Phys. 88, 045006 (2016).
C. Maes, Séminaire Poincaré 2, 29 (2003).
U. Seifert, Rep. Prog. Phys. 75, 126001 (2012).
F. Ritort, Adv. Chem. Phys. 137, 31–123 (2008).
S. Ciliberto, Phys. Rev. X 7, 021051 (2017).
I. A. Martínez et al., Nat. Phys. 12, 67–70 (2016).
G. T. Landi, M. Paternostro, Rev. Mod. Phys. 93, 035008
(2021).
9. P. Martin, A. J. Hudspeth, F. Jülicher, Proc. Natl. Acad. Sci. U.S.A.
98, 14380–14385 (2001).
10. C. Battle et al., Science 352, 604–607 (2016).
11. H. Turlier et al., Nat. Phys. 12, 513–519 (2016).

5 of 6

,

1.
2.
3.
4.
5.
6.
7.
8.

y

REFERENCES AND NOTES

y g

The agreement between mechanical and bulk
calorimetric estimates of the RBC metabolic energy turnover suggests that the heat produced
in the glycolytic pathway is tightly coupled with
membrane flickering due to active kickers.
Tight mechanochemical coupling is critical to
an efficient free-energy chemical transduction.
It has been observed in processive enzymes
(e.g., polymerases, transport motors, etc.) (46)
and in allosteric coupling in ligand binding
(47). Tightly coupled processes are related to
emergent cycles in cellular metabolism and
chemical reaction networks, particularly for
the relevant glycolytic cycle of RBCs (48). A
clarifying example of weak versus tight coupling is the effect of the trap stiffness in deriving s (Fig. 4D). Unless the probe stiffness
is smaller than the RBC stiffness, the probe’s
passive fluctuations mask the system’s activity and s. In addition to molecular motors and
living cells, the VSR should apply to timeresolved photoacoustic calorimetry (49) and
enzyme catalysis, where the effective diffusion
constant of the enzyme increases linearly with
the heat released (50), a consequence of Eq. 1.
Moreover, spatially resolved maps of partial
measurements of s for weak mechanochemical coupling provide insight into the structural features underlying heat dissipation
in biological cells. In a wider context, the VSR
applies to nonlinear systems, from non-Gaussian
active noise to nonlinear potentials (materials
and methods S13). Finally, we stress that different models can fit the experimental data.
However, the power of the VSRs, Eqs. 1 and
8, is given by the constraint imposed by the
sum of variances over the experimental timescales. By fitting the experimental data to a
single function, the total variance V T ðt Þ over
several decades, the contribution of dissipative processes over multiple timescales is appropriately weighted in the sum balance. This
distinguishes our approach from plain model
fitting of the experimental power spectrum
to derive the model parameters (35) that may
lead to inaccurate estimations (materials and
methods S14). In this regard, the VSR links
modeling with energetics.

y

1 March 2024

active noise, ∼100 s1 for the RBC experiments
(tables S2 to S4).

g

Di Terlizzi et al., Science 383, 971–976 (2024)

ta → 0, sðta Þ saturates to a finite value whereas
aðta Þ ∼ ta , decreasing V x (fig. S6). We hypothesize that the anticorrelation observed
in the s map derives from the highly heterogeneous ta (mean 0.05 s and standard deviation 0.2 s) but nearly constant D (mean 4.4 pN
and standard deviation 0.2 pN) across all pixel
units. A constant-noise amplitude D with a
heterogeneous ta suggests a uniform density
of kickers but a heterogeneous ATP concentration cATP across the RBC surface, which
modulates the ATP binding rate of the kickers,
ta1 ∼ kbind º cATP .
The s map of a single RBC determines the
finite correlation length x for the spatially varying s field, a main prediction of active field theories (37, 38) and stochastic hydrodynamics
(39). For healthy RBCs, x has been estimated
from the spatial correlation functionCss ðd Þ, and
Cxx ðd Þ of the traces at a curvilinear distance d
along the RBC contour, Fig. 4H. Functions can
be fitted to an exponential ∼expðd=xÞ with
xss ∼ 0:35 T 0:05 mm and xxx ∼ 0:82 T 0:02 mm,
giving the median x ∼ 0:6 T 0:2 mm. This value
is larger than the lateral resolution of the
microscope (200 nm). The structure factor of
the s field along the cell contour shows a characteristic peak at a domain length l ∼ 1:3 mm,
which is larger than xss , possibly due to the
heterogeneous cortex-membrane bindingunbinding dynamics that produce differently
active s domains (materials and methods S11).
A two-layer active model in a ladder with an
interlayer coupling kxx further corroborates
the value obtained for xxx (materials and methods S12). The average heat flux density can
be estimated as js ¼ s=x2 ¼ ð2 T 1Þ 104 kB T =
ðs  mm2 Þ with x2 the typical area of an entropyproducing region. In summary, for an RBC
of typical surface area A ∼ 130 mm2 , one obtains sRBC ¼ js  A ¼ ð2 T 1Þ  105 kB T =s (OT
stretching, at lowest kb ); sRBC ¼ ð2 T 1Þ
105 kB T =s (OT sensing); and sRBC ¼ ð3 T 1Þ
106 kB T =s (OM). These values are compatible
with calorimetric bulk measurements of packed
6
RBCs, sbulk
RBC ¼ ð2 T 1Þ  10 kB T =s (40, 41) and
are larger than indirect measures based on
the breakdown of the fluctuation-dissipation
theorem and effective temperatures (11, 42).
The significantly low s values obtained for
passive RBCs (blue data in Fig. 4, E to G and
I) validate our approach. Our sRBC ∼ 105 
106 kB T =s is higher than the values obtained
through information-theoretic measures based
on the breakdown of detailed balance (12, 14).
Intuitively, the VSR (Eqs. 1 and 8) sets an
energy balance between fluctuating positions
and forces, both conjugated energy variables,
a missing feature in the thermodynamic uncertainty relation and coarse-graining models (43–45). In general, the VSR captures most
of s because sampling rates, 40 kHz for OT
stretching, 25 kHz for OT sensing, and 2 kHz
for OM, are higher than the frequency of the

p

reaching s ¼ ð3 T 1Þ  103 kB T =s averaged over
RBCs, for the lowest kb . This value is compatible with OT-sensing measurements, s ¼
ð2 T 1Þ  103 kB T =s for healthy RBCs, which
is larger than for passive RBCs (red and blue
symbols in Fig. 4E). Moreover, s appears correlated with the variance of the flickering
signal as measured from the position traces,
V x ¼ x2  x 2 (Fig. 4D). The apparent correlation demonstrates that the probe stiffness
kb must be lower than the stiffness of the
RBC, kRBC ∼ 5  103 pN/nm, to measure s;
otherwise, the active flickering of the RBC
membrane is suppressed by the passive fluctuations of the bead. The correlation between
s and Vx is also explicitly shown in fig. S4,
where a color-map plot of the stiffness shows
that we can detect active flickering and s
only for kb < kRBC. Indeed, for the largest trap
stiffness kb ∼ 5  102 pN/nm, one obtains
s ∼ 10 kB T =s (rightmost points in Fig. 4D), a
value almost constant if the RBC is stretched
up to 30 pN (fig. S4). The measured s is extensive with the bead–RBC contact area. Estimations from video images (Fig. 4A and
materials and methods) yield circular contact areas of a ¼ 0:8 T 0:2 mm2 for both OTtype experiments giving the heat flux density
js ¼ s=a ¼ ð3 T 1Þ  103 kB T =ðs  mm2 Þ for OT
stretching at low kOT and js ¼ ð1:8 T 0:6Þ
103 kB T =ðs  mm2 Þ for OT sensing. Such estimations are subject to uncertainty in the
actual diameter and shape of the contact area.
Furthermore, we have analyzed the simulation
data of the OT-sensing experiments based
on the three-dimensional numerical model
of (11). The active and passive trajectories
for the sensing bead give s ∼ 104 kB T =s and
s ∼ 20 kB T =s, respectively (materials and methods S10).
For the OM experiments, we show in Fig.
4C the color map of the position variance V x
(healthy, top; passive, bottom), and in Fig. 4F
we show the color map of s (circles, healthy;
diamonds, passive), measured over pixels of
area 50 nm by 50 nm along the RBC contour.
For the healthy RBCs, both s and V x reveal an
RBC heterogeneous activity with average values s ¼ ð7 T 1Þ  103 kB T =s and V x ¼ 400 T
10 nm2 . Molecular maps of heterogeneous
RBC deformability have been previously reported (36). In contrast to OT experiments
(Fig. 4, D and E), for OM experiments s and V x
are anticorrelated in the active regime (Pearson
coefficient ∼  0:4 ) with high-variance regions showing lower s (Fig. 4G). Results for
other RBCs are shown in fig. S5. This counterintuitive result demonstrates the critical
role of the active timescale ta, which, for fixed D,
determines the active contribution to the total
þ aðta Þsðta Þ (eq. S41a)
variance, V x ¼ V passive
x
with aðta Þ positive and monotonically increasing with ta and sðta Þ given in eq. S42. It
can be shown that in the high-activity limit

RES EARCH | R E S E A R C H A R T I C L E

32. T. J. Byers, D. Branton, Proc. Natl. Acad. Sci. U.S.A. 82,
6153–6157 (1985).
33. M. Mell, F. Monroy, PLOS ONE 13, e0207376 (2018).
34. P. Martin, A. J. Hudspeth, Proc. Natl. Acad. Sci. U.S.A. 96,
14306–14311 (1999).
35. G. Tucci et al., Phys. Rev. Lett. 129, 030603 (2022).
36. D. E. Discher, N. Mohandas, E. A. Evans, Science 266,
1032–1035 (1994).
37. C. Nardini et al., Phys. Rev. X 7, 021007 (2017).
38. T. GrandPre, K. Klymko, K. K. Mandadapu, D. T. Limmer, Phys.
Rev. E 103, 012613 (2021).
39. T. Markovich, É. Fodor, E. Tjhung, M. E. Cates, Phys. Rev. X 11,
021057 (2021).
40. U. Bandmann, M. Monti, I. Wadsö, Scand. J. Clin. Lab. Invest.
35, 121–127 (1975).
41. P. Bäckman, Thermochim. Acta 205, 87–97 (1992).
42. E. Ben-Isaac et al., Phys. Rev. Lett. 106, 238103 (2011).
43. D. M. Busiello, S. Pigolotti, Phys. Rev. E 100, 060102 (2019).
44. D.-K. Kim, Y. Bae, S. Lee, H. Jeong, Phys. Rev. Lett. 125,
140604 (2020).
45. P. Bilotto, L. Caprini, A. Vulpiani, Phys. Rev. E 104, 024140
(2021).
46. A. I. Brown, D. A. Sivak, Chem. Rev. 120, 434–459 (2020).
47. N. V. Dokholyan, Chem. Rev. 116, 6463–6487 (2016).
48. A. Wachtel, R. Rao, M. Esposito, J. Chem. Phys. 157, 024109
(2022).
49. K. S. Peters, G. J. Snyder, Science 241, 1053–1057 (1988).
50. C. Riedel et al., Nature 517, 227–230 (2015).
51. I. Di Terlizzi et al., Variance sum rule for entropy production.
(2023); https://doi.org/10.5061/dryad.h44j0zpsw

D.H.-A. and F.M. are supported by the Spanish Research Council
(grant PID2019-108391RB-100 and grant TED2021-132296B). T.B. is
supported by the European Research Council (consolidator grant
771201 and the Deutsche Forschungsgemeinschaft under Germany’s
Excellence Strategy - EXC 2067/1-390729940). M.B. is supported
by research grant BAIE_BIRD2021_01 from the University of Padova.
F.R. is supported by ICREA Academia 2018. Author contributions:
I.D.T., M.B., and F.R. conceptualized the study; M.G., D.H.-A., T.B.,
and F.M. collected and curated the data. I.D.T. wrote the software
for data analysis and performed visualization. I.D.T. and M.G.
analyzed the data. F.R. administered the project. I.D.T., M.B., and
F.R. wrote the original draft. All authors discussed the results and
implications of the methodology and commented on the
manuscript. Competing interests: The authors declare no
competing financial interests. Data and materials availability: All
data needed to evaluate the conclusions in the paper, and the
code for fitting the VSR, are available at Dryad (51). Figures 1A
and 2A, fig. S1, and the inset of Fig. 3A were created with
BioRender.com. License information: Copyright © 2024 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.sciencemag.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adh1823
Materials and Methods
Supplementary Text
Figs. S1 to S15
Tables S1 to S5
References (52–60)

p

ACKN OWLED GMEN TS

Funding: M.G. and F.R. are supported by the Spanish Research
Council (grant PID2019-111148GB-100 and PID2022-139913NB-100).

Submitted 16 February 2023; accepted 9 January 2024
10.1126/science.adh1823

g

12. C. W. Lynn, E. J. Cornblath, L. Papadopoulos, M. A. Bertolero,
D. S. Bassett, Proc. Natl. Acad. Sci. U.S.A. 118, e2109889118
(2021).
13. J. Li, J. M. Horowitz, T. R. Gingrich, N. Fakhri, Nat. Commun.
10, 1666 (2019).
14. É. Roldán, J. Barral, P. Martin, J. M. Parrondo, F. Jülicher,
New J. Phys. 23, 083013 (2021).
15. A. C. Barato, U. Seifert, Phys. Rev. Lett. 114, 158101
(2015).
16. J. M. Horowitz, T. R. Gingrich, Nat. Phys. 16, 15–20
(2020).
17. G. Bisker, M. Polettini, T. R. Gingrich, J. M. Horowitz, J. Stat.
Mech. 2017, 093210 (2017).
18. G. Teza, A. L. Stella, Phys. Rev. Lett. 125, 110601 (2020).
19. D. J. Skinner, J. Dunkel, Proc. Natl. Acad. Sci. U.S.A. 118,
e2024300118 (2021).
20. A. Dechant, S.-i. Sasa, Phys. Rev. X 11, 041061 (2021).
21. C. Dieball, A. Godec, Phys. Rev. Lett. 129, 140601
(2022).
22. E. Dieterich, J. Camunas-Soler, M. Ribezzi-Crivellari, U. Seifert,
F. Ritort, Nat. Phys. 11, 971–977 (2015).
23. R. Rodríguez-García et al., Biophys. J. 108, 2794–2806
(2015).
24. C. Manzo, M. F. Garcia-Parajo, Rep. Prog. Phys. 78, 124601
(2015).
25. S. Scott et al., Phys. Chem. Chem. Phys. 25, 1513–1537 (2023).
26. W. W. Ahmed et al., Biophys. J. 114, 1667–1679 (2018).
27. S. Salinas-Almaguer et al., Sci. Rep. 12, 933 (2022).
28. R. Garcia-Millan, G. Pruessner, J. Stat. Mech. 2021, 063203
(2021).
29. H. Turlier, T. Betz, Annu. Rev. Condens. Matter Phys. 10,
213–232 (2019).
30. Y.-Z. Yoon et al., Biophys. J. 97, 1606–1615 (2009).
31. T. Betz, M. Lenz, J.-F. Joanny, C. Sykes, Proc. Natl. Acad. Sci.
U.S.A. 106, 15320–15325 (2009).

y
y g
y
,

Di Terlizzi et al., Science 383, 971–976 (2024)

1 March 2024

6 of 6

RES EARCH

MARINE CONSERVATION

Divergent responses of pelagic and benthic fish
body-size structure to remoteness and protection
from humans
Tom B. Letessier1,2,3*, David Mouillot4, Laura Mannocci4,1, Hanna Jabour Christ3,
Elamin Mohammed Elamin5, Sheikheldin Mohamed Elamin6, Alan M. Friedlander7,8, Alex Hearn9,10,
Jean-Baptiste Juhel11, Alf Ring Kleiven12, Even Moland12,13, Nicolas Mouquet1,4, Portia Joy Nillos-Kleiven12,
Enric Sala7, Christopher D. H. Thompson3, Laure Velez4, Laurent Vigliola11, Jessica J. Meeuwig3,14

1 March 2024

1 of 7

,

Letessier et al., Science 383, 976–982 (2024)

Our surveys, conducted from January 2006 to
May 2020, recorded a total 823,849 individual
fish (pelagic: 106,424, benthic: 717,425; Fig. 2),
representing 139 families and 1460 species of
fishes and sharks (211 pelagic, 1376 benthic,
and 127 species recorded in both systems) and
weighing a combined 744 metric tons (pelagic:
325 tons; benthic: 418 tons). Our dataset lacked
representation from the North Pacific, and representation in the central Pacific and in most
of the Atlantic was limited to pelagic systems
only. Size-frequency distributions were generated by aggregating sizes within six broad
brackets of absolute latitude (Fig. 3), which
revealed distinct patterns within each system
that were robust to an unbalanced survey design (17). Benthic median sizes were generally
larger than pelagic medians (range of medians:
pelagic 4 to 134 g, benthic 27 to 120 g) owing
to the greater representation of smaller size
classes (<30 g). Upper size classes were better
represented in pelagic size distributions than
in benthic size distributions (range of 95th
percentiles, pelagic 0.4 to 83.3 kg, benthic 1.3 to
2.9 kg; Fig. 3A). Size spectra slopes, a measure
of the proportion of large to small individuals
(17), were contrasted between systems by regressing normalized size-frequency distributions
on the log10-log10 scale (Fig. 3B). Slope values
were consistently steeper (more negative) in benthic systems than in pelagic systems (table S1),

y

*Corresponding author. Email: tom.letessier@ioz.ac.uk

Body-size structure across systems

y g

CESAB – FRB, Montpellier, France. 2Institute of Zoology, Zoological Society of London, Regent’s Park, London, UK. 3Marine
Futures Lab, School of Biological Sciences, University of Western Australia, Perth, WA, Australia. 4MARBEC, Université de
Montpellier, CNRS, Ifremer, IRD, Montpellier, France. 5Red Sea Fisheries Research Station, P.O. Box 730, Port Sudan, Red Sea
State, Sudan. 6Faculty of Marine Science and Fisheries, Red Sea State University, P.O. Box 24, Port Sudan, Red Sea State,
Sudan. 7National Geographic Society, Washington, DC 20036, USA. 8Hawai‘i Institute of Marine Biology, University of
Hawai‘i, Kāne‘ohe, Hawai‘i, USA. 9Galapagos Science Center, Universidad San Francisco de Quito, Quito, Ecuador. 10MigraMar,
Olema, CA, USA. 11ENTROPIE, Institut de Recherche pour le Développement, IRD-UR-UNC-IFREMER-CNRS, Centre IRD de
Nouméa, Nouméa Cedex, New-Caledonia, France. 12Institute of Marine Research, Nye Flødevigveien 20, 4817 His, Norway.
13
Centre for Coastal Research (CCR), Department of Natural Sciences, University of Agder, P.O. Box 422, N-4604 Kristiansand,
Norway. 14Oceans Institute, University of Western Australia, Perth, WA, Australia.
1

y

the next, for example, plankton, planktivorous
fishes, and piscivorous fishes (10, 11). However,
assessing such size-structured variation across
marine habitats and regulations is particularly
challenging because dedicated survey methodologies with different size selectivity are used in
pelagic and benthic systems. For instance, whereas pelagic fishes are conventionally sampled
through longlines and midwater trawls or
acoustic techniques (12, 13), benthic fishes
are mainly surveyed by underwater visual
census (14) or with bottom trawls and other
habitat-specific gears (15), which makes intersystem comparisons difficult.
Stereo baited remote underwater video
stations (BRUVS) represent a unifying, nondestructive, and fisheries independent method that can estimate relative abundance and
body size across virtually any marine system
(16). In this work, we conducted a widespread
size-based assessment of marine pelagic and
benthic nekton fishes (>1 g), spanning six orders
of magnitude in body size, from zooplankton
size classes (~3 to 4 cm) to large oceanic predators (~1000 kg; Fig. 1). We combined records
from multiple surveys inside and outside MPAs,
resulting in 6701 BRUVS deployed in pelagic
systems and 10,710 BRUVS deployed in ben-

g

B

ody size is a universal biological property that influences ecological processes
at the individual, population, and ecosystem levels (1). Measuring size spectra
(size frequencies plotted on a log-log
scale) is therefore a useful framework through
which to understand and predict overexploitation (2), nutrient cycling (3), and productivity (4). Moreover, understanding how body
sizes are distributed in the oceans has ramifications for conservation and fisheries science and is highly relevant to several of the
United Nations (UN) Sustainable Development Goals. In particular, effective biodiversity
conservation (5) and 30% protection coverage
by 2030 (“30 by 30” goal) (6) require understanding of how successful marine protected
areas (MPAs) are likely to be in different socioenvironmental contexts (7). Within a given
pelagic or benthic system, size spectra typically show consistent alternations between
overrepresented and underrepresented sizes,
resulting in regular peaks and troughs (8, 9).
When slopes of size spectra are shallow and
peaks are prominent, the spread between peaks
is generally considered to reflect predator-prey
relationships, with each peak representing a
different trophic group that is preyed upon by

p

Animal body-size variation influences multiple processes in marine ecosystems, but habitat heterogeneity has
prevented a comprehensive assessment of size across pelagic (midwater) and benthic (seabed) systems
along anthropic gradients. In this work, we derive fish size indicators from 17,411 stereo baited-video
deployments to test for differences between pelagic and benthic responses to remoteness from human
pressures and effectiveness of marine protected areas (MPAs). From records of 823,849 individual fish, we
report divergent responses between systems, with pelagic size structure more profoundly eroded near
human markets than benthic size structure, signifying greater vulnerability of pelagic systems to human
pressure. Effective protection of benthic size structure can be achieved through MPAs placed near markets,
thereby contributing to benthic habitat restoration and the recovery of associated fishes. By contrast,
recovery of the world’s largest and most endangered fishes in pelagic systems requires the creation of highly
protected areas in remote locations, including on the High Seas, where protection efforts lag.

thic systems, which corresponds to 13,402 and
10,710 hours of footage, respectively, across the
Atlantic, Indian, and Pacific Oceans. This database yielded length measurements for individual fish, which were converted to weights
using taxa-specific allometric conversion parameters (17, 18) (Fig. 2 and fig. S1).
To better understand how MPAs may effectively protect fish size structure in the context
of the “30 by 30” goal, we tested two competing and mutually exclusive hypotheses regarding the influence of human pressures on
fish size structure in pelagic and benthic systems. First, we hypothesized a greater human
footprint in pelagic systems compared with
benthic systems because the larger body size
and longer life of many oceanic species renders them more vulnerable to fisheries (19).
Therefore, we expect that pelagic fish size
structure is more sensitive to protection
status and human pressures than benthic
fish size structure. As an alternative hypothesis,
the migratory capacity of many large pelagic
species and the widespread activities of highsea fishing fleets (20) result in a comparatively
low human footprint and low MPA effectiveness in pelagic systems in contrast to benthic
systems, where local human pressure has acted
longer (21) and where fish size structure would
therefore be more affected and sedentary species would benefit more from MPAs (22).

RES EARCH | R E S E A R C H A R T I C L E

y g

slope theoretically reflecting the steepness of
the trophic pyramid (25). We then built explanatory generalized least-square (GLS) models (35) to test the two competing hypotheses
by identifying how human pressure and protection status affected pelagic and benthic fish
size indicators. In addition to controlling for
spatiotemporal autocorrelation and socioenvironmental conditions that are known to influence the effectiveness of spatial protection
status (36) (fig. S3 and table S2), our models considered interactions between systems
(pelagic or benthic) and protection status, as

1 March 2024

represented by three different categories of
spatial protection (37) (not protected, partially
protected, or highly protected) (17), and human pressure, as represented by travel time to
human markets (38) (log10 minutes).
GLS models of relatively small and relatively
large fishes achieved moderate explanatory
power [R2 adjusted for nonsignificant explanatory variables (adjR2), small individuals: 0.257,
large individuals: 0.343], revealing an effect of
market proximity and protection status, which
was consistent in direction but specific in magnitude to each system (P < 0.05; Fig. 4A, figs. S4
2 of 7

,

Fig. 1. Body-size variability in pelagic and benthic systems, recorded by stereo BRUVS. Pelagic
systems are shown on the left and benthic systems on the right. (A) Great white shark (Carcharodon carcharias).
(B) Grey reef shark (Carcharhinus amblyrhynchos). (C) Yellowfin tuna (Thunnus albacares). (D) Horse-eye jack
(Caranx latus). (E) Juvenile jack (Carangidae sp). (F) Tiger shark (Galeocerdo cuvier). (G) Two-spot red snapper
(Lutjanus bohar). (H) Spiny dogfish (Squalus acanthias). (I) Goldband fusilier (Caesio chrysozona). (J) Creole
wrasse (Clepticus parrae). [Credits: Photos were taken by the authors].

y

Letessier et al., Science 383, 976–982 (2024)

y

We tested our hypotheses concerning the difference in relative sensitivity of pelagic and
benthic size structure by extracting three size
indicators (33) from frequency-size distributions
of nekton fishes aggregated by survey date (17)
(fig. S2); the typical body sizes (log10, kg) of
relatively small individuals and of relatively
large individuals, as represented by the values
at the first and second modal frequency peaks;
and the exponent b of the size spectra slope (34).
These three indicators capture the main dimensions of size structure within each system, at
the scale of the survey day, with the size of
relatively small and large individuals representing relatively lower and higher trophic
levels, respectively (10), and the size spectra

g

Human footprint on size structure

p

reflecting the greater absolute and relative number of large individuals in pelagic systems (17).
Both the spread between peaks and sizespectra slope values were distinct between
pelagic and benthic systems across biogeographical scales, which suggests that each
system supports distinct food webs and energy pathways (23). The presence of prominent peaks in pelagic systems is consistent
with previous reports (10) and suggests that
each peak reflects a trophic group that is
preyed upon by the next, with shallower slopes
reflecting carnivorous feeding (11). In benthic
systems, peaks were less clearly defined and
slopes steeper, consistent with greater levels
of herbivorous feeding (11) likely stemming
from greater dependence on seabed algae
compared with in the midwater (24). Greater
prevalence of carnivory in pelagic systems implies that the proportion of production retained
between trophic levels is higher (25) as a result of more-direct energy transfer in these
systems than in benthic systems. Overrepresentation of intermediate size classes (30 to
500 g) in benthic systems is consistent with
complex habitat structure in coastal ecosystems such as kelp forests and coral reefs (26)
that provide size-selective refugia (27). Elevated benthic productivity within these size
classes is further promoted through system
connectivity and benthic-pelagic coupling (28),
whereby passively drifting plankton are consumed by planktivorous and piscivorous fishes
near the seabed (29). Conversely, pelagic productivity and energetic needs in upper trophic
levels are promoted by more-direct energy transfer (11) and are facilitated by greater home ranges
such that individuals in upper trophic levels can
forage from the top of multiple benthic food
webs (30) or from more productive geographical
regions such as those in temperate latitudes (31).
Mobile strategies in upper trophic levels typically involve pelagic foraging incursions or are associated with fully pelagic lifestyles (32), which
results in a greater prevalence of upper trophic
levels in pelagic systems.

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,
Fig. 2. Body sizes of pelagic and benthic fishes identified on BRUVS. (A) Survey effort of BRUVS, showing the outlines of the world’s Economic Exclusive Zones
in gray contours (some of which are contested). Each circle represents a single expedition, with the circle diameter being proportional to the number of BRUVS
deployed. Circles are jittered to minimize overplotting. (B) Pelagic and benthic fish body sizes (kg, n = 823,849) categorized by species identity (n = 1460) and rankordered by median species body size. (C) Marginal density distribution plots of body sizes.

and S5, and tables S3 and S4). In both systems,
individuals were larger if highly protected and
remote from markets, consistent with our
present understanding re-garding how vulnerability and exploitation vary with protection and accessibility (36). However, relatively
small and large individuals in pelagic systems
Letessier et al., Science 383, 976–982 (2024)

were both consistently more sensitive to protection status and to market remoteness, with a cumulative impact of protection status and market
remoteness. In benthic systems, relatively small
individuals were less sensitive to protection than
large individuals, in keeping with expectations
on how vulnerability to exploitation varies with

1 March 2024

differences in life history (14, 19). Moreover, the
effect of protection status saturated with remoteness, with remoteness having increasingly less
relative impact under higher protection.
GLS models of size spectra (adjR2, sizespectra slope: 0.273; Fig. 4B, fig. S6, and
table S5) showed divergent effects in each
3 of 7

RES EARCH | R E S E A R C H A R T I C L E

p

Fig. 3. Pelagic size spectra
are shallower than those of
their benthic counterparts
across biogeographical
scales. (A) Frequency density
distribution of fish body sizes
aggregated into six absolute
latitude brackets (0 to 10, 10 to
15, 15 to 20, 20 to 23, 23 to
33, 33 to 65) of equal numbers
of body sizes (n = 137,308),
with vertical line and number
showing median and 95th
percentile values, respectively.
(B) Abundance size spectra,
normalized by dividing the
frequency counts by the
width of the bin, with lines
representing fit of linear
regressions (pelagic slope
mean: −1.38, range: −1.47 to
−1.29; benthic mean: −1.58,
range: −1.63 to −1.54).

g
y
4 of 7

,

1 March 2024

Our results suggest that size-structure resilience to human pressure is lower in pelagic
systems than in benthic systems. In theory,
size spectra slopes are expected to steepen
with increasing human exploitation as a consequence of predator depletion, leading to a
commensurate decline in mean trophic level
(39). However, reports of human pressure responses in benthic systems are conflicting, with
both a steepening size spectra slope (39) and
a modest increase in mean trophic level reported (14, 40). This apparent conflict may
stem from difficulties in establishing appropriate baselines in “pristine” benthic systems,
which show wide-ranging size spectra slope
values (39) (i.e, −1.95 to −1.13) and both inverse
and concave trophic pyramids (14, 30). Our
observations of only a marginal effect on benthic slopes are, in any case, consistent with
reports of a comparatively modest impact of
human pressure on mean trophic level, which
has been corroborated from across a wide
range of benthic systems and arguably by a

y

Letessier et al., Science 383, 976–982 (2024)

inclusion of the Indian Ocean data (fig. S9).
Our main findings concerning the direction of
both remoteness and protection in pelagic and
benthic systems remained largely unchanged
from those derived using the full dataset.
Taken together, our models support our first
hypothesis, that pelagic fish size structures are
more vulnerable to human pressure than their
benthic counterparts. That both relatively small
and relatively large individuals in pelagic systems were consistently affected near markets
means that greater sensitivity in pelagic systems cannot be attributed solely to the greater
occurrence of larger (and therefore more vulnerable) individuals. In benthic systems, the
magnitude of protection effect declined with
market distance, in contrast to a cumulative
effect with market distance in pelagic systems.
This contrasting result means that high protection status can, even near markets, mitigate
human pressures in benthic systems, whereas
effective protection in pelagic systems requires
market remoteness.

y g

system, with size spectra slopes in pelagic
systems showing a pronounced and rapid
steepening with market proximity under high
protection and marginal effects of protection
status and market proximity after that. By contrast, slopes in benthic systems were marginally affected, becoming less negative (shallower)
near markets, independently of protection status. Without protection, steepening of pelagic
slopes and shallowing of benthic slopes resulted in converging size structure between
systems with considerable overlap in slope
values near markets in unprotected locations.
A sensitivity analysis testing the model robustness to the unbalanced survey reported similar
effects of market proximity, with minor differences between models rerun with 10% of
randomly dropped data points (17). Greater
differences were observed between model
reruns with ocean-specific data dropped. Notably, the results that showed that pelagic
systems were highly responsive to highly protected remote areas were conditional on the

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,
Fig. 4. Human influences on fish body-size structure in pelagic and benthic systems. Marginal plots of the influence of increased travel time to market (log10, min)
on fish size indicators under different levels of protection status (not protected, partly protected, and highly protected). (A) Mean body size of relatively small and relatively
large fishes (log10, kg). (B) Slopes of fish size spectra. Lines indicate predictions from GLS models, and shaded areas indicate 95% confidence intervals.

greater range of survey methods, including
underwater visual censuses, scientific trawl
surveys, and stock assessments (14, 40). Our
confidence that human pressure results in
only marginally shallower benthic size spectra
as a reflection of a comparatively minor change
in relative proportion of larger size classes is
strengthened by the observed consistency of
this shallowing across protection status but is
Letessier et al., Science 383, 976–982 (2024)

in contrast with expectation from “fishing down
the food web” and other predictions from sizestructured biodiversity loss (41).
Our results add to a body of evidence that
suggests that benthic systems are relatively
resilient compared with their pelagic counterparts. The emergence of benthic resilience is
not fully understood, and any proposed mechanism in support is speculative. However, one

1 March 2024

possible explanation may be related to the
emergence of alternative energy pathways when
heavy exploitation triggers trophic cascades
(42). Prey releases are generally predicted to
occur as a consequence of trophic cascades
under predator depletion (43). However, in
benthic systems such as coral reefs, prey releases can be counteracted through size-based
redundancy and feeding flexibility, which exist
5 of 7

RES EARCH | R E S E A R C H A R T I C L E

Policy implications

Our size-based assessment has enriched our
understanding of ongoing marine biodiversity loss, revealing divergent impacts across
pelagic and benthic communities, which may,
as a result, converge toward a common intermediate and artificial size structure. Many processes that are important for maintaining
productivity across trophic levels are supported by size-structured association within
coupled benthic-pelagic systems. Convergence
of pelagic and benthic communities toward an
artificial size structure should be of concern if
this results in a decoupling of pelagic and
benthic ecosystem components, thereby disrupting fundamental processes that underpin
functionality. Alternatively, it is plausible that
these processes are buffered by the emergence
of previously unknown benthic-pelagic associations, thereby ensuring resilience under sizestructured biodiversity loss. To help address
the uncertainty concerning the functional consequence of size structure erosion, we recommend that future research efforts explore the
link between size structure, ecosystem functioning, and connectivity, particularly in the
context of coupled benthic-pelagic systems.
Such knowledge would also have application
within biodiversity conservation and ecosystem restoration.
REFERENCES AND NOTES

,

6 of 7

y

1. R. H. Peters, The Ecological Implications of Body Size,
Cambridge Studies in Ecology (Cambridge Univ. Press, 1983).
2. C. F. Clements, M. A. McCarthy, J. L. Blanchard, Nat. Commun.
10, 1681 (2019).
3. P. Le Mézo, J. Guiet, K. Scherrer, D. Bianchi, E. Galbraith,
Biogeosciences 19, 2537–2555 (2022).
4. A. Rogers, J. L. Blanchard, P. J. Mumby, J. Appl. Ecol. 55,
1041–1049 (2018).
5. E. Sala et al., Nature 592, 397–402 (2021).
6. Convention on Biological Diversity, Conference of the Parties to
the Convention on Biological Diversity (CBD/COP/15/Part-II/
L.1) (United Nations Environment Programme, 2022).
7. J. E. Cinner et al., Proc. Natl. Acad. Sci. U.S.A. 115,
E6116–E6125 (2018).
8. P. M. Yurista et al., Can. J. Fish. Aquat. Sci. 71, 1324–1333 (2014).
9. F. J. Heather, R. D. Stuart-Smith, J. L. Blanchard, K. M. Fraser,
G. J. Edgar, Ecol. Lett. 24, 2146–2154 (2021).
10. A. G. Rossberg, U. Gaedke, P. Kratina, Nat. Commun. 10, 4396
(2019).
11. J. P. W. Robinson, J. K. Baum, Can. J. Fish. Aquat. Sci. 73,
496–505 (2016).
12. M. A. Peck et al., Prog. Oceanogr. 191, 102494 (2021).
13. S. Medoff, J. Lynham, J. Raynor, Science 378, 313–316 (2022).
14. N. A. J. Graham et al., Curr. Biol. 27, 231–236 (2017).
15. M. McLean et al., Glob. Change Biol. 25, 3972–3984 (2019).
16. N. E. Bosch et al., Conserv. Biol. 36, e13807 (2022).
17. See supplementary materials and methods.
18. R. Froese, P. Pauly, FishBase (2022); https://www.fishbase.org.
19. N. Pacoureau et al., Nature 589, 567–571 (2021).
20. D. A. Kroodsma et al., Science 359, 904–908 (2018).
21. D. Tickler, J. J. Meeuwig, M.-L. Palomares, D. Pauly, D. Zeller,
Sci. Adv. 4, eaar3279 (2018).
22. M. Di Lorenzo et al., Nat. Commun. 13, 4381 (2022).
23. J. L. Blanchard, R. F. Heneghan, J. D. Everett, R. Trebilco,
A. J. Richardson, Trends Ecol. Evol. 32, 174–186 (2017).
24. C. M. Duarte et al., Glob. Ecol. Biogeogr. 31, 1422–1439 (2022).
25. R. Trebilco, J. K. Baum, A. K. Salomon, N. K. Dulvy, Trends Ecol. Evol.
28, 423–431 (2013).

y g

1 March 2024

Conclusions

y

International policy, including the KunmingMontreal Global Biodiversity Framework (GBF)
COP15 declaration of 30% of the ocean to be
protected by 2030 (6), requires that extensive
areas of the oceans are set aside for protection
in order to enhance biodiversity, ecosystem
function, and ecological integrity and connectivity. To meet multiple GBF targets and address
several of the UN Sustainable Development
Goals, our analysis addressed two questions that
are critical to the implementation of MPAs, related to ecological indicators and MPA placement, and one question concerning sustainable
fisheries practices more broadly.
1) Particular characteristics of pelagic systems result in size structure that is highly
sensitive to human pressure and render size
indicators a powerful guide for priority placements of spatial protection, monitoring, and
ecosystem-based management. In benthic systems, size indicators are comparatively less sensitive, so decisions should be informed through
other indicators such as biomass (7) or functional diversity (49).
2) Pelagic vulnerability across multiple size
classes reinforces the need for protection to
provide refugia and rebuild depleted populations. A reversal of ongoing marine megafauna
loss (19) is possible but requires intervention
efforts that include implementation of highly
protected MPAs in remote locations, including
on the High Seas, consistent with the new
High Seas Treaty (50). Homogenization of
pelagic and benthic size structures signals
the extent of already-experienced human impacts on benthic systems. For benthic systems,
we confirm that protection would offer greater
relative benefits in accessible locations (7),
which should also be prioritized in order to
rebuild coastal habitats and ecosystems.
3) Human impact across pelagic size classes
indicates that it is not just the large predators
that are vulnerable but also smaller-sized species, which underpin major fisheries, such as
the anchoveta and sardines (12). Whether for
single species or “balanced harvesting” strategies that target the entire size spectra, pelagic
fisheries remain attractive to the commercial
industry (12, 19). However, top-down control
and low body-size redundancy are characteristics that render pelagic ecosystems inherently
dynamic and vulnerable to overexploitation.
We therefore caution against further expansion in pelagic fisheries, many of which are
already overexploited or fully exploited, particularly as long as pelagic megafauna and the

top-down control they exert remain threatened (19).

g

Letessier et al., Science 383, 976–982 (2024)

habitat degradation scenarios, benthic size
spectra are in fact expected to adopt characteristics more reminiscent of those of pelagic
systems, with more pronounced peaks and
greater spread (4), which reflects loss in sizestructured refugia at intermediate sizes.

p

as a result of high species richness (14). For example, increases in the relative proportion of
trigger fish and wrasse are observed to counteract prey release of sea urchin, after depletion in high trophic levels (14), which results
in greater food web flexibility and resilience.
Benthic habitat complexity, which offers refugia for fish of intermediate sizes (30 to 500 g),
may act further to moderate top-down control
(4). Conversely, pelagic systems are associated
with lower species richness and carnivorous
feeding strategies with larger movement scales
(19) across a wider range of body sizes, which
results in low size-based redundancy. Trophic
replacements have been reported in a pelagic
food web (44): In the Benguela upwelling, a
benthic species (the bearded goby Sufflogobius
bibarbatus) was discovered to thrive after the
depletion of sardines (Sardinops sagax) as a result
of distinctive foraging behavior and physiological adaptations to anoxia. This replacement,
which involved the emergence of a previously
unknown benthic-pelagic association in response to external pressure, suggests that lack
of resilience in pelagic food webs is associated
with low size-based redundancy and limited
alternative energy pathways.
Disentangling ecological processes from human pressures is notoriously complicated by
the correlated and often confounding nature
of human activities. In this work, potentially
confounding differences in exploitation histories and fisheries practices exists between
pelagic and benthic ecosystems. Benthic trawl
fisheries were some of the first to be developed
after industrialization (45), whereas pelagic
fisheries developed comparatively later (21),
under rising profit requirements (46). As such,
a loss of baseline and a preselection of particular sizes likely occurred before our surveys
(47). However, potentially confounding histories in each system is unlikely to explain the
distinction in size-structured characteristics
or the divergent responses to human pressure.
This is because human pressure near markets
resulted in pelagic and benthic systems that
are more similar in size structure than their
remote and more pristine counterparts, with
greater overlap in size spectra slope values
and convergent size structure. If the effect of
market distance on size spectra or the general
distinction between pelagic and benthic systems were confounded by historical size preselection, we would expect to see remote pelagic
and benthic systems with greater overlap in
size spectra value than those near market, as a
reflection of more pristine and therefore less
distinct states in those remote locations, in
contrast to our results. Moreover, that historical baselines in pelagic and benthic systems
are likely more characteristic and dissimilar to
each other than they are in their present state
is consistent with hypothesized preselection
from historical habitat loss (45, 48): Under

RES EARCH | R E S E A R C H A R T I C L E

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adi7562
Materials and Methods
Supplementary Text
Figs. S1 to S9
Tables S1 to S5
References (53–80)
MDAR Reproducibility Checklist

y

This work and associated analyses were only made possible
through the dedicated video-processing efforts from numerous

la république en Polynésie Française; Galapagos National Park
Directorate, El rol de los islotes oceánicos para la protección de
especies marinas migratorias; Government of French Polynesia
Declaration 01/03/2013; Government of New Caledonia Convention
120325; Marine Fisheries Administration, Port Sudan, Red Sea State;
Ministerio de Medio Ambiente y Desarrollo Sostenible, Ministry of
Fisheries, Marine Resources and Agriculture; Environment Protection
Agency Permit for route through MPA work; Ministry of Foreign
Affairs of the Kingdom of Tonga; Niue Department of Fisheries,
Agriculture and Forestry; Regional Government of the Azores; Palau
National Government; Sistema Nacional de Áreas de Conservación
(SINAC); Ministerio de Medio Ambiente y Energía; Serviço do
Parque Natural da Madeira; Norwegian Directorate of Fisheries;
Southern and Northern Province of New Caledonia; Tristan da
Cunha Government. Author contributions: Authors 4 to 17 appear
in alphabetical order and contributed equally to this work.
Conceptualization: T.B.L., D.M., L.Vi., and J.J.M.; Data curation:
T.B.L., J.J.M., H.J.C., L.M.; Formal analysis: T.B.L., D.M., L.M.; Funding
acquisition: J.J.M., E.S., T.B.L., L.Vi., D.M., N.M., E.M.; Visualization:
T.B.L.; Writing – original draft: T.B.L., D.M., L.M.; Writing – review
and editing: D.M., L.M., H.J.C., E.M.E., S.M.E., A.M.F., A.H., J.-B.J.,
A.R.K., E.M., N.M., P.J.N.-K., E.S., C.D.H.T., L.Ve., J.J.M. Competing
interests: The authors declare no competing interests. Data
and materials availability: Data from New Caledonia, Tonga, and
French Polynesia for seabed BRUVS (51) and for midwater BRUVS
(52) are available at Zenodo. The remaining data from the
other 77 locations and reproducible code for this analysis are
available at https://github.com/LauraMannocci/sizespectra and
can be found on the FishBase BRUVS portal (www.fishbase.org).
License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journalarticle-reuse

g

ACKN OW LEDG MEN TS

colleagues, technicians, and students. In particular, we thank
P. Bouchet, T. Langlois for Global Archive, S. Weber, A. López,
G. Kendrick, J. Clough, N. Casajus, J. Monk, D. Tickler, and
G. Boussarie. We are also grateful for the assistance of the master
and crew of the numerous vessels from which the field work was
conducted. T.B.L. acknowledges OKEANOS, Department of
Oceanography and Fisheries, University of the Azores, for hosting
him and thus facilitating collegial discussions around the analysis.
Other aspects of this work also benefited from discussions
with experts and colleagues at the Institute of Marine and Antarctic
Studies, University of Tasmania, and at the Australian Antarctic
Program. Funding: This work was funded by the Australian
Institute of Marine Science; PTT Exploration and Production PLC;
Australian Academy of Science; Chevron; Darwin Initiative (grant
no. DPLUS063); European Union’s BEST initiative (grant no. 1599);
Fisheries Research and Development Corporation; Ian Potter
Foundation; Jock Clough Foundation; MERL; National Geographic Pristine Seas; Natural Heritage Trust; National Environmental
Research Program (UK); National Environmental Science Program
(AUS); Pilbara Marine Conservation Partnership (AUS); Rottnest
Island Authority; TeachGreen; Totale (Fr); Vermilion Oil and Gas
Australia; Waitt Institute; WA Marine Science Institute; Woodside
Energy; Galapagos Conservation Trust; Galapagos Science Center;
MigraMar (EC); European Union, MARHAB (grant no. 101135307);
Norwegian Agency for Development Cooperation (NORAD); Norwegian
Embassy in Khartoum (SD) through UNIDO (SAP ID 130130); Norway
county municipality of Trøndelag; municipalities of Hitra, Frøya,
and Tvedestrand; French Oceanographic fleet through Pristine and
Apex campaigns; and the IMR Coastal Ecosystems Programme.
T.B.L. was funded by the synthesis center CESAB of the French
Foundation for Research on Biodiversity (FRB), the Mediterranean
Centre for Environment and Biodiversity Laboratory of Excellence
(CeMEB LabEx) (https://www.labex-cemeb.org), and the Bertarelli
Foundation. Permits: All research activities were conducted
under national authority permits issued by Ascension Island
Government (ERP-2017-08); Australian Commonwealth Government
[PKNP_2016_1, PA2018-00091-1 (variation PA2018-00091-3),
PA2018-00091-2 (variation PA2018-00091-4),CMR-16-000426,
CMR-18-000550, CMR-17-000526, PA2018-00036-1, PA201800079-1, CMR-17-000526, CMR-17-000526]; Australian
Government Great Barrier Reef Marine Park Authority G17/39150.1,
DBCA 01-000049-8; Delegation regionale à la recherche et à la
technologie - Haut-commissariat de la Republique en Polynesie
Francaise, 05/08/2014; Department of Parks and Wildlife 01000049-7; Directorate of Fisheries 23/4532; Foreign and
Commonwealth Office, British Indian Ocean Territory Directorate;
France’s Ministry of Ecological Transition, Le Haut commissariat de

p

26. K. L. Nash, N. A. J. Graham, S. K. Wilson, D. R. Bellwood,
Ecosystems 16, 478–490 (2013).
27. A. Rogers, J. L. Blanchard, P. J. Mumby, Curr. Biol. 24,
1000–1005 (2014).
28. J. L. Blanchard et al., J. Anim. Ecol. 78, 270–280 (2009).
29. R. A. Morais, D. R. Bellwood, Curr. Biol. 29, 1521–1527.e6 (2019).
30. J. Mourier et al., Curr. Biol. 26, 2011–2016 (2016).
31. C. S. Bird et al., Nat. Ecol. Evol. 2, 299–305 (2018).
32. P. D. van Denderen, M. Lindegren, B. R. MacKenzie,
R. A. Watson, K. H. Andersen, Nat. Ecol. Evol. 2, 65–70 (2018).
33. Y.-J. Shin, M.-J. Rochet, S. Jennings, J. G. Field, H. Gislason,
ICES J. Mar. Sci. 62, 384–396 (2005).
34. A. M. Edwards, J. P. W. Robinson, M. J. Plank, J. K. Baum,
J. L. Blanchard, Methods Ecol. Evol. 8, 57–67 (2017).
35. A. C. AitkenIV, Proc. R. Soc. Edinb. 55, 42–48 (1936).
36. G. J. Edgar et al., Nature 506, 216–220 (2014).
37. K. Grorud-Colvert et al., Science 373, eabf0861 (2021).
38. E. Maire et al., Ecol. Lett. 19, 351–360 (2016).
39. J. P. W. Robinson et al., Glob. Change Biol. 23, 1009–1022
(2017).
40. T. A. Branch et al., Nature 468, 431–435 (2010).
41. D. Pauly, V. Christensen, J. Dalsgaard, R. Froese, F. Torres Jr.,
Science 279, 860–863 (1998).
42. W. J. Ripple et al., Trends Ecol. Evol. 31, 842–849 (2016).
43. T. Fung, K. D. Farnsworth, D. G. Reid, A. G. Rossberg,
Nat. Commun. 6, 6657 (2015).
44. A. C. Utne-Palm et al., Science 329, 333–336 (2010).
45. C. Roberts, The Unnatural History of the Sea: The Past and
Future of Humanity and Fishing (Gaia, 2007).
46. S. A. Sethi, T. A. Branch, R. Watson, Proc. Natl. Acad. Sci. U.S.A.
107, 12163–12167 (2010).
47. D. Pauly, Trends Ecol. Evol. 10, 430 (1995).
48. N. A. J. Graham et al., Conserv. Biol. 21, 1291–1300 (2007).
49. A. Dalongeville et al., J. Appl. Ecol. 59, 2803–2813 (2022).
50. United Nations General Assembly, “Draft agreement under the
United Nations Convention on the Law of the Sea on the
conservation and sustainable use of marine biological diversity
of areas beyond national jurisdiction” (United Nations, 2023).
51. L. Vigliola, G. Mou Tham, J.-B. Juhel, BRUVS_Seabed_APEX.
Zenodo (2024); https://doi.org/10.5281/zenodo.7793637.
52. L. Vigliola, G. Boussarie, T. Letessier, J. Meeuwig,
BRUVS_Midwater_APEX. Zenodo (2024); https://doi.org/10.5281/
zenodo.7793697.

Submitted 16 May 2023; accepted 24 January 2024
10.1126/science.adi7562

y g
y
,

Letessier et al., Science 383, 976–982 (2024)

1 March 2024

7 of 7

RES EARCH

EVOLUTION

A morphological basis for path-dependent evolution
of visual systems
Rebecca M. Varney1*, Daniel I. Speiser2, Johanna T. Cannon1, Morris A. Aguilar1,
Douglas J. Eernisse3, Todd H. Oakley1*
Path dependence influences macroevolutionary predictability by constraining potential outcomes after
critical evolutionary junctions. Although it has been demonstrated in laboratory experiments, path
dependence is difficult to demonstrate in natural systems because of a lack of independent replicates.
Here, we show that two types of distributed visual systems recently evolved twice within chitons,
demonstrating rapid and path-dependent evolution of a complex trait. The type of visual system
that a chiton lineage can evolve is constrained by the number of openings for sensory nerves in its shell
plates. Lineages with more openings evolve visual systems with thousands of eyespots, whereas those
with fewer openings evolve visual systems with hundreds of shell eyes. These macroevolutionary
outcomes shaped by path dependence are both deterministic and stochastic because possibilities
are restricted yet not entirely predictable.

,

1 March 2024

y

Varney et al., Science 383, 983–987 (2024)

y g

*Corresponding author. Email: varney@ucsb.edu (R.M.V.);
oakley@ucsb.edu (T.H.O.)

y

University of California, Santa Barbara, Santa Barbara, CA,
USA. 2University of South Carolina, Columbia, SC, USA.
California State University, Fullerton, Fullerton, CA, USA.

3

g

1

could be identified as critical junctions. These
functionally similar outcomes may still differ in form, so critical junctions may act to
direct evolutionary pathways toward functionally similar but morphologically distinct
outcomes. Second, even if convergent evolution
reveals potential critical junctions, convergent
evolution of traits occurs most commonly in
organisms with very different body plans and
ecologies, which are likely to exert different
selective pressures on traits. Therefore, most
instances of convergent evolution are not effective replicates for establishing path dependence (12). Finally, accurately reconstructing
the evolutionary histories of convergent traits
requires understanding of ancestral conditions
and knowledge of the timing of key transitions
in character states. This requires a detailed fossil
record and/or a robust phylogenetic history beyond that which is available for many lineages.
Together, these obstacles make identification
of critical junctions and path dependence in
natural systems enormously challenging.
By overcoming the challenges imposed by
other natural traits, the visual systems of chitons
(Mollusca; Polyplacophora) provide a compelling
case to test hypotheses about path-dependent
evolution. First, morphologically distinct visual
systems may have evolved separately in different lineages of chitons (21). Chiton visual systems likely evolved from aesthetes, which are
numerous, microscopic sensory organs embedded in the eight articulating shell plates
of these heavily armored mollusks (22). Aesthetes likely have multiple sensory functions,
including sensitivity to light, but they do not
confer vision (21, 23, 24). In most chitons, nerves
from aesthetes run through narrow channels
in the shell plates before exiting through slits
at the edges of each plate (Fig. 1, A and B)
(25, 26). In some lineages, pigmented clusters of photoreceptors (20 to 35 mm wide), hereafter referred to as eyespots, are attached to

p

E

stablishing the extent to which the evolutionary trajectories of complex systems
are contingent on historical events, a
phenomenon called path dependence
(1, 2), is fundamental for understanding the predictability of evolution. If there is
a single, locally optimal solution to an environmental problem, then evolution will tend
to be predictable; if many functionally similar solutions exist, then evolution will tend
to be unpredictable (3–6). Path dependence
occurs when evolutionary trajectories contain “critical junctions,” which we define as
events that commit lineages to one of multiple
possible evolutionary pathways, thereby constraining the suite of possible outcomes. Path
dependence is well established for the evolution of particular proteins in unicellular laboratory systems [e.g., (7–13)]. Although specific
evolutionary outcomes are restricted by earlier
events in some natural systems [e.g., (14–16)],
path dependence is very difficult to establish
in any system outside of the laboratory.
Demonstrating path dependence in natural
systems is challenging because it requires the
identification of critical junctions and elucidation of the constraints that those junctions
impose on future evolutionary paths (17–20).
First, critical junctions are difficult to identify
because alternative evolutionary pathways are
often not observable along the singular history
of life. Path dependence may be inferred from
convergent origins of complex traits because
these events illustrate multiple evolutionary
pathways, analogous to replicates in controlled
laboratory experiments. If splits in evolutionary
trajectories lead to functionally similar outcomes in separate lineages, then those splits

aesthetes (Fig. 1, D, G, and J) (27, 28). In other
lineages, the aesthetes are interspersed with
camera-type eyes with image-forming lenses
made of shell material (up to 145 mm wide),
hereafter referred to as shell eyes [(22, 29–32);
see Fig. 1, C, F, I]. If eyespots and shell eyes
evolved separately in chitons, then these distributed visual systems may represent distinct
evolutionary paths to a convergent functional
outcome: spatial vision. Indeed, computational
modeling and behavioral experiments indicate
that both the eyespot- and shell eye–based distributed visual systems of chitons provide spatial vision (26, 31–33). Second, chitons have a
relatively rich fossil record, permitting timecalibrated phylogenetic analyses (34). If the
distributed visual systems of chitons have recent
origins, then their evolutionary histories may
be reconstructed with greater confidence than
those of other visual systems, which largely
have ancient histories (35). Finally, fossil and
extant chitons are found in similar environments and thus tend to be ecologically similar:
Most species live (or lived) on hard substrates
in intertidal or shallow subtidal habitats. The
body plan present in both fossil and extant
chitons is consistent across clades and evolutionary time (36–39). Species exhibiting the full
range of shell-embedded sensory organs can
even be found living on the same rock (40).
Here, we investigated whether the evolution
of distributed visual systems in chitons is path
dependent by mapping the origins of eyespots
and shell eyes onto the most comprehensive
chiton phylogeny produced to date. We then
used the rich fossil record of chitons to timecalibrate our phylogeny, and graphed outcomes onto a phylomorphospace to identify
specific points where earlier events committed
some lineages to one or another specific evolutionary pathway, resulting in multiple solutions
to the evolution of spatial vision. Our discoveries
show that the evolution of complex visual systems, which are often portrayed as deterministic
[e.g., (41)], is path dependent: Events at specific
points (critical junctions) constrain lineages to
one of a subset of possible pathways.
Chitons rapidly evolved visual systems
four times in two distinct forms

To characterize patterns of visual system evolution in chitons, we used genomic target capture and Bayesian inference to produce the
most complete phylogeny of chitons to date,
with emphasis on Chitonina, the suborder that
includes more than half of all extant chiton species and most species with eyespots or shell
eyes. We found that distributed visual systems
evolved separately in chitons at least four times:
two lineages through eyespots and two other
lineages through shell eyes (Fig. 2, A and B).
The two lineages that contain species with
eyespots, Callochitonida (28) and Chitonidae:
Chitoninae (27, 42), are distantly related to
1 of 5

RES EARCH | R E S E A R C H A R T I C L E

A

B

C

D

E

F

G

H

I

J

i.s.
s.r.
p
g
y

1 March 2024

placophorans (38) have or had aesthetes, but, as
ASR reveals, eyespots and shell eyes are recent
and nonhomologous additions to chiton sensory systems. Eyespots evolved independently in
Callochitonida and Chitonidae: Chitoninae, and
likewise shell eyes evolved independently within Chitonidae: Acanthopleurinae + Toniciinae
and in Schizochitonidae. Within Chitonidae,
eyespots evolved in a subclade in Chitoninae and
shell eyes evolved in the last common ancestor
of the other two subfamilies, Acanthopleurinae
and Toniciinae. ASR showed that these visual
systems evolved separately: The last common
ancestor of Chitonidae only had aesthetes (95%
2 of 5

,

Varney et al., Science 383, 983–987 (2024)

imens available. Nevertheless, we are confident
in our results because tests of branch stability
do not indicate phylogenetic uncertainty in the
placement of S. incisus in any of our analyses
(see the supplementary materials, section
5.3, “Leaf instability testing”).
Next, to assess support for independent origins of distributed visual systems in chitons, we
performed ancestral state reconstruction (ASR).
Using ASR, we found high support (≥95% proportional marginal likelihood) for all four instances of visual system evolution in chitons
occurring independently. Not only all Chitonina,
but all living chitons and even ancient fossil poly-

y

one another, consistent with previous molecular phylogenies (43–45). Likewise, distributed visual systems based on shell eyes also
evolved twice separately, once in Chitonidae:
Acanthopleurinae + Toniciinae and once in
Schizochitonidae, which in our phylogeny is a
sister to the rest of Chitonina, making it a distant relative of Acanthopleurinae + Toniciinae.
The placement of the sole extant genus of
Schizochitonidae, Schizochiton, has been uncertain across studies of chitons (43, 46, 47) in part
because Schizochiton contains only two accepted
species, S. incisus (46) and S. jousseaumei
(Dupuis, 1917), and S. incisus were the only spec-

from it. (F) Internal morphology of a shell eye, with photoreceptor cells
forming a retina (orange) beneath a lens (blue) and a large optic nerve running
through the upper layer of the shell plate (gray). (G) Internal morphology
of an aesthete with an attached eyespot, with a patch of photoreceptor cells
(orange) beneath a clear portion of the shell plate (red) and a nerve running
through the upper layer of the shell plate (gray). Note that panels (B) to (G)
are not to scale. (H) SEM image of the surface of an anterior shell plate from
Katharina tunicata, a chiton with aesthetes only. The location of a single
macraesthete is circled in green. Scale bar, 100 mm. (I) SEM of the surface of an
anterior shell plate from Acanthopleura brevispinosa, a chiton with shell eyes. The
location of a single macraesthete is circled in green, and a single shell eye is
circled in blue. Scale bar, 100 mm. (J) SEM image of the surface of an anterior shell
plate from Chiton marmoratus, a chiton with eyespots. The location of a single
macraesthete is circled in green, and a single eyespot is circled in red. Note
that eyespots are connected to macraesthetes but appear as open regions of shell
plate on SEM. The pigment of eyespots is only visible through decalcification
of shell plates. Scale bar, 100 mm.

y g

Fig. 1. Sensory organs embedded in chiton shell plates differ in morphology.
(A) A chiton (top) with the most anterior shell plate outlined in orange and
popped out below. On the ventral side (middle), slit rays (s.r.) are visible leading
to each insertion plate slit (i.s.). On the dorsal side (bottom), only insertion
slits are visible. (B) Distribution of aesthetes (green) on the shell plate of a chiton
with only aesthetes. Nerves (green lines) from aesthetes run through channels
in the tegmentum, the visible outer layer of the shell (gray), exit the shell plate
through an insertion slit, and then join the lateral neuropil (black circle), a part of the
chiton nervous system. (C) Distribution of aesthetes (green) and shell eyes (blue) on
the shell plate of a chiton with shell eyes. The density of aesthetes is higher, and
nerves from both aesthetes and shell eyes travel through the shell plate to exit
through an insertion slit. (D) Distribution of aesthetes (green) and eyespots (red) on
the shell plate of a chiton with eyespots. Each eyespot is paired with one aesthete,
and nerves from both aesthetes and eyespots travel through the shell plate to
exit through several insertion slits. (E) Internal morphology of an aesthete, the
simplest sensory structure embedded in the upper layer of the shell plate (gray).
Here, a macraesthete (green) is depicted with a micraesthete (pink) branching

RES EARCH | R E S E A R C H A R T I C L E

A

B
Aesthetes only
Eyespots
Shell eyes

p

slit #
300

Outer ring indicates broader phylogenetic groups of chitons (gray boxes).
(B) Time-calibrated phylogeny generated with Bayesian inference showing
four independent origins of visual systems in chitons. Divergence times
correspond to the geologic time scale below. ASR implies that all origins of
eyespots or shell eyes in chitons come from an aesthete-only starting point,
where all proportional marginal likelihoods are >95%. Additional support
metrics are included in the supplemental materials (tables S2 and S8 and
figs. S2, S3, and S9).

Number of slits in shell plates is a critical
junction in the evolution of chiton visual systems

To demonstrate path dependence in natural
systems, it is first necessary to identify critical
junctions. To discover critical junctions during
the evolution of visual systems in chitons, we
examined morphological differences between
the shell plates of species with only aesthetes,
species with eyespots, and species with shell eyes,
and examined ancestral states for these traits.
Most chitons integrate their shell-embedded
sensory organs into their nervous system by
passing nerves through slits along the edges of
3 of 5

,

1 March 2024

visual system in 2 million years, within an order of magnitude of the 7 million years that we
estimate for visual system evolution within
Chiton. For comparison, the only published
estimate of the time required to evolve an eye
is from vertebrates, in which eyes evolved in
~30 million years (35, 54). Recent origins of
shell eyes and eyespots in chitons allow us to
calibrate the timing of visual system evolution
with greater confidence from fossils. Thus,
the recent origins and rapid evolution of visual systems in chitons make them particularly valuable for understanding how complex
traits evolve.

y

from the Lower Paleocene (66 to 59.2 Ma ago)
(48–50). Time-calibrated phylogeny and fossil
evidence together indicate that both instances
of shell eyes in chitons represent the most recent
origins of camera-type eyes known. By comparison, the more ancient camera-type eyes of vertebrates and cephalopods originated at least
500 and 425 Ma ago, respectively (35, 51). Distributed visual systems based on eyespots may
have evolved even more recently than those
based on shell eyes. Time-calibrated phylogeny
and fossil evidence places the origin of eyespots
in Chiton between 25 and 75 Ma ago. By contrast, eyespots in Callochiton could be as old as
260 Ma ago, but this date is less certain.
To determine how rapidly chitons evolved
visual systems based on eyespots, we quantified the time between the origin of eyespots
in Chiton and the most recent ancestor in
Chitonina lacking pigmented eyespots. Using
our fossil-calibrated time tree, we found that
eyespots in Chiton originated within a period
of seven million years. Theoretical models estimate that eyes can evolve within 363,992 generations (3, 52), so if we assume that chitons
have a generation time of 3 years [based on
available studies of other chiton genera, e.g.,
(53)], then a lineage of chitons could evolve a

y g

Varney et al., Science 383, 983–987 (2024)

Ma ago

y

proportional marginal likelihood). Thus, each
of the four separate origins of visual systems in
chitons is inferred to have evolved independently, including convergent origins of two
different types of distributed visual systems,
one based on eyespots and the other on shell
eyes (Fig. 2A). Further, we found that eyespots
and shell eyes evolved from aesthetes separately in Chitonidae, rather than eyes evolving
from eyespots that evolved from aesthetes, a
stepwise pattern that would have corresponded
to the relative levels of morphological complexity demonstrated by these sensory organs.
To understand the timing of separate origins of visual systems in chitons, we used fossil
occurrence data to time-calibrate the phylogeny
and found that all four convergent visual systems
in chitons evolved within the past 260 million
years (Fig. 1B and fig. S8). Our time-calibrated
phylogeny indicates shell eyes evolved in the
last common ancestor of Acanthopleurinae +
Toniciinae between 150 and 100 million years
(Ma) ago and in Schizochitonidae between 250
and 200 Ma ago, estimates considerably older
than the earliest verified fossil evidence of shell
eyes in fossil Toniciinae from the middle Eocene
(48 to 38 Ma ago) and fossil Schizochitonidae
represented by two species of Incisiochiton

100

g

Fig. 2. Two types of distributed visual systems evolved convergently
in chitons, one based on eyespots (red) and the other on shell eyes
(blue). (A) The full maximum likelihood phylogeny of chitons produced by
this study (outgroups not shown). Branch coloration indicates ASR of the
number of slits in the anterior shell plate, where dark green represents
0 slits and pink represents >10 slits. Inner ring indicates the sensory
organs embedded in the shell plates of taxa (aesthetes only, aesthetes and
eyespots, or aesthetes and shell eyes), as well as phylogenetic affiliations.

200

RES EARCH | R E S E A R C H A R T I C L E

800
700
600
500
400

Macraesthetes/mm²

300
200
100

15

20

Slits (head valve)

,

4 of 5

y

1 March 2024

In light of the four separate origins of visual
systems in chitons and of the different morphological characters associated with them,
we hypothesized that evolutionary outcomes
in chiton visual systems were constrained by
path-dependent evolution. From this hypothesis, we predicted that critical junctions have
resulted in gaps in the morphospace of chiton
visual systems due to the absence of intermediate forms: Chitons can have either eyespots
or shell eyes, but not visual systems that share
morphological characters with both.
To test our hypotheses about path-dependent
evolution in chitons, we constructed a phylomorphospace based on morphological traits
associated with visual systems: aesthete density and number of slits in anterior shell plates.
From museum specimens representing taxa
across our phylogeny (fig. S4), we removed anterior shell plates and counted slits. We then
quantified macraesthete densities (Fig. 1, H to
J, green circles) from scanning electron microscope (SEM) images of these same plates and
plotted these values alongside the number of
slits in each species. Consistent with our prediction, we found a pronounced gap in the phylomorphospace, which suggests that slit number
acts as a constraint on the type of visual system that a lineage of chitons can evolve (Fig. 3).
The absence of intermediates in the phylomorphospace of chiton visual systems shows that
chitons evolved vision through one of two distinct paths and suggests that the morphological characters that define each type of visual
system are mutually exclusive. As in our ASR
analysis, our phylomorphospace analysis indicated that increases in slit number preceded the
evolution of eyespots in chitons. The number
of slits in the anterior shell plates of chitons
with eyespots was consistently higher than the
ancestral number of slits in those lineages, and
only those lineages of chitons that increased the
number of slits in their plates evolved eyespots.
Slit number therefore acts as a critical junction,
an event that commits a lineage of chitons to
one or the other evolutionary pathway toward
spatial vision despite being unrelated to vision
at the time of change. These critical junctions
make the evolution of visual systems in chitons
path dependent, and they are visible as a gap
in phylomorphospace created by mutually exclusive solutions for spatial vision that are
morphologically divergent yet functionally
convergent.

y g

Fig. 3. Path dependence of visual system evolution in chitons is indicated by a phylomorphospace of vision-related morphological traits,
showcasing two mutually exclusive solutions
to vision. Phylomorphospace of chiton visual
systems with number of slits on the anterior
shell plate on the x axis and aesthete density
(in aesthetes/mm2) on the y axis (37 species).
Lepidopleurida are excluded for visualization, but
a complete phylomorphospace is available in
fig. S12 (39 species). There are two separate origins
of shell eyes (blue squares indicate members of
Acanthopleurinae + Toniciinae, and the blue
diamond indicates the position of Schizochiton
incisus) and two separate origins of eyespots (red
triangles indicate the members of Chitoninae with
eyespots and red inverted triangles indicate the
members of Callochitonidae). Green circles indicate
chitons with only aesthetes. As predicted, there
is a gap in the morphospace, indicating that slit
number is a critical junction in chiton visual
system evolution. The gap results from the absence
of intermediate forms of visual systems in chitons.
Lineages are committed to one or the other
visual system at a critical junction (the central split
in lineages). Ancestral states of each lineage of
chitons with either shell eyes or eyespots are
included as open polygons of shapes corresponding
to those denoting extant species.

A phylomorphospace supports path-dependent
evolution of chiton visual systems

y

10

and Chitonida had eight or nine slits, indicating
an independent origin of the increased number
of slits in Callochitonida (fig. S8). In chitons, the
evolution of eyespots appears to follow an increase in slit number, but the evolution of shell
eyes does not, indicating that slit number is a
critical junction during visual system evolution.

g

Varney et al., Science 383, 983–987 (2024)

which we infer lacked eyespots, likely had 12
to 13 slits (fig. S8). Therefore, slits became
more numerous in the Chiton + Radsia +
Sypharochiton clade before the evolution of
eyespots within Chiton. In Callochitonida, the
other clade of chitons that evolved eyespots,
an eyeless sister species or clade is unavailable because all Callochitonida examined to
date appear to have eyespots (28). All species
of Callochitonida in our analysis had >16 slits.
Ancestral state reconstruction suggests that
the last common ancestor of Callochitonida

p

their shell plates [(22, 25, 46, 55, 56); fig. S5]. In
species with shell eyes, these slits are a vital
part of the distributed visual system, because
they make space for optic nerves to exit shell
plates and make synaptic contact with the central nervous system (21). We predicted that as
lineages of chitons added new types of sensory
organs to their shell plates (like eyespots or
shell eyes), they would require larger or more
numerous slits for additional nerves to pass
through. We compiled data on the number of
slits in the anterior shell plates of all chitons
in our phylogeny (table S2). Chitons may add
slits to their shell plates as they grow, but slit
number is not correlated with body size in
chitons (larger chiton species do not have more
slits) (fig. S14). Members of the clade sister to
all remaining extant chitons, Lepidopleurida,
lack slits. They innervate their aesthetes by running nerves through pores in their relatively
thin shell plates, and outgroup comparisons
indicate this is the ancestral state of crowngroup chitons [table S2; (34, 57)]. In the
remaining orders of chitons, Chitonida and
Callochitonida, species have thicker shell plates,
so slits are necessary to innervate aesthetes; all
of these chitons have at least five slits on their
anterior shell plates, and most have fewer
than 10 (37). Like most chitons that only have
aesthetes, most extant species with shell eyes
have 10 or fewer slits on their anterior shell
plates, including S. incisus, which has seven, and
the inferred ancestor of Acanthopleurinae +
Toniciinae (fig. S8), which has eight. In contrast
to species with shell eyes, all extant chitons
with eyespots that we examined have anterior
shell plates with between 14 and 21 slits. We
thus hypothesized that an increased number
of slits is a critical junction in the evolution of
visual systems in chitons, favoring the evolution of eyespots but not shell eyes.
If an increased number of slits is a critical
junction that imposes a functional constraint
favoring the evolution of eyespots instead of
shell eyes, then an increase in the number of
slits will predate the origin of eyespots themselves. Within Chitoninae (Fig. 2B), we compared the number of slits between species with
eyespots and those with only aesthetes. All
species with eyespots that we examined had
≥14 slits across their anterior shell plates. Chiton
cumingsii, sister to the remaining members of
Chiton in our phylogeny, does not have eyespots but has 14 slits (table S2 and personal
observation by D.E. and D.I.S.). Sister to Chiton,
species in the Radsia + Sypharochiton clade, all
of which lack eyespots, have 13 to 16 slits. Sister
to Chiton + Radsia + Sypharochiton, species in
the Rhyssoplax + Tegulaplax clade have eight
to 10 slits. We performed ancestral state reconstruction and found that the ancestors of
Chitonidae and of Chitoninae each most likely
had eight or nine slits, whereas the last common
ancestor of Chiton + Radsia + Sypharochiton,

RES EARCH | R E S E A R C H A R T I C L E

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adg2689
Materials and Methods
Supplementary Text
Figs. S1 to S14
Tables S1 to S9
References (61–100)
MDAR Reproducibility Checklist
Submitted 21 December 2022; resubmitted 16 June 2023
Accepted 11 January 2024
10.1126/science.adg2689

5 of 5

,

We thank P. Bouchet, B. Buge, and N. Puillandre at MNHN, Paris;
G. Paulay, J. Slapcinsky, and A. Bemis at UF, Florida Museum
of Natural History; D. Geiger and V. Delnavaz at SBMNH,
Santa Barbara, CA; and T. Gosliner and C. Piotrowski at CASIZ,
San Francisco, CA for help with their respective collections. We are
grateful to the following people for providing specimens in
accordance with each country’s collection regulations: I. Shita
Arlyza, P. Barber, L. Brooker, C. Cáceres Martínez, B. Dell’Angelo,
R. Emlet, A. França, M. Hendrickx, A. Hodgson, C. Ibáñez,
J. Whelpley, M. Langdon, S. Lockhart, P. Marko, T. Nakano,
R. Noseworthy, J. Noseworthy, R. Sagarin, B. Sirenko, J. Sigwart,
H. W. Detrich, C. Starger, S. Wiedrick, M. Weber, D. Willette,
and C. Young. We also thank J. Wolfe, N. Hensley, and members of
the Oakley lab for insightful remarks on the manuscript and
L. Brooker and K. Kocot for providing additional chiton transcriptome
data and support. J.T.C. thanks A. Swafford for assistance with code.
Funding: This work was supported by the National Science
Foundation (grant DEB 1354831 to D.I.S. and T.H.O.; grant DEB
1355230 to D.J.E.; grant EAGER 1045257 to T.H.O.; and grant
IOS 1754770 to T.H.O.). The Center for Scientific Computing (CSC)
is supported by the California NanoSystems Institute and the
Materials Research Science and Engineering Center (supported by
NSF grant DMR 1720256) at UC Santa Barbara. Use was made
of computational facilities purchased with funds from the National
Science Foundation (CNS-1725797) and administered by the CSC.
Author contributions: Conceptualization: D.I.S., D.J.E., T.H.O.;
Funding acquisition: D.I.S., D.J.E., T.H.O.; Investigation: R.M.V.,
D.I.S., J.T.C., M.A.A., D.J.E.; Methodology: R.M.V., D.I.S., J.T.C., M.A.A.,
D.J.E., T.H.O.; Project administration: T.H.O.; Supervision: T.H.O.;
Visualization: R.M.V.; Writing – original draft: R.M.V., D.I.S., T.H.O.;
Writing – review & editing: R.M.V., D.I.S., J.T.C., M.A.A., D.J.E., T.H.O.
Competing interests: The authors declare no competing interests.
Data and materials availability: All data and code used in the
analyses are available in the main text, the supplementary
materials, and in the Dryad repository (60). License information:
Copyright © 2024 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse

y

1 March 2024

AC KNOWLED GME NTS

y g

Varney et al., Science 383, 983–987 (2024)

1. J. B. Losos, T. R. Jackman, A. Larson, K. Queiroz,
L. Rodriguez-Schettino, Science 279, 2115–2118 (1998).
2. B. C. Dickinson, A. M. Leconte, B. Allen, K. M. Esvelt, D. R. Liu,
Proc. Natl. Acad. Sci. U.S.A. 110, 9007–9012 (2013).
3. D. E. Nilsson, S. Pelger, Proc. Biol. Sci. 256, 53–58 (1994).
4. D. Arendt, Int. J. Dev. Biol. 47, 563–571 (2003).
5. T. K. Suzuki, S. Tomita, H. Sezutsu, BMC Evol. Biol. 14, 229 (2014).
6. U. Schlüter, A. P. M. Weber, Plant Cell Physiol. 57, 881–889 (2016).
7. J. R. Meyer et al., Science 335, 428–432 (2012).
8. A. Spor et al., Evolution 68, 772–790 (2014).
9. M. J. Harms, J. W. Thornton, Nature 512, 203–207 (2014).
10. E. R. Jerison, M. M. Desai, Curr. Opin. Genet. Dev. 35, 33–39 (2015).
11. M. S. Johnson et al., eLife 10, e63910 (2021).
12. V. C. Xie, J. Pu, B. P. Metzger, J. W. Thornton, B. C. Dickinson,
eLife 10, e67336 (2021).
13. K. Karkare, H.-Y. Lai, R. B. R. Azevedo, T. F. Cooper, Mol. Biol.
Evol. 38, 2869–2879 (2021).
14. N. J. Dominy, J. C. Svenning, W. H. Li, J. Hum. Evol. 44, 25–45 (2003).
15. R. S. Waples, D. J. Teel, J. M. Myers, A. R. Marshall, Evolution
58, 386–403 (2004).
16. B. P. Williams, I. G. Johnston, S. Covshoff, J. M. Hibberd, eLife
2, e00961 (2013).
17. Z. D. Blount, R. E. Lenski, J. B. Losos, Science 362, eaam5979 (2018).
18. J. Beatty, J. Philos. 103, 336–362 (2006).
19. J. Beatty, I. Carrera, Journal of the Philosophy of History 5,
471–495 (2011).
20. E. Desjardins, Biol. Philos. 26, 339–364 (2011).
21. D. R. Chappell, D. I. Speiser, D. J. Eernisse, A. C. N. Kingston, in
Distributed vision: From simple sensors to sophisticated combination
eyes, E. Buschbeck, M. Bok, Eds. (Springer, 2023), pp. 147–167.
22. H. N. Moseley, J. Cell Sci. S2, 37–60 (1885).
23. D. J. Eernisse, P. D. Reynolds, in Microscopic Anatomy of
Invertebrates, F. W. Harrison, A. J. Kohn, Eds. (Wiley-Liss,
1994), vol. 5, Mollusca 1, pp. 56–110.
24. M. J. Vendrasco, C. Z. Fernandez, D. J. Eernisse, B. Runnegar,
Am. Malacol. Bull. 25, 51–69 (2008).
25. H. von Knorre, Jena Z. Naturwiss 61, 469–632 (1925).
26. D. R. Chappell, D. I. Speiser, J. Exp. Biol. 226, jeb244710 (2023).
27. W. Haas, K. Kriesten, Zoomorphologie 90, 253–268 (1978).
28. J. M. Baxter, M. G. Sturrock, A. M. Jones, J. Zool. 220, 447–468
(1990).
29. M. Nowikoff, Zoologie 88, 154–186 (1907).
30. P. R. Boyle, Nature 222, 895–896 (1969).
31. D. I. Speiser, D. J. Eernisse, S. Johnsen, Curr. Biol. 21, 665–670 (2011).
32. L. Li et al., Science 350, 952–956 (2015).
33. A. C. N. Kingston, D. R. Chappell, D. I. Speiser, J. Exp. Biol. 221,
jeb.183632 (2018).
34. S. S. Puchalski, D. J. Eernisse, C. C. Johnson, Am. Malacol. Bull.
25, 87–95 (2008).
35. T. D. Lamb, E. N. Pugh Jr., S. P. Collin, Evolution (N. Y.) 1,
415–426 (2008).
36. P. Kaas, A. M. Jones, K. L. Gowlett-Holmes, in Mollusca:
The Southern Synthesis, vol. 5, P. L. Beesley, G. J. B. Ross,
A. Wells, Eds. (CSIRO, 1998), pp. 161–194.
37. C. Z. Fernandez, M. J. Vendrasco, B. Runnegar, The Veliger 49,
51–69 (2007).
38. M. J. Vendrasco, T. E. Wood, B. N. Runnegar, Nature 429,
288–291 (2004).
39. J. Vinther, E. A. Sperling, D. E. G. Briggs, K. J. Peterson,
Proc. Biol. Sci. 279, 1259–1268 (2012).
40. P. Glynn, “On the ecology of the Carribean chitons Acanthopleura
granulata Gmelin and Chiton tuberculatus Linné: Density,
mortality, feeding, reproduction, and growth” in Smithsonian
Contributions to Zoology (Smithsonian, 1970).
41. R. Dawkins, Climbing Mount Improbable (Norton, 1997).
42. M. Nowikoff, Z. Wiss. Zool. 93, 668–680 (1909).
43. A. Okusu, Org. Divers. Evol. 3, 281–302 (2003).

y

Convergent evolution is often portrayed as an
inevitable feature of lineages moving toward
an optimal solution to an environmental problem (58). Such arguments dismiss contingency
as a lesser force than selective pressure and
assert that, given sufficient time, an optimized
trait will evolve in a deterministic manner. However, chiton visual systems present a morphospace with multiple solutions: Networks of
either eyespots or shell eyes provide chitons
with spatial vision. We found evidence for a
critical junction that defines convergent evolutionary pathways to spatial vision in chitons,
where lineages split into two discrete trajectories that led to mutually exclusive types of visual systems. A gap in the phylomorphospace
of chiton visual systems suggests that critical
junctions have constrained lineages of chitons
to particular evolutionary paths such that they
can evolve one type of visual system but not
another. Thus, no intermediate visual systems
appear in the phylomorphospace. The two
types of distributed visual systems of chitons
rely on differing morphological innovations,
like slit number in shell plates, that predate
the evolution of spatial vision. Therefore, evolutionary outcomes are constrained by earlier
evolutionary events. Previous studies, which

RE FERENCES AND NOTES

44. I. Irisarri, J. E. Uribe, D. J. Eernisse, R. Zardoya, BMC Evol. Biol.
20, 22 (2020).
45. J. Moles, T. J. Cunha, S. Lemer, D. J. Combosch, G. Giribet,
J. Molluscan Stud. 87, eyab019 (2021).
46. B. Sirenko, Venus (Tokyo) 65, 27–49 (2006).
47. E. Schwabe, Ruthenica 30, 55–68 (2020).
48. B. Dell’Angelo, M. Sosso, A. Kroh, A. Dulai, Bull. Geosci. 90,
359–370 (2015).
49. B. Dell’Angelo, J.-F. Lesport, A. Cluzaud, M. Sosso, Boll.
Malacol. 54, 1–47 (2018).
50. B. I. Sirenko, B. Dell’Angelo, Tr. Zool. Inst. 327, 128–134 (2023).
51. Y. Zhang et al., Nat. Ecol. Evol. 5, 927–938 (2021).
52. D. Berlinski, “A scientific scandal” (Discovery Institute, 2002);
https://www.discovery.org/a/1408/.
53. N. M. Otway, Mar. Biol. 121, 105–116 (1994).
54. T. D. Lamb, S. P. Collin, E. N. Pugh Jr., Nat. Rev. Neurosci. 8,
960–976 (2007).
55. P. R. Boyle, Cell Tissue Res. 153, 383–398 (1974).
56. F. P. von Fischer, Am. Malacol. Bull. 6, 153–159 (1988).
57. J. D. Sigwart, E. Schwabe, H. Saito, S. Samadi, G. Giribet,
Invertebr. Syst. 24, 560–572 (2011).
58. S. C. Morris, Philos. Trans. R. Soc. Lond. B Biol. Sci. 365,
133–145 (2010).
59. P. Shah, D. M. McCandlish, J. B. Plotkin, Proc. Natl. Acad.
Sci. U.S.A. 112, E3226–E3235 (2015).
60. Data and code for: R. M. Varney, D. I. Speiser, J. Cannon,
M. Aguilar, D. J. Eernisse, T. H. Oakley, A morphological basis
for path- dependent evolution of visual systems, Dryad
(2024); https://doi.org/10.25349/D9FC8D.

g

Conclusions

were confined to molecular experiments in laboratory environments, demonstrated that path
dependence can dictate the order of adaptations and the persistence of changes across
evolutionary time (2, 9, 59). Here, we have demonstrated path dependence in a naturally evolving system. Evolution is as much historical
as biological, so clarifying the role of history
in shaping evolutionary outcomes is critical
to our understanding of the extent to which
complex systems evolve in predictable ways.

p

Increases in aesthete density are associated
with the evolution of visual systems in chitons,
but aesthete density is not a critical junction
because it does not constrain the type of visual
system that a lineage can evolve. Most chitons
with either shell eyes or eyespots have a greater
density of aesthetes than chitons with only aesthetes (Fig. 3). The ancestral states predicted for
both clades of chitons with shell eyes (Fig. 3,
open blue square and open blue diamond, and
fig. S11) and the clade in Chitoninae with eyespots
(Fig. 3, open red triangle, and fig. S11) indicate
increases in aesthete density before or concurrent
with the evolution of distributed visual systems.
Because species in all four lineages that gained
visual systems tend to have denser arrays of
aesthetes than their sister lineages, an increase
in aesthete density may be a preadaptation for
a lineage of chitons evolving a visual system.
Slit number is a critical junction and not a
preadaptation because slit number constrains
the type of visual system that may evolve, but slit
number can increase without the subsequent
evolution of a visual system. When slit number
increased in a lineage of chitons but aesthete
density did not (Fig. 3, green circles), neither eyespots nor eyes evolved, emphasizing the role of
increased aesthete density as a preadaptation.
The evolution of visual system type in chitons is
thus constrained, rather than being deterministic or stochastic: When aesthete density increases
alongside slit number, lineages can evolve eyespots, and when aesthete density increases but slit
number does not, lineages can evolve shell eyes.

RES EARCH

STAR FORMATION

A far-ultraviolet–driven photoevaporation flow
observed in a protoplanetary disk
Olivier Berné1*, Emilie Habart2, Els Peeters3,4,5, Ilane Schroetter1, Amélie Canin1, Ameek Sidhu3,4,
Ryan Chown3,4, Emeric Bron6, Thomas J. Haworth7, Pamela Klaassen8, Boris Trahin2,
Dries Van De Putte9, Felipe Alarcón10, Marion Zannese2, Alain Abergel2, Edwin A. Bergin10,
Jeronimo Bernard-Salas11,12, Christiaan Boersma13, Jan Cami3,4,5, Sara Cuadrado14,
Emmanuel Dartois15, Daniel Dicken2, Meriem Elyajouri2, Asunción Fuente16, Javier R. Goicoechea14,
Karl D. Gordon9,17, Lina Issa1, Christine Joblin1, Olga Kannavou2, Baria Khan3, Ozan Lacinbala2,
David Languignon6, Romane Le Gal1,18,19, Alexandros Maragkoudakis13, Raphael Meshaka2,
Yoko Okada20, Takashi Onaka21,22, Sofia Pasquini3, Marc W. Pound23, Massimo Robberto9,17,
Markus Röllig20, Bethany Schefter3, Thiébaut Schirmer2,24, Thomas Simmer2, Benoit Tabone2,
Alexander G. G. M. Tielens23,25, Sílvia Vicente26, Mark G. Wolfire23, PDRs4All Team†

Berné et al., Science 383, 988–992 (2024)

1 March 2024

1 of 5

,

*Corresponding author. Email: olivier.berne@irap.omp.eu
†PDRs4All Team authors and affiliations are listed in the supplementary materials.

y

1
Institut de Recherche en Astrophysique et Planétologie, Université de Toulouse, Centre National de la Recherche Scientifique (CNRS), Centre National d’Etudes Spatiales, 31028 Toulouse,
France. 2Institut d’Astrophysique Spatiale, Université Paris-Saclay, CNRS, 91405 Orsay, France. 3Department of Physics and Astronomy, The University of Western Ontario, London, ON N6A 3K7,
Canada. 4Institute for Earth and Space Exploration, The University of Western Ontario, London, ON N6A 3K7, Canada. 5Carl Sagan Center, Search for ExtraTerrestrial Intelligence Institute,
Mountain View, CA 94043, USA. 6Laboratoire d’Etudes du Rayonnement et de la Matière, Observatoire de Paris, Université Paris Science et Lettres, CNRS, Sorbonne Universités, F-92190
Meudon, France. 7School of Physics and Astronomy, Queen Mary University of London, London E1 4NS, UK. 8UK Astronomy Technology Centre, Royal Observatory Edinburgh, Blackford Hill EH9
3HJ, UK. 9Space Telescope Science Institute, Baltimore, MD 21218, USA. 10Department of Astronomy, University of Michigan, Ann Arbor, MI 48109, USA. 11ACRI-ST, Centre d’Etudes et de
Recherche de Grasse, F-06130 Grasse, France. 12Innovative Common Laboratory for Space Spectroscopy, 06130 Grasse, France. 13NASA Ames Research Center, Moffett Field, CA 94035, USA.
14
Instituto de Física Fundamental, Consejo Superior de Investigaciones Científicas, 28006 Madrid, Spain. 15Institut des Sciences Moléculaires d’Orsay, Université Paris-Saclay, CNRS, 91405 Orsay,
France. 16Centro de Astrobiología, Consejo Superior de Investigaciones Científicas, and Instituto Nacional de Técnica Aeroespacial, 28850 Torrejón de Ardoz, Spain. 17Johns Hopkins University,
Baltimore, MD 21218, USA. 18Institut de Planétologie et d’Astrophysique de Grenoble, Université Grenoble Alpes, CNRS, F-38000 Grenoble, France. 19Institut de Radioastronomie Millimétrique,
F-38406 Saint-Martin d’Hères, France. 20I. Physikalisches Institut, Universität zu Köln, 50937 Köln, Germany. 21Department of Astronomy, Graduate School of Science, The University of
Tokyo, Tokyo 113-0033, Japan. 22Department of Physics, Faculty of Science and Engineering, Meisei University, Hino, Tokyo 191-8506, Japan. 23Astronomy Department, University of Maryland,
College Park, MD 20742, USA. 24Department of Space, Earth and Environment, Chalmers University of Technology, Onsala Space Observatory, SE-439 92 Onsala, Sweden. 25Leiden Observatory,
Leiden University, 2300 RA Leiden, Netherlands. 26Instituto de Astrofísica e Ciências do Espaço, P-1349-018 Lisboa, Portugal.

y g




y

(FUV) photons, those with energies below the
Lyman limit (energy E < 13.6 eV), dominate
the photoevaporation process. The effect affects
the disk mass, radius, and lifetime (7, 10, 12–18);
its chemical evolution (19–21); and the growth
and migration of any planets forming within
the disk (22). However, these processes have
not been directly observed.
Most observational constraints on the massloss rates associated with photoionization have
been obtained for objects in the Orion Nebula
known as proplyds, in which the ionization of
FUV-driven photoevaporation flows from disks
produces comet-shaped ionization fronts (23, 24).
Modeling of the observed ionization fronts of
proplyds has indicated mass-loss rates M ≈ 10−8
to 10−6 solar masses per year (M⊙ year−1) (25–27).
However, those observations did not determine the physical conditions (radiation field,
gas temperature, and density) at the locations
where the photoevaporation flows are launched.

Figure 1 shows optical and near-infrared images of the Orion Bar, a ridge in the Orion
Nebula (31) situated about 0.25 pc southeast
of the Trapezium Cluster of massive stars.
The western edge of the bar constitutes the
ionization front (Fig. 1B), which separates
regions where the gas is fully ionized and at
temperature T ~ 10 4 K from the neutral
atomic region at T ~ 500 to 1000 K. We investigated the source [BOM2000] d203-506
(hereafter d203-506) (32, 33), a protoplanetary
disk seen in absorption against the bright
background, which is located at the following
coordinates: right ascension 5h35m20s.357 and
declination −5°25′05″.81 (J2000 equinox). Previous observations of d203-506 found no sign
of an ionization front (32–34), indicating that
the radiation field reaching the disk is dominated by FUV photons.
We obtained near-infrared and submillimeter
observations of d203-506, with the James Webb
Space Telescope (JWST) and the Atacama Large
Millimeter Array (ALMA), respectively, both
at angular resolution ~0.1″ (corresponding to
~40 au at the distance of the Orion Nebula).
JWST images were taken in multiple broad and
narrow band filters using the Near-Infrared
Camera (NIRCam) instrument (35). We also obtained near-infrared spectroscopic observations using the integral field unit (IFU) of the
Near-Infrared Spectrograph (NIRSpec) instrument on JWST (35). The ALMA interferometric
data cubes provided maps of rotational emission lines from the molecules HCN and HCO+,
with a velocity resolution of 0.2 km s−1 (35).
Figure 2 compares the JWST and ALMA images to archival optical images from the Hubble

g

Y

oung low-mass stars are surrounded by
protoplanetary disks of gas and dust,
which have lifetimes of a few million
years (1–3) and are the sites of planet
formation (4). Planet formation is limited by processes that remove mass from the
disk, such as photoevaporation (5). This occurs
when the upper layers of protoplanetary disks
are heated by x-ray or ultraviolet photons. Radiative heating increases the gas temperature,
which brings the local sound speed above the
escape velocity of the disk, causing the gas to
escape from the system. Those photons could
be from the central star (6) or from nearby
massive stars (7). Because most low-mass stars
form in clusters that also contain massive stars,
most protoplanetary disks are exposed to external radiation, and so they are expected to
experience photoevaporation driven by ultraviolet photons during their lifetime (7–11).
Theoretical models predict that far-ultraviolet

Images of a photoevaporation flow

p

Most low-mass stars form in stellar clusters that also contain massive stars, which are sources of farultraviolet (FUV) radiation. Theoretical models predict that this FUV radiation produces photodissociation
regions (PDRs) on the surfaces of protoplanetary disks around low-mass stars, which affects planet
formation within the disks. We report James Webb Space Telescope and Atacama Large Millimeter
Array observations of a FUV-irradiated protoplanetary disk in the Orion Nebula. Emission lines are detected
from the PDR; modeling their kinematics and excitation allowed us to constrain the physical conditions
within the gas. We quantified the mass-loss rate induced by the FUV irradiation and found that it is
sufficient to remove gas from the disk in less than a million years. This is rapid enough to affect giant
planet formation in the disk.

In the regions where FUV photons penetrate
the disk, a photodissociation region (PDR) (28)
forms at the disk surface. Most observational
tracers of PDR physics (spectral lines of H2, O,
and C+) are in the near- and far-infrared wavelength ranges. The spatial scale of PDRs in externally illuminated disks is a few hundred
astronomical units (au), which corresponds to
angular sizes <1 arc sec (″) for the closest starforming clusters (12, 29, 30).

RES EARCH | R E S E A R C H A R T I C L E

A

B

C

d203-506

50 au

p

4000 au

g

Fig. 1. Optical and near-infared images of the
Orion Bar region. (A) HST optical image centered at
coordinates right ascension 5h35m20s.183 and
declination −5°25′06″.14. [O III] filter at 502 nm is in
blue, Ha filter at 656 nm is in green, and [N II] filter
at 658 nm is in red. [Credit: NASA/STScI/Rice
University/C. O’Dell et al. (47)]. (B) JWST near-infrared
image of the same region at the same scale. Filters
centered at 1.4 and 2.0 mm are in blue; at 2.77, 3.00,
3.23, 3.35, and 3.32 mm in green; at 4.05 mm in orange;
and at 4.44, 4.80, and 4.70 mm in red. In (A) and
(B), the Orion Bar separates regions where the gas is
fully ionized (upper right) from those where it is in
neutral form (lower left). (C) Zoomed-in view of the
d203-506 disk. Red is the NIRCam image in the
2.12-mm filter, which traces molecular hydrogen; blue is
the 1.64-mm filter, which traces [Fe II] emission lines;
and green is the emission in the 1.40-mm broad-band
filter, which traces scattered light.

A

B

C

D

E

F

G

H

I

y
y g
y
,

Fig. 2. Multiwavelength
images of the d203-506
disk. (A) HST optical image
in a Ha filter (23). (B to
F) JWST near-infrared
images (35). [(E) is reproduced with permission from
(33)]. (G to I) ALMA submillimeter images (35). In all
panels, the white-filled
ellipse indicates the size
and shape of the point
spread function or reconstructed telescope beam,
and the horizontal bar is
100 au. The white-dashed
ellipse in (D) indicates the
shape of the aperture used
for the extraction of the
NIRSpec spectrum in Fig. 4.
The wavelength and
physical assignment of
each image is labeled above
each panel; in (H) and (I),
(4-3) is an abbreviation of
(v = 0, J = 4 → 3). 1 Jy =
1 × 10−26 Wm−2 Hz−1.

and NIRCam H2 (v = 1 → 0, J = 3 → 1) at 2.12 mm,
respectively, which both show emission from
the PDR surrounding the disk and absorption
at the center. Both the H2 rovibrational and
HCO+ rotational emission lines trace warm
(gas kinetic temperatures Tgas ~ 500 to 1000 K)

(v = 0, J = 4 → 3) line, where v and J denote
the vibrational and rotational quantum numbers, respectively, at 354.505 GHz (845.664 mm),
which traces cold molecular gas. Figure 2, I
and E, shows emission maps of ALMA HCO+
(v = 0, J = 4 → 3) at 356.734 GHz (840.381 mm)

Tow
maard Tr
ssiv ape
e s ziu
tars m

Space Telescope (HST). The nearly edge-on (35)
dusty disk of d203-506 is visible in absorption
in all the HST and JWST images (Fig. 2, A to F)
but in emission in the 344-GHz (870 mm) dust
continuum (Fig. 2G). It is also seen in emission
with ALMA in Fig. 2H, which shows the HCN

Berné et al., Science 383, 988–992 (2024)

1 March 2024

2 of 5

RES EARCH | R E S E A R C H A R T I C L E

molecular gas in PDRs (31). The PDR is also
bright in the 3.35-mm NIRCam filter (Fig. 2F),
which is dominated by aromatic infrared band
(AIB) emission from ultraviolet-excited polycyclic aromatic hydrocarbon (PAH) molecules.
PAHs are known to be tracers of PDRs (36)
and have been previously mapped in a proplyd
in the Orion Nebula (37). The PDR in d203506 is spatially resolved and extends south
from the disk, forming a lobe shape. A jet is
visible in the NIRCam [Fe II] filter at 1.62 mm
(Fig. 2C). A bright emission spot is present in
the H2 and HCO+ images in the northwestern part of the PDR and is also visible in the
broad-band filter at 1.4 mm (Fig. 2B). The loca-

tion of this bright spot coincides with the region of interaction between the jet and the
PDR, which is visible only on the side facing
the Trapezium Cluster. There is also AIB emission in the 3.35-mm NIRCam filter at this location (Fig. 2F), which indicates ultraviolet
excitation. Figure 3 shows a schematic diagram of our interpretation of the morphological features in d203-506.
Physical properties of the PDR

Figure 4 shows the NIRSpec spectrum of d203506 (35). Numerous rovibrational emission lines
of CO (v = 1 → 0 and v = 2 → 1), OH (v = 1 → 0),
CH+ (v = 1 → 0), and H2 (up to J = 15) are

p
g
y










y g

Fig. 3. Schematic diagram of
our interpretation of d203-506.
Dark brown is the edge-on disk of
cold molecular gas, which appears
in absorption in the NIRCam
images but in emission in the
ALMA HCN and dust continuum
images (Fig. 2). Brown arrows
indicate molecular gas escaping
from this disk, which feeds the
photoevaporation flow. This
produces an envelope around the
disk (light brown shading), which
is delimited by the dissociation
front (orange band), where
molecular hydrogen is dissociated
into hydrogen atoms by FUV
photons (pink arrows) coming
from the Trapezium Cluster. The
transition from molecular gas in
the disk to atomic gas under
ultraviolet irradiation constitutes the PDR. Blue shows the jet from the central star, which corresponds to the [Fe II]
emission (Fig. 2C). The jet interacts with the envelope, producing a bright emission spot (yellow). The
surroundings of d203-506 (gray) consist of diffuse atomic gas.

detected. We interpreted these lines as coming from the PDR, so we traced the physical
conditions of gas in that region. We modeled
(35) the H2 lines using the Meudon PDR code
(38), which calculates the H2 excitation in
PDRs (Fig. 5). We derived the hydrogen number density nH and temperature of the gas in
the H2 emitting layer (compare with Figure
2E) as nH = 5.5 × 105 to 1.0 × 107 cm−3 and
Tgas = 1240 to 1260 K.
We fitted a Keplerian orbit model to the
HCN emission map (fig. S2) and used it to
set an upper limit on the mass of d203-506’s
central star M∗ < 0:3 M⊙ (35). Taking Tgas ~
1250 K asqdetermined
above, the speed of
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
B Tgas
soundcS ≡ 7=5k
¼ 3:3 km s1, where kB is the
mmH
Boltzmann constant, mH is the mass of the
hydrogen atom, and m is the ratio of total mass
to hydrogen mass of the gas, which we assume
equals the interstellar value [m = 1.4 (39)].
This value of cS exceeds the escape velocity
at distances from the central star above a critical value, defined as the gravitational radius
∗
rg ≡ GM
cS 2 (40), where G is Newton’s gravitational
constant. For M∗ < 0:3 M⊙ and Tgas ~ 1250 K,
we found rg < 26 au. This is much smaller than
the observed radial extent of the H2 emission
rH2 ¼ 132T13 au [and its height hH2 ¼ 56T13 au
(35)]. Therefore, the gas in this layer is not
gravitationally bound and flows outward from
the disk, at roughly the speed of sound. The
associated mass flux through the PDR is j =
mmHnHcS, and the total mass loss rate is M =
j × S, where S is the surface area of the H2emitting layer (35). Including the uncertainties
on rH2 , hH2 , nH, and Tgas (table S1), we calculated M = 1.4 × 10−7 to 4.6 × 10−6 M⊙ year−1.
We also investigated an alternative method
of determining the mass-loss rate, using 1D
dynamical models, and found a consistent
value of M (35). All the physical quantities
that we derived are listed in table S1.
Implications for planet formation

1 March 2024




3 of 5

,

Berné et al., Science 383, 988–992 (2024)




y

Fig. 4. JWST NIRSpec spectrum of d203-506. Colored tick marks indicate detected species, as labeled.
The broad emission bands at 3.3 and 3.4 mm are C-H vibrational emission from PAH molecules. Other
unlabeled lines are mostly atomic (e.g., [O I] or [Fe II]). There are no data at wavelengths 2.40 to 2.50 mm and
4.05 to 4.18 mm owing to gaps in the NIRSpec detectors.

Gas in protoplanetary disks is the raw material
from which giant planets form, so mass loss due
to photoevaporation can limit the formation of
such planets. The effects of FUV radiation depend
on the stellar mass, which sets the strength of the
gravitational field that acts to retain the gas.
Theoretical models of planet growth under
the influence of external FUV photoevaporation have predicted that FUV radiation fields
with intensityG0 ≳ 500 [where G0 is the ratio of
the ultraviolet radiation field to the average
value in the interstellar medium (41)] suppress giant planet formation around stars
with masses≲ 0:5 M⊙ (22). Our results for d203506 are consistent with this prediction: We
found that M∗ < 0:3 M⊙ and the radiation
field is G0 ≲ 105 (35), whereas the mass-loss
rate M = 1.4 × 10−7 to 4.6 × 10−6 M⊙ year−1
implies a disk depletion timescale t ≡ Md/M <
0.13 million years, where Md is the disk mass

RES EARCH | R E S E A R C H A R T I C L E

Best model (small grains)
Best model (large grains)
Observed intensities

p

38. F. Le Petit, C. Nehme, J. Le Bourlot, E. Roueff, Astrophys. J.
Suppl. Ser. 164, 506–529 (2006).
39. B. T. Draine, Physics of the Interstellar and Intergalactic
Medium (Princeton Univ. Press, 2011).
40. D. Hollenbach, D. Johnstone, S. Lizano, F. Shu, Astrophys. J.
428, 654 (1994).
41. H. J. Habing, Bull. Astron. Inst. Neth. 19, 421 (1968).
42. A. Stolte et al., Astron. Astrophys. 578, A4 (2015).
43. P. D. Sheehan, J. A. Eisner, Astrophys. J. 857, 18 (2018).
44. J. A. Johnson, K. M. Aller, A. W. Howard, J. R. Crepp,
Publ. Astron. Soc. Pac. 122, 905–915 (2010).
45. S. Reffert, C. Bergmann, A. Quirrenbach, T. Trifonov,
A. Künstler, Astron. Astrophys. 574, A116 (2015).
46. E. A. Bergin, C. Alexander, M. Drozdovskaya, M. Gounelle,
S. Pfalzner, arXiv:2301.05212 [astro-ph.EP] (2023).
47. NASA Hubblesite, Star Factory at the Center of the Orion
Nebula, Space Telescope Science Institute (2005);
https://hubblesite.org/contents/media/images/2005/12/
1674-Image.html.
48. J. Champion, ALMA cubes and integrated maps of the
d203-506 protoplanetary disk. Zenodo (2023); https://doi.org/
10.5281/zenodo.8196030.
49. I. Schroetter, JWST NIRSpec spectrum of the d203-506
protoplanetary disk. Zenodo (2023); https://doi.org/10.5281/
zenodo.10260214.
50. The Meudon PDR Team, Meudon PDR Code, version used for
Berné et al. Science 2024. Zenodo (2023); https://doi.org/
10.5281/zenodo.10488834.
AC KNOWLED GME NTS

1 March 2024

4 of 5

,

Berné et al., Science 383, 988–992 (2024)

y

1. J. P. Williams, L. A. Cieza, Annu. Rev. Astron. Astrophys. 49,
67–117 (2011).
2. K. E. Haisch Jr., E. A. Lada, C. J. Lada, Astrophys. J. 553,
L153–L156 (2001).
3. S. M. Andrews, Annu. Rev. Astron. Astrophys. 58, 483–528
(2020).
4. M. Keppler et al., Astron. Astrophys. 617, A44 (2018).
5. U. Gorti, R. Liseau, Z. Sándor, C. Clarke, Space Sci. Rev. 205,
125–152 (2016).
6. U. Gorti, D. Hollenbach, Astrophys. J. 690, 1539–1552
(2008).
7. F. C. Adams, D. Hollenbach, G. Laughlin, U. Gorti, Astrophys.
J. 611, 360–379 (2004).
8. C. J. Lada, E. A. Lada, Annu. Rev. Astron. Astrophys. 41, 57–115
(2003).
9. M. Fatuzzo, F. C. Adams, Astrophys. J. 675, 1361–1374
(2008).
10. A. J. Winter, J. D. Kruijssen, M. Chevance, B. W. Keller,
S. N. Longmore, Mon. Not. R. Astron. Soc. 491, 903–922
(2020).
11. A. J. Winter, T. J. Haworth, Eur. Phys. J. Plus 137, 1132
(2022).

y g

RE FE RENCES AND N OT ES

12. H. Störzer, D. Hollenbach, Astrophys. J. 515, 669–684
(1999).
13. A. Scally, C. Clarke, Mon. Not. R. Astron. Soc. 325, 449–456
(2001).
14. C. J. Clarke, Mon. Not. R. Astron. Soc. 376, 1350–1356
(2007).
15. F. Concha-Ramírez, M. J. C. Wilhelm, S. Portegies Zwart,
T. J. Haworth, Mon. Not. R. Astron. Soc. 490, 5678–5690
(2019).
16. A. J. Winter et al., Mon. Not. R. Astron. Soc. 478, 2700–2722
(2018).
17. R. B. Nicholson et al., Mon. Not. R. Astron. Soc. 485,
4893–4905 (2019).
18. G. A. L. Coleman, T. J. Haworth, Mon. Not. R. Astron. Soc. 514,
2315–2332 (2022).
19. C. Walsh, T. J. Millar, H. Nomura, Astrophys. J. Lett. 766, L23
(2013).
20. T. J. Haworth, Mon. Not. R. Astron. Soc. 503, 4172–4182
(2021).
21. R. D. Boyden, J. A. Eisner, Astrophys. J. 947, 7 (2023).
22. A. J. Winter, T. J. Haworth, G. A. Coleman, S. Nayakshin,
Mon. Not. R. Astron. Soc. 515, 4287–4301 (2022).
23. C. R. O’Dell, Z. Wen, Astrophys. J. 436, 194 (1994).
24. C. R. O’Dell, Astron. J. 115, 263–273 (1998).
25. D. Johnstone, D. Hollenbach, J. Bally, Astrophys. J. 499,
758–776 (1998).
26. W. Henney, S. Arthur, Astron. J. 116, 322–335 (1998).
27. W. J. Henney, C. R. O’Dell, Astron. J. 118, 2350–2368
(1999).
28. A. Tielens, D. Hollenbach, Astrophys. J. 291, 722 (1985).
29. H. Chen et al., Astrophys. J. 492, L173–L176 (1998).
30. J. Champion et al., Astron. Astrophys. 604, A69 (2017).
31. J. R. Goicoechea et al., Nature 537, 207–209 (2016).
32. J. Bally, C. R. O’Dell, M. J. McCaughrean, Astron. J. 119,
2919–2959 (2000).
33. O. Berné et al., Nature 621, 56–59 (2023).
34. E. Habart et al., Astron. Astrophys. 673, A149 (2023).
35. Materials and methods are available as supplementary
materials.
36. E. Peeters et al., Astron. Astrophys. 390, 1089–1113 (2002).
37. S. Vicente et al., Astrophys. J. Lett. 765, L38 (2013).

y

(35). This is faster than even very early planet
formation (42, 43).
A positive correlation has been observed
between stellar mass and the frequency of
Jupiter-mass exoplanets orbiting those stars
(44, 45), which we suggest could be due to
FUV radiation in stellar clusters during the
planet formation process. Dynamical and compositional studies of the Solar System indicate
that it formed in a stellar cluster that contained
one or more massive stars (46), and so it might
have been affected by FUV radiation.

g

Fig. 5. Observed and modeled H2. Line intensities for d203-506. Blue squares show the observed line
intensities (table S2), with error bars enclosing 50% uncertainty, estimated from the c2 (35). The
instrumental uncertainties are smaller than the symbols. Colored lines are the best-fitting models (35)
using small dust grains (orange) or large dust grains (green); circles show the model intensities of each
line. H2 line identifications are abbreviated; the full quantum levels corresponding to this notation are
listed in table S2.

We thank the anonymous referees for helpful feedback and
comments and C. O’Dell for providing comments. The JWST
helpdesk provided support with data reduction. Part of this work
used the DiRAC Data Intensive service at Leicester, operated by
the University of Leicester IT Services, which forms part of the
STFC DiRAC HPC Facility. Funding: A.F. thanks the Spanish MICIN
for funding support from grant PID2019-106235GB-I00. O.B.,
I.S., and A.C. are funded by the Centre National d’Etudes Spatiales
(CNES) through the APR program. J.R.G. and S.C. thank the
Spanish MCINN for funding support under grant PID2019106110GB-I00. P. Guillard thanks the University Pierre and Marie Curie,
the Institut Universitaire de France, the CNES, the Programme
National de Cosmologie and Galaxies (PNCG), and the Physique
Chimie du Milieu Interstellaire (PCMI) programs of Centre National
de la Recherche Scientifique/INSU for financial support. E.P. and
J.C. acknowledge support from the Institute for Earth and
Space Exploration, the Canadian Space Agency, and the Natural
Sciences and Engineering Research Council of Canada.
C.B. acknowledges funding from the San José State University
Research Foundation (grant 80NSSC22M0107) and the Internal
Scientist Funding Model (ISFM) Laboratory Astrophysics Directed
Work Package at NASA Ames. T.O. is supported by JSPS Bilateral
Program grant 120219939. T.J.H. is funded by a Royal Society
Dorothy Hodgkin Fellowship. A.F. was supported by the Spanish
program Unidad de Excelencia María de Maeztu CEX2020-001058M, financed by grant MCIN/AEI/10.13039/501100011033. NN is
funded through UAEU Program for Advanced Research (UPAR)
grant G00003479. Y. Zhang is funded by the National Science
Foundation of China (NSFC) grant 11973099 and the China
Manned Space Project grants CMS-CSST-2021-A09 and -A10.
M. Rö and Y.O. are supported by the Collaborative Research Centre
956, sub-project C1, funded by the Deutsche Forschungsgemeinschaft
(DFG) project ID 184018867. P. Merino acknowledges grants
EUR2021-122006, TED2021-129416A-I00 and PID2021-125309OA-I00
funded by MCIN/AEI/10.13039/501100011033 and European Union
NextGenerationEU/PRTR. A. Pathak acknowledges financial support
from the Department of Science and Technology Core Research
Grant (DST-CRG) SERB-CRG/2021/000907, and the Institutes
of Eminence (IoE) grant BHU incentive/2021-22/32439. M. Buragohain
acknowledges a DST INSPIRE Faculty fellowship. J. He is sponsored
by the Chinese Academy of Sciences (CAS), through a grant to the
CAS South America Center for Astronomy (CASSACA) in Santiago,
Chile. H. Zettergren acknowledges support from the Swedish Research
Council contract 2020-03437. Author contributions: O.B., E.H., and E.P.
led the JWST observing program. O.B. led the study and drafted
the manuscript and produced Figs. 1, 3, and 4 and figs. S1, S3, and
S4. E. B. and F.L.P. produced the PDR models and Fig. 5 and figs.
S4 to S6. T.J.H. produced the 1D dynamical models and fig. S7. F.A.
produced the disk 3D radiative transfer models and fig. S2. A. Canin
produced Fig. 2. J. Champion was principal investigator of the ALMA
observing program. J. Champion and E. Chapillon performed the
ALMA data reduction. P.K. analyzed the ALMA data. I.S., A.S., R.C., D.V.D.P.,
and F.A. reduced the NIRSpec data. A.C. and B.Tr. reduced the

RES EARCH | R E S E A R C H A R T I C L E

NIRCam data. I.S. extracted the NIRSpec spectrum. I.S. and M.Z.
measured the line intensities. M.Z. produced the LTE radiative
transfer models of CH+ and OH. J.C. provided the radiative transfer
models for CO. All other authors contributed to the research
presented in this paper or provided detailed feedback on
the manuscript. All authors meet the journal’s authorship
requirements. Competing interests: The authors declare no
competing interests. Data and material availability: The JWST
and HST data are available on the MAST portal http://mast.
stsci.edu under the proposal IDs 1288 and 6603, respectively. The

ALMA raw data are available on the ALMA archive at https://
almascience.eso.org/aq under project code 2017.1.01478.S. Our
reduced ALMA data cubes and maps are archived at Zenodo (48),
as is our reduced NIRSPec spectrum of d203-506 (49). The
modified Meudon PDR code used in this paper is available at
Zenodo (50). License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/sciencelicenses-journal-article-reuse

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adh2861
PDRs4All Team Authors and Affiliations
Materials and Methods
Figs. S1 to S7
Tables S1 and S2
References (51–78)
Submitted 24 February 2023; accepted 12 January 2024
10.1126/science.adh2861

p
g
y
y g
y
,

Berné et al., Science 383, 988–992 (2024)

1 March 2024

5 of 5

RES EARCH

NEUROSCIENCE

Touch sensation requires the mechanically gated ion
channel ELKIN1
Sampurna Chakrabarti1, Jasmin D. Klich1, Mohammed A. Khallaf1,2, Amy J. Hulme3,
Oscar Sánchez-Carranza1, Zuzanna M. Baran1,4, Alice Rossi1, Angela Tzu-Lun Huang1,
Tobias Pohl4, Raluca Fleischer1, Carina Fürst1,5, Annette Hammes5, Valérie Bégay1,
Hanna Hörnberg4,6, Rocio K. Finol-Urdaneta3, Kate Poole7, Mirella Dottori3, Gary R. Lewin1,8,9*
Touch perception is enabled by mechanically activated ion channels, the opening of which excites
cutaneous sensory endings to initiate sensation. In this study, we identify ELKIN1 as an ion channel likely
gated by mechanical force, necessary for normal touch sensitivity in mice. Touch insensitivity in Elkin1−/−
mice was caused by a loss of mechanically activated currents (MA currents) in around half of all sensory
neurons activated by light touch (low-threshold mechanoreceptors). Reintroduction of Elkin1 into
sensory neurons from Elkin1−/− mice restored MA currents. Additionally, small interfering RNA–mediated
knockdown of ELKIN1 from induced human sensory neurons substantially reduced indentation-induced
MA currents, supporting a conserved role for ELKIN1 in human touch. Our data identify ELKIN1 as a core
component of touch transduction in mice and potentially in humans.

The absence of ELKIN1 did not alter the
neurochemical makeup of the sensory ganglia
1 of 7

,

1 March 2024

Sensory deficits in Elkin1−/− mice

y

Chakrabarti et al., Science 383, 992–998 (2024)

We hypothesized that ELKIN1 could also be
involved in mammalian touch sensation. We
generated a CRISPR-Cas9–mediated genomic
deletion of the mouse Tmem87a/Elkin1 gene
locus spanning sequences coding for transmembrane domains 2 to 6, which includes the
proposed ion-conduction pathway (fig. S3A)
(11). Mice homozygous for the genomic deletion (Elkin1−/− mice) were viable and born at
the expected Mendelian ratios [wild type (WT),
25.7%; Elkin1+/−, 46.2%; Elkin1−/−, 28%; n =
132] (table S2). Single-molecule fluorescent in
situ hybridization (smFISH) and immunohistochemistry with an antibody against ELKIN1
showed that Elkin1−/− mice were complete
null mutants (Fig. 2A). Our validated ELKIN1
antibody revealed that ELKIN1 protein levels
appeared to be especially high in sensory neurons of the DRGs. ELKIN1 protein was robustly
detected in all subsets of DRG neurons, which is
consistent with single-cell expression data from
mice, macaques, and humans (17–21); around
60% of neurons showed high ELKIN1 expression (fig. S3, B and C). Sensory neurons expressing high amounts of ELKIN1 made up 34% of
neurofilament heavy chain positive (NF200+)
large neurons with myelinated axons. High
ELKIN1 expression was also found in 75% of
isolectin-B4 positive (IB4+) nonpeptidergic small
neurons (22) and 45% of small neurons positive
for the capsaicin-gated transient receptor potential channel, TRPV1 (fig. S3, B and C). Both
of these two neurochemically distinct nociceptor types are reported to be responsive
to mechanical force (23).

y g

*Corresponding author. Email: glewin@mdc-berlin.de

Mouse sensory neurons express ELKIN1

y

Molecular Physiology of Somatic Sensation Laboratory,
Max Delbrück Center for Molecular Medicine in the
Helmholtz Association (MDC), 13125 Berlin-Buch, Germany.
2
Department of Zoology and Entomology, Faculty of Science,
Assiut University, Assiut 71516, Egypt. 3School of Medical,
Indigenous and Health Sciences, Molecular Horizons,
University of Wollongong, Wollongong, NSW 2522, Australia.
4
Molecular and Cellular Basis of Behavior, Max Delbrück
Center for Molecular Medicine in the Helmholtz Association
(MDC), 13125 Berlin-Buch, Germany. 5Molecular Pathways in
Cortical Development, Max Delbrück Center for Molecular
Medicine in the Helmholtz Association (MDC), 13125
Berlin-Buch, Germany. 6NeuroCure Cluster of Excellence,
Humboldt-Universität zu Berlin, 10117 Berlin, Germany.
7
School of Biomedical Sciences, Faculty of Medicine &
Health, University of New South Wales, Sydney, NSW 2052,
Australia. 8Charité-Universitätsmedizin Berlin, 10117 Berlin,
Germany. 9German Center for Mental Health (DZPG), partner
site Berlin, 10117 Berlin, Germany.

We previously identified ELKIN1 (TMEM87A)
as a protein that is both necessary and sufficient to confer mechanosensitivity to highly
metastatic human melanoma cells (9). Cryo–
electron microscopy (cryo-EM) structures of
human ELKIN1 recently revealed a monomeric seven-transmembrane protein with an
N-terminal extracellular Golgi-dynamics domain fold (GOLD domain) (10). A second,
higher-resolution structure recently identified a cation-conduction pathway through the
protein (11). We overexpressed Elkin1 in human embryonic kidney (HEK) 293T cells lacking PIEZO1 channels (HEK293TPiezo1−/− cells)
(12) and found large indentation-induced mechanically activated currents (MA currents)
in a majority of transfected cells (Fig. 1, A
and B). Cells were plated on laminin 511 and
poly-L-lysine (PLL), a substrate that supports
increased mechanosensitivity (13); untransfected cells showed no indentation-induced
currents. ELKIN1-dependent currents were
rapidly-adapting (RA) with fast inactivation
time constants (<10 ms), similar to those of
PIEZO2 ion channels (5, 14) (Fig. 1, A to C). Using
substrate deflection by means of pillar arrays
(9, 14), we also found robust, mechanically activated currents in all HEK293TPiezo1−/− cells
transfected with Elkin1-expression constructs,
but also in cells transfected with Elkin1 lacking the N-terminal GOLD-domain (Elk1D170)
(Fig. 1, C and D, and fig. S1A) (9). Most of the
pillar-evoked currents were RA (inactivation
<10 ms) or intermediately adapting [(IA), inactivation between 10 and 50 ms]. Measurements
of pillar-gated currents at different holding potentials revealed a linear current-voltage relationship with a reversal potential of 0 mV for
both Elkin1- and Elk1D170-transfected cells
(Fig. 1E). Therefore, our results suggest that
the GOLD domain is not necessary for me-

g

1

ELKIN1 can detect mechanical force

p

T

ouch sensation is fundamental to our
sense of self, our social interactions, and
our exploration of the tactile world (1, 2).
Sensation is initiated at specialized end
organs in the skin, innervated by lowthreshold mechanoreceptors (LTMRs) with
their cell bodies in the dorsal root ganglia
(DRGs). The peripheral endings of LTMRs are
equipped with mechanically gated ion channels that can be opened by very small forces to
initiate and enable touch perception (3, 4).
The mechanically gated ion channel PIEZO2
is expressed by most sensory neurons in the
DRGs (5), and in the absence of PIEZO2, around
half of LTMRs no longer respond to mechanical
stimuli (6–8). The DRGs also contain so-called
nociceptors, sensory neurons specialized to detect potentially damaging and painful stimuli,
including intense mechanical force (3). Many
nociceptors express PIEZO2 channels but remain
mechanosensitive in its absence. The preservation of mechanosensitivity in many LTMRs
in the absence of PIEZO2 channels (6–8) led us
to search for other mechanically gated ion channels that could account for PIEZO2-independent
sensory mechanotransduction.

chanical activation of ELKIN1. ELKIN1 currents showed a distinctive pharmacological
profile, being sensitive to the nonspecific pore
blocker Gd3+(30 mM) but barely affected by
ruthenium red (30 mM), a compound that completely blocks other mechanosensitive channels,
such as PIEZO1 and PIEZO2 (5, 15) (Fig. 1F and
fig. S1B). Additionally, in agreement with recent
reports (11, 15), we found that cells expressing
mouse Elkin1 display prominent leak currents
at very positive (>+60 mV) and very negative
potentials (<−100 mV) (Fig. 1G and fig. S1C).
ELKIN1 reconstituted into proteoliposomes reportedly show single-channel activity at very
positive potentials (16). We also found that
Elkin1-transfected HEK293TPiezo1−/− cells showed
currents, which were substantially potentiated
by application of mild positive-pressure pulses
(20 mm of Hg) applied via the cell-attached
pipette (fig. S2, A to C). Therefore, we provide
multiple lines of evidence that ELKIN1 is likely an ion channel that can detect mechanical
force.

RES EARCH | R E S E A R C H A R T I C L E

B

C

D

800

Elkin1

500

12
12
7

50

-400

El

1
kin

l3
R
dC
+R µM +G µM
30
30

0

8

-200

100

-100

100
Voltage (mV)
-100

-50

-100

-150

-200
n1
ki
El

,

2 of 7

y

Large sensory neurons of the DRGs are predominantly mechanoreceptors required for
touch (4, 24). We therefore recorded MA currents from isolated sensory neurons evoked
by mechanical indentation and substrate deflection (Fig. 3, A and D) (24, 25). Normally,
almost all large neurons exhibit robust, predominantly RA MA currents to both cell indentation and substrate deflection (14, 25), which
we confirmed in WT animals (Fig. 3, A to F).
However, only half of the large neurons (diameter >30 mm, fig. S4A) from Elkin1−/− mice
displayed any MA current (Fig. 3, B and E).
The insensitivity to mechanical stimuli was
therefore concomitant to a loss of RA MA
currents in Elkin1−/− mice (Fig. 3, C and F).
The current amplitude of MA currents in the
remaining mechanosensitive neurons was sim-

y g

Sensory neuron mechanically activated
currents are lost in Elkin1−/− mice

ilar in WT and Elkin1−/− mice; however, there
was a small but significant (P = 0.01, unpaired
Student’s t test) increase in the deflection
threshold in neurons from Elkin1−/− mice as
assessed through the pillar assay (fig. S4, B
and C). Large sensory neurons recorded from
Elkin1−/− mice also showed a slightly depolarized resting membrane potential (P = 0.001,
unpaired Student’s t test), which is consistent
with the idea that this channel may contribute to membrane leak (fig. S4D). A significant
loss of MA currents was even detectable after
loss of just one Elkin1 allele (P = 0.005, unpaired Student’s t test) (fig. S4E). To show
that the loss of MA currents was not due to
indirect effects of Elkin1 gene inactivation,
we conducted an acute rescue experiment.
Using an adeno-associated virus neurotropic
for sensory neurons (AAV-PHP.S-hSyn-dtommElkin1), we reintroduced Elkin1 back into
acutely isolated sensory neurons from Elkin1−/−
mice. ELKIN1 protein was detected in infected
sensory neurons from Elkin1−/− mice, and there
was a significant (P = 0.01, c2 test) rescue of
MA currents: Only 40% of neurons had MA

y

cells. (F) Quantification of pillar-deflection threshold and properties of ELKIN1dependent currents in the presence of pore blockers ruthenium red (RR)
and GdCl3. Each dot represents a cell, and numbers in bars are number of
MA-pillar stimuli. (G) Current-voltage relationship of Elkin1-transfected cells
(each dot is a mean of n = 14 mock and n = 15 Elkin1 cells) as assessed through a
series of voltage steps from −150 to 90 mV. Three group comparisons were
made with one-way ANOVA followed by multiple comparison test, and two
group comparisons were made with Student’s t test. Proportions were compared
using c2 test. *P < 0.05, **P < 0.01. Error bars indicate SEM.

as heat withdrawal thresholds, were unaltered
in Elkin1−/− mice (fig. S3D). Elkin1−/− mice also
showed no deficits in open-field locomotion
(fig. S3E).

1 March 2024

200

g

Chakrabarti et al., Science 383, 992–998 (2024)

17

1
Cl 3
kin +RR µM +Gd M
µ
El
30
30
I
A
RA
SA

Fig. 1. ELKIN1 forms a mechanically gated channel. (A) Pie charts show total
number of mechanically responsive transfected cells. Shown in (A) (top) and
(B) are indentation-induced currents in HEK293TPiezo1−/− cells upon transfection
with mouse Elkin1 cDNA. Dots represent individual cells. (C) Representative
MA currents evoked by pillar deflection at −60 mV along with (D) quantification
of MA-current amplitude and properties at 100- to 250-nm force bin (each dot
represents a cell, and numbers in bars are number of MA-pillar stimuli). (E) Currentvoltage relationship of MA currents evoked in cells transfected with Elkin1 or
Elk1D170. Dots are mean of n = 10 in Elkin1 and n = 15 in Elk1D170-transfected

because the percentage of DRG neurons positive for markers such as NF200, IB4, TRPV1,
and tyrosine hydroxylase (TH) was unchanged
in the Elkin1−/− mice as compared with WT (Fig.
2B). Ultrastructural analysis of the saphenous
nerve revealed no pathology or loss of myelinated or unmyelinated axons in Elkin1−/− mice
(Fig. 2C and table S1). However, behavioral
indicators of touch sensitivity, such as percentage of responses to a cotton swab, were
profoundly reduced in Elkin1−/− mice as compared with WT animals (WT, 90% paw withdrawal versus Elkin1−/−, 47.5% paw withdrawal;
P < 0.0001, unpaired Student’s t test) (Fig.
2D). Paw-withdrawal thresholds to von Frey
filaments were also significantly elevated [P <
0.0001, two-way analysis of variance (ANOVA)
with Sidak post hoc test], with substantial
deficits observed across a range of von Frey
force filaments in Elkin1−/− mice (Fig. 2E). However, responses to brush stimuli in Elkin1−/−
were similar to those for WT mice (fig. S3E).
These results confirm reduced sensitivity to
mechanical forces in Elkin1−/− mice. However, nonmechanosensory modalities, such

14
6

29

0

Current (pA/pF)

1000

Elkin1

k

-200

50
Voltage (mV)

300

0

m

-50

mock
8

% MA current

200

1500

Threshold (nm)

400

*

100

Elkin1
Elk1 Δ170

1

kin

El

37

Elkin1 Elk1Δ170
IA
RA
SA

p

MA current (pA)

600

G

**

*

22

0
m

F

E

19
14

50

oc

Elk1 Δ170

1
0

Elkin1

35 nm

% of MA currents

2

300
250
200
150
100
50
0

Current (pA/pF) @ -150 mV

0

10 ms

70

200

3

Δ1

400

Elkin1

50 pA

600

El
k1

Elkin1
Untransfected
Mechanically responsive

11
11

7

oc
k

8/15

100

*

600

40 nm
Inactivation Tau (ms)

0/12

2s

MA current (pA)

200 pA

Untransfected
Elkin1

**

-60 mV

5 μm

4 μm

MA current (pA)

A

RES EARCH | R E S E A R C H A R T I C L E

p
g
y

Fig. 2. Elkin1−/− mice are touch insensitive. (A) Representative images of
ELKIN1 expression pattern, obtained using a smFISH probe against Elkin1
(top panel; scale bar, 20 mm) and antibody staining against ELKIN1 (bottom
panel; scale bar, 50 mm) from WT and Elkin1−/− DRGs. (B) (Top panels)
Representative images of NF200 (red), IB4 (blue), TRPV1 (magenta), and TH
(yellow) staining in WT and Elkin1−/− DRGs and (bottom) quantification of percent
of positive neurons in each group from three male mice. More than 500 neurons

were counted in each category. (C) Ultrastructural analysis of saphenous nerve.
Scale bar, 1 mm. (D) Percent response of WT and Elkin1−/− mice (n = 16) to brushing
of a cotton swab. (E) Paw withdrawal threshold (left, n = 16) and ascending-dose
response (right, n = 10) of WT and Elkin1−/− mice to von Frey filaments. Two group
comparisons were made with Student’s t test and two-way ANOVA with Sidak
post hoc test (for von Frey ascending-dose response). **P < 0.01, ***P < 0.001.
Error bars indicate SEM. Data from both male and female mice.

y g

currents in mock transfected cells as compared
with 75% of cells 48 hours after infection with
AAV-PHP.S-hSyn-dtom-mElkin1 (Fig. 3, G and
H, and fig. S5A).

Chakrabarti et al., Science 383, 992–998 (2024)

ELKIN1 and PIEZO2 share roles in
sensory mechanotransduction

The phenotype that we observed in Elkin1−/−
sensory neurons was similar to the knockdown or genetic ablation of the PIEZO2 mechanosensitive ion channel (5–7). Using smFISH,
we detected colocalization of Elkin1 and Piezo2
mRNA in WT DRG neurons, but no change
in Piezo2 mRNA expression was observed

1 March 2024

3 of 7

,

Human stem cells can be differentiated into sensory neuron–like cells that have characteristic
electrophysiological properties of DRG neurons, including MA currents (26, 27). We could
detect ELKIN1 protein in NEUROGENIN2induced human sensory neurons (26), the staining for which was abolished by small interfering
RNA (siRNA)–mediated mediated knockdown
of ELKIN1 (Fig. 3I and fig. S6A). Our induced
human sensory neurons also had robust MA
currents, which increased in size with increasing indentation amplitudes. MA currents required higher indentation amplitudes and
attained much smaller peak amplitudes in
induced sensory neurons transfected with

in Elkin1−/− sensory neurons (fig. S5B). Thus,
Elkin1 ablation does not affect Piezo2 expression. As shown previously (5), we found
that in WT DRG neurons, siRNA-mediated
knockdown of Piezo2 approximately halved
the number of neurons with MA currents
(fig. S5C). If Elkin1 exerts its function through
Piezo2, knockdown of Piezo2 in Elkin1−/− neurons should not cause a further decrease in
MA currents. By contrast, our results show
that MA currents in Elkin1−/− neurons could
be reduced further after Piezo2 knockdown
(fig. S5C). Thus, neurons retaining MA currents in Elkin1−/− mice appear to have predominantly PIEZO2-dependent MA currents. We
next asked whether there is a functional interaction between these two proteins by expressing Piezo2 or Piezo2 and Elkin1 in
N2aPiezo1−/− cells. We found no differences in the
amplitude or kinetics of MA currents found
in single- and double-transfected cells, indicating no substantial functional interaction

y

Mechanically activated currents in human
sensory neurons depend on ELKIN1

ELKIN1 siRNA, as compared with control
scrambled siRNA (Fig. 3I and fig. S6, B and
C). Thus, ELKIN1 is required for normal MAcurrent expression in both mouse and human
sensory neurons. Additionally, in these induced human sensory neurons, knockdown
of PIEZO2 with siRNA also decreased MA
currents, and very few MA currents remained
after the knockdown of both ELKIN1 and PIEZO2
(fig. S6C). We therefore postulated that there
may be some functional interaction between
PIEZO2 and ELKIN1.

RES EARCH | R E S E A R C H A R T I C L E

SA

4

80
60

12

40
20
19

20 pA

3

80

*

RA

60
5

40

4

WT
SA
RA

H

IA

20
0

WT
Elkin1-/Non-responders
Responders

200 ms

1

15

10

E

Elkin1-/IA

SA

1

80

4

9

60
40
20
17

0

100
80
60

32

0

6

14

+mock +AAV-Elkin1
Non-responders
Responders

siCTRL

4
2
0
+mock +AAV-Elkin1

***
80
MA current (pA/pF)

20

Inactivation Tau (ms)

% responsive large neurons

10

siELKIN1

2s

9 μm

40

90

siCTRL
siELKIN1

70
60
50
40
30
20

g

50 pA

8 μm

60

DAPI
ELKIN1

6

Elkin1-/IA

p

7 μm

4

10

WT
SA
RA

Induced human sensory neurons

8

80

6

0

*
100

13

16

20

11

WT
Non-responders
Responders

2s

15

40

Elkin1-/-

I

Elkin1-/-

F

**
100

% MA large neurons

2

0

G

1

% responsive large neurons

IA

D
100

% MA large neurons

RA

C

**
100

200 pA

B
% responsive large neurons

A

10
0
0

1 2 3 4
Indentation (µm)

5

y

Severe mechanoreceptor deficits
in Elkin1−/− mice

We next investigated whether ELKIN1 is required for the mechanosensory function of
identified mouse mechanoreceptors. First, we
used an ex vivo saphenous skin-nerve preparation to trace the trajectory of single units by
means of an electrical stimulus until the point
of exit from the nerve branch. Using a mechanical stimulus, we then searched for the
mechanosensitive receptive field of the same
unit, which was usually located close to the
exit point from the nerve branch (6, 7, 24, 31).
Single identified Ab fibers with the fastest
conduction velocities (>10 m/s) usually al-

1 March 2024

ways have a mechanosensitive receptive field,
as confirmed for WT mice (Fig. 4A). However,
blinded recordings made from Elkin1−/− mice
revealed that 40% of the Ab fibers had no
detectable mechanosensitive receptive field
(9/26 Ab fibers) (Fig. 4A). Next, we examined
the stimulus-response properties of the remaining identified mechanoreceptors in the hairy
skin. Ab-fiber LTMRs innervating Merkel cells
are classified as slowly adapting mechanoreceptors (SAMs) with a dynamic and static
response to ramp-and-hold force stimuli (Fig.
4B). Around 50% of Ab fibers are typically
classified as SAMs in the hairy skin (24, 32),
and this was the case in both WT and Elkin1−/−
mice (fig. S8A). However, the firing rates of
SAMs to the static constant force phase of the
stimulus were strongly reduced at all stimulus strengths in Elkin1−/− mice as compared
4 of 7

,

Chakrabarti et al., Science 383, 992–998 (2024)

creased mechanical thresholds and increased
current amplitude in the presence of STOML3
(fig. S7B).

y

between these proteins in a heterologous
expression system (fig. S5D).
A known PIEZO2 modulator is the MEC-2–
related mechanotransduction protein STOML3,
which was shown to sensitize PIEZO2 channels to substrate deflection (14, 24, 28–30). We
next asked whether there is also a molecular
interaction between STOML3 and ELKIN1.
Using a tripartite green fluorescent protein
(GFP)–based protein complementation assay, we observed a robust green fluorescent
signal, indicating close association between
STOML3 and ELKIN1 (fig. S7A). The protein
complementation signal for a STOML3/ELKIN1
interaction was also blocked in the presence
of the STOML3 oligomerization blocker OB1
(fig. S7A) (30). Coexpression of Stoml3 with
Elkin1 in HEK293TPiezo1−/− cells revealed that
ELKIN1-dependent MA currents displayed de-

of Elkin1−/− DRG neurons stained by anti-RFP after being transduced by
AAV-PHP.S-hSyn-dtom-mElkin1. Scale bar, 20 mm. (Bottom) Representative
traces of indentation-induced currents from a transduced neuron. (H) (Left)
Percent of mechanically active large-diameter neurons upon transduction with
AAV-PHP.S-hSyn-eGFP (mock) or AAV-PHP.S-hSyn-dtom-mElkin1. (Right) Quantification
of inactivation time constants of the measured currents. Each data point
represents a single cell measured. (I) Representative images of NEUROGENIN2induced human sensory neurons before (top) and after (bottom) transfection with
ELKIN1 siRNA (left, scale bars, 50 mm; right, two-way ANOVA). Proportions were
compared using c2 test; four group comparisons were made using ANOVA followed
by multiple comparison test. *P < 0.05, ** P < 0.01, ***P < 0.001. Error bars indicate
SEM. Data from both male and female mice. DAPI, 4′,6-diamidino-2-phenylindole.

y g

Fig. 3. ELKIN1 is necessary and sufficient for mechanically gated currents
in mouse and human sensory neurons. (A) Representative traces of currents
generated by large-diameter neurons from Elkin1−/− mice. Scale bar, 20 mm.
(B) Percent of large-diameter neurons with an MA current in WT and Elkin1−/−
mice. (C) Percent of rapidly adapting (RA), intermediately adapting (IA), and
slowly adapting (SA) MA currents in large-diameter neurons. Number of cells is
denoted in the bars. (D) Representative traces of currents generated by pillar
deflection of a large-diameter neuron from an Elkin1−/− mouse. Scale bar,
20 mm. (E) Percent of mechanically sensitive large-diameter neurons in WT
and Elkin1−/− mice in pillar assay. Number of cells is denoted in the bars.
(F) Percent of RA, IA, and SA MA currents in large-diameter neurons. Number
of MA-pillar stimulations is denoted in the bars. (G) (Top) Representative image

RES EARCH | R E S E A R C H A R T I C L E

Discussion

,

5 of 7

y

Here we show that ELKIN1 is necessary for
the mechanosensory function of most LTMRs.
Elkin1−/− mice have electrically excitable sensory axons in the skin that were completely
unable to respond to mechanical stimuli. The
sustained firing of SAMs to constant force partly
depends on PIEZO2 expressed in mechanosensory Merkel cells (39). Our data suggest that
ELKIN1 is required for the PIEZO2 independent
transduction in SAMs because sustained responses were severely reduced in Elkin1−/−
mice. The ability of induced human sensory
neurons to transduce mechanical forces was severely diminished after knockdown of ELKIN1.
Thus, ELKIN1 is an ion channel gated by mechanical force that likely has a conserved role
in the transduction of light touch in mice and
humans.
Consistent with the expression pattern of
ELKIN1, maintained firing of C-fiber nociceptors to constant force was also impaired in
the absence of the ELKIN1 protein. The loss of
mechanically gated currents, impaired touchdriven behavior, and deficits in LTMR function
are reminiscent of mice lacking Stoml3 and of
conditional Piezo2 mutants (6, 7, 24, 28). Our
data support a model in which ELKIN1 and
PIEZO2 channels share roles in sensory mechanotransduction in LTMRs and in which both
channels can be modulated by STOML3. There
is evidence that STOML3 can also modulate
MA currents in nociceptors, which is consistent

y g

1 March 2024

mice (Fig. 5, F and G, and fig. S10). Reduced
firing rates toward the end of the stimulus
were observed for all intensities of stimulation
(fig. S10). Thus, unlike in mechanoreceptors,
ELKIN1 has a limited role in nociceptors but
may be necessary for maintaining sensitivity
to constant forces in these neurons.

y

Chakrabarti et al., Science 383, 992–998 (2024)

of MA currents in small- and medium-diameter
neurons that displayed broad-humped action potentials characteristic of nociceptors (22, 25, 37)
(Fig. 5A and fig. S9). Many nociceptors in WT
mice lack MA currents to cell indentation
(~40%) (23, 25), but this was not different in
neurons recorded from Elkin1−/− animals (Fig.
5, A and B). However, when the MA currents
were classified as RA (inactivation time <10 ms),
IA (inactivation time constant 10 to 50 ms),
and SA (slowly adapting, inactivation time
constant >50 ms), we identified a significant
(P = 0.03, c2 test) reduction in the proportion of RA MA currents in Elkin1−/− mice as
compared with those in WT animals (Fig.
5C). In addition, we found a small but significant (P = 0.01, unpaired Student’s t test)
elevation in the amplitude of indentation
needed to evoke the first MA current in nociceptors from Elkin1−/− mice (Fig. 5D). MA
currents evoked through substrate deflection
showed no change between WT and Elkin1−/−
mice (fig. S9B). We next focused our analysis on nonpolymodal C fibers that respond
exclusively to mechanical stimuli and not to
thermal stimuli because this population shows
robust firing to mechanical force (38). In
the ramp-and-hold force protocol, the firing
of mechanosensitive C fibers from Elkin1−/−
mice was no different from that in WT animals (Fig. 5, E and F). We next analyzed the
time course of C-fiber activation during a
10 s–long constant-force stimulus. C fibers
generally show a moderate degree of adaptation during constant-force stimuli (7). However, we found that even though initial firing
rates were similar between genotypes at a
stimulus strength of 100 mN, firing rates
dropped significantly more (P = 0.004, twoway ANOVA with multiple comparison test)
during the stimulus in C fibers from Elkin1−/−

g

with those in controls (Fig. 4B). A plot of the
peristimulus time histogram for SAMs stimulated with 150 mN of force reveals that firing
rates decrease to almost zero just 3 s into a
10-s stimulus (Fig. 4C). However, the same SAMs
from WT and Elkin1−/− mice showed similar
dynamic phase responses (fig S8C). The remaining Ab mechanoreceptors were classified as
rapidly adapting mechanoreceptors (RAMs),
which only respond to skin movement and
code the velocity of skin movement (3, 4, 33).
As a population, RAMs still coded the velocity
of ramp stimuli, but the overall firing rates
were significantly lower in Elkin1−/− mice as
compared with those in controls (P = 0.03, twoway ANOVA with multiple comparison test)
(Fig. 4D). These results could easily reflect loss
of MA currents in mechanoreceptors but could
also be due to morphological disruption of sensory endings. However, an analysis of mechanoreceptor endings in the skin of Elkin1−/− mice did
not reveal any obvious deficits (fig. S8C). These
results show that around half of the LTMRs are
insensitive to mechanical forces in Elkin1−/−
mice, but in addition, the remaining LTMRs
showed profound functional deficits in their
ability to detect mechanical force.
We also made recordings from single mechanosensitive nociceptors in the saphenous
nerve. Sensory neurons with thinly myelinated
Ad axons can be classified as A-fiber mechanonociceptors (AMs), which signal fast pain (34),
or as down hair (D-hair) receptors, which are
specialized LTMRs with directional sensitivity
(35, 36). We found no change in the stimulus
properties of D-hair or AM afferents in Elkin1−/−
mice as compared with those of controls (fig.
S8D). Many DRGs with high amounts of ELKIN1
also appear to be nociceptors, an assertion based
on the presence of markers such as IB4 and
TRPV1 (22). We thus made a focused analysis

fibers). (D) Representative spikes evoked from rapidly adapting mechanoreceptors
to a moving ramp stimulus and quantification of the mean firing rates to
ramps of increasing speed. Proportions were compared using c2 test with results
from (A). All other group comparisons were made using two-way ANOVA followed
by multiple comparison test. *P < 0.05, **P < 0.01, ***P < 0.001. Error bars
indicate SEM. Data from both male and female mice.

p

Fig. 4. ELKIN1 is required for LTMR function. (A) Percent of mechanosensitive
fast-conducting Ab fibers in the saphenous nerve assessed with an electrical
search protocol. (B) (Top) Representative spikes evoked from SA mechanosensitive Ab fibers in WT and Elkin1−/− mice and (bottom) quantification of mean
spike rates with increasing force. (C) Absolute number of spikes over a 10-s time
period in the 150-mN force bin (each dot represents average response from

RES EARCH | R E S E A R C H A R T I C L E

p
g
y

1. A. R. Sobinov, S. J. Bensmaia, Nat. Rev. Neurosci. 22, 741–757
(2021).
2. M. Mikkelsen, E. L. Wodka, S. H. Mostofsky, N. A. J. Puts, Dev.
Cogn. Neurosci. 29, 140–150 (2018).
3. G. R. Lewin, R. Moshourab, J. Neurobiol. 61, 30–44 (2004).
4. S. G. Lechner, G. R. Lewin, Physiology 28, 142–150 (2013).
5. B. Coste et al., Science 330, 55–60 (2010).
6. S. S. Ranade et al., Nature 516, 121–125 (2014).
7. S. E. Murthy et al., Sci. Transl. Med. 10, eaat9897 (2018).
8. B. U. Hoffman et al., Proc. Natl. Acad. Sci. U.S.A. 119,
e2115821119 (2022).
9. A. Patkunarajah et al., eLife 9, e53308 (2020).

Chakrabarti et al., Science 383, 992–998 (2024)

1 March 2024

25. J. Hu, G. R. Lewin, J. Physiol. 577, 815–828 (2006).
26. A. J. Hulme et al., Front. Cell. Neurosci. 14, 600895
(2020).
27. K. Schrenk-Siemens et al., Nat. Neurosci. 18, 10–16
(2015).
28. M. Huang, G. Gu, E. L. Ferguson, M. Chalfie, Nature 378,
292–295 (1995).
29. R. O’Hagan, M. Chalfie, M. B. Goodman, Nat. Neurosci. 8,
43–50 (2005).
30. C. Wetzel et al., Nat. Neurosci. 20, 209–218 (2017).
31. M. Kress, M. Koltzenburg, P. W. Reeh, H. O. Handwerker,
J. Neurophysiol. 68, 581–595 (1992).
32. M. Koltzenburg, C. L. Stucky, G. R. Lewin, J. Neurophysiol. 78,
1841–1850 (1997).
33. F. Schwaller et al., Nat. Neurosci. 24, 74–81 (2021).
34. A. Arcourt et al., Neuron 93, 179–193 (2017).
35. J. Walcher et al., J. Physiol. 596, 4995–5016 (2018).
36. M. Rutlin et al., Cell 159, 1640–1651 (2014).
37. H. R. Koerber, R. E. Druzinsky, L. M. Mendell, J. Neurophysiol.
60, 1584–1596 (1988).
38. N. Milenkovic, C. Wetzel, R. Moshourab, G. R. Lewin,
J. Neurophysiol. 100, 2771–2783 (2008).
39. S.-H. Woo et al., Nature 509, 622–626 (2014).

6 of 7

,

RE FE RENCES AND N OT ES

10. C. M. Hoel, L. Zhang, S. G. Brohawn, eLife 11, e81704
(2022).
11. A. Han et al., bioRxiv 2023.01.03.522544 [Preprint] (2023).
https://doi.org/10.1101/2023.01.03.522544.
12. A. E. Dubin et al., Neuron 94, 266–270.e3 (2017).
13. L.-Y. Chiang et al., Nat. Neurosci. 14, 993–1000 (2011).
14. K. Poole, R. Herget, L. Lapatsina, H.-D. Ngo, G. R. Lewin,
Nat. Commun. 5, 3520 (2014).
15. L. Pope, M. Lolicato, D. L. Minor Jr., Cell Chem. Biol. 27,
511–524.e4 (2020).
16. H. Kang et al., bioRxiv 2023.01.03.522543 [Preprint] (2023).
https://doi.org/10.1101/2023.01.03.522543.
17. D. Usoskin et al., Nat. Neurosci. 18, 145–153 (2015).
18. N. Sharma et al., Nature 577, 392–398 (2020).
19. J. Kupari et al., Nat. Commun. 12, 1510 (2021).
20. H. Yu et al., bioRxiv 2023.03.17.533207 [Preprint] (2023).
https://doi.org/10.1101/2023.03.17.533207.
21. P. Ray et al., Pain 159, 1325–1345 (2018).
22. C. L. Stucky, G. R. Lewin, J. Neurosci. 19, 6497–6505
(1999).
23. L. J. Drew, J. N. Wood, P. Cesare, J. Neurosci. 22, RC228
(2002).
24. C. Wetzel et al., Nature 445, 206–209 (2007).

y

with a role for ELKIN1 in conferring robustness
to the C-fiber responses to force (14). The identification of ELKIN1 as a mechanically gated
ion channel necessary for somatosensory function increases our understanding of the entirety of touch transduction.

mechanosensitive C fibers in WT and Elkin1−/− mice. (F) Quantification of the
firing rates to increasing forces. (G) Mean spiking rate over a 10-s time period to
a 100-mN force stimulus. Dots in (F) and (G) represent average of all fibers.
Proportion was compared using c2 test. Two group comparisons were made
using Student’s t test. All other group comparisons were made using two-way
ANOVA followed by multiple comparison test. *P < 0.05, **P < 0.01, ***P <
0.001. Error bars indicate SEM. Data from both male and female mice.

y g

Fig. 5. C mechanonociceptors show reduced firing to sustained mechanical
force in Elkin1−/− mice. (A) Representative indentation-induced current
evoked from WT and Elkin1−/− small DRG neurons. Scale bar, 20 mm. (B) Percent
of mechanically sensitive small- and medium-diameter neurons in WT and
Elkin1−/− mice. (C) Percent of RA, IA, and SA MA currents found in small- and
medium-diameter neurons and their threshold for activation. (D) The number of
cells is denoted as dots in the bar graph. (E) Representative spikes from

RES EARCH | R E S E A R C H A R T I C L E

40. S. Chakrabarti et al., Touch sensation requires the mechanically
gated ion channel ELKIN1, Dryad (2024); https://doi.org/10.5061/
dryad.0cfxpnw8s.
ACKN OW LEDG MEN TS

We thank F. Bartelt and L. Dalmasso for help with mouse
genotyping and B. Purfürst for electron microscopy. We thank
J. Poulet, S. Lechner, and members of the Lewin lab for constructive
comments on the manuscript. M.D. thanks D. J. Adams for
access to the patch-clamp rig. The H9 cell line (WA09) used in this
study is available under a material transfer agreement with WiCell.
Funding: This research was funded by ERC grant Sensational
Tethers 789128 and ERA-NET NEURON Sensory Disorders project
TRANSMECH to G.R.L., Deutsche Forschungsgemeinshaft grant
CRC958 to A.H. and G.R.L., and University of Wollongong,
Friedreich’s Ataxia Research Alliance (USA), and Friedreich Ataxia
Research Association (Australia) to M.D. S.C. and A.R. were

recipients of Alexander von Humboldt research fellowships. A.T.-L.H.
was a recipient of a Ministry of Science and Technology (Taiwan)
fellowship (111-2917-I-564-011). Author contributions:
Conceptualization: S.C. and G.R.L. Mouse model design and
validation: G.R.L. and V.B. Patch clamp physiology and anatomy
and cell biology: S.C. with help from O.S.-C., K.P., and Z.M.B.
Antibody validation: C.F. and A.H. Human induced sensory neurons
and analyses: A.J.H. with help from R.K.F.-U. and M.D. Tripartite-GFP
design and implementation: R.F., A.R., and A.T.-L.H. Molecular
biology: A.T.-L.H. Skin nerve electrophysiology: J.D.K. (nociceptors),
M.A.K. (mechanoreceptors, A-fiber mechanonociceptors, and D-hair
receptors), and G.R.L. (electrical search). Behavioral assessment:
S.C. with help from T.P. and H.H. Electron microscopy and
analysis: J.D.K. and S.C. Writing: S.C. and G.R.L., with input from
all authors. Supervision and funding: G.R.L., A.H., and M.D.
Competing interests: The authors declare that they have no
competing interests. Data and materials availability: All data are

available in the manuscript or the supplementary materials.
Prism files are available in Dryad (40). License information:
Copyright © 2024 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adl0495
Materials and Methods
Figs. S1 to S10
Tables S1 and S2
References (41–45)
MDAR Reproducibility Checklist
Submitted 27 September 2023; accepted 26 January 2024
10.1126/science.adl0495

p
g
y
y g
y
,

Chakrabarti et al., Science 383, 992–998 (2024)

1 March 2024

7 of 7

RES EARCH

CATALYSIS

Stable anchoring of single rhodium atoms by indium
in zeolite alkane dehydrogenation catalysts
Lei Zeng1†, Kang Cheng1†, Fanfei Sun2†, Qiyuan Fan1,3†, Laiyang Li1†, Qinghong Zhang1, Yao Wei2,
Wei Zhou1, Jincan Kang1, Qiuyue Zhang1, Mingshu Chen1, Qiunan Liu4, Liqiang Zhang4, Jianyu Huang4,
Jun Cheng1, Zheng Jiang2*, Gang Fu1*, Ye Wang1*
Maintaining the stability of single-atom catalysts in high-temperature reactions remains extremely
challenging because of the migration of metal atoms under these conditions. We present a strategy for
designing stable single-atom catalysts by harnessing a second metal to anchor the noble metal atom
inside zeolite channels. A single-atom rhodium-indium cluster catalyst is formed inside zeolite silicalite-1
through in situ migration of indium during alkane dehydrogenation. This catalyst demonstrates
exceptional stability against coke formation for 5500 hours in continuous pure propane dehydrogenation
with 99% propylene selectivity and propane conversions close to the thermodynamic equilibrium value
at 550°C. Our catalyst also operated stably at 600°C, offering propane conversions of >60% and
propylene selectivity of >95%.

1 March 2024

1 of 7

,

We chose Rh as the active metal for PDH because theoretical studies have predicted that

y

Zeng et al., Science 383, 998–1004 (2024)

Catalyst design and preparation

y g

*Corresponding author. Email: jiangzheng@ustc.edu.cn (Z.J.);
gfu@xmu.edu.cn (G.F.); wangye@xmu.edu.cn (Y.Wa.)
†These authors contributed equally to this work.

y

State Key Laboratory of Physical Chemistry of Solid
Surfaces, Innovation Laboratory for Sciences and
Technologies of Energy Materials of Fujian Province (IKKEM),
Collaborative Innovation Center of Chemistry for Energy
Materials, College of Chemistry and Chemical Engineering,
Xiamen University, Xiamen 361005, China. 2Shanghai
Synchrotron Radiation Facility, Shanghai Institute of Applied
Physics, Chinese Academy of Sciences, Shanghai 201210,
China. 3School of Chemistry and Chemical Engineering,
Shanxi University, Taiyuan 030006, China. 4Clean Nano
Energy Center, State Key Laboratory of Metastable
Materials Science and Technology, Yanshan University,
Qinhuangdao, China.

g

1

Despite these intensive research efforts, there
is no reported catalyst demonstrating stability
exceeding 500 hours in continuous PDH under industrially relevant conditions. Most of
the state-of-the-art catalysts show continuous
decreases in activity or selectivity even during
short-term PDH reactions (3). The synthesis of
confined bimetallic sites or alloys could extend
the lifetime to between 100 and 200 hours, but
their synthetic procedures are too intricate
to be suitable for large-scale applications, and
these catalysts also undergo deactivation during long-term operation (8, 12–14). The decline
in performance of the alloy-based catalysts can
be attributed to the separation of the bimetallic active phases fabricated by intricate procedures that lose their initial efficiency (1, 6, 15).
This deactivation results from the dynamic
nature of metals at high PDH temperatures.
Preventing the active single-atom catalysts,
even those with bimetallic sites or alloys, from
deterioration under harsh reaction conditions
remains a big challenge.
Here, we present a strategy to use the dynamic migration characteristic of indium for facile
construction of a single-atom rhodium-indium
alloy catalyst that is highly efficient and robust
for long-term alkane dehydrogenation without regeneration. By combining an In precursor
with rhodium-bearing silicalite-1 (S-1), a pure
siliceous zeolite with MFI structure, through
simple physical mixing or impregnation, we prepared a PDH catalyst that achieved a propylene
yield near the thermodynamic limit at 450° to
600°C and was stable for at least 5500 hours in
the conversion of pure propane at 550°C. This
catalyst could also work for the dehydrogenation of ethane and n-butane to ethylene and
butenes, respectively.

p

T

he escalating demand for propylene has
driven studies of catalytic nonoxidative
propane dehydrogenation (PDH) for onpurpose production of propylene with
polymer-quality purity. Two types of catalysts, chromium oxide–based and platinumbased catalysts (CrOx/Al2O3 and PtSn/Al2O3,
respectively), are currently used in the commercial processes, but both suffer from fast
deactivation caused by coke deposition or
catalyst-structure deterioration at the high
temperatures (>500°C) needed to achieve high
propane conversions (1–5) and thus require
frequent catalyst regeneration. Generally, Ptbased catalysts show better performance and
higher stability than do CrOx-based catalysts,
which typically deactivate within minutes.
Further improvements in Pt catalyst stability
and coking suppression have been achieved
through the use of metal additives (typically
Sn, Zn, Ga, Cu, or even a rare-earth metal)
to form nanoscale bimetallic sites or alloys
(6–8). The beneficial effect of atomic dispersion of noble metals on activity and selectivity was disclosed in PDH (9–11). A few
studies have reported that the stability could
be enhanced by confining the metal active
sites inside the pore channels of microporous
materials (8, 12–14).

it is the most active metal for C−H activation
(16, 17). However, the Rh-based catalysts reported to date have not shown better performances than the Pt-based catalysts and
underwent deactivation within tens of hours
even in the presence of bimetallic improvement (10, 18). Hannagan et al. (10) found an
enhancement in PDH, especially at low temperatures, by dispersing Rh onto Cu surfaces
and observed a stable formation of propylene at 350°C for 50 hours, but the propane
conversion is low under such conditions owing
to the thermodynamic limitation. We performed
CH4–D2 exchange reactions to probe the abilities of a series of S-1–confined noble metal
(denoted as NM@S-1, where NM = Rh, Pt, Pd,
Ir, or Ru) catalysts toward the C–H bond activation. The Rh@S-1 catalyst showed the
highest activity in the formation of hydrogen
deuteride (HD) and deuterated methane molecules (Fig. 1A) and thus was experimentally confirmed to be the most efficient in terms of the
C−H bond activation. However, only a very low
propane conversion was observed over Rh@S-1
even at 600°C, probably because of the quick
catalyst deactivation at the very initial stage. The
reactions with propane pulses to monitor the
conversion of propane at such a stage revealed
that Rh@S-1 mainly catalyzed the cracking of
propane to methane (breakage of all C–C bonds)
instead of the dehydrogenation to propylene,
and it also quickly deactivated after several
propane pulses (Fig. 1B).
We succeeded in constructing a second metalmodified Rh@S-1 catalyst for selective PDH by
harnessing the dynamic migration of a metal
under reaction conditions. Dynamic migration of elements such as Zn and In in a catalyst
under reaction conditions has been observed
in a few studies (19–22), but harnessing this
effect to construct an efficient heterogeneous
catalyst has received less attention (23, 24).
Zhao et al. (24) demonstrated that reductively
treating a mixture of ZnO- and SiO2-based
porous materials caused the migration of Zn0
with a low melting point onto the surface of
porous material to generate a supported single
ZnOx catalyst active for PDH. We note that the
thermal treatment of In2O3 particles in contact
with a zeolite at temperatures of 350° to 570°C
led to the migration of In species into its micropores (20, 25, 26). The solid-state ion exchange
was proposed for such a migration in an early
work (26). Our Fourier transform infrared
(FTIR) studies using pyridine as a probe molecule (27) revealed that the thermal treatment
(600°C) of a physical mixture of In2O3 and
S-1 powders (denoted bulk-In2O3+S-1) caused
a decrease in silanol groups in S-1, and such
a decrease became more pronounced with
hydrogen treatment (Fig. 1C and fig. S1).
However, this decrease was not observed after heat treatment of the mixtures of S-1 and
Ga2O3 or CuO (Fig. 1C and fig. S1), which were

RES EARCH | R E S E A R C H A R T I C L E

A

Rh

Pt

Pd

Ir

Ru

B

Blank

Rh

H2

Intensity (a.u.)

Intensity (a.u.)

HD
CH3D
CH2D2

CH4
C2H6
C3H6

CHD3
CD4
1 2 3

1 2 3

1 2 3

1 2 3

1 2 3

1

D

2

3

4

5

6

60

40

Conv. Rh@S-11
Select. Rh@S-11
Conv. bulk-In2O3 + Rh@S-1
Select. bulk-In2O3 + Rh@S-1

20

y

Conversion or selectivity (%)

80

g

Intensity (a.u.)

bulk-In2O3+S-1 (H2 600 oC)

1

p

S-1 (600 oC)
bulk-In2O3+S-1 (600 oC)

3

100

bulk-CuO+S-1 (600 oC)
bulk-Ga2O3+S-1 (600 oC)

Si-OH

2

Pulse reaction: C3H8 → C3H6 + H2 + [ CH4 + Cx ]

Pulse reaction: CH4 + D2 → CH4-iDi (1≤i≤4) + HD

C

C3H8

fully
converted

0
3800

3700

3600

3500

0

Wave number (cm−1)

12

16

quantity of silanol groups and to mitigate potential interference from water species.
(D) Catalytic performance for PDH over Rh@S-1 and bulk-In2O3+Rh@S-1. Reaction
conditions: C3H8/N2 = 1:3, WHSV = 10 hour–1, temperature (T) = 600°C. a.u.,
arbitrary units; Conv., conversion; Select., selectivity.

2 of 7

,

1 March 2024

pylene selectivity and the shortest induction period (<3 hours). We also synthesized RhM@S-1
catalysts (M = Zn, Ga, Sn, and Cu) with Rh and
the second metal both confined inside S-1. The
subsequent catalytic studies demonstrated the
distinctive role of In in enhancing the propane
conversion, propylene selectivity, and catalyst
stability (fig. S6). Unlike the RhIn@S-1 catalyst,
which displayed stable propane conversions
of >60% and propylene selectivity of >95%, the
other RhM@S-1 catalysts showed not only lower
propane conversions but also monotonic decreases in performances.
The RhIn/S-1 catalyst prepared by coimpregnation of Rh and In species onto S-1 exhibited much lower propane conversion and rapid
deactivation (Fig. 2A). Thus, the confinement
of Rh inside S-1 appeared to be necessary
for obtaining high PDH activity and stability,
whereas the location of In species in the fresh
catalyst determined the induction period. A

y

We studied the effect of other introduction
methods of In precursors on PDH performance.
A physically mixed catalyst of In/S-1, which
was preliminarily prepared by impregnation
and contained nanosized In2O3 (fig. S3), and
Rh@S-1 (denoted as In/S-1+Rh@S-1; see table
S1 for the loading amounts of Rh and In) was
also efficient for PDH to propylene, but a longer induction period of ~30 hours was required
(Fig. 2A). An In/Rh@S-1 catalyst, prepared by
simple impregnation and containing nanosized
In2O3 largely located on the outer surface of
Rh@S-1 (fig. S4), showed an induction period
of ~7 hours.
We next synthesized a RhIn@S-1 catalyst
with Rh and In species introduced simultaneously into S-1 during hydrothermal synthesis. The confinement of both species inside
S-1 was imaged with transmission electron microscopy (TEM) (fig. S5). This catalyst showed
the highest initial propane conversion and pro-

y g

Zeng et al., Science 383, 998–1004 (2024)

8

Time on stream (hours)

Fig. 1. Catalytic reactivity of Rh@S-1 and In-modified Rh@S-1. (A) CH4‒D2
exchange reactions over NM@S-1 catalysts. (B) Pulse reactions for propane
conversion over Rh@S-1 catalyst. (C) Pyridine-adsorbed FTIR spectra for physical
mixtures of metal oxides and S-1. Pyridine was used to monitor alterations in the

previously used for modifying Rh catalysts
(10, 18).
The change in IR spectra for the In2O3-based
system suggests that In species may migrate into
pure siliceous S-1 zeolite without ion-exchange
sites by occupying the silanol groups at high
temperatures and that the reductive atmosphere
expedites migration. Our catalytic PDH assessment showed improvement in performance
when In2O3 powders were in contact with
Rh@S-1 (Fig. 1D). Compared with the Rh@S-1,
bulk-CuO+Rh@S-1, and bulk-Ga2O3+Rh@S-1
catalysts, which had negligible activity in PDH
during the entire reaction period, the bulkIn2O3+Rh@S-1 catalyst (the physical mixture
of In2O3 and Rh@S-1) had an induction period (Fig. 1D and fig. S2). Both propane conversion and propylene selectivity increased
during the initial 6 hours and then remained
above 50 and 95%, respectively, over the bulkIn2O3+Rh@S-1 catalyst.

4

RES EARCH | R E S E A R C H A R T I C L E

C3H6 select.
80
C3H8 conv.
60

RhIn@S-1
In/Rh@S-1
In/S-1 + Rh@S-1
RhIn/S-1

40
20
0
0

10

20

30

C

100

1

C3H6 select.

80
60

C3H8 conv.

40

In/Rh@S-11
PtSnK/Al2O3

20

PtSnK@S-1

0
1000

1100

T = 550 oC, pure propane, WHSV = 8 h−1

40

Equilibrium yield of propylene
C3H8 conv.

0
2000

3000

4000

5000

10 13

a

16
17

1*

Time on stream (hours)
100

80

Select.

C=4 (mixturea)

Conv.

n-Co4

Co3

C=3

C=2

Co2

60

40

20

0
450

500

550

600

650

catalytic results are listed in table S2. (D) Stability evaluation of the
In/Rh@S-1 catalyst for the dehydrogenation of pure propane. Reaction
conditions: T = 550°C, pure C3H8, WHSV = 8 hour–1. (E) The dependence
of conversion and selectivity on temperature for the dehydrogenation of
ethane, propane, and n-butane over the In/Rh@S-1 catalyst. The dotted lines
represent the conversion levels of thermodynamic equilibrium. Reaction
conditions: WHSV = 10 hour−1, alkane/N2 = 1:3.

over some catalysts, they are transient and decrease rapidly because of the poor catalyst duration (Fig. 2C and table S2). A few studies
have demonstrated long-term stability, but the
catalyst is operated at a low space velocity or a
low temperature that leads to low propylene
STY. For the PDH catalysts reported, the propylene STY values have not exceeded 10 mol
gnoble metal–1 hour–1 after 500 hours of reaction
(Fig. 2C). Thus, simultaneously achieving both
high propylene STY and catalytic stability has
remained a challenging goal.
Our present In/Rh@S-1 catalyst had propylene
STYs of 45 mol gRh–1 hour–1 on the basis of Rh
weight or 6.0 gC3H6 gcat–1 hour–1 on the basis of
the entire catalyst weight after continuous operation at 600°C for 1200 hours (Fig. 2C and
table S2) and avoided the trade-off between
propylene STY and stability. We further performed rigorous PDH reactions with pure propane as the feed. Propane conversions near
the thermodynamic equilibrium values and
propylene selectivity of 99% were achieved
at 550°C, and we observed no catalyst deactivation during a reaction period exceeding
5500 hours (Fig. 2D), representing a major
step forward in PDH under such stringent
3 of 7

,

to 650°C, and frequent catalyst regeneration is
required owing to deactivation mainly by coke
deposition (1). Our In/Rh@S-1 catalyst was
surprisingly stable during a 1200-hour conversion of propane diluted by N2 with a weight
hourly space velocity (WHSV) of 10 hour−1 at
600°C (Fig. 2B); propane conversion was sustained at as high as 63 ± 2.0%, and propylene
selectivity was 98 ± 2.0%. The present catalyst
not only exhibits high PDH stability (table S2)
but also can work without the co-feeding of H2
that sacrifices propane conversion to alleviate
catalyst deactivation by coke deposition. In contrast, the PtSnK/Al2O3 catalyst used in the current
commercial process deactivated rapidly under
the same reaction conditions. The PtSnK@S-1
catalyst, which displayed improved stability
through the confinement of active bimetallic
sites (12), also underwent deactivation under
our reaction conditions (Fig. 2B).
We further evaluated the propylene spacetime yield (STY)—which is the product yield
per unit volume or weight of catalyst per unit
time and the metric usually adopted for comparison of PDH reaction rates among different
catalysts—versus the catalyst duration. Although
high propylene STY values have been reported

1 March 2024

17

60 120 500 1500 2500 3500 4500 5500

y

We performed intensive studies with the In/
Rh@S-1 catalyst. The optimization reveals that at
a fixed Rh loading of 0.35 wt %, the catalyst with
an In loading of 1.93 wt % that was typically
used in this work shows high performance. A
lower In loading led to not only a lower propane conversion and propylene selectivity but
also poorer catalyst stability, whereas a higher
In loading decreased the propane conversion
(fig. S7). Next, the long-term stability was tested.
A typical Pt catalyst–based commercial PDH
process usually uses propane diluted by H2
with a space velocity of 4 to 13 hour−1 at 525°

15

y g

Long-term activity and stability testing

10

9b

y

shorter distance or closer proximity between
In and Rh species results in a shorter induction period, and thus the induction period
may correspond to the generation of RhIn
bimetallic active sites for PDH by migration
of In species into the pore of zeolite S-1 for
subsequent interaction with the confined Rh
species. These results also offer a simpler
protocol of physical mixing or impregnation
for the fabrication of efficient bimetallic PDH
catalysts despite relatively longer induction
periods.

Pure
C3H8

g

Fig. 2. Alkane dehydrogenation performance of In-modified Rh@S-1
catalysts. (A) Effect of the manner of introduction of In into Rh@S-1 on catalytic
performances. (B) Stability of In/Rh@S-1 catalyst as well as typical Pt catalysts
for PDH. Reaction conditions: T = 600°C, WHSV = 10 hour−1, C3H8/N2 = 1:3.
(C) Comparison of STY versus time on stream of In/Rh@S-1 with some
typical Pt- and Rh-based catalysts. STY is defined as the number of moles of
propylene produced per gram of NMs per hour. All reaction conditions and

PtCu/Al2O3
PtCuSiO3
PtCu/SiO2
PtGa/SiO2
PtGaPb/SiO2
PtLa/mz-deGa
PtMn/SiO2
Pt@Ge-UTL

9a,b PtZn@S-1

6 14
8 2

3

10
11
12
13
14
15
16
17

Reaction temperature (°C)

Time on stream (hours)

Zeng et al., Science 383, 998–1004 (2024)

4

0

Conversion or selectivity (%)

80

1000

9
20 10 7

Time on stream (hours)

C3H6 select.

0

30

1200

100

20

12

1 In/Rh@S-1
(This work)
2 RhCu/SiO2
3 RhGa/Al2O3
4 Rh/ZrO2
5 PtSn/CeO2
6 PtSnK@S-1
7 PtSn/Al2O3-S
8 PtSn/SiO2

p

Conversion or selectivity (%)

900

E

60

40

0

100

Time on stream (hours)

D

11

5

0

40

50

STY (molC3H6 gNMs -1 h -1)

B

100

Conversion or selectivity (%)

Conversion or selectivity (%)

A

RES EARCH | R E S E A R C H A R T I C L E

C

Absorbance (a.u.)

Rh Rh

Rh@S-1(used)
CO

CO
CO

In/Rh@S-1(used)
CO

B

CO

A

Rh Rh

Rh Rh
Rhx

2 nm

2 nm

2150

Rh@S-1 (used)

Rh@S-1 (fresh)

D

1950

1850

1750

Wave number (cm-1)

F

In edge

1.2

In0

0.8

0.4

g

20 nm

In3+

p

Normalized absorption (a.u.)

E

20 nm

2050

In2O3
In foil
In/Rh@S-1 (fresh)
In/Rh@S-1 (used)
0.0
27960

27975

27990

Energy (eV)

H

I
In edge

In-O In-Rh In-In

Rh-O Rh-Rh

Rh edge

Rh-In

0.8

0.4

Rh2O3
Rh foil
Rh@S-1 (fresh)
Rh@S-1 (used)
In/Rh@S-1 (fresh)
In/Rh@S-1 (used)

In2O3
In foil

In/Rh@S-1 (fresh)
In/Rh@S-1 (used)

FT magnitude (a.u.)

FT magnitude (a.u.)

Rh0
Rh2O3
Rh foil

y g

Normalized absorption (a.u.)

1.2

Rh edge

Rh3+

27945

y

G

27930

In/Rh@S-1 (used)

In/Rh@S-1 (fresh)

Rh@S-1 (fresh)
Rh@S-1 (used)

In/Rh@S-1 (fresh)
In/Rh@S-1 (used)

0.0
23232

23254

23276

1

23298

2

3

5

conditions. The selectivity of propylene was
≥95% for pure propane conversions at 600°C
and ≥90% at 630°C, and only a slight deactivation was observed after 120 hours at 630°C
(fig. S8). The high stability in pure propane
conversions was also observed at conversions
below thermodynamic equilibrium at a high
WHSV, whereas the two reference catalysts,

2

3

4

5

R (Å)

image) of S-1. (F) Normalized In K-edge XANES spectra of In/Rh@S-1 and
reference samples. (G) Normalized Rh K-edge XANES spectra of Rh@S-1,
In/Rh@S-1, and reference samples. (H) k2-weighted In K-edge EXAFS spectra
of In/Rh@S-1 and reference samples. k denotes the wave vector of the
photoelectron. (I) k2-weighted Rh K-edge EXAFS spectra of Rh@S-1, In/Rh@S-1,
and reference samples.

PtSnK/Al2O3 and PtSnK@S-1, underwent quick
deactivation under the same conditions (fig. S9).
To maximize the utilization of Rh, we investigated the effect of Rh loadings. At a fixed In
loading of 1.93 wt %, the In/Rh@S-1 catalyst
with a Rh loading of 0.010 wt % could still
work for PDH with pure propane at a high
WHSV of 150 hour−1 in spite of a low propane

1 March 2024

1

,

Fig. 3. Migration of indium oxide and formation of RhIn clusters. (A and
B) High-angle annular dark-field scanning transmission electron microscopy
(HAADF-STEM) images of Rh@S-1 (fresh) and Rh@S-1 (used). (C) CO-adsorbed
FTIR spectra for Rh@S-1 (used) and In/Rh@S-1 (used). (D and E) HAADFSTEM images for In/Rh@S-1 (fresh) and In/Rh@S-1 (used). The red arrows
point to white aggregates of In2O3 species on the outer surface (edge in the

Zeng et al., Science 383, 998–1004 (2024)

4

R (Å)

Energy (eV)

y

23210

conversion (fig. S10), offering a propylene STY
of up to 1860 mol gRh–1 hour–1 (table S2). On
the other hand, an increase in the Rh loading
from 0.35 to 0.62 wt % was rather unbeneficial
to the propane conversion and propylene selectivity. An In/Rh@S-1 catalyst with 0.95 wt %
In and 0.05 wt % Rh provided a propylene STY
of >600 mol gRh–1 hour–1 at a propane conversion
4 of 7

RES EARCH | R E S E A R C H A R T I C L E

of >45% at 600°C (fig. S11), which is one or two
orders of magnitude higher than those reported for Pt-based catalysts at comparable
propane conversions (table S2). These results
further suggest the existence of a suitable In/Rh
ratio to achieve optimum performance, which
may imply a specific configuration of the RhIn
active site.
We also found that the In/Rh@S-1 catalyst
was very efficient for the dehydrogenation of
ethane and n-butane to ethylene and butenes,
respectively. As the temperature increased from
450° to 650°C, as was the case for propane, the
conversions of ethane and n-butane increased
and were near the thermodynamic equilibrium
values (Fig. 2E), and the selectivity of corresponding alkenes (see fig. S12 for the distribution of butenes) was 98 ± 2%. The catalyst also
displayed high stability for the dehydrogenation of ethane and butanes (fig. S13).

y
,

5 of 7

y g

1 March 2024

Density functional theory (DFT) calculations
for the formation energies of some monometallic and bimetallic clusters composed of Rh
and In atoms reveal that the RhIn clusters are
energetically more stable than the monometallic Rh or In clusters (fig. S22). Taking Rh5 as
a model of Rh clusters, we studied the dynamic
evolution of a Rh5 cluster in the presence of In
atoms at 550°C by an ab initio molecular dynamics (AIMD) simulation approach (30). The
result suggests that the Rh atoms begin to disperse when the Rh5 cluster is combined with
In5 or In80 ensembles (fig. S23). The presence
of sufficient In atoms could enable the complete isolation of Rh atoms, forming RhIn clusters with single-atom Rh sites. The Bader charge
analysis indicates a negative charge on Rh atoms
(fig. S24, A and B), which would contribute to
the dispersion and isolation of Rh atoms in the
RhIn clusters because of the electrostatic repulsion. The total energy profile analysis shows
energy gains from the formation of Rh–In
bonds that can compensate for the energy cost
for breaking Rh–Rh bonds in the Rh5/In20 or
Rh5/In80 system (fig. S24, C and D). To validate the interactions between Rh and In and
to mitigate possible biases arising from the distinctive nature of cluster models, we extended
our studies to the evolution of Rh clusters on
the In(100) plane and observed a comparable
phenomenon of Rh–Rh separation (fig. S25).
However, such a Rh–Rh separation was not
observed on the Cu(100) plane (fig. S26).
The role of In in preventing Rh from aggregation in zeolite was also studied with the
AIMD simulation by taking the structure evolution of a Rh–Rh dimer inside S-1 as a model of

y

Zeng et al., Science 383, 998–1004 (2024)

Theoretical results

g

We characterized the as-synthesized and postreaction Rh@S-1 catalyst using aberrationcorrected scanning transmission electron
microscopy (AC-STEM). The Rh species in the
as-synthesized Rh@S-1 catalyst existed as single atoms (Fig. 3A and fig. S14), whereas after
PDH reaction, the Rh atoms tended to aggregate and form clusters with sizes close to or
even larger than the channel dimension of S-1
(∼0.55 nm) (Fig. 3B and fig. S15). The presence
of aggregated Rh clusters in the Rh@S-1 catalyst after PDH was further confirmed by the
CO-adsorbed IR spectrum. The IR bands attributable to geminal CO adsorption (twin
bands at around 2100 and 2000 cm–1), linear
CO adsorption (at around 2050 cm–1), and bridge
adsorption (at around 1960 and 1800 cm–1) were
observed (Fig. 3C), and the appearance of intense bridge CO adsorption bands provided
evidence for the formation of Rh clusters in
the Rh@S-1 catalyst after PDH.
For the fresh In/Rh@S-1 catalyst, nanoparticles or nanoaggregates could be observed
(Fig. 3D), and with a combination of x-ray
photoelectron spectroscopy (XPS) and argon
ion sputtering, we identify these species as
nanosized In2O3 largely located on the outer
surface of S-1 (fig. S4). The In/Rh@S-1 catalyst
after reaction contained fewer and smaller
In2O3 clusters (Fig. 3E and fig. S4), whereas the
In loading did not change notably (table S1).
The disappearance of In2O3 aggregates was
also clearly observed for the bulk-In2O3+Rh@S-1
catalyst after reaction (fig. S16), which also
showed an induction period during the PDH
reaction (Fig. 1D). These observations suggest
that In species migrated into the channel of
S-1 in these catalysts under reaction conditions, becoming finely dispersed. Small clusters of several angstroms were observed in
the used RhIn@S-1 catalyst with the shortest
induction period (fig. S17), suggesting that the

spectra, we ascribed the peak at around 2.5 Å
for the working In/Rh@S-1 catalyst to the Rh–In
coordination.
Neither the In K-edge nor the Rh K-edge
EXAFS spectra showed the appearance of the
In–Rh or Rh–In coordination in the fresh In/
Rh@S-1. For the working In/Rh@S-1catalyst,
CNs of Rh–In bonds (Rh-edge) of 4.2 ± 0.6 and
In–O bonds (In-edge) of 2.0 ± 0.2 (fig. S20 and
table S3) were obtained from the EXAFS fitting analysis. Similar CNs were obtained for
the RhIn@S-1 catalyst under reaction conditions (fig. S21 and table S3). The combination of
operando EXFAS, AC-STEM, and CO-adsorbed
FTIR results enables us to infer that the active
structure for PDH is the RhIn4 site, which is anchored onto the framework of S-1 through the
In–O linkage and is analogous to a ligandstabilized intermetalloid cluster (29). The
operando XAS results supported our idea
that the In species in a low valance state
under reaction conditions migrated into S-1,
resulting in a single-atom Rh site coordinated
with In atoms and stabilized by the framework
of S-1.

p

Catalyst characterization

active site is the RhIn cluster composed of
several atoms. In the CO-adsorbed IR spectra
for the used In/Rh@S-1, In/S-1+Rh@S-1, and
RhIn@S-1 catalysts, the bridge CO adsorption
at around 1960 and ~1800 cm–1 was absent (Fig.
3C and fig. S18), unlike for the used Rh@S-1
catalyst, which suggests that the Rh species in
the In-modified Rh@S-1 catalysts exist as single atoms. Thus, the migration of In species
would prevent the aggregation of Rh single
atoms into clusters during reaction.
We performed operando x-ray absorption
spectroscopy (XAS) studies on the structure
of active sites and the dynamic migration of
In species in the In/Rh@S-1 catalyst under
PDH conditions. The In K-edge x-ray absorption near-edge structure (XANES) spectrum
indicated that the oxidation state of In was
reduced from +3 to between +1 and 0 (Fig. 3F).
The operando Rh K-edge XANES spectrum
showed a decrease in the white line and the
position shift of the absorption edge toward
the left side as compared with the Rh foil
(Fig. 3G). This observation not only indicated the complete reduction of Rh3+ in the
working catalyst but also the enrichment of
electrons at the Rh site that likely occurred
through interaction with In species. We infer
that the reduced In species migrated into zeolite S-1 and formed bimetallic RhIn clusters
with Rh, leading to the electron transfer from
In to Rh (28).
The operando extended x-ray absorption
fine structure (EXAFS) spectroscopic analysis provides information on evolution of the
RhIn cluster structure in the In/Rh@S-1 catalyst
during the reaction. The real-space In K-edge
EXAFS spectra (Fig. 3H) showed that the peak
at 1.52 Å, which we assigned to the In–O bond,
decreased, confirming the reduction of In species during the reaction. Despite its low intensity, the remaining In–O peak was the one
dominant oxygen linkage and suggested the
incomplete reduction of In3+ to Ind+. A peak at
around 2.5 Å, which we attributed to the In–Rh
coordination, was observed in the working catalyst and provided evidence for the direct
interaction between In and Rh atoms. The
operando Rh K-edge EXAFS spectra (Fig. 3I)
confirmed the complete reduction of Rh species because of the lack of Rh–O coordination.
The Rh–Rh coordination with peaks at around
1.9 and 2.5 Å could be observed in the used
Rh@S-1 catalyst, and the EXAFS fitting analysis
offered a Rh–Rh coordination number (CN) of
5.5 ± 0.8 (fig. S19 and table S3), which agreed
well with the AC-STEM and CO-adsorbed FTIR
results suggesting that this catalyst consisted
of Rh clusters. The peak at around 1.9 Å of the
Rh–Rh coordination did not appear in the
In/Rh@S-1 catalyst. The AC-TEM and COadsorbed FTIR results for this catalyst indicate
the atomic dispersion of Rh. By combining additional information from the In K-edge EXAFS

RES EARCH | R E S E A R C H A R T I C L E

B

3.2

Rh2In8 evolution
t = 0 ps

Two RhIn4 cluster in vacuum
Two RhIn4 cluster in S-1

4

t = 15 ps

3.0

3

Free energy (eV)

Average Rh-Rh distance (Å)

A

2.8

2.6

2.4

2

1

0

Initial

-1

2.2

Rh2

Rh2In2

Rh2In4

Rh2In6

Rh2In8

Rh2In10

9

8

7

AIMD models (inside the silicalite-1)

C

1

TS6u

C3H8(g)
-1

(2)

TS6d

0

*CH2+*CHCH3+2*H
TS3

(1)
TS7u

(3)
*C3H7+*H

TS7

d

(4)

C3H6(g)+2*H
TS8

*C3H4+4*H

TS11
0

TS12

*C3H5+3*H

*C3H5+3*H
−1
*C3H6+2*H

C2H4(g)+*CH2+2*H
TS9

*C2H4+*CH2+2*H

-3

*C3H4+4*H

TS10

*C2H4+*CH3+*H

y

Cracking on Rh 5
Dehydrogenation on Rh 5

C3H4(g)+4*H

C3H4(g)+4*H 1

*CH2CH2*CH2+2*H
-2

2

Deep dehydrogenation on Rh5

TS4

TS2

TS1

3

Rh-Rh distance (Å)
TS5

Dehydrogenation on RhIn4
Deep dehydrogenation on RhIn4
Cracking on RhIn4

4

g

Free energy (eV)

2

5

p

3

6

*C2H4+CH4(g)

Reaction coordinate
Fig. 4. Dynamics simulations of RhIn sites and DFT calculation results. (A) AIMD simulation of the evolution of Rh2Inx (x = 0 to 10) clusters with free In atoms in
S-1. (B) Slow-growth method to simulate the agglomeration process of two RhIn4 clusters. (C) Free energy profiles for C3H8 dehydrogenation on Rh5 and RhIn4
clusters. (1), 2-*C3H7+*H; (2), *C3H6+2*H; (3), 2-*C3H7+*H (dashed line) and 1-*C3H7+*H (solid line); and (4), *C3H6+2*H.

6 of 7

,

1 March 2024

of the system decreased when Rh atoms in two
separate sites converged toward each other in a
vacuum, forming a Rh2In8 cluster with a Rh–Rh
distance in the range of 5.7 to 2.7 Å (Fig. 4B).
A further decrease in the Rh–Rh distance to
≤2.7 Å, the diameter of a Rh atom, increased
the free energy as a result of the repulsion. In
contrast, when the RhIn4 site was confined in
S-1, the free energy increased continuously with
the approach of the two RhIn4 moieties to form
a dimeric complex and exceeded 2.00 eV when
the Rh–Rh distance became shorter than 4 Å.
The simulation result indicated that the RhIn4
site confined in zeolite S-1 had very little driving force to aggregate, hence providing an
effective means of constructing highly stable
PDH catalysts.
DFT calculations were performed to gain
insights into the effect of Rh dispersion on
PDH catalysis. We adopted Rh5 and RhIn4 as
models of Rh@S-1 and In/Rh@S-1 catalysts,
respectively, for better computational efficiency
(fig. S30). As indicated by the Gibbs free energy,

y

Zeng et al., Science 383, 998–1004 (2024)

catalyst with such a low In content underwent deactivation within 5 hours (fig. S7). An
increase in In content up to 1.93 wt % (In/Rh
ratio ≈ 5) further increased the propane conversion at the steady state and achieved excellent catalyst stability. These results are in
agreement with the AIMD simulation suggesting that an In/Rh ratio of 4 or more is
necessary to acquire atomic dispersion of Rh.
An overly high In content led to a decreased
propane conversion (fig. S7), probably because
of the covering of RhIn4 clusters or the blocking of micropores by excessive InOx. Thus, both
the AIMD simulation and the catalytic results
suggest the key function of In in separating Rh
atoms, leading to an efficient catalyst for PDH.
We further investigated the role of zeolite
S-1 in enhancing the catalyst stability by simulating the aggregation of the RhIn4 site in
vacuum and in S-1 using a slow-growth method
(30) (Fig. 4B and fig. S29). The Gibbs free energy of the isolated RhIn4 site was set to zero.
The simulation showed that the free energy

y g

the confined catalyst (fig. S27). In the stabilized
structure, the Rh dimer had a bond distance
of 2.3 Å and was located in the intersection of
the straight and sinusoidal 10-membered ring
channels. By adding In atoms to the original
Rh2 cluster, the Rh–Rh bond rapidly elongated
to ~2.8 Å at an In/Rh ratio of 4, which was
even larger than the length of the Rh–Rh bond
(2.67 Å) in Rh foil (31) (Fig. 4A). A snapshot of
simulation progress at 15 ps illustrates the
separation of the Rh–Rh bond by the presence
of In atoms in particular at an In/Rh ratio of ≥4
(Fig. 4A and fig. S27). Moreover, the AIMD simulation for a RhIn4 cluster inside S-1 at 550°C
shows that the Rh–In bond lengths in the cluster at the equilibrium position are close to those
obtained by the fitting analysis of the operando
EXAFS result (fig. S28 and table S3), confirming
the reliability of the simulation results.
Our catalytic experiments showed that the
addition of In with a low loading (0.48 wt %)
into Rh@S-1 could enhance both the propane
conversion and propylene selectivity, but the

RES EARCH | R E S E A R C H A R T I C L E

Discussion

RE FERENCES AND NOTES

AC KNOWLED GME NTS

We thank L. Cui, H. Chen, J. Cheng, W. Wang, W. Li, and X. Xiong
at Xiamen University for help with partial theoretical calculations,
in situ XPS characterization, and PDH evaluation. We thank X. Chen
from Tsinghua University and M. Tang from Utrecht University
for help with AC-STEM characterization. We thank the BL14W1 XAFS
beamline of Shanghai Synchrotron facilities (SSRF) for providing
beamtime. Funding: This work was supported by the National
Key Research and Development Program of the Ministry of Science
and Technology of China (2022YFA1504603 and 2022YFA1504500)
and the National Natural Science Foundation of China (22121001,
22222206, U22A20392, 22132004, 92045303, 91945301, and
92145301). Author contributions: L.Ze. conducted the catalyst
exploration, catalytic reactions, basic catalyst characterizations,
data analysis, and drafted the paper. K.C. and Qin.Z. guided
the experimental design and data analysis and revised the paper.
F.S. and Y.We. performed operando XAS studies and analyzed
the data. Q.F. and L.L. contributed to DFT calculations and drafted the
DFT section of the paper. W.Z. and J.K. assisted in the long-term
stability evaluation and analyzed the data. Qiu.Z. and M.C. conducted
part of the characterizations and analyzed the data. Q.L., L.Zh.,
and J.H. conducted major TEM characterizations and analyzed
relevant images. J.C. provided suggestions and guidance about
the DFT calculations. Z.J. guided the operando XAS measurements,
analyzed the data, and drafted part of the paper. G.F. guided
the simulations and DFT calculations, analyzed the data, and
drafted part of the paper. Y.Wa. initiated the project, guided the
study, and co-wrote the paper. All authors approved the final
version of the manuscript. Competing interests: J.K., L.Ze., W.Z.,
Qin.Z., and Y.Wa. are inventors listed on Chinese patent
(ZL201910260429.7) filed by Xiamen University, which covers
the Rh-based catalysts reported in this paper. The authors declare
that they have no other competing interests. Data and materials
availability: All data are available in the main text or the
supplementary materials. License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse

y

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk5195
Materials and Methods
Figs. S1 to S30
Tables S1 to S3
References (32–51)

y g

1. J. J. Sattler, J. Ruiz-Martinez, E. Santillan-Jimenez,
B. M. Weckhuysen, Chem. Rev. 114, 10613–10653 (2014).
2. Z. Nawaz, Rev. Chem. Eng. 31, 413–436 (2015).
3. S. Chen et al., Chem. Soc. Rev. 50, 3315–3354 (2021).
4. Y. Dai et al., Chem. Soc. Rev. 50, 5590–5630 (2021).
5. M. Monai, M. Gambino, S. Wannakao, B. M. Weckhuysen,
Chem. Soc. Rev. 50, 11503–11529 (2021).
6. A. H. Motagamwala, R. Almallahi, J. Wortman, V. O. Igenegbai,
S. Linic, Science 373, 217–222 (2021).
7. M. D. Marcinkowski et al., Nat. Chem. 10, 325–332 (2018).
8. R. Ryoo et al., Nature 585, 221–224 (2020).
9. Y. Nakaya, J. Hirayama, S. Yamazoe, K. I. Shimizu, S. Furukawa,
Nat. Commun. 11, 2838 (2020).
10. R. T. Hannagan et al., Science 372, 1444–1447 (2021).
11. K. Searles et al., J. Am. Chem. Soc. 140, 11674–11679
(2018).
12. L. Liu et al., Nat. Mater. 18, 866–873 (2019).
13. Q. Sun et al., Angew. Chem. Int. Ed. 59, 19450–19459 (2020).
14. Y. Ma et al., Nat. Catal. 6, 506–518 (2023).
15. J. Wang et al., ACS Catal. 11, 4401–4410 (2021).
16. C.-T. Au, C.-F. Ng, M.-S. Liao, J. Catal. 185, 12–22 (1999).
17. Y. Wang, P. Hu, J. Yang, Y.-A. Zhu, D. Chen, Chem. Soc. Rev.
50, 4299–4358 (2021).
18. N. Raman et al., ACS Catal. 9, 9499–9507 (2019).
19. C. W. Andersen et al., Angew. Chem. Int. Ed. 56, 10367–10372
(2017).
20. J. L. Weber et al., Catal. Today 342, 161–166 (2020).
21. Y. Wang et al., Angew. Chem. Int. Ed. 60, 17735–17743 (2021).
22. Y. Ding et al., ACS Catal. 11, 9729–9737 (2021).
23. L. Xu et al., Science 380, 70–76 (2023).
24. D. Zhao et al., Nature 599, 234–238 (2021).
25. P. Gao et al., ACS Catal. 8, 571–578 (2018).
26. M. R. Mihályi, H. K. Beyer, Chem. Commun. 2001, 2242–2243
(2001).
27. B. Chakraborty, B. Viswanathan, Catal. Today 49, 253–260
(1999).
28. S. Nishimura, A. T. N. Dao, D. Mott, K. Ebitani, S. Maenosono,
J. Phys. Chem. C 116, 4511–4516 (2012).

29. K. Mayer, J. Weßing, T. F. Fässler, R. A. Fischer, Angew. Chem.
Int. Ed. 57, 14372–14393 (2018).
30. T. K. Woo, P. M. Margl, P. E. Blöchl, T. Ziegler, J. Phys. Chem. B
101, 7877–7880 (1997).
31. M. Harada, K. Asakura, Y. Ueki, N. Toshima, J. Phys. Chem. 97,
10742–10749 (1993).

g

We succeeded in constructing a highly efficient and highly stable alkane-dehydrogenation
Rh single-atom catalyst by using the dynamic
migration of In species under reaction conditions. The catalyst demonstrated C2–C4 alkane
conversions approaching thermodynamic equilibrium values with excellent alkene selectivity
in a wide temperature range of 450° to 650°C.
The catalyst stability exceeds 1200 hours in
PDH to propylene with a propylene yield of
>60% at 600°C. No catalyst deactivation was
observed for 5500 hours under a harsh atmosphere of pure propane, simulating industrial
dehydrogenation conditions. A propylene STY
of >600 mol gRh–1 hour–1 was attained at a propane conversion of >45%, which is one or two
orders of magnitude larger than those reported

for Pt-based catalysts under comparable propane conversions. Our studies revealed that
In species function like solvents and alloying
agents, which effectively created in situ dilute
Rh atoms in the form of RhIn4 sites attached
onto the zeolite framework through In–O linkages. This transformation turned the unselective and unstable Rh-based catalysts into highly
selective and ultrastable PDH catalysts. Our
findings also offer a simple protocol for synthesizing robust single-atom catalysts that
are operated at harsh reaction conditions.

p

the dehydrogenation of propane to adsorbed
propylene (*C3H6) was energetically favorable
on the Rh5 cluster. However, the subsequent
desorption of *C3H6 was endothermic, with an
energy requirement of 1.2 eV (Fig. 4C). The
deep dehydrogenation of *C3H6 into C3H4 was
impeded by a substantial free energy barrier
for the desorption of *C3H4 (2.1 eV). Alternatively, the formation of cracking products by
means of *CH2CH2*CH2 on the Rh5 cluster was
more feasible, characterized by a lower free
energy barrier of 0.8 eV. For the RhIn4 cluster,
although the dehydrogenation step of C3H8
encountered a higher free energy barrier of
1.3 eV, the energy required for the desorption of
*C3H6 was largely reduced (0.24 eV). The deep
dehydrogenation and cracking on the RhIn4
cluster presented free energy barriers of 1.5 and
2.8 eV, respectively (Fig. 4C). Thus, the propylene
formation path became more favored. These
DFT calculation results clearly demonstrate the
contribution of atomic dispersion of Rh by In
species to achieving high propylene selectivity
during PDH.

Submitted 26 August 2023; accepted 22 January 2024
10.1126/science.adk5195

y
,

Zeng et al., Science 383, 998–1004 (2024)

1 March 2024

7 of 7

RES EARCH

ATMOSPHERIC DYNAMICS

Large-scale self-organization in dry
turbulent atmospheres
Alexandros Alexakis1*, Raffaele Marino2, Pablo D. Mininni3, Adrian van Kan4,
Raffaello Foldes2, Fabio Feraco2,5
How turbulent convective fluctuations organize to form larger-scale structures in planetary atmospheres
remains a question that eludes quantitative answers. The assumption that this process is the result of an
inverse cascade was suggested half a century ago in two-dimensional fluids, but its applicability to
atmospheric and oceanic flows remains heavily debated, hampering our understanding of the energy
balance in planetary systems. We show using direct numerical simulations with spatial resolutions of
122882 × 384 points that rotating and stratified flows can support a bidirectional cascade of energy,
in three dimensions, with a ratio of Rossby to Froude numbers comparable to that of Earth’s
atmosphere. Our results establish that, in dry atmospheres, spontaneous order can arise through an
inverse cascade to the largest spatial scales.

Set up

@t u þ u  ∇u þ 2W  u ¼ ∇P  ez N f þ n∇2 u þ f

ð1Þ
@t f þ u  ∇f ¼ N ez  u þ k∇2 f

ð2Þ

where W is the solid body rotation rate, N is
the Brunt-Väisälä frequency, P is the pressure,
n is the viscosity, k is the density diffusivity,
and f is an external forcing acting at scales
1 of 5

,

We consider a fluid in a Cartesian, triply periodic domain of vertical height H and horizontal dimension L = 32H, in the presence of
gravity, a stable mean density gradient, and
solid body rotation in the vertical direction
(30). The dynamics of the system are described
by the incompressible velocity field u and the
normalized density variation f, governed by
the Boussinesq equations (31, 32)

y

1 March 2024

y g

Alexakis et al., Science 383, 1005–1009 (2024)

y

*Corresponding author. Email: alexakis@phys.ens.fr

g

1
Laboratoire de Physique de l’Ecole normale supérieure, ENS,
Université PSL, CNRS, Sorbonne Université, Université ParisDiderot, Sorbonne Paris Cité, Paris, France. 2Université de Lyon,
CNRS, École Centrale de Lyon, INSA Lyon, Université Claude
Bernard Lyon 1, Laboratoire de Mécanique des Fluides et
d’Acoustique, UMR5509 - F-69134, Écully, France. 3Universidad
de Buenos Aires, Facultad de Ciencias Exactas y Naturales,
Departamento de Física, and CONICET - Universidad de Buenos
Aires, Instituto de Física Interdisciplinaria y Aplicada (INFINA),
Ciudad Universitaria, 1428 Buenos Aires, Argentina. 4Department
of Physics, University of California Berkeley, Berkeley, CA 94720,
USA. 5Leibniz-Institute of Atmospheric Physics at the University
of Rostock, 18225 Kühlungsborn, Germany.

ferring enstrophy to smaller scales, whereas energy is transferred to larger scales. This process
takes place on a continuum of scales, forming
a constant flux of energy from small to large
scales in what is known as an inverse energy
cascade, as opposed to the disordered forward
energy cascade observed in 3D turbulence that
is directed to small scales.
Although planetary atmospheres are often
very thin (Earth’s atmosphere has horizontal
synoptic scales of the order of 1000 km and a
pressure scale height of 7.6 km), the corresponding flows are far from being 2D. Nonetheless, two-dimensionality is not imperative
for the appearance of self-organization. 3D
rotating and stratified flows (two key ingredients of atmospheric dynamics) conserve a
different invariant—the potential vorticity—
that can also lead to an inverse cascade. This
happens in the quasi-geostrophic limit, where
rotation and stratification are asymptotically
strong (6) and where gravito-inertial waves
are filtered out. Inverse cascades can also be
present in rotating Rayleigh-Bénard convection (7–11), where in this case, a generalized
quasi-geostrophic limit can be considered that
partially preserves gravito-inertial modes.
However, for most planetary flows, the quasigeostrophic limit is, at best, a crude approximation, with gravito-inertial waves composing a
substantial part of the energy budget cascading
energy forward (12–15). Thus, an inverse cascade
in planetary atmospheres caused either by
two-dimensionality or quasi-geostrophy remains
conjectural.
Could atmospheric dynamics display an inverse cascade away from these limits? In recent
years it has been demonstrated that a hybrid
state can be reached such that larger scales
cascade energy inversely, whereas smaller scales
cascade energy forward in what is now known
as a bidirectional cascade (16). Bidirectional
cascades were observed early on with direct
numerical simulations (DNS) (17–22). In ro-

p

F

low structures thousands of kilometers
wide are not uncommon in the atmosphere of Earth and that of other planets. The energy of these structures could
originate from processes associated with
the global atmospheric circulation but could
also originate from smaller-scale convective
turbulence. In the latter case, small-scale eddies conspire to self-organize into larger structures. Such a process goes against our daily
life experience, where turbulence generates
smaller-scale erratic structures, such as those
observed when pouring milk into a cup of
coffee. It is therefore necessary to come up
with convincing mechanisms for how such
large-scale organization can take place in planetary atmospheres.
One of the most important theoretical discoveries in the 20th century in the field of nonequilibrium physics is the phenomenon of
self-organization, which spontaneously creates
large-scale order out of small-scale disorder. One
of the first examples of this process was given by
Onsager with the statistical mechanics of a gas
of point vortices (1) that was later generalized
to two-dimensional (2D) turbulent flows (2–5). A
2D flow conserves an additional invariant, the
enstrophy, given by the mean squared vorticity.
The relation between energy and enstrophy
leads to an incompatibility for the simultaneous
bulk transfer of both quantities to the small
scales. As a result, vortices self-interact, trans-

tating and stratified flows, simulations also
indicate the presence of bidirectional cascades
(23–25), though in a regime where rotation
and stratification are comparable in strength,
which is typical for the ocean but not for the
atmosphere.
Nonetheless, the existence of self-organization
processes through a bidirectional cascade in
planetary atmospheres has become a compelling possibility as recent research using satellite images with cloud tracking analysis and in
situ aircraft measurements has estimated the
flux (and thus also the direction) of the energy
cascades in planetary flows in Earth’s atmosphere (26, 27), the ocean (28), and the Jovian
atmosphere (11, 29). These studies have affirmed
the presence of both inverse and forward energy
cascades depending on the scale examined or on
the altitude. However, satellite images constrain
the measurements to 2D slices, thus ignoring
any processes occurring in the third direction.
Up to now, there is no definite evidence of
whether planetary atmospheric flows satisfy
the necessary conditions for a bidirectional
cascade to establish itself. The difficulty in
answering such questions lies, on the one hand,
in the fact that information from satellite images
is limited and, on the other hand, in the extreme
parameter values that are met in planetary atmospheres, which are hard to obtain in DNS.
However, not only has the technology to perform high-cadence high-resolution observations
of the atmosphere just started to come along,
but the computational power to perform DNS
of stratified atmospheres in a realistic parameter space has also become available. In this
work, with the use of DNS in a large grid using
40,000,000 central processing unit (CPU) hours,
we establish that the fluid model of a rotating
and stably stratified dry atmosphere described
by the nonhydrostatic Boussinesq equations can
generate a bidirectional cascade leading to largescale organization of the flow.

RES EARCH | R E S E A R C H A R T I C L E

2 of 5

,

1 March 2024

y

Alexakis et al., Science 383, 1005–1009 (2024)

y g

Fig. 1. Visualization of density fluctuations ϕ and of the velocity field in the computational domain.
Structures at scales much larger than the forcing (i.e., at the scale of the domain height, with wave number
kF = 2p/H) are abundant in the visible horizontal plane (left), indicative of an efficient transfer of the energy toward
the lowest modes along the perpendicular direction in Fourier space. The large-scale patterns are visible in the
flow visualization, shown by arrows in a zoomed-in view (top right). At the same time, 3D instability patterns
and small-scale features are detectable in both horizontal and vertical cuts of the zoomed-in simulation domain
(bottom right), which suggests the action of a forward turbulent cascade. See the supplementary materials for
a movie of the density fluctuations in the entire domain. Visualizations were done with VAPOR (35).

y

Figure 1 shows visualizations of the flow and
the density field made using VAPOR (35). Structures with horizontal widths 10 times as large as

them by their argument—i.e., Ei(k), Ei(k⊥), and
Ei(k‖) where i is T, K, P, GW, or QG, which
stand for total, kinetic, potential, gravito-inertial
wave, and quasi-geostrophic, respectively. In
addition, we define Ei(k⊥, k‖) that shows the
spectral energy density for a given pair k⊥, k‖.
The left panels in Fig. 2 show the energy
spectra Ei(k), Ei(k⊥), and Ei(k‖) with the energy
component i as indicated in the legend. The
inset also shows the ratios RGW = EGW/ET and
RQG = EQG/ET. In the top panel, for k > kH, the
spectra have been averaged over shells of width
kH because otherwise large peaks of period kH
are observed due to the strong domain anisotropy, shown by the light gray lines for the total
energy spectrum. For wave numbers larger
than kH and smaller than the viscous wave
number kn, the spherically averaged spectrum
displays a power-law behavior with an exponent very close to Kolmogorov’s prediction
k−5/3 for homogeneous isotropic turbulence.
This power-law behavior, composed 70% by
gravito-inertial waves, is indicative of a forward energy cascade. At k smaller than kH, a
similar power law is observed (albeit with a
smaller prefactor). This indicates the presence
of an inverse cascade. This energy at small k is
almost exclusively kinetic, dominated by 2D
quasi-geostrophic modes.
For Ei(k⊥), three different power laws can be
observed. First, in the range kL < k⊥ < kH, a
k⊥−5/3 law is observed, where kL = 2p/L. This is
consistent with Earth’s atmospheric spectrum
between ≈10 and 500 km (36). Second, in the
range kH < k⊥ < kB = N/U, a steeper power law
close to k⊥−3 is observed, where kB is the buoyancy wave number. Finally, at larger k⊥, a
shallower power-law slope starts to appear
with exponent close to −5/3. Finally, the last
panel of Fig. 2 shows Ei(k‖) with k−5/3 and k−3
power laws indicated as references, the latter
observed in the atmosphere at vertical scales
near 1 km caused by gravity waves (37).
The right panels of Fig. 2 show the energy
fluxes across different surfaces in wave number
space: across constant k spheres P i(k), constant k⊥ cylinders P i(k⊥), and constant k‖
planes P i(k‖). As with the spectra, we distinguish between fluxes based on their arguments.
Here, i is T, K, or P, which stand for total,
kinetic, and potential energies, respectively.
Positive values imply a flux of energy toward
larger wave numbers, whereas negative values
indicate a flux toward smaller wave numbers.
P T(k) flux is positive for k > kH, which
indicates a forward cascade. However, a small
fraction, corresponding to 5% of the total
energy injection rate, cascades toward larger
scales. This is seen in the negative flux observed at k < kH. This flux is also constant for
more than a decade of wave numbers almost
up to kL. The inset in Fig. 2 shows the amplitude of this negative flux measured from different simulations varying only Re. The flux

g

Results

that of H can be seen by visual inspection. At the
same time, looking at the zoomed-in cross sections, it is obvious that these structures are far
from being 2D. In the larger scales, pancake
structures of alternating sign of f along the
vertical and the emergence of macroscopic
cyclones and anticyclones are visible in Fig. 1 (top
right). These features are observed in weather
maps and are a landmark of larger-scale, energycontaining structures in Earth’s atmosphere. At
the same time, smaller-scale overturning events
can be seen in the zoomed-in section that are
one-tenth the size of H (Fig. 1, bottom right) and
are also detectable sometimes in the sky as
Kelvin-Helmholtz billows. Thus, even at this
qualitative level, the presence of a bidirectional
cascade is evident.
To become more quantitative, we note that
the inviscid Boussinesq equations conserve the
total energy ET given by the sum of kinetic
energy EK and potential energy EP. Alternatively
we decomposed it into the energy of gravitoinertial modes EGW and the energy of quasigeostrophic modes EQG, where ET = EGW + EQG.
Gravito-inertial modes are dispersive wave
modes due to the combined restoring force of
gravity and Coriolis, whereas quasi-geostrophic
modes balance Coriolis and gravity forces with
pressure (see materials and methods for their
exact definitions). These energies are distributed differently in the Fourier space, among
vertical wave numbers k‖ and horizontal wave
numbers k⊥. We define three different energy
spectra
averaged
over fixed k ‖ , k ⊥ , and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
k ¼ k2∥ þ k2⊥ . We do not define a new symbol for each spectrum but distinguish between

p

‘F ∼ H injecting energy at a rate D. Although
this model has some strong simplifications,
like periodicity or a simplified forcing mechanism, it is the most elementary model capturing
the necessary physics to reproduce atmospheric
dynamics.
This system has four independent nondimensional control parameters: (i) the Reynolds
4=3
number ReD ¼ D1=3 kH =n; (ii) the Prandtl
number Pr ¼ n=k, that here is set to unity; (iii)
2=3
=W ; and
the Rossby number RoD ¼ D1=3 kH
1=3 2=3
(iv) the Froude number FrD ¼ D kH =N (with
kH = 2p/H). We can also define dimensionless
parameters based on the domain size L and the
flow root mean square velocity U as, for example,
Re ¼ UL=n,Ro ¼ U =ðHWÞ, and Fr ¼ U =ðNH Þ,
which are closer to the definitions used in atmospheric measurements.
Simulations were performed at resolutions
of 122882 × 384 grid points (30, 33). As a
reference, a domain height H = 15 km (equal
to twice the pressure scale height in Earth’s
atmosphere) corresponds to a domain length
of 480 km (corresponding to atmospheric mesoscales) and a vertical and horizontal resolution
of 39 m. Our simulations are characterized
by ReD ¼ 2000, RoD ¼ 1, and FrD ¼ 0:025, or
alternatively Re ≈ 2 × 106, Ro ≈ 0.4, and Fr ≈
0.01. These values are also compatible, for
example, with that of the mesosphere–lower
thermosphere (MLT) (34).

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
Fig. 2. Energy spectra and fluxes. (Left) Spherically (top), cylindrically (middle), and plane (bottom) averaged energy spectra, for all energy components. Insets show the
ratios of energy components RGW = EGW/ET (pink) and RQG = EQG/ET (green). (Right) Energy fluxes across spheres (top), cylinders (middle), and planes (bottom) in
spectral space. Total energy flux (black line), energy flux of kinetic energy (blue line), and energy flux of potential energy (red line) are shown. The forcing wave number kF ≈ kH
(where energy is injected), the buoyancy wave number kB = N/U, and the dissipation wave number kn (where energy is dissipated) are indicated by vertical dashed lines.

y g

Conclusions

We have shown that dry turbulent atmospheres
modeled by the nonhydrostatic Boussinesq
equations can lead to a bidirectional energy
cascade. The results showed that there is a
flux of energy directed to the small wave numbers k, corresponding to 5% of the total energy
injection rate at the largest Reynolds number. This flux, albeit small, is shown to per-

1 March 2024

sist up to the largest scales of the system and
is Re-independent as large values of Re are
reached.
Our analysis provides a detailed description
of how energy is transferred across scales and
between different modes. These transfers, indicated by the arrows sketched in Fig. 3, summarize the results in this work. Stratification,
rotation, and the geometric constraint of finite
H all play a role in the formation of this inverse cascade. In physical terms, at the scale of
the forcing, stratification is dominant, constraining a large fraction of the energy to QG
modes. This leads to the formation of pancake
structures, known in stratified turbulence (38),
that move energy to smaller k⊥ and larger k‖.
This process ceases at wave numbers where
stratification is comparable to rotation, Nk⊥ º
2Wk‖. Rotation, which tends to bidimensionalize
the flow (39), prevents larger k‖ modes from
appearing, and energy is converted to gravitoinertial mode energy that cascades back to
3 of 5

,

Alexakis et al., Science 383, 1005–1009 (2024)

k⊥ scaling at small k⊥ in Fig. 2. When rotationdominated scales are reached at 2Wk∥ ≃ Nk⊥
(black dotted line), QG modes transfer their
energy to GW modes that cascade it back to
small scales. Of the energy that has moved to
smaller k⊥, a finite amount is transferred
(cyan arrow) below the smallest dashed white
line (k2⊥ þ k2∥ ¼ k2H ). The energy in these modes
forms the E(k) º k−5/3 spectrum for k < kH. This
component of the energy is the only one that
escapes to the largest scales k → 0 and corresponds to a true inverse cascade.

y

increases with Re and saturates at the largest
Re. P T (k⊥) is also positive for k⊥ > kH and
negative for k⊥ < kH. However, in this case, the
fraction of energy that cascades toward smaller
k⊥ is five times as large as P T (k). P T (k‖) is
positive everywhere.
Although in 1D spectra and fluxes it is easier
to identify power laws, the energy distribution
depends on k‖ and k⊥ independently. In Fig. 3,
we show color-shaded plots of ET (k⊥, k‖) and
RGW ¼ EGW ðk⊥ ;k∥ Þ=ET ðk⊥ ;k∥ Þ. The arrows indicate the direction of the energy transfers
based on the fluxes in Fig. 2. A part of the injected energy is transferred to larger k⊥ (purple
arrows), producing the k⊥−3 spectrum observed
in the kH < k⊥ < kB range. RGW (k⊥, k‖) indicates
that this forward transfer takes place through
GW modes (green arrow). The peak of E(k⊥, k‖)
is observed at k⊥ ≃ 2kL and k∥ ≃ 2kH , formed by
an inverse transfer indicated by the black
arrow and dominated by QG modes. This
energy is responsible for the formation of the

RES EARCH | R E S E A R C H A R T I C L E

,

4 of 5

y

1 March 2024

1. L. Onsager, Nuovo Cim. 6, 279–287 (1949).
2. R. H. Kraichnan, Phys. Fluids 10, 1417–1423 (1967).
3. R. H. Kraichnan, D. Montgomery, Rep. Prog. Phys. 43, 547–619
(1980).
4. C. E. Leith, Phys. Fluids 11, 671–672 (1968).
5. G. K. Batchelor, Phys. Fluids 12, II-233–II-239 (1969).
6. J. G. Charney, J. Atmos. Sci. 28, 1087–1095 (1971).
7. P. H. Roberts, Philos. Trans. Royal Soc. A 263, 93–117
(1968).
8. F. H. Busse, J. Fluid Mech. 44, 441–460 (1970).
9. A. M. Rubio, K. Julien, E. Knobloch, J. B. Weiss, Phys. Rev. Lett.
112, 144501 (2014).
10. B. Favier, L. J. Silvers, M. R. E. Proctor, Phys. Fluids 26, 096605
(2014).
11. L. Siegelman et al., Nat. Phys. 18, 357–361 (2022).
12. J. Vanneste, I. Yavneh, J. Atmos. Sci. 61, 211–223 (2004).
13. J. Vanneste, Annu. Rev. Fluid Mech. 45, 147–172 (2013).
14. J. Thomas, D. Daniel, J. Fluid Mech. 911, A60 (2021).
15. J. Thomas, D. Daniel, J. Fluid Mech. 902, A7 (2020).
16. A. Alexakis, L. Biferale, Phys. Rep. 767–769, 1–101 (2018).
17. L. M. Smith, J. R. Chasnov, F. Waleffe, Phys. Rev. Lett. 77,
2467–2470 (1996).
18. A. Celani, S. Musacchio, D. Vincenzi, Phys. Rev. Lett. 104,
184506 (2010).
19. K. Seshasayanan, S. J. Benavides, A. Alexakis, Phys. Rev. E 90,
051003 (2014).
20. S. J. Benavides, A. Alexakis, J. Fluid Mech. 822, 364–385
(2017).
21. A. Sozza, G. Boffetta, P. Muratore-Ginanneschi, S. Musacchio,
Phys. Fluids 27, 035112 (2015).
22. A. van Kan, A. Alexakis, J. Fluid Mech. 899, A33 (2020).
23. R. Marino, P. D. Mininni, D. Rosenberg, A. Pouquet, Europhys.
Lett. 102, 44006 (2013).
24. A. Pouquet, R. Marino, Phys. Rev. Lett. 111, 234501 (2013).
25. R. Marino, A. Pouquet, D. Rosenberg, Phys. Rev. Lett. 114,
114504 (2015).
26. D. Byrne, J. A. Zhang, Geophys. Res. Lett. 40, 1439–1442
(2013).
27. G. P. King, J. Vogelzang, A. Stoffelen, J. Geophys. Res. Oceans
120, 346–361 (2015).
28. D. Balwada, J.-H. Xie, R. Marino, F. Feraco, Sci. Adv. 8,
eabq2566 (2022).
29. R. Young, P. L. Read, Nat. Phys. 13, 1135–1140 (2017).
30. Materials and methods are available as supplementary
materials.
31. J. Pedlosky, Geophysical Fluid Dynamics (Springer, 1987).
32. G. K. Vallis, Atmospheric and Oceanic Fluid Dynamics:
Fundamentals and Large-Scale Circulation (Cambridge Univ.
Press, 2017).
33. P. D. Mininni, D. Rosenberg, R. Reddy, A. Pouquet, Parallel
Comput. 37, 316–326 (2011).
34. H.-L. Liu, P. B. Hays, R. G. Roble, J. Atmos. Sci. 56, 2152–2177
(1999).
35. J. Clyne, P. Mininni, A. Norton, M. Rast, New J. Phys. 9, 301
(2007).
36. G. Nastrom, K. Gage, W. Jasperson, Nature 310, 36–38
(1984).
37. T. VanZandt, Geophys. Res. Lett. 9, 575–578 (1982).
38. P. Billant, J.-M. Chomaz, Phys. Fluids 13, 1645–1651 (2001).
39. H. P. Greenspan, The Theory of Rotating Fluids (Cambridge
Univ. Press, 1968).
40. W. C. Skamarock, S.-H. Park, J. B. Klemp, C. Snyder, J. Atmos. Sci.
71, 4369–4381 (2014).
41. K. S. Gage, G. D. Nastrom, J. Atmos. Sci. 43, 729–740
(1986).
42. D. K. Lilly, J. Atmos. Sci. 40, 749–761 (1983).
43. U. Schumann, J. Atmos. Sci. 76, 3847–3862 (2019).
44. J. Callies, O. Bühler, R. Ferrari, J. Atmos. Sci. 73, 4853–4872
(2016).
45. P. D. Mininni et al., Dataset for large-scale self-organisation in
dry turbulent atmospheres, version 1.0, Zenodo (2023);
https://doi.org/10.5281/zenodo.8091949.
46. D. Rosenberg, P. Mininni, P. C. Di Leoni, pmininni/GHOST:
GHOST master, version 2.0.0, Zenodo (2023); https://doi.org/
10.5281/zenodo.8015308.
47. S. Pearse et al., NCAR/VAPOR: Vapor 3.8.1, version 3.8.1,
Zenodo (2023); https://doi.org/10.5281/zenodo.7779648.

y g

Alexakis et al., Science 383, 1005–1009 (2024)

REFERENCES AND NOTES

y

Fig. 3. 2D energy spectra in log-log scale. White dashed lines indicate isotropic contours (i.e., modes with
constant wave number k). The solid white line indicates the maximum resolved wave numbers. (Left) Total
energy spectrum. (Right) Gravity wave energy spectrum ratio. The black dotted lines marks 2Wk|| = Nk?,
where inertial wave frequency matches gravity wave frequency. The arrows indicate the direction of the flux
of energy (see text for a description).

vides a link between theory and field observations that can help validate or discard
theoretical explanations.

g

which are not part of our domain. At scales
smaller than 500 km, which are considered in
our model, and down to ≈10 km, the observed
spectrum follows k−5/3 scaling, which agrees
with our simulation. The dynamics producing
this observed scaling, and whether it originates
from a forward or an inverse cascade, have
been long debated (40). An inverse cascade
acting on those scales was proposed as an
explanation (41, 42), but it was later discarded
(23, 40) because purely stratified turbulence
develops no inverse cascade. Our simulation
shows that the combination of a realistic
aspect ratio with realistic parameters gives rise
to upscaling of energy with a k−5/3 spectrum
even at scales as small as 15 km. Although weak,
the persistence of the inverse cascade makes a
few percent of the flux enough to account for
the observed mesoscale energy (42). But unlike
the transfer to gravity waves hypothesized by
Lilly (42), the inverse transfer we observe feeds
the QG modes. At even smaller scales (<10 km),
model parameterizations and the lack of
observations with simultaneous high spatial
coverage and resolution, especially in the upper
atmosphere, prevent us from drawing definitive conclusions about spectral slopes and their
origins. The spectrum in this range is still
anisotropic, displays variability, and is sensitive to atmospheric conditions and to the local
energy dissipation rate (43). It is difficult to
make direct comparisons with the simulation
in this range, but some common features can
be identified, such as the prevalence of gravity
waves and the occurrence of two horizontal
Kolmogorov subranges above 10 km and below
1 km with different amplitudes mediated by a
steeper spectrum, as has sometimes been reported in observational campaigns (43, 44).
Our work presents a spectral energy distribution from the mesoscales down to the smallest
turbulent scales as a direct outcome of the
physical mechanisms in the model and the
parameter range examined. As such, it pro-

p

larger k. This is true for all k except for the k‖ =
0 modes that are unaffected by rotation. These
modes, which follow 2D dynamics, cascade the
energy to ever smaller k⊥. Their stability against
3D perturbations is assured by rotation and the
finiteness of H that leads to the k‖ = 0 modes
being isolated (16). As a result, they cascade
energy to ever smaller k⊥ with no channel to
return this energy back to the small scales.
An important outcome from this picture is
that energy fluxes obtained from horizontal
averages can significantly overestimate the
inverse energy flux. In the presently examined
simulation, P T(k⊥) was five times as large as
the true inverse flux P T(k). This result limits
the observational estimates of P T(k). Most
present estimates of the inverse flux are based
on averages of 2D slices obtained by satellite
images. They thus contain no information on
the fields’ variations in the third direction, and
as a result, it is P T(k⊥) that is measured and
not P T(k), which represents the true inverse
flux. As a reference, and for comparisons with
observations, we provide as supplementary
materials [fig. S1 (30)] spectra and third-order
structure functions along horizontal tracks, such
as those resulting from airplane or satellite
measurements. The structure functions display
a change in sign indicative of an inverse cascade but overestimate the inverse flux, just as
horizontal averages do. Thus, these estimates
of the inverse flux could be significantly larger
than their true values.
Although our model is not designed to capture the full complexity of Earth’s atmosphere,
which also has energy sources at planetary scales
larger than our computational domain, it reproduces some known features in the mesoscale
range. Global models indicate that the atmospheric spectrum varies with altitude, with
observations showing that planetary and
synoptic scales follow a k−3 spectrum at scales
larger than 500 km. This is attributed to a
direct cascade originating at planetary scales,

RES EARCH | R E S E A R C H A R T I C L E

ACKN OW LEDG MEN TS
We thank an anonymous reviewer for the suggestion to calculate
data along horizontal tracks, allowing for comparisons with
observational data. The simulation output was analyzed on HPC
facilities at the École Normale Superieure in Paris (France), at École
Centrale de Lyon (PMCS2I) in Ecully (France), and at Departamento
de Fisica (FCEN, UBA) in Argentina. Funding: Computer resources
in Joliot-Curie at CEA were provided by PRACE (research project ID
2020235566) and by GENCI (allocation no. A0110506421). This work
was supported by the projects “DYSTURB” (no. ANR-17-CE30-0004)
and “EVENTFUL” (no. ANR-20-CE30-0011) funded by the French
Agence Nationale de la Recherche (ANR), by the Studienstiftung
des deutschen Volkes, by the National Science Foundation
(grants DMS-2009563 and DMS-2308337), and by the German

Research Foundation (no. 522026592). Author contributions:
A.A., R.M., P.D.M., and A.v.K. conceived and planned the
numerical experiments and wrote the manuscript with input
from all authors. A.A. and P.D.M. performed necessary
modifications to the GHOST code. A.A. and A.v.K. performed the
computations. R.F. and F.F. performed visualizations. All authors
analyzed the data and discussed the results. Competing
interests: The authors declare that they have no competing
interests. Data and materials availability: Data in all figures are
available in the supplementary materials and on Zenodo (45).
The GHOST code used for the simulations is available on Zenodo
(46). 3D renderings were done with VAPOR, which is available
on Zenodo (47). License information: Copyright © 2024 the
authors, some rights reserved; exclusive licensee American

Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adg8269
Materials and Methods
Supplementary Text
Fig. S1
References (48–56)
Movie S1
Submitted 24 January 2023; accepted 12 January 2024
10.1126/science.adg8269

p
g
y
y g
y
,

Alexakis et al., Science 383, 1005–1009 (2024)

1 March 2024

5 of 5

RES EARCH

RIVER FLOW

Anthropogenic climate change has influenced global
river flow seasonality
Hong Wang1,2, Junguo Liu1,3,4*, Megan Klaar2, Aifang Chen1, Lukas Gudmundsson5, Joseph Holden2
Riverine ecosystems have adapted to natural discharge variations across seasons. However, evidence
suggesting that climate change has already impacted magnitudes of river flow seasonality is limited
to local studies, mainly focusing on changes of mean or extreme flows. This study introduces the use of
apportionment entropy as a robust measure to assess flow-volume nonuniformity across seasons,
enabling a global analysis. We found that ~21% of long-term river gauging stations exhibit significant
alterations in seasonal flow distributions, but two-thirds of these are unrelated to trends in annual
mean discharge. By combining a data-driven runoff reconstruction with state-of-the-art hydrological
simulations, we identified a discernible weakening of river flow seasonality in northern high latitudes
(above 50°N), a phenomenon directly linked to anthropogenic climate forcing.

1 March 2024

1 of 6

,

Wang et al., Science 383, 1009–1014 (2024)

For better interpretation of seasonal variations, AE trends were combined with annual
mean river flow trends (fig. S3A) to divide stations into six types (Fig. 2). For each station,
we define low-flow months as the three calendar months with the lowest long-term monthly mean river flow, whereas high-flow months
are the three calendar months with greatest
monthly mean flow. Therefore, trends in lowand high-flow months (TLH) contribute to annual mean and seasonal variations of river flow.
We found seasonal variations in the northern
high latitudes to be dominated by increasing
low flows and decreasing high flows (L+H−;
increasing trends of low flows and decreasing trends of high flows), accounting for
~46% of sites, and increasing low flows (L+H*;
increasing trends of low flows and nonprominent changes in high flows) solely accounting for ~14%. These two patterns result in a
significant decrease in RFS with no significant
annual mean trends (L+H−) or significantly

y

*Corresponding author. Email: liujg@sustech.edu.cn

The northern high latitudes (above 50°N) showed
discernible weakening in the seasonal cycle of
river flow [increasing AE (fig. S1)] (Fig. 1). In
northern North America (N. NA), ~40% of stations showed significantly decreasing RFS (P <
0.05). In comparison, only ~2% showed significantly increasing trends of RFS. Similar re-

y g

Changes in the trend of river flow seasonality

School of Environmental Science and Engineering, Southern
University of Science and Technology, Shenzhen 518055, China.
2
water@leeds, School of Geography, University of Leeds, Leeds
LS2 9JT, UK. 3North China University of Water Resources
and Electric Power, Zhengzhou 450046, China. 4Henan Provincial
Key Lab of Hydrosphere and Watershed Water Security, North
China University of Water Resources and Electric Power,
Zhengzhou 450046, China. 5Institute for Atmospheric and
Climate Science, ETH Zürich, 8092 Zürich, Switzerland.

y

1

Monthly high- and low-flow changes
mediate seasonality

g

is growing evidence that ACC has influences on
seasonal river flow, as indicated by climate and
hydrological model simulations (14, 15). Moreover, human-induced seasonal changes were
detected in the western United States (16). However, the question of whether ACC is detectable
in the magnitude of RFS at global scales remains unanswered, and it is challenging to
transfer the evidence across regions, as river
flow may vary locally because of atmospheric,
oceanic, and terrestrial modulation (17).
In this study, we used in situ observations
of monthly average river flow (18) from 10,120
gauging stations with either a minimum data
length of 35 years or uncertainty-controlled
shorter records from 1965 to 2014. In addition, we used recently published datasets of
observation-based global gridded runoff (GRUN)
reconstructed from data-driven models calibrated with observations to achieve global-scale
coverage (19). Multimodel results from the InterSectoral Impact Model Intercomparison Project
phase 2b (ISIMIP2b) are presented for detecting and attributing observed changes to human
influences on the climate system (20). To inform
our global analysis of seasonal flow regimes, a
generalized seasonality index was developed
by using apportionment entropy (AE) (21). AE
nonparametrically quantifies how evenly flow
rates are distributed across months; high AE
indicates a uniform distribution (low RFS), and
low AE points to large month-to-month differences (high RFS). AE can be used to characterize
changes in a hydroclimatological context because
it is statistically well suited to the characteristics of highly variable flow regimes (21, 22).

p

A

nthropogenic climate warming, which
could drive changes in the hydrological
cycle, has received increasing attention
(1). Water availability, a key concern, is
directly related to ecosystem functions
and societal interactions (2, 3). However, human activities are altering river flow patterns worldwide, both directly through flow
regulations and indirectly through land-use
change and the impacts of anthropogenic climate change (ACC) on air temperature, precipitation, soil moisture, and snowmelt regimes
(4–6). Consequently, more than two-thirds of
the world’s rivers have been altered even without considering the indirect impacts from ACC,
which is characterized by human-induced alterations in greenhouse gases and aerosols (7).
River flow seasonality (RFS) plays a critical
role in floods and droughts, threatening water
security and freshwater biodiversity (4, 8). For
example, a substantial portion of the early meltwater from snowpack depletion may ultimately
flow into oceans without being available for
human use (9). In addition, weakening RFS
(e.g., flood frequency reduction) can greatly
simplify community-wide riparian plant networks (10, 11). Riparian vegetation subsequently
influences freshwater biota to make seasonal
use of riparian areas and floodplains for feeding or breeding purposes (12).
Recent studies have shown a tendency for
changes in RFS from timing or magnitude perspectives (5, 6, 13). However, these studies either
lack global representativeness or fail to consider
the impact of ACC explicitly. Nonetheless, there

sults were also observed in south Siberia (S. SI),
with ~32% of stations showing significantly decreasing seasonality and only ~1% of stations
showing significantly increasing seasonality.
We further found a comparable pattern in
Europe (EU), with ~19% of the stations experiencing significantly decreasing RFS, mainly
located in northern Europe (N. EU), western
Russia (W. RU), and the European Alps, whereas ~4% showed significantly increasing RFS,
mainly concentrated downstream of the Alps
(Fig. 1C). In addition, regions in the contiguous
United States (CONUS) present predominantly
decreasing trends of RFS overall, except for
rivers in the Rocky Mountains and Florida in
the western and eastern CONUS, respectively.
In central North America (C. NA), significantly
decreasing RFS trends account for ~18% of
stations, in contrast with 4% that show significantly increasing RFS trends (Fig. 1D). W. RU,
upper Midwest (U. Midwest), and S. SI display
greater changes in magnitude for RFS trends.
Although more than half of the stations in the
northern high latitudes showed no statistically
significant (P > 0.05) trends, both the significant and nonsignificant trends tell the same
story, that is, RFS decreased in the northern
high latitudes. Contrary to the above results,
increasing RFS was observed in ~25% of the
stations in southeast Brazil (S. BR) versus
~4% with decreasing RFS. Global patterns in
seasonality trends during the more recent period of 1970 to 2019 generally agree with the
trends computed for the 1965 to 2014 time frame
(fig. S2). However, it should be noted that some
spatial resolution was lost when we used the
1970 to 2019 time window, for example, in
W. RU, because data are not available for a
sufficient length of time.

RES EARCH | R E S E A R C H A R T I C L E

p
g

To understand the change of RFS trends regionally, we focused on annual mean and monthly river flow trends and normalized flow
regimes in nine hotspots (fig. S5). Increasing
trends were most pronounced in low-flow
months, except in S. BR, which was in agreement with a large proportion of L+H− and
L+H* stations (Fig. 2). This suggests that the
upper limit of the environmental flow envelopes is increasingly being exceeded during
low-flow months in high latitudes (26). To interpret the potential mechanism of RFS trends,
we chose subspaces of nine hotspots for a finer
analysis (fig. S6). In snowmelt-dominated regions, decreasing snow fraction corresponding
to snow-rain transition and snowpack depletion plays a more important role in shaping RFS
than precipitation. Warmer temperatures can
deplete snowpacks, contributing to greater
frequency of high-flow events and lower frequency of low-flow events prior to the normal
flood season, and hence reducing monthly differences in river flow (5, 23). Early spring greening, closely related to early spring snowmelt,
2 of 6

,

1 March 2024

Regional seasonality changes
and potential mechanisms

y

Wang et al., Science 383, 1009–1014 (2024)

flow variations (L*H+ and L*H−; ~9%), highlighting the key role that low-flow changes
played in the AE shifts. Trends of annual low
and high flows were analyzed separately (fig.
S4), supporting our findings that increasing
river flow in low-flow months is contributing
to weakening RFS in the snowmelt-dominated
areas.
Our results are consistent with those of other
studies that have explored seasonal trends at
regional scales. For example, earlier timing
and reduced flood magnitude have been observed in N. EU, W. RU, and the European Alps
(5, 6). Previous studies also support findings
of stations facing increasing RFS. For instance,
the frequency of low-flow events increased
in the low-flow season (23), corresponding to
the spreading of L−H+ and L−H* categories.
In the Rocky Mountains region (CONUS),
early snowmelt can reduce river flow in lowflow months owing to less extensive spring
snow cover (L−H*), which aligns with (11). Moreover, stations characterized by L−H+, such
as those in S. BR, suggest a high risk of hazards
from both drought and flood events, which
aligns with (24, 25).

y g

increasing annual mean trends (L+H*). The distribution of the two patterns agrees with
broad-scale climate trends in snowmelt-dominated regions (N. NA, N. EU, W. RU, S. SI,
higher-elevation European Alps, U. Midwest,
and northeast CONUS) (Fig. 2). Out of 1137 stations in the snowmelt-dominated areas (Fig. 2A,
gray regions), 979 were experiencing significant weakening of RFS. Around 30% of river
gauges displayed significantly increasing RFS
and prominent low-flow decreases, herein represented as L−H* (~12%) and L−H+ (~18%),
such as in Florida and the Rocky Mountains in
the CONUS, lower catchments of the European
Alps, and S. BR. A smaller number of sites experienced changes in high-flow months, including L*H+ (~2%) and L*H− (~8%), suggesting
that high flows alone play a minor role in influencing RFS trends.
Overall, most stations showed L+H− and
L−H+ (~65%), indicating that low and high
flows interact to affect seasonality and mask
annual mean trends of river flow (fig. S3B).
Moreover, the proportion of sites (L−H* and
L+H*; ~26%) where low-flow variations are
the predominant factor is double that of high-

(P < 0.05, hatched) and nonsignificant (P > 0.05, solid) trends, corresponding to
the direction of increasing (green) and decreasing (brown) RFS trends in five
boxed regions. (C and D) Subareas in (C) EU and (D) C. NA that were dominated
by the same AE change direction are delimited by dashed gray lines: N. EU,
W. RU, the high-elevation European Alps, Pacific Northwest, upper Midwest, and
northeast CONUS. Significance was estimated by the Mann-Kendal trend test.

y

Fig. 1. RFS trends represented by AE (% decade−1) over 50 years (1965 to
2014). (A) Map shows stations with significant RFS trends (P < 0.05); green represents
increasing RFS and decreasing AE trends, and brown represents decreasing RFS
and increasing AE trends. Stations with nonsignificant changes (P > 0.05) are
represented by smaller gray dots. The five boxes mark the regions of interest: N. NA,
EU, S. SI, C. NA, and S. BR. (B) Pie charts show the distribution of significant

RES EARCH | R E S E A R C H A R T I C L E

p

Fig. 2. Classification of
potential reasons for changes
in RFS. (A) Spatial distribution
of TLH indicated by change
directions of AE and annual mean
river flow (NS, not significant).
*, no predominant changes
of low or high flows. Labels show
illustrations of flow-regime
changes from period t1 to
t2, corresponding to six types of
TLH. Regions where snow
fraction in precipitation was >0.2
are shown in gray as snowmeltdominated areas. The five
boxes mark the regions of interest:
N. NA, EU, S. SI, C. NA, and
S. BR. (B) Pie charts show the
proportion of TLH in the five boxed
regions in (A). (C and D) Subareas
in (C) EU and (D) C. NA with
the same AE change direction
are delimited by dashed gray lines:
N. EU, W. RU, the high-elevation
European Alps, Pacific Northwest, upper Midwest, and northeast CONUS.

g

Climate change detection and attribution
analysis of RFS trends

3 of 6

,

1 March 2024

y

To investigate whether ACC has caused the
consistent decreasing trends of RFS in northern high latitudes, we augmented the assessment of in situ observations with an analysis
of the GRUN to obtain a comprehensive spatial and temporal representation of RFS trends
(19). The reconstructed spatial trend patterns
of AE were compared with the corresponding trend pattern estimated by a multimodel
ensemble mean of 27 simulations from global hydrological models (GHMs) (20). These
GHMs considering human water and land
use (HWLU) are driven by atmospheric data
from climate models that account for historical radiative forcing (HIST) (scenario abbreviated as HIST&HWLU) (20). The simulated
trends were consistent with the reconstruction, highlighting that the GHM simulations
generally capture the observed changes (Fig. 3,
A, B, and D). Simulations from GHMs also
showed a general agreement above 50°N that
supports the spatial pattern of RFS changes
(fig. S7). Some differences are expected between the multimodel mean and observations.

For example, the magnitude of AE trends was
weaker in the multimodel mean. This weaker
magnitude is most likely due to the averaging
across the ensemble that reduces the effects
of internal variability in the climate forcing,
whereas the GRUN reconstruction represents
a single observed evolution of the system (32).
GRUN does not account for the effects of HWLU,
which possibly caused some differences in the
magnitude of AE trends. In addition, the high
uncertainties of GRUN reconstruction and
multimodel simulations possibly contribute
to the disagreement in the Arctic region of
northern Canada. However, the simulated trends
in AE that are derived from GHMs that account
for HWLU and are driven with atmospheric
variables from preindustrial control climate
models simulations (Picontrol&HWLU) failed
to capture the observed changes (Fig. 3, C and
D), indicating that HWLU is not contributing
to the weakening pattern of RFS. Analyses
from 1970 to 2019 showed the same trends in
RFS, indicating that simulations are consistent
with observations only when ACC is considered (figs. S8 and S9).
To quantitatively assess the influence of ACC
on the observed spatial pattern and temporal
evolution of RFS across northern high latitudes,
correlation-based (17, 32, 33) and optimal fingerprinting methods (17, 33, 34) were used to test
against the null hypothesis that there is no
detectable pattern of AE trends in the observations resulting from ACC. The correlation
approach uses all available AE trends and

y g

Wang et al., Science 383, 1009–1014 (2024)

The reason for the difference between results
may be due to varied El Niño Southern Oscillation (ENSO) impacts in different study periods (31). Temperature anomalies in ENSO
phases strongly influence precipitation and
snow accumulation and in turn affect spring
and summer river flow.

y

can exacerbate soil moisture deficits in spring
and summer [high-flow months in snowmeltdominated areas; e.g., the Alps (fig. S6)] (27).
This can indirectly dampen river flow regimes
by reducing runoff generation in high-flow
months. However, decreasing RFS can also
coincide with increasing soil moisture in highflow months [e.g., W. RU (fig. S6)]. More precipitation falling as rain when air temperatures
are around freezing is associated with shallower snowpacks and likely increased infiltration resulting from less frozen upper soil layers
and therefore leads to a rise in soil moisture
and smaller floods in the spring flood season
(5). Soil moisture initially decreased before increasing in N. NA, northeast CONUS, and S. SI,
indicating a shift of primary driving factor (fig.
S6). Additionally, permafrost mass loss may
continue to generate runoff thereafter [e.g.,
S. SI (fig. S5E)] (28).
Precipitation plays a more important role
in RFS in non–snowmelt-dominated regions.
For instance, seasonality of precipitation and
river flow are positively correlated (table S1;
Spearman rank correlation coefficients, r = 0.65)
owing to the dominance of rain in the coast
of the Pacific Northwest (fig. S6) (23). Similarly,
increased RFS is associated with increased
precipitation seasonality in S. BR (table S1; r =
0.93) (fig. S6), in agreement with (29). We noticed that most stations in the Pacific Northwest show increasing monthly river flow in
the late spring and summer, contradicting the
findings of previous studies (fig. S5F) (4, 30).

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y

[corrtemporary(HIST, GRUN)] across 50°N to 90°N. Spearman correlations
between the multimodel mean from HIST&HWLU simulations and 216 chunks of
Picontrol simulations with 50-year segments are shown as an empirical
probability density function in gray. Vertical blue lines mark the 95 and 99%
cumulative probabilities of an assumed normal distribution for the correlations.
(Inset) The confidence interval of the scaling factor plot from optimal
fingerprinting method with an uncertainty range of 0.5 to 99.5%. A signal
was detected if the lower confidence bound was >0 (the solid line). The
amplitude of the mean response was consistent with the observations
if the confidence interval included 1 (the dashed line). The RCT passed (P > 0.1),
indicating the consistency between the regression residuals and the modelsimulated variability.

y g

Fig. 3. Comparison of AE trends from observation-based reconstructions
and global hydrological models for 1965 to 2014 (% decade−1) in the
northern high latitudes (above 50°N). (A) Reconstruction from GRUN.
(B and C) Simulated changes based on the multimodel mean that account
for HWLU under the effects of either HIST (B) or Picontrol (C). Areas with
annual precipitation <100 mm and Greenland are masked in gray. (D) Multimodel
(mdl) mean time series of annual AE anomalies for HIST&HWLU and
Picontrol&HWLU responses and GRUN observations averaged for the
northern high latitudes (above 50°N). The red spread is ensemble standard
deviation of HIST&HWLU, and thin gray lines are 27-model results of
Picontrol&HWLU. (E) Correlations of AE anomalies between simulations with and
without ACC [corrtemporary(Picontrol, HIST)] or observation-based reconstruction

,

anomalies of simulations under HIST&HWLU
and Picontrol (table S2). The spatial correlations
between HIST simulations and a large ensemble of the Picontrol simulations were calculated from the multimodel simulation mean
of AE trends forced with HIST and every chunk
(216 estimates) of AE trends with Picontrol conditions (fig. S10). There was a 95% probability
that the spatial Spearman correlations between
multimodel mean of HIST simulations and
GRUN [corrspatial(HIST, GRUN)] were greater
than what was expected from Picontrol simulations [corrspatial(Picontrol, HIST]). This strongly
suggests that ACC is the underlying cause for
the spatial pattern of RFS. In addition, the
temporal correlation of AE anomalies between
Wang et al., Science 383, 1009–1014 (2024)

multimodel mean of HIST simulations and
GRUN [corrtemporary(HIST, GRUN)] was larger
than for all correlations from Picontrol versus
HIST [corrtemporary(Picontrol, HIST)] at 99%
confidence level (Fig. 3E). These findings further confirm that human-induced emissions
contribute to the decrease in RFS in the northern high latitudes.
An additional test used regularized optimal fingerprinting (34). Observations (y) are
regressed on the multimodel mean of the simulation forced with HIST&HWLU (x) while
considering sampling error v and natural variability derived from Picontrol&1860soc (e):
y = b (x–v) + e. The inset in Fig. 3E displays
the scaling factor b above zero at 99% con-

1 March 2024

fidence level in the experiment of HIST, suggesting that the simulations capture the observed
AE changes when human-influenced climate
changes are considered. Furthermore, the confidence interval of scaling factor b under HIST
includes one, which implies that the magnitude of the simulated trend is consistent with
the observations. Moreover, a residual consistency test (RCT) did not indicate an inconsistency between the regression residuals and
the model-simulated variability (P > 0.1). An
overall climate change detection and attribution analysis for 1970 to 2019 provided
further evidence that human-induced emissions continue to contribute to decreased
RFS in the northern high latitudes (fig. S11).
4 of 6

RES EARCH | R E S E A R C H A R T I C L E

5 of 6

,

Funding: This study was supported by the National Natural
Science Foundation of China (grant no. 42361144001), the
Strategic Priority Research Program of the Chinese Academy of
Sciences (grant no. XDA20060402), the Shenzhen Science and
Technology Program (KCXFZ20201221173601003), the Henan
Provincial Key Laboratory of Hydrosphere and Watershed Water
Security, and the School of Geography and water@leeds at
the University of Leeds. Author contributions: J.L., H.W., M.K.,
and J.H. conceived and designed the study. H.W. conducted the
analyses and drafted the paper under the supervision of J.L.,
M.K., and J.H. All authors contributed to data interpretation and
provided substantive revisions on the manuscript. Competing
interests: The authors declare no competing interests. Data
availability: The river flow time series from GSIM can be
downloaded from https://doi.org/10.1594/PANGAEA.887470
(46). The runoff reconstruction dataset GRUN is available from

y

1 March 2024

AC KNOWLED GME NTS

y g

1. P. Wu, N. Christidis, P. Stott, Nat. Clim. Chang. 3, 807–810
(2013).
2. C. J. Vörösmarty, P. Green, J. Salisbury, R. B. Lammers,
Science 289, 284–288 (2000).
3. P. C. D. Milly, K. A. Dunne, A. V. Vecchia, Nature 438, 347–350
(2005).
4. N. K. Singh, N. B. Basu, Nat. Sustain. 5, 397–405 (2022).
5. G. Blöschl et al., Nature 573, 108–111 (2019).
6. G. Blöschl et al., Science 357, 588–590 (2017).
7. M. Dynesius, C. Nilsson, Science 266, 753–762 (1994).
8. M. Palmer, A. Ruhi, Science 365, eaaw2087 (2019).
9. T. P. Barnett, J. C. Adam, D. P. Lettenmaier, Nature 438,
303–309 (2005).
10. J. D. Tonkin, D. M. Merritt, J. D. Olden, L. V. Reynolds,
D. A. Lytle, Nat. Ecol. Evol. 2, 86–93 (2018).
11. S. B. Rood et al., J. Hydrol. (Amst.) 349, 397–410 (2008).
12. H. L. Bateman, D. M. Merritt, Glob. Ecol. Conserv. 22, e00957
(2020).
13. C. Wasko, R. Nathan, M. C. Peel, Water Resour. Res. 56,
e2020WR027233 (2020).
14. S. Eisner et al., Clim. Change 141, 401–417 (2017).
15. K. Marvel et al., Earths Futur. 9, e2021EF002019 (2021).
16. T. P. Barnett et al., Science 319, 1080–1083 (2008).
17. L. Gudmundsson, S. I. Seneviratne, X. Zhang, Nat. Clim. Chang.
7, 813–816 (2017).
18. H. X. Do, L. Gudmundsson, M. Leonard, S. Westra, Earth Syst.
Sci. Data 10, 765–785 (2018).
19. G. Ghiggi, V. Humphrey, S. I. Seneviratne, L. Gudmundsson,
Earth Syst. Sci. Data 11, 1655–1674 (2019).
20. K. Frieler et al., Geosci. Model Dev. 10, 4321–4345
(2017).
21. X. Feng, A. Porporato, I. Rodriguez-Iturbe, Nat. Clim. Chang. 3,
811–815 (2013).

y

RE FERENCES AND NOTES

22. G. Konapala, A. K. Mishra, Y. Wada, M. E. Mann, Nat. Commun.
11, 3044 (2020).
23. E. N. Dethier, S. L. Sartain, C. E. Renshaw, F. J. Magilligan,
Sci. Adv. 6, eaba5939 (2020).
24. M. R. Viola, C. R. de Mello, S. C. Chou, S. N. Yanagi,
J. L. Gomes, Int. J. Climatol. 35, 1054–1068 (2015).
25. D. Bartiko, D. Y. Oliveira, N. B. Bonumá, P. L. B. Chaffe,
Hydrol. Sci. J. 64, 1071–1079 (2019).
26. V. Virkki et al., Hydrol. Earth Syst. Sci. 26, 3315–3336
(2022).
27. X. Lian et al., Sci. Adv. 6, eaax0255 (2020).
28. E. Gautier et al., J. Hydrol. (Amst.) 557, 475–488
(2018).
29. V. B. P. Chagas, P. L. B. Chaffe, G. Blöschl, Geophys. Res. Lett.
49, e2021GL096754 (2022).
30. P. W. Mote, S. Li, D. P. Lettenmaier, M. Xiao, R. Engel,
NPJ Clim. Atmos. Sci. 1, 2 (2018).
31. D. Lee, P. J. Ward, P. Block, Environ. Res. Lett. 13, 044031
(2018).
32. R. S. Padrón et al., Nat. Geosci. 13, 477–481
(2020).
33. L. Grant et al., Nat. Geosci. 14, 849–854 (2021).
34. A. Ribes, S. Planton, L. Terray, Clim. Dyn. 41, 2817–2836
(2013).
35. B. Arheimer, C. Donnelly, G. Lindström, Nat. Commun. 8, 62
(2017).
36. L. Gudmundsson et al., Science 371, 1159–1162 (2021).
37. IPCC, in Climate Change 2013: The Physical Science Basis
(Cambridge Univ. Press, 2013), pp. 659–740.
38. C. E. Iles, G. C. Hegerl, Nat. Geosci. 8, 838–842 (2015).
39. M. Huss, R. Hock, Nat. Clim. Chang. 8, 135–140
(2018).
40. L. Gudmundsson, J. Kirchner, A. Gädeke, J. Noetzli,
B. K. Biskaborn, Environ. Res. Lett. 17, 095014 (2022).
41. W. R. Berghuijs, R. A. Woods, M. Hrachowitz, Nat. Clim. Chang.
4, 583–586 (2014).
42. M. G. Floriancic, W. R. Berghuijs, P. Molnar, J. W. Kirchner,
Water Resour. Res. 57, e2019WR026928 (2021).
43. M. Mudelsee, M. Börngen, G. Tetzlaff, U. Grünewald, Nature
425, 166–169 (2003).
44. J. D. Hunt et al., Nat. Commun. 11, 947 (2020).
45. N. L. Poff, J. D. Olden, D. M. Merritt, D. M. Pepin, Proc. Natl.
Acad. Sci. U.S.A. 104, 5732–5737 (2007).
46. L. Gudmundsson, H. X. Do, M. Leonard, S. Westra, The Global
Streamflow Indices and Metadata Archive (GSIM) - Part 2:
Time Series Indices and Homogeneity Assessment, PANGAEA
(2018); https://doi.org/10.1594/PANGAEA.887470.
47. G. Ghiggi, L. Gudmundsson, V. Humphrey, G-RUN: Global
Runoff Reconstruction, Figshare, version 2, (2019);
https://doi.org/10.6084/m9.figshare.9228176.v2.
48. U. Schneider, S. Hänsel, P. Finger, E. Rustemeier, M. Ziese,
GPCC Full Data Monthly Product Version, 2022 at 2.5°:
Monthly Land-Surface Precipitation from Rain-Gauges
built on GTS-based and Historical Data, Deutscher
Wetterdienst, (2022); https://doi.org/10.5676/DWD_GPCC/
FD_M_V2022_250.
49. M. Cucchi et al., Near surface meteorological variables
from 1979 to 2019 derived from bias-corrected reanalysis,
Copernicus Climate Change Service (C3S) Climate
Data Store (CDS), version 2.1, (2020); https://doi.org/
10.24381/cds.20d54e34.
50. J. Muñoz Sabater, ERA5-Land hourly data from 1950
to present, Copernicus Climate Change Service (C3S)
Climate Data Store (CDS), (2019); https://doi.org/10.24381/
cds.68d2bb30.

g

Wang et al., Science 383, 1009–1014 (2024)

ACC is acknowledged to far surpass natural
forcing in dominating a warming future (37).
Our findings present changes in the seasonal
cycles of river flow by adapting an AE perspective and clearly demonstrate that decreased
RFS is attributable to ACC in the northern
high latitudes. Possible climatic mechanisms
that might drive flow-regime dampening under
ACC include early snowpack depletion (23),
loss of glacier extent (39), permafrost loss
(40), increasing proportion of precipitation
as rainfall (41), and shorter freezing periods
(42, 43) interacting with ocean-atmosphere
oscillations (31). Depending on the region, some
of these drivers can be more important than
others in explaining RFS changes.
This study provides a standpoint for understanding changing seasonal patterns of river
flow. There is an increasing need for accelerated climate adaptation efforts to safeguard
freshwater ecosystems, achieved through, for
example, use of managed environmental flows
(8). Additionally, these efforts are essential for
establishing sustainable water resource management by identifying and mitigating risks
related to flood and drought, exploring seasonal
storage opportunities, and optimizing allocations for irrigation or hydropower generation
(4, 44). It should be noted that water management might synergistically contribute to RFS
dampening (35, 45). Therefore, it is essential
to develop mitigation strategies and adaptation planning to alleviate the future homogenization of seasonal river flow, particularly in
locations such as European Russia, Scandinavia,
and Canada.

p

The primary climate change detection and
attribution assessment that focuses on the
northern high latitudes to optimize the signalto-noise ratio is complemented with regional
assessment (fig. S12). Changes of RFS were
captured with 10 to 90% confidence intervals
in Alaska, northern Europe, and northern Asia,
defined by the Intergovernmental Panel on Climate Change Special Report on Extreme Events,
only if ACC is considered. These results confirm
the robustness of our conclusions regarding the
influence of ACC on the temporal evolution of
RFS in the northern high latitudes. Seasonality
changes were also detected by model simulations that account for anthropogenic emissions
in central America, southern Africa, and east
Asia. This finding implies that human-induced
emissions potentially exert an influence on the
seasonality of monsoon precipitation and consequent runoff dynamics.
We acknowledge that human interference,
such as flow regulation through reservoirs, may
also contribute to RFS changes (35). Notably,
however, more than three-fifths of the in situ
observations, which are free from reservoir flow
regulation (located in the subbasins with zero
degree of regulation), exhibited the same spatial pattern of RFS trends as identified in our
global dataset (fig. S13). Moreover, an observational reconstruction runoff derived from
GRUN, which is free from human interference
(including reservoirs, human water management, and land-use change), demonstrated a
similar trend to that observed at the stations,
though with smaller magnitudes of RFS trends
(fig. S14). Additionally, simulations replicating
preindustrial climate conditions but considering
historical human activities (Picontrol&HWLU)
failed to reproduce the trend of RFS in the
northern high latitudes. Combining the climate change detection and attribution analysis for grid cells where direct observation
data are available robustly showed that ACC
contributes to the weakening of RFS in the
northern high latitudes (fig. S15). We note that
historical natural climate forcing (i.e., solar
and volcanic activity) was not excluded when
using ISIMIP2b to undertake the climate change
detection and attribution analysis (36). Nonetheless, natural climate forcing has a limited
impact on river flow owing to much smaller
solar changes compared with ACC (37) and
the short-lived influence of volcanic eruptions
(38). Furthermore, no significant trends of precipitation seasonality have been observed in
the northern high latitudes, demonstrating
that precipitation seasonality change cannot
account for our results (fig. S16). It is likely that
observed rain-snow transition and increasing snowmelt under global warming led to a
weakening trend of RFS in the northern high
latitudes (fig. S6 and table S1). The underlying
physics behind this assertion is temperature
driven rather than precipitation driven, and

RES EARCH | R E S E A R C H A R T I C L E

https://doi.org/10.6084/m9.figshare.9228176.v2 (47). The model
results are available from the ISIMIP2b project in Water (global)
sector (https://www.isimip.org/outputdata/). Other data are
presented in the supplementary materials. The streamflow time
series from the GRDC are available at https://www.bafg.de/GRDC.
The observed monthly GPCC precipitation is available at https://
doi.org/10.5676/DWD_GPCC/FD_M_V2022_250 (48). The
CRUTEM5 dataset is available at https://www.metoffice.gov.uk/
hadobs/crutem5/data/CRUTEM.5.0.1.0/download.html. The CPC

soil moisture data can be downloaded from https://www.esrl.noaa.
gov/psd. The bias-corrected reanalysis of WFDE5 and ERA5
monthly land data can be downloaded from the CDS website [WFDE5,
https://doi.org/10.24381/cds.20d54e34 (49); ERA5, https://doi.
org/10.24381/cds.68d2bb30 (50)]. License information: Copyright
© 2024 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adi9501
Materials and Methods
Figs. S1 to S18
Tables S1 to S3
References (51–68)
Submitted 28 May 2023; accepted 17 January 2024
10.1126/science.adi9501

p
g
y
y g
y
,

Wang et al., Science 383, 1009–1014 (2024)

1 March 2024

6 of 6

RES EARCH

FRAMEWORK MATERIALS

Fast growth of single-crystal covalent organic
frameworks for laboratory x-ray diffraction
Jing Han1, Jie Feng1, Jia Kang1, Jie-Min Chen1, Xin-Yu Du1, San-Yuan Ding1, Lin Liang1,2*, Wei Wang1*
The imine-exchange strategy makes single-crystal growth of covalent organic frameworks (COFs) with
large size (>15 microns) possible but is a time-consuming process (15 to 80 days) that has had limited
success (six examples) and restricts structural characterization to synchrotron-radiation sources for
x-ray diffraction studies. We developed a CF3COOH/CF3CH2NH2 protocol to harvest single-crystal COFs
within 1 to 2 days with crystal sizes of up to 150 microns. The generality was exemplified by the feasible
growth of 16 high-quality single-crystal COFs that were structurally determined by laboratory singlecrystal x-ray diffraction with resolutions of up to 0.79 angstroms. The structures obtained included
uncommon interpenetration of networks, and the details of the structural evolution of conformational
isomers and host-guest interaction could be determined at the atomic level.

1 of 6

,

1 March 2024

y

Han et al., Science 383, 1014–1019 (2024)

y g

*Corresponding author. Email: wang_wei@lzu.edu.cn (W.W.);
lianglin@lzu.edu.cn (L.L.)

Initially, we replaced CH3COOH, which has
the negative logarithm of the acid dissociation
constant (pKa) of 4.76 (23), with a stronger acid,
CF3COOH, which has a pKa of 0.23 (23). This
change accelerated the imine-exchange process (24–26) for the fast growth of single-crystal
COF-300 from the condensation of benzene1,4-dicarboxaldehyde (BDA, 12 mg, 0.089 mmol)
and tetrakis(4-aminophenyl)methane (TAM,
20 mg, 0.052 mmol) in 1,4-dioxane (Fig. 2A) (4).
When we used CF3COOH (6 M, 0.1 ml) as the
catalyst and C6H5NH2 (81 ml, 10 equiv.) as the
modulator, COF-300 was rapidly crystallized as
uniform rodlike crystals with the average size of
10 mm in 2 hours (Fig. 2D and scheme S1). However, the crystal size could not be further increased
by prolonging the reaction time because C6H5NH2
(pKa C6H5NH3+ = 4.62) (23) was not a suitable
nucleation inhibitor when CH3COOH was replaced
by CF3COOH as a more acidic catalyst (table S1).
Accordingly, we screened a series of organic
bases as the compatible modulator (table S2)
and optimized the concentration ratios of the
acid and the modulator (table S3). We found
that, in the presence of CF3COOH (6 M, 0.1 ml)
as the catalyst and CF3CH2NH2 (pKa CF3CH2NH3+ =
5.66, 70 ml, 10 equiv.) (23) as the modulator,
single-crystal COF-300 could be harvested within
2 days (Fig. 2A and scheme S2) with the uniform
size of 60 mm by 30 mm by 30 mm (Fig. 1B and
Fig. 2E). The growth rate of single-crystal COF-300
reached 1.25 mm/hour, which is 21 times as fast as
the rate of 0.06 mm/hour previously reported (Fig.
2C). Using a laboratory single-crystal x-ray diffractometer, we could detect the nascent COF-300
and the hydrated COF-300 (COF-300-H2O), and
the single-crystal structures could be directly
solved as sevenfold-interpenetrated dia-c7 topology (13, 27) and anisotropically refined with
resolutions of 0.83 and 0.81 Å (tables S4 and
S5), respectively.

y

State Key Laboratory of Applied Organic Chemistry,
Lanzhou Magnetic Resonance Center, College of Chemistry
and Chemical Engineering, Lanzhou University, Lanzhou,
Gansu 730000, China. 2Institute of Nanoscience and
Nanotechnology, School of Materials and Energy, Lanzhou
University, Lanzhou, Gansu 730000, China.

Fast synthesis of known COF single crystals

g

1

and dynamic nature of COFs at the atomic
level.

p

C

ovalent organic frameworks (COFs) are
extended porous crystals formed by the
reaction of organic precursors as building
blocks, which form two-dimensional (2D)
or 3D arrays (1–10). Under typical conditions, the reaction products are small crystallites
(powders). The growth of high-quality COF
single crystals (11–22) must avoid misassembly
of the building blocks. Specifically, the growth
of large-sized (>15-mm) single-crystal COFs
amenable for x-ray diffraction (XRD) analysis
usually requires slow crystallization (at least
15 days) (13). In our previous studies, to construct imine-linked single-crystal COFs from
covalent polymerization of amines and aldehydes
(13, 16), we employed acetic acid (CH3COOH)
as the catalyst and aniline (C6H5NH2) as the
modulator. The use of aniline has efficiently
converted COF crystallization from imine formation to imine-exchange reactions (Fig. 1A).
This approach yielded single-crystal COFs suitable for XRD studies with sizes of 15 to 100 mm
but required growth times of 15 to 80 days.
In this study, we report the fast synthesis of
large-sized single-crystal COFs. In the presence
of 2,2,2-trifluoroacetic acid (CF3COOH) as the catalyst and 2,2,2- trifluoroethylamine (CF3CH2NH2)
as the modulator, 16 different COFs with
crystal sizes ranging from 50 to 150 mm were
synthesized in 1 to 2 days (Fig. 1B). The quality
of these single crystals was enough for their
single-crystal structures to be directly determined by laboratory XRD with resolutions up
to 0.79 Å. These high-resolution XRD data
revealed the indeterminate topology, conformational evolution, host-guest interaction,

Using the CF3COOH/CF3CH2NH2 protocol,
we successfully synthesized the previously reported single-crystal COFs (13, 16)—LZU-111,
LZU-79, COF-303, and LZU-306—as high-quality
single crystals within 2 days (Fig. 1B and schemes
S3 to S6). The sizes of LZU-111 (~50 mm) and
LZU-79 (~100 mm) obtained were comparable
to those achieved previously but required 25
to 40 days for synthesis. The sizes of COF-303
(~100 mm) and LZU-306 (~150 mm) were larger
than those previously reported (~15 mm in
15 days for COF-303 and ~50 mm in 25 days
for LZU-306).
Taking the noninterpenetrated pts-structured
LZU-306 (Fig. 2B) as the example, similar to the
case of COF-300, the crystallization of LZU-306
occurred rapidly with the CF3COOH/C6H5NH2
protocol (Fig. 2F) but resulted in irregular
crystals of poor quality. Using the CF3COOH/
CF3CH2NH2 protocol, after 4 hours, crystallization led to the appearance of uniform microcrystals with the size of ~10 mm that could be
observed with optical microscopy. After 12 hours,
the crystal size reached ~30 mm (Fig. 2F and
fig. S85). After 36 hours, large single crystals
(150 mm by 100 mm by 100 mm) had grown
(Fig. 2F and fig. S87). The growth rate reached
4.17 mm/hour, which is 52 times as fast as that
previously reported for the CH3COOH/C6H5NH2
protocol (0.08 mm/hour) (Fig. 2C). We obtained XRD data with a resolution of 1.15 Å
with the laboratory light source. The noninterpenetrated single-crystal structure of LZU-306
was directly solved, and all of the nonhydrogen atoms could be anisotropically refined. In
the previous work, the resolution for the XRD
data reached 1.80 Å (16) with a synchrotronradiation light source, and the direct determination of the single-crystal structures was
unattainable.
We further verified the generality of this
CF3COOH/CF3CH2NH2 protocol by rapidly growing COF structures as high-quality single crystals.
The increase in the growth rate enabled us to
optimize the experimental conditions efficiently.
As a result, 10 different single-crystal COFs were
harvested by simple screening of the suitable
solvents and the equivalent of CF3CH2NH2
(Fig. 1B, scheme S5, and schemes S8 to S12).
The single crystals reached sizes of 60 to 150 mm
in 1 to 2 days. Structures from laboratory XRD
were directly solved and refined with resolution
of up to 0.79 Å (tables S8 to S19). Among
these structures, we found an uncommon 3D
framework with the complicated fourfold
[2+2]–interpenetrated pts structure (Fig. 3).
We also followed the structural evolution among
a series of conformational COF isomers that
directly correlated with the subtle changes in
the local conformation of the linkages (Fig. 4
and fig. S31). Lastly, we accurately located
guest molecules within the pores and further
evaluated host-guest interactions in COFs (Fig. 5
and fig. S31).

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 1. Fast growth of large-sized single-crystal COFs. (A) Imine-exchange
strategy that used CH3COOH/C6H5NH2 in the previous work for the growth of
single-crystal COFs in 15 to 80 days. (B) Protocol developed using CF3COOH/
CF3CH2NH2 in this work for fast growth of single-crystal COFs in 1 to 2 days.
The optical microscopic images for 16 kinds of single-crystal COFs obtained in
1 to 2 days with sizes of 50 to 150 mm are shown. Diversified monomers used
in this study for the growth of single-crystal COFs are as follows: TAM,

Unknown COF structures revealed by
CF3COOH/CF3CH2NH2 protocol

The CF3COOH/CF3CH2NH2 protocol revealed
three previously unknown single-crystal COFs
with the isoreticular pts topology that were
Han et al., Science 383, 1014–1019 (2024)

1 March 2024

tetrakis(4-aminophenyl)methane; ADA-CHO, adamantane-1,3,5,7tetracarbaldehyde; ADAT-CHO, 1,3,5,7-tetrakis(4-formylphenyl)adamantane;
TFM, tetrakis(4-formylphenyl)methane; TFS, tetrakis(4-formylphenyl)silane;
BDA, benzene-1,4-dicarboxaldehyde; DABP, 4,4′-diaminobiphenyl; PDA,
phenylenediamine; BFBZ, 4,7-bis(4-formylbenzyl)-1H-benzimidazole;
TPE-NH2, tetrakis(4-aminophenyl)ethene; TPB-NH2, 1,2,4,5-tetrakis(4-aminophenyl)
benzene; and TPE-CHO, tetrakis(4-formylphenyl)ethene.

synthesized in 1 day (Fig. 1B and Fig. 3A).
LZU-308, constructed from adamantane1,3,5,7-tetracarbaldehyde (ADA-CHO) and
1,2,4,5-tetrakis(4-aminophenyl)benzene (TPBNH2), was crystallized with the size of ~60 mm

in 1 day (scheme S8). LZU-309, formed by
TAM and tetrakis(4-formylphenyl)ethylene
(TPE-CHO), was crystallized with the size
reaching ~80 mm in 1 day (scheme S9). LZU-307,
produced by 1,3,5,7-tetrakis(4-formylphenyl)
2 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g
~10 mm–sized COF-300 obtained with the CF3COOH/C6H5NH2 protocol. (E) The
SEM image of ~60 mm–sized COF-300 obtained with the CF3COOH/CF3CH2NH2
protocol. (F) Average sizes of single-crystal LZU-306 along with the reaction time
obtained by using the CF3COOH/CF3CH2NH2 (red dots, figs. S85 to S87), CH3COOH/
C6H5NH2 (black filled dots), and CF3COOH/C6H5NH2 (black empty dots, fig. S84) protocols.

y

Fig. 2. Fast growth of single-crystal COF-300 and LZU-306. (A and B) Fast
growth of single-crystal COF-300 in 2 days and LZU-306 in 1.5 days with the
CF3COOH/CF3CH2NH2 protocol. (C) Comparison of the data for the crystallization
time, crystal size, growth rate, and resolution of XRD for COF-300 and LZU-306
reported in the previous work (13, 16) and in this work. (D) The SEM image of

y g
y
,
Fig. 3. Fast growth of pts-structured single-crystal LZU-308, LZU-309, and LZU-307 with the CF3COOH/CF3CH2NH2 protocol. (A) Growth of single crystals
of noninterpenetrated LZU-308, twofold-interpenetrated LZU-309, and fourfold [2+2]–interpenetrated LZU-307 in 1 day. (B) Crystal structures and topological
structures of LZU-307 viewed along the c axis and b axis.
Han et al., Science 383, 1014–1019 (2024)

1 March 2024

3 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,
Fig. 4. Synthesis and structural analysis of conformational isomers of single-crystal COFs. (A) Fast growth of single-crystal LZU-311, COF-303, and LZU-310
within 2 days. (B) The space groups, unit-cell parameters, unit-cell volumes, and linkage conformations of the single-crystal isomers. (C) Single-crystal structures
and skeleton geometries of COF-303, COF-303-p, COF-303-a, and COF-303-BnOH. The bottom illustrations show the angle of the tetrahedral node and the length of
the linker in each structure.

adamantane (ADAT-CHO) and tetrakis(4aminophenyl)ethene (TPE-NH2), was crystallized with the size of ~80 mm in 1 day (scheme
S10). We used tetrahydrofuran as the universal
solvent for the growth of single-crystal COFs
with high quality. Laboratory XRD analysis
directly identified the structures of LZU-308,
Han et al., Science 383, 1014–1019 (2024)

1 March 2024

LZU-309, and LZU-307 as non-, twofold-, and
fourfold-interpenetrated pts frameworks (Fig.
3A and figs. S10, S12, and S14), respectively. These
results served as the experimental evidence
that the degree of interpenetration in COFs
could be progressively increased with the elongation of the linkers (28). Crystallized with a

rhombohedral morphology, LZU-307 had an
uncommon fourfold [2+2]–interpenetrated
structure. The space group of LZU-307 was determined as Cmma, with unit-cell parameters
of a = 21.947(3) Å, b = 33.671(6) Å, and c =
23.432(3) Å (numbers in parenthesis are the
error in the last digit) and a large unit-cell volume
4 of 6

RES EARCH | R E S E A R C H A R T I C L E

Fig. 5. Host-guest structure of single-crystal COF-303-BnOH. (A) Singlecrystal structure of COF-303-BnOH, viewed from the c axis. (B) Arrangement of
BnOH molecules in the COF-303 channels. The dashed lines in red and blue
represent the O–H‧‧‧O and C–H‧‧‧p distances between the adjacent BnOH

5 of 6

,

It has been acknowledged that the nucleation
barrier for crystallization could be reduced by
adding catalysts into the system (32, 33). The
formation of single-crystal COFs in our work
was based on covalent polymerization through
imine-exchange reactions that can be effectively
catalyzed by acids (25). When CH3COOH was
replaced with the stronger acid, CF3COOH,
the growth rates of single-crystal COFs were
significantly enhanced (83 and 57 times for
COF-300 and LZU-306, respectively, table S1).
Its synergy with CF3CH2NH2 as the compatible
modulator ensured the universal harvest of
large-sized (50- to 150-mm) single crystals of
3D COFs with high quality in 1 to 2 days. We
further found that the CF3COOH/CF3CH2NH2
protocol also enabled the growth of a 2D
single-crystal COF (34), LZU-115, reaching
a size of ~10 mm within 2 days (scheme S13 and
figs. S124 and S125). Accordingly, the fast
growth of single-crystal COFs for laboratory
XRD analysis would probably renew the research paradigm for precise assembly across
the length scale through covalent bonding.
This finding challenges the traditional belief

y

1 March 2024

Discussion

y g

Han et al., Science 383, 1014–1019 (2024)

y

The high-resolution XRD data provided key
information on the linkage conformation and
guest location at the atomic level, through which
the structural evolution and dynamic nature
of COFs have been clarified. For example, the
pores of COF-303 were fully occupied with
1,4-dioxane. The structure was identified as the
space group of I41/a with unit-cell parameters
of a = b = 25.8651(15) Å and c = 7.7001(6) Å
(Fig. 4B and table S12). The adjacent –C=N–
and –C=N– linkages exhibited a trans conformation. Upon evaporation at 300 K for 10 min,
1,4-dioxane as the guest molecule was partially
removed, resulting in the formation of COF-303-p.
The crystal structure changed to the space group
of I42d with unit-cell parameters of a = b =

host-guest interaction that has been demonstrated in Fig. 5.
The conformational transformation triggered
by guest molecules was also observed for the
ninefold interpenetrated single-crystal COF,
LZU-310 (Fig. 4B, fig. S31, and tables S17 to
S19). Unlike the case of BnOH as the guest
molecule, a dynamic contraction occurred upon
the aggregation of water guests within the LZU310 channels. This contraction was caused by
the stronger interaction between water molecules
and nitrogen atoms of the LZU-310 framework,
as visualized by the shorter distance of 1.99 Å
(fig. S31, D and E). Accordingly, the XRD information, with high accuracy, not only rendered
single-crystal COFs as candidates for crystalline
sponge (30, 31) but also provided in-depth
understanding of the structural adaptability
and responsiveness of dynamic COFs.

g

Structural transformations and
host-guest interactions

23.6776(10) Å and c = 7.9219(4) Å (Fig. 4B and
table S13). In this case, the adjacent –C=N– and –
C=N– linkages changed to a semi-cis conformation.
Upon the complete removal of 1,4-dioxane,
the activated COF-303 (COF-303-a) underwent
an extensive structural transformation in
which adjacent linkages were converted to the
cis form [a = b = 20.177(5) Å and c = 8.783(2) Å]
(Fig. 4B and table S14). Analysis on the skeleton geometries indicated that, as the angles
of the tetrahedral nodes were decreased from
87.6° (COF-303) to 81.0° (COF-303-p) and 66.6°
(COF-303-a), the unit-cell volumes decreased
from 5151.4(7) to 4441.2(4) and 3576(2) Å3 (Fig. 4,
B and C), whereas the lengths of the organic
linkers remained almost unchanged (from 18.68
to 18.23 and 18.39 Å, Fig. 4C). Further experiments indicated that the structural transformation among these conformational isomers
was reversible (table S16 and fig. S46). Thus, the
emergence effect (29) exemplified here showed
that changes in the global frameworks were
governed by subtle but oriented alternation
on the conformation of imine linkages.
The laboratory XRD data had sufficient resolution to accurately locate guest molecules within
the COF frameworks. For example, COF-303 with
BnOH as bulky guests (named COF-303-BnOH)
reached an XRD resolution of 0.79 Å that
enabled the explicit determination of all of the
nonhydrogen atoms in the host-guest structure
(table S15). BnOH molecules were arranged
into four columns with an interlaced manner
through hydrogen bonding (with the O–H‧‧‧O
distance of 1.88 Å, red line) and C–H‧‧‧p interactions (with the C–H‧‧‧p distance of 3.04 Å,
blue line) (Fig. 5, A and B). In addition, the
T-shaped p interaction in the host-guest structure was identified in four types with the C–H‧‧‧p
distances of 2.96, 3.25, 3.27, and 3.47 Å, respectively (Fig. 5C). Compared with the COF-303-a
structure, COF-303-BnOH was expanded with
a 50% increase in the unit-cell volume (Fig. 4,
B and C). Accordingly, this dynamic expansion
was induced by the aggregation of bulky BnOH
guests within the COF-303 channels through the

p

of 17316(5) Å3 (table S10). The interpenetration
pattern introduced the structural complexity
of LZU-307 and led to the low crystallographic
symmetry. Specifically, every two of the four
independent networks were interlocked with
each other along the three different axes
through the interpenetration vectors of [0,1/
2,1/3], [1/2,0,1/3], and [1/2,1/2,0] (Fig. 3B). The
translation vectors along the crystallographic
a axis (10.97 Å) and b axis (16.84 Å) exhibited a
common shift of 1/2, while the translation
vectors along the c-axis (7.81 Å) displayed a
distinct shift of 1/3.
The COFs that we synthesized (Fig. 4A)
exhibited excellent crystallinity. For example, the laboratory XRD data for LZU-311
(table S11), with a sixfold-interpenetrated dia
structure, reached a resolution of 0.84 Å. The
data for COF-303, COF-303-p, COF-303-a,
and COF-303-BnOH (tables S12 to S15) as
sevenfold-interpenetrated conformational isomers reached resolutions of 0.81, 0.79, 0.88,
and 0.79 Å, respectively (BnOH, benzyl alcohol); those for LZU-310, LZU-310-H2O, and
LZU-310-BnOH (tables S17 to S19) as ninefoldinterpenetrated conformational isomers reached
resolutions of 0.81, 0.79, and 0.84 Å, respectively.

molecules, respectively. (C) Local structure of COF-303-BnOH, highlighting the
C–H‧‧‧p distances between the COF-303 framework as the host and BnOH as the
guest molecule. C atoms of BnOH (light blue); C atoms of COF-303 skeleton
(gray); N atoms (blue); O atoms (red); H atoms (white).

RES EARCH | R E S E A R C H A R T I C L E

(35) that the growth of high-quality single crystals requires slow crystallization with the cost
of time consumption.
RE FE RENCES AND N OT ES

A. P. Côté et al., Science 310, 1166–1170 (2005).
H. M. El-Kaderi et al., Science 316, 268–272 (2007).
C. S. Diercks, O. M. Yaghi, Science 355, eaal1585 (2017).
F. J. Uribe-Romo et al., J. Am. Chem. Soc. 131, 4570–4571 (2009).
S. Y. Ding et al., J. Am. Chem. Soc. 133, 19816–19822 (2011).
S. Kandambeth et al., J. Am. Chem. Soc. 134, 19524–19527 (2012).
S.-Y. Ding, W. Wang, Chem. Soc. Rev. 42, 548–568 (2013).
E. Jin et al., Science 357, 673–676 (2017).
X. Wang et al., Nat. Chem. 10, 1180–1189 (2018).
W. Zhang et al., Nature 604, 72–79 (2022).
Y.-B. Zhang et al., J. Am. Chem. Soc. 135, 16336–16339 (2013).
D. Beaudoin, T. Maris, J. D. Wuest, Nat. Chem. 5, 830–834 (2013).
T. Ma et al., Science 361, 48–52 (2018).
J. A. R. Navarro, Science 361, 35–35 (2018).
A. M. Evans et al., Science 361, 52–57 (2018).
L. Liang et al., Angew. Chem. Int. Ed. 59, 17991–17995 (2020).
H.-S. Xu et al., Nat. Commun. 11, 1434 (2020).
L. Peng et al., Nat. Commun. 12, 5077 (2021).
C. Kang et al., Nat. Commun. 13, 1370 (2022).
A. Natraj et al., J. Am. Chem. Soc. 144, 19813–19824 (2022).
S. Wang et al., J. Am. Chem. Soc. 145, 12155–12163 (2023).
Z. Zhou et al., Nat. Chem. 15, 841–847 (2023).
J.-D. Yang, X.-S. Xue, P. Ji, X. Li, J.-P. Cheng, Internet
Bond-energy Databank (pKa and BDE): iBonD Home Page
(2024); http://ibond.nankai.edu.cn.

ACKN OWLED GMEN TS

J.H., L.L., and W.W. thank Y.-L. Shao, H. Wang, and G.-H. Xi for
assisting with the XRD data collection and S. Chen and S. Guo for
assisting with CO2 adsorption-desorption experiments. Insightful
discussions with T. Ma and W. Yu, Y. Li, O. M. Yaghi, and J.-L. Sun
are much appreciated. Funding: This work was financially
supported by the National Key R&D Program of China (no.
2022YFA1503300), the National Natural Science Foundation of
China (no. 92056202), the China Postdoctoral Science Foundation
(no. 2021M691373), and the China Postdoctoral Innovation Talents
Support Program (no. BX2021116). Author contributions: W.W.

led the project. J.F., J.H., L.L., and W.W. conceived the idea. J.H.,
L.L., J.K., and J.-M.C. conducted the synthesis and crystal growth
of 3D COFs. X.-Y.D. conducted the growth of 2D single-crystal
COFs. L.L. and J.H. carried out the crystallographic studies. J.H.,
L.L., J.-M.C., and J.K. carried out the characterizations. J.H., L.L.,
J.K., and X.-Y.D. took the crystal images and photos. J.H., L.L.,
J.F., J.-M.C., S.-Y.D., and W.W. discussed the results. L.L., J.H.,
and W.W. interpreted the results and wrote the manuscript.
Competing interests: The authors declare that they have no
competing interests. Data and materials availability: Crystallographic
data reported in this paper are tabulated in the supplementary
materials and archived at the Cambridge Crystallographic Data Centre
(CCDC) under reference nos. CCDC 2294453 to 2294464 and 2294641
to 2294643. All other data needed to evaluate the conclusions in the
paper are present in the paper or the supplementary materials.
License information: Copyright © 2024 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk8680
Materials and Methods
Supplementary Text
Figs. S1 to S125
Tables S1 to S19
References (36–42)
Data S1 to S27
Checkcif files for Data S1 to S15

p

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.

24. S. J. Rowan, S. J. Cantrill, G. R. L. Cousins, J. K. M. Sanders,
J. F. Stoddart, Angew. Chem. Int. Ed. 41, 898–952 (2002).
25. M. Ciaccia, S. Di Stefano, Org. Biomol. Chem. 13, 646–654
(2015).
26. N. Giuseppone, J.-L. Schmitt, E. Schwartz, J.-M. Lehn, J. Am.
Chem. Soc. 127, 5528–5539 (2005).
27. T. Ma et al., J. Am. Chem. Soc. 140, 6763–6766 (2018).
28. H.-L. Jiang, T. A. Makal, H.-C. Zhou, Coord. Chem. Rev. 257,
2232–2249 (2013).
29. P. L. Luisi, Found. Chem. 4, 183–200 (2002).
30. Y. Inokuma et al., Nature 495, 461–466 (2013).
31. S. Lee, E. A. Kapustin, O. M. Yaghi, Science 353, 808–811
(2016).
32. E. Borisenko, Ed., Crystallization and Materials Science of
Modern Artificial and Natural Crystals, (InTech, 2012), p. 253.
33. L. Li et al., New J. Chem. 47, 20703–20707 (2023).
34. L. Wang, J. Liu, J. Wang, J. Huang, Chem. Eng. J. 473, 145405
(2023).
35. B. R. Pamplin, Crystal Growth (Pergamon Press, 1980).

Submitted 15 September 2023; accepted 17 January 2024
10.1126/science.adk8680

g
y
y g
y
,

Han et al., Science 383, 1014–1019 (2024)

1 March 2024

6 of 6

RES EARCH

ORGANIC CHEMISTRY

Aminative Suzuki–Miyaura coupling
Polpum Onnuch, Kranthikumar Ramagonolla, Richard Y. Liu*
The Suzuki–Miyaura and Buchwald–Hartwig coupling reactions are widely used to form carbon-carbon
(C–C) and carbon-nitrogen (C–N) bonds, respectively. We report the incorporation of a formal nitrene
insertion process into the Suzuki–Miyaura reaction, altering the products from C–C–linked biaryls
to C–N–C–linked diaryl amines and thereby joining the Suzuki–Miyaura and Buchwald–Hartwig coupling
pathways to the same starting-material classes. A combination of a bulky ancillary phosphine ligand
on palladium and a commercially available amination reagent enables efficient reactivity across
aryl halides and pseudohalides, boronic acids and esters, and many functional groups and heterocycles.
Mechanistic insights reveal flexibility on the order of bond-forming events, suggesting potential for
expansion of the aminative cross-coupling concept to encompass diverse nucleophiles and electrophiles
as well as four-component variants.

1 of 6

,

At the outset of our investigations, we examined the reaction between 4-methoxyphenyl
triflate (1a) and 4-(trifluoromethyl)phenylboronic
acid (3a) in the presence of a variety of electrophilic amination reagents as formal precursors of parent nitrene (“NH”). Using catalysts
supported by typical phosphine ligands (such
as RuPhos) (Fig. 1C, entry 2), complete conversion
to Suzuki–Miyaura coupling products was observed after 12 hours, with no apparent participation of the amine reagent O-diphenylphosphinyl
hydroxylamine (DPPH, 2a) (29). By contrast,
when a t-BuBrettPhos-modified Pd catalyst was
used under optimized conditions (Fig. 1C, entry 1), the desired aminative coupling product
(4a) was obtained in 96% after 12 hours with
only trace Suzuki–Miyaura product (5a). The
use of tBuXPhos, a ligand with similar steric
properties and scaffold to t-BuBrettPhos, was
found to be nearly equally effective (Fig. 1C
entry 3). However, BrettPhos, a ligand typically
used in Buchwald–Hartwig cross-coupling
between (het)aryl (pseudo)halides and anilines

y

1 March 2024

Reaction development
y g

Onnuch et al., Science 383, 1019–1024 (2024)

y

*Corresponding author. Email: richardliu@chemistry.harvard.edu

g

Department of Chemistry and Chemical Biology, Harvard
University, Cambridge, MA 02138, USA.

by means of diverted pathways that generate
alternative, high-value products. In the past
decade, late-stage insertion and deletion reactions have attracted tremendous attention
as a strategy to generate structural diversity
(27). By analogy, we asked whether the insertion of a bridging atom between the nucleophilic and electrophilic components could
be a universal modification to cross-coupling
reactions that generates new products from
existing partners. The well-established carbonylative Stille cross-coupling (28) represents an
example of this approach, but the generalization
of the concept to insertions of other ambiphilic
components, especially heteroatomic ones, appears to have escaped systematic consideration.
To demonstrate the proposed concept, we
pursued the introduction of a formal nitrene
insertion into the Pd-catalyzed Suzuki–Miyaura
cross-coupling pathway, rerouting its endpoint
from biaryl products (C–C linkage) toward diaryl amines (C–N–C linkage), a privileged substructure class among bioactive compounds
(Fig. 1B). Many industrial research operations
maintain extensive libraries of custom aryl
halides (or pseudohalides) and boronic acids
(or esters), and we envisioned that through
the addition of a simple reagent, these building blocks could be conveniently repurposed
to furnish amines. This type of scaffold change
from biaryl to diaryl amine, previously inaccessible in a single operation, could be useful
for fine tuning the geometry, polarity, and Hbonding ability of many candidates. Achieving
this goal would effectively unite the two most
prominent metal-catalyzed coupling manifolds
(Suzuki–Miyaura and Buchwald–Hartwig) by
connecting their products to common precursor pools. Without requiring the separate
synthesis and purification of new reagents, the
chemical space accessible from existing functionalized intermediates could be multiplicatively increased.
We anticipated that realization of the intended three-component coupling might be
met with several distinct challenges. First,

p

T

ransition metal–catalyzed cross-coupling
reactions have become indispensable tools
for the synthesis of important organic
compounds, such as therapeutics (1–3),
agrichemicals (3, 4), energy-storage materials (5, 6), and functional polymers (7, 8).
Over the past half century in medicinal chemistry, three of the 20 most frequently practiced
transformations are palladium (Pd)–catalyzed
cross-couplings (Suzuki–Miyaura, Sonogashira,
and Buchwald–Hartwig) (9). Over time, the popularization of cross-coupling has considerably
influenced which sectors of chemical space are
heavily emphasized during drug discovery and,
therefore, the structures of recently approved
small-molecule pharmaceuticals. For example, there has been a proliferation of (hetero)
biaryl and aryl amine motifs because of the
reliability and generality of Suzuki–Miyaura
(10–13) and Buchwald–Hartwig catalysis (Fig. 1A)
(14–16). These examples suggest that new, general strategies to expand the product space of
essential cross-coupling schemes can enhance
structural diversity during candidate generation and improve the speed and success rate of
pharmaceutical development.
Traditionally, research aimed at broadening
the synthetic utility of cross-coupling methodology has focused on the development of catalysts and reaction conditions that engage
distinct reactive partners (electrophiles or nucleophiles), a campaign punctuated by major
recent achievements such as the fluorination
(17, 18) and trifluoromethylation of aryl electrophiles (19–21), carbon-carbon (C–C) coupling
from alkyl electrophiles (22, 23), reductive crosselectrophile couplings (24, 25), and activation
of carbon-hydrogen (C–H) bonds for crosscoupling (26). An attractive but rarely explored
research strategy involves the repurposing of
classical, widely available coupling partners

modern Pd-based catalysts perform Suzuki–
Miyaura coupling so efficiently that for an
amine insertion to intercede, the original process would likely need to be decelerated, either
through deactivation of the catalyst toward
reductive elimination or inhibition of transmetallation with the aryl boron nucleophile.
However, any attenuation of reactivity must be
carefully balanced; after N-insertion, the metal
center must still be capable of achieving the
second C–N bond formation. Likewise, the
reactivity of the nitrene reagent must be precisely adjusted: It must be electrophilic enough
to efficiently insert, yet it should avoid reacting prematurely with Pd(0) before oxidative
addition of the aryl halide. Last, there remains
the task of avoiding homocoupling processes
that install two of the same aryl group on the
product rather than one derived from each
coupling partner.
We report that the combination of a bulky
catalyst and commercially available amination
reagent affords a convenient and highly general solution. Our protocol is effective for all
common classes of electrophiles (aryl chlorides,
bromides, triflates, and tosylates) and compatible with an exceptionally broad scope of
polar functional groups, substitution patterns,
and heterocyclic partners relevant to medicinal
chemistry. The strategy is easily used on latestage intermediates to prepare amine-inserted
variants of drug candidates, and preliminary
results suggest that the insertion concept can be
extended to other reaction classes (allylic substitution) and even four-component variants
(ArX + CO + NH + Ar’M). The cross-selectivity
of this reaction is notable because mechanistic
experiments indicate that multiple competing
mechanisms likely operate simultaneously.

RES EARCH | R E S E A R C H A R T I C L E

A

Me

SO2Me

SMC

O
N

Me

N

O

OCF3

Me

SMC
Cl

SMC

N

Sonidegib

B

OMe

MeO
Me

N
i-Pr

N
N SMC
BHC
HN
N

BHC

i-Pr
N

N
H

Et

N

N

X

N
CONMe2

N

N
HN

Fenofibrate

Etoricoxib

N

CO2i-Pr

O

CO-SMC

Me

F
F

H BHC
N
N

N

N

N
H

O

Cl

Ribociclib

(HO)2B

SMC

+

Me
N
N

• repurpose SMC partners to
make BHC products

SMC
N

H
N

NH2

X

N

“NH”

BHC

+
Erdafitinib

Abemaciclib

C

OTf

1a

entry

(HO)2B

+
CF3

2a
1.5 equiv

CF3

H
N

t-BuBrettPhos Pd G3 (3 mol%)

+

K3PO4 (4.0 equiv), 2-MeTHF (0.2 M), 80 °C

CF3

MeO

3a
2.0 equiv

MeO

4a

variation from standard conditions

1

p

+
MeO

O
P
H2 N
O
Ph
Ph

NO2

None

96

RuPhos Pd G3 (3 mol%)

3

t-BuXPhos Pd G3 (3 mol%)

4

BrettPhos Pd G3 (3 mol%)

70

5*

K3PO4 (3.5 equiv)

71

6*

KOH (3.5 equiv)

7

Reagent 2d (1.5 equiv)

8*

KOH (3.5 equiv), ArBpin instead of ArB(OH)2

H2 N

107
93

MsO
L Pd
H2 N

O

g

2

3

5a

NO2

13

2d

36

L Pd G3
R2

y

21
79

5

40

20

PR12
i-Pr

PCy2
Oi-Pr

i-PrO

63
79

0

R2
i-Pr

40
60
Yield (%)

5
80

i-Pr
R1 = t-Bu, R2 = OMe: t-BuBrettPhos
R1 = Cy, R2 = OMe: BrettPhos
100
R1 = t-Bu, R2 = H: t-BuXPhos

RuPhos

Substrate scope

The generality of this optimized protocol is
notable because the method is not only effective across a diverse set of (hetero)aryl triflates
(Fig. 2) but also—with only slight modifications
to the conditions (optimization data is provided
in table S3)—is compatible with (hetero)aryl
bromides and chlorides (Fig. 3). Empirically,
we found that bases other than potassium phosphate are necessary: The reaction can proceed
with a weak, soluble, organic base such as 1,8diazabicyclo(5.4.0)undec-7-ene (DBU), but the
use of KOH is more broadly applicable. Because
of the relatively mild reaction conditions, the
cross-coupling exhibits substantial functional
group compatibility. Electron-deficient and
electron-rich examples of both aryl triflates

1 March 2024

and aryl boronic acids were coupled in high
yields (4a to 4d). Other electron-withdrawing
substituents that were well tolerated include
nitro (4x) and nitrile (4ab). A variety of ortho
substituents—including methyl (4e, 4f, 4g,
4aj, and 4al), alkoxy (4 g, 4h, 4l, 4q, 4y, 4z,
4af, and 4ak), and halides (4i, 4r, and 4s)—
on either or both coupling partners did not
adversely affect the reactivity, resulting in moderate to high yields. A substrate containing
a primary alcohol was coupled in 36% yield
(4j). The potential for side reactions such as
O-arylation and oxidation of the alcohol through
b-hydride elimination of a Pd(II) alkoxide could
be responsible for the low yield in this case.
Substrates containing carboxylic acid derivatives such as amide (4k); a-acidic aryl ketones
(4k and 4aa) and alkyl ketones (4l); and
esters including ethyl (4d), methyl (4 g, 4h,
4k, and 4q), and isopropyl (4am) all reacted
2 of 6

,

Onnuch et al., Science 383, 1019–1024 (2024)

provided that KOH was used as the base (Fig. 1C,
entry 8).

y

(30), could only catalyze the reaction with
diminished yield (70%) (Fig. 1C, entry 4) and
selectivity (36% 5a was formed). Relative to the
optimized conditions, decreasing equivalents
of 2a and 3a were associated with incomplete
conversion and lower yield of 4a (Fig. 1C, entry
5). A stronger base, such as potassium hydroxide (KOH), could be used to restore high yield
and full conversion, but in polar solvents—such
as N,N′-dimethylformamide (DMF) (table S1,
entry 4), acetonitrile (MeCN) (table S1, entry 3),
and 2-methyltetrahydrofuran (2-MeTHF) (Fig.
1C, entry 6)—triflate decomposition to phenol
was a competing side reaction. Other ambiphilic
aminating agents such as 2d (Fig. 1C, entry 7),
which have been previously reported (31) to
effect amination of aryl boronic acids, were ineffective in this context (table S2, entries
10 to 13). Instead of boronic acid 3a, its pinacol ester displayed equally efficient reactivity

y g

Fig. 1. Background and concept. (A) Suzuki–Miyaura (SMC) and Buchwald–Hartwig (BHC) cross-coupling in drug development. (B) This work: SMC with NH
insertion to access BHC products. (C) Challenges for single-heteroatom insertion. (D) Selected results from reaction optimization (yields shown as determined with
gas chromatography analysis). Asterisk indicates 2a (1.1 equiv), 3a (1.5 equiv).

RES EARCH | R E S E A R C H A R T I C L E

O
P
O
Ph
Ph

OTf
R1

H 2N

+

H
N

t-BuBrettPhos Pd G3 (1 to 3 mol%), K 3PO 4 (4.0 equiv)

(HO)2B
R2

+

R2

R1

2-MeTHF (0.2 M), 80 °C

1.5 to 2.0 equiv

2a
1.1 to 2.0 equiv

(Het)ArOTf
H
N
R2

R1

Me

H
N

Me

MeO

4a: R1 = 4-OMe, R2 = 4-CF3, 81%
4b: R1 = 4-OMe, R2 = 4-F, 93%
4c: R1 = 4-OMe, R2 = 4-OTBS, 81%
4d: R1 = 4-CO2Et, R2 = 4-CF3, 70%
4e: R1 = 4-OMe, R2 = 2-Me, 75%
O

4f: 36%
using ArBpin

O
S
N
O

Boc

N

MeO2C

Me

4h: 68%, 91% ee
from L-Boc-tyrosine-OMe

O
Me

H
N

N

OH

4j: 36%

N

H

4l: 44%, from estrone

H
N

H
N

N

N

N
Me

Me

S

N

Me

N

N

S

4o: 84%
F

O

F

MeO

F

4p: 56%

H
N
F

O

Ph

H
N

Oi-Pr
N
EtO2C

Me

4s: 45% using ArBpin
9% using ArB(OH)2

4r: 79% using ArBpin
0% using ArB(OH)2

F

4t: 44%

y

4q: 42%

F

N

g

H
N

OMe

F3C

N

4n: 45%

4m: 76%
H
N

Me

H
N

N

Me

N

p

H
N

OMe

Me

O

4k: 86%
Me

4i: 56%

MeO2C

H

CO2Me

O

N

H

O
Ph

OMe

H
N

H
N

O

OBn

H
N

NH

MeO2C

4g: 67%

O

Cl

H
N

OMe

H
N

Fig. 2. Scope of coupling with aryl triflates. Asterisk indicates 2a (2.0 equiv), ArB(OH)2 (2.0 equiv), DBU (3.0 equiv), PhMe (0.2 M), 80°C. “†” indicates 2a
(1.5 equiv), ArB(OH)2 (1.5 equiv), DBU (3.0 equiv), PhMe (0.2 M), 60°C. “‡” indicates 2a (1.1 equiv), ArBPin instead of ArB(OH)2 (1.5 equiv), KOH (3.5 equiv).
“§” indicates KOH (3.0 equiv), 2a (1.1 equiv), ArB(OH)2 (1.5 equiv), PhMe (0.2 M). “¶” indicates ArOTs instead of ArOTf, DBU (4.0 equiv), MeCN (0.2 M).

X
+

X = Br or Cl

O
P
O
Ph
Ph

R2

+

H
N

t-BuBrettPhos Pd G3 (1 to 3 mol%), t-BuBrettPhos (0 to 1 mol%)

(HO)2B

R2

R1

KOH (3.0 to 4.0 equiv), MeCN (0.2 M), 80 °C

y g

R1

H 2N

1.2 to 2.0 equiv

2a
1.1 to 2.0 equiv

(Het)ArBr
H
N
F

H
N

PhO

H
N
Me

4z: 84%

4aa: 67%
H
N
O

N

N

H
N

OMe
O

N

N

Me

O

4ae: 42%

S

O

4af: 63%

NO2

N
O

H
N

H
N
CN

4ac: 51%

H
N

H
N

4ad: 82%

O
F

4ag: 70%

O

MeO

S

4ab: 56%

N

N

4y: 72%

4x: 68%

H
N

MeO

N

S

4w: 69%

4v: 45%

OBn

H
N

H
N

Me
F

F3C

4u: 62%

N

,

O

H
N

H
N

y

O

N

F3C

4ah: 60%

Me

H
N

N

O

4ai: 78%

Fig. 3. Scope of coupling with aryl halides. Asterisk indicates 2a (1.5 equiv), ArB(OH)2 (1.5 equiv), DBU (3.0 equiv). “†” indicates 2a (2.0 equiv), ArB(OH)2
(2.0 equiv), DBU (3.0 equiv).
Onnuch et al., Science 383, 1019–1024 (2024)

1 March 2024

3 of 6

RES EARCH | R E S E A R C H A R T I C L E

successfully, as did those containing strained
cyclopropane rings (4k and 4ad) and carbamates (4h and 4al). Nitrogen-containing heterocycles such as pyridines activated at the 2
(4p), 3 (4 g, 4n, 4o, 4p, and 4q), and 4 positions (4s) and pyrimidine (4l) were compatible
with the reaction conditions. 6-(Morpholinyl)
pyridine, a common moiety in small-molecule
drug development, was also tolerated (4y and
4ae). Fused six- (4n, 4ah, and 4ai) and fivemembered (4m, 4ab, 4ac, 4af, 4ag, and 4am)
heterocycles were coupled in moderate to excellent yields. The N–N bond in 4o was left
intact under our conditions; however, substrates
with the (pseudo)halide directly attached to a
five-membered ring proved difficult to couple
(fig. S7), which is consistent with prior observations that had been attributed to catalyst
deactivation, slow reductive elimination, and
instability toward base-promoted ring fragmentations (32). Last, several examples of sulfurcontaining five-membered heterocycles reacted
successfully (4q, 4ab, and 4ac).
Prior studies have shown that highly electrondeficient boronic acids with ortho heteroatoms
and polyfluorinated systems are particularly
challenging to couple under Suzuki–Miyaura
conditions owing to rapid competing protodeboronation (33, 34). We found that examples 4r and 4s were difficult to prepare from
boronic acids, resulting in poor or undetectable
yield of product and substantial formation of
protodeboronation side products. However,
if instead the pinacol esters of the requisite
boronic acids were used, the desired products
were obtained in considerably improved yields
of 79% (4r) and 45% (4s). Showcasing the
mild and versatile conditions, the reaction was
used to derivatize some simple natural products:
We obtained 4h from L-tyrosine (with 91% retention of enantiopurity), 4l from estrone, and
4i and 4s from flavone. Compounds 4aa and
4ad were synthesized in excellent yields from an
aryl bromide intermediate from the synthesis of
adapalene, a topical treatment for acne vulgaris.
The reaction conditions can be effective for aryl
tosylates as well (4t), although further optimization may be needed to increase the yield.

A

Applications and extensions

Fig. 4. Applications of aminative Suzuki–Miyaura coupling. (A) NH insertion into drugs synthesized by
means of SMC. (B) Late-stage modification of ArCl-containing drugs. (C) Four-component coupling involving
sequential NH and CO insertion. (D) Aminative Tsuji-Trost allylation. Asterisk indicates 0.4 mmol scale,
[Pd] (3 mol %), 2a (1.1 equiv), ArB(OH)2 (1.5 equiv), KOH (3.5 equiv), 2-MeTHF (0.2 M), 80°C. “†” indicates
1.0 mmol scale, [Pd] (3 mol %), 2a (1.1 equiv), ArB(OH)2 (1.2 equiv), KOH (3.5 equiv), MeCN (0.2 M), 80°C, yield
reported is an average between 2 runs. “‡” indicates [Pd] (2 mol %), t-BuBrettPhos (1 mol %), 2a (1.1 equiv),
ArB(OH)2 (1.2 equiv), KOH (3.0 equiv), MeCN (0.2 M), 80°C. “§” indicates [Pd] (2 mol %), dppf (0.6 equiv), DBU
(3.0 equiv), MeCN (0.2 M), 80°C. “¶” indicates [Pd] (2 mol %), DBU (3.0 equiv), MeCN (0.2 M), 80°C.

Me
Me
O

SMC

O
N

N

Me

N

Me

N

O

OCF3

Me

Me

O

N
H

Br

N
H

Sonidegib

1aj
+
H 2N

(HO)2B
OCF3

O
P
Ph
Ph
2a

Me

O

O

“NH” Insertion

N

Me

N

O

Me

3aj
N
H

H
N
OCF3

4aj: 33%
SO2Me
Cl

SMC

N

OBn

Etoricoxib Intermediate
Cl

(HO)2B

Br

p

+
N

SO2Me

OBn

1ak

H2N

3ak

O
P
Ph
Ph
2a

O

H
N

Cl

“NH” Insertion

N

OBn

SO2Me

B

Cl

H
N

(HO)2B
+

Drug

“NH” insertion

O

Me

O

O
P
Ph
Ph

N

i-PrO

Me Me
O

O

EtO

Drug

4al: 58%, from Loratadine

Br

+

Fe(CO)5

F3C

0.5 equiv

1w

N
N Me

N

H
N

C

H
N

O

y

H2N

g

4ak : 50%

+

O
H2 N
P
O
Ph
Ph

4am: 75%, from Fenofibrate
4-Component
Coupling

(HO)2B

+
F

N
H
F3C

6ao: 55%

3b
1.5 equiv

2a
1.2 equiv

F

O

y g

D
OAc
F

7ap
OAc

+

O
H 2N
P
O
Ph
Ph

Allylation

+
F

3b
1.2 equiv

N
H

9ap
from 7ap: 54%
from 8ap: 36%

,

8ap

Because of its broad reliability and functionalgroup compatibility, this method allows for
the convenient reuse of complex Suzuki–Miyaura
partners to access NH-inserted variants of druglike molecules through inclusion of the DPPH
reagent during the coupling reaction (Fig. 4A).
For example, synthesis of Sonidegib, a Smoothened (SMO) inhibitor used for the treatment of
basal cell carcinoma, relies on Suzuki–Miyaura
coupling of 1aj and 3aj (35). Following our
protocol, these same reagents can be repurposed
to afford analog 4aj in modest yield (33%) (Fig.
4A, top). Similarly, Etoricoxib, a cyclooxygenase-2
(COX-2) inhibitor used to treat arthritis pain,
Onnuch et al., Science 383, 1019–1024 (2024)

y

2a
1.1 equiv

(HO)2B

can be synthesized by means of Suzuki–Miyaura
coupling reactions from 1ak (Fig. 4B, bottom)
(36). We subjected 1ak and 3ak to aminative
coupling conditions to afford the product 4ak,
a modified Etoricoxib intermediate, in 50%
yield on a 1-mmol scale. The haloselectivity

1 March 2024

under these conditions completely favors aryl
bromides and triflates over chlorides (for example, 4i). Last, we demonstrated several examples of functionalization of pharmaceutical
intermediates and small-molecule drugs containing aryl chlorides, which can be less reactive
4 of 6

RES EARCH | R E S E A R C H A R T I C L E

A
Ar

H
N

Ar X

Ar´

H
N

Ar

Ar´

HN

LPd
X

Ar
H 2N

Ar

O

O
P
Ph
Ph

LPd
HN

Base

Ar´ NH

LPd0

LPd0

LPd

B(OH)2

Ar X
Ar´

Base-H+

Ar
LPd
X

Ar´

O
H
P
N
B
Ph
O
Ph
HO OH

Ar´

Ar

Ar´ NH2
Base
Base

O
Ar´ B(OH)2
LPd
HN

P Ph
Ph
O

LPd
O
HN
P O
Ph Ph

Ar

Base

Ar
Ar
LPd
O HN
Ar´
B OH
P O
Ph Ph
OH

Electrophile First

B

Ar´ B(OH)2

Nucleophile First

CO2Et

t-BuBrettPhos-Pd

H 2N

+

O
P
O
Ph
Ph

C

+

2a
1.1 equiv

OTf
H2 N

MeO

1a

O

O
P
Ph
Ph

2-MeTHF (0.05 M), 80 °C, 15 min
then HCl quench

O

+

O
B

Coupling *

F

H
N

3-Component
F

MeO

F
+

F

3r
1.5 equiv

F

4r

No [Pd]

CO2Et

7d
25%

6d
3%

F

F

2a
1.1 equiv

H2N
F

F

6r

55

0.5 mol% 4
Pd loading

EtO2C

EtO2C

66
62
87

2 mol%
0

22

44

66

26
88

110

Yield (%)

Mechanistic insights

1 March 2024

REFERENCES AND NOTES

1. J. Rayadurgam, S. Sana, M. Sasikumar, Q. Gu, Org. Chem. Front.
8, 384–414 (2021).
2. J. Yin, in Applications of Transition Metal Catalysis in Drug
Discovery and Development, M. L. Crawley, B. M. Trost, Eds.
(John Wiley & Sons, 2012), pp. 97–163.
3. C. Torborg, M. Beller, Adv. Synth. Catal. 351, 3027–3043 (2009).
4. P. Devendar, R.-Y. Qu, W.-M. Kang, B. He, G.-F. Yang, J. Agric.
Food Chem. 66, 8914–8934 (2018).
5. A. Omidvar, ACS Appl. Energy Mater. 3, 11463–11469 (2020).
6. M. Quant et al., Chem. Sci. 13, 834–841 (2021).
7. A. K. Leone, E. A. Mueller, A. J. McNeil, J. Am. Chem. Soc. 140,
15126–15139 (2018).
8. J. Zhu et al., Acc. Chem. Res. 51, 3191–3202 (2018).

5 of 6

,

A plausible pathway for initial C–N bond formation to take place from the aryl electrophile
(Fig. 5A, mechanism I, “electrophile-first”) could
involve a 1,2-shift of the aryl ligand from Pd(II) to
a coordinated and deprotonated 2a, which generates the amido complex LPd(NHAr)OPOPh2.
This process resembles that proposed by Knochel
and coworkers for the electrophilic amination
of organozinc reagents with organic azides (38).
The resulting Pd(II) phosphinate can undergo
transmetallation with a boronic acid and reductive elimination to afford the desired product.
The idea that amination of the aryl electrophile
might precede C–N bond formation with the
aryl nucleophile is consistent with observation
of aniline derived from the former in some
cases. For example, 8% of 1-aminonaphthalene
was isolated in the reaction to produced 4j in
Fig. 2. As further evidence, a stoichiometric
reaction between Pd-1, a likely on-cycle postoxidative-addition complex, with reagent 2a in
the presence of base resulted in 3% of aniline
6d and 25% diaryl amine 7d, showing that C–N

y

Onnuch et al., Science 383, 1019–1024 (2024)

proceeds through a Pd p-allyl intermediate, 8ap
was subjected to the same reaction conditions to
afford the same product, although in a slightly
diminished yield (36%). In the absence of a Pd
catalyst, no product was observed.

y g

than aryl bromides in cross-coupling. Late-stage
modification of Loratidine and Fenofibrate provided 4al and 4am, respectively, in good yields.
These examples illustrate the power of this reaction in providing direct access to new drug
candidates without introducing additional operations or intermediates on all stages of drug
synthesis.
Formal nitrogen insertion into cross-coupling
reactions is a concept readily generalizable beyond the Suzuki–Miyaura couplings shown
above. For example, by tandem insertion of NH
and a carbonyl (C=O) group, Suzuki–Miyaura
coupling partners can be used to make amides
as an alternative to traditional amide-bond formation, which is one of the most frequently
used reactions in medicinal chemistry (9). As an
example, 1w and 3b were coupled in the presence of 2a and iron pentacarbonyl [Fe(CO)5]
as the carbonyl source to produce 6ao in good
yield (55%) (Fig. 4C; further optimization data
is available in table S4). We also observed the
successful application of this NH insertion concept in the context of Pd-catalyzed allylation
(Tsuji–Trost) chemistry (Fig. 4D) (37). Under
unoptimized conditions, substrate 7ap undergoes coupling with reagent 2a and boronic
acid 3b to furnish the linear product 8ap in
good yield (54%) and selectivity (only branched
product observed). As evidence that the reaction

y

Fig. 5. Mechanistic insights. (A) Proposed mechanisms. (B) Stoichiometric reaction between Pd-1 and
2a. (C) Effects of Pd loading on total C–N bond formation from boronate ester 3r. Asterisk indicates
t-BuBrettPhos Pd G3, KOH (1.5 equiv), 2-MeTHF (0.2 M), 80°C.

g

16

1 mol%

p

+

H
N

NH2

KOH (2.0 equiv)

OTf

Pd-1

O
P
H2N
O
Ph
Ph
+

formation from (pseudo)halide is possible under the reaction conditions (Fig. 5B). The formation of 7d is expected from the reaction of
6d with Pd-1 in the absence of any other competitive nucleophile (a boronic acid was not included in these experiments).
During the course of our studies, it became
clear that initial C–N bond formation from the
aryl nucleophile side could also viable (Fig. 5A,
mechanism II, “nucleophile-first”). Electrophilic
amination of boronic acids by reagents such as
2d has been reported (37), although such transformations are typically limited to electron-rich
substrates, and many reagents competent for this
process are not effective in our three-component
coupling. Under our optimized conditions, for
some substrate combinations, Pd-independent
formation of aniline from arylboron can occur.
In one possibility, this aniline could then be
arylated through a typical Buchwald–Hartwig
pathway to form the three-component coupling product. However, the total amount of
C–N bond formation from the boronate (sum
of yields for 4r and 6r) increases with catalyst
loading, as does the relative ratio of 4r to 6r
(Fig. 5C). The ability of the catalyst to affect
total C–N bond formation from the boronate
implies that there is also a Pd-dependent mechanism for creating this bond. A proposed explanation is illustrated in Fig. 5A, mechanism
II, in which arylpalladium(II), boronic acid,
and DPPH form a complex, in which the Lewis
acidity of the metal increases the electrophilicity
of the nitrogen atom and facilitates aryl migration from boron. Because of this pathway, yields
obtained in aminative Suzuki reactions can exceed those of Buchwald–Hartwig amination from
pregenerated anilines under the same conditions (supplementary materials, section 8.4).
Work continues in our laboratory to achieve
a more detailed characterization of available
pathways and to elucidate the effect of substrate structure on which of the nucleophile-first
or electrophile-first mechanisms is preferred.
However, at this point, our studies have definitively established the possibility of forming insertive cross-coupling products through either
order of events. Looking forward, we argue that
this mechanistic flexibility portends favorably
for the extension of the aminative cross-coupling
concept to diverse classes of both nucleophiles
and electrophiles.

RES EARCH | R E S E A R C H A R T I C L E

9. D. G. Brown, J. Boström, 59, 4443–4458 (2016).
10. J. Almond-Thynne, D. C. Blakemore, D. C. Pryde, A. C. Spivey,
Chem. Sci. 8, 40–62 (2017).
11. M. A. Düfert, K. L. Billingsley, S. L. Buchwald, J. Am. Chem. Soc.
135, 12877–12885 (2013).
12. K. Billingsley, S. L. Buchwald, J. Am. Chem. Soc. 129,
3358–3366 (2007).
13. M. J. Buskes, M.-J. Blanco, Molecules 25, 3493 (2020).
14. H. H. Nguyen et al., ACS Med. Chem. Lett. 12, 1605–1612 (2021).
15. A. T. Garrison et al., J. Med. Chem. 61, 3962–3983 (2018).
16. C. Fischer, B. Koenig, Beilstein J. Org. Chem. 7, 59–74 (2011).
17. H. G. Lee, P. J. Milner, S. L. Buchwald, J. Am. Chem. Soc. 136,
3792–3795 (2014).
18. D. A. Watson et al., Science 325, 1661–1664 (2009).
19. R. J. Lundgren, M. Stradiotto, Angew. Chem. Int. Ed. 49,
9322–9324 (2010).
20. E. J. Cho et al., Science 328, 1679–1681 (2010).
21. Y. Ye, M. S. Sanford, J. Am. Chem. Soc. 134, 9034–9037 (2012).
22. A. W. Dombrowski et al., ACS Med. Chem. Lett. 11, 597–604 (2020).
23. J. Choi, G. C. Fu, Science 356, eaaf7230 (2017).
24. D. A. Everson, D. J. Weix, J. Org. Chem. 79, 4793–4798 (2014).
25. D. J. Weix, Acc. Chem. Res. 48, 1767–1775 (2015).
26. N. Y. S. Lam, K. Wu, J.-Q. Yu, Angew. Chem. Int. Ed. 60,
15767–15790 (2021).

27. J. Jurczyk et al., Nat. Synth. 1, 352–364 (2022).
28. A. M. Echavarren, J. K. Stille, J. Am. Chem. Soc. 110, 1557–1565
(1988).
29. W. Klötzer, J. Stadlwieser, J. Raneburger, Org. Synth. 64, 96 (1986).
30. B. T. Ingoglia, C. C. Wagen, S. L. Buchwald, Tetrahedron 75,
4199–4211 (2019).
31. S. Voth, J. W. Hollett, J. A. McCubbin, J. Org. Chem. 80,
2545–2553 (2015).
32. E. C. Reichert, K. Feng, A. C. Sather, S. L. Buchwald, J. Am.
Chem. Soc. 145, 3323–3329 (2023).
33. H. L. D. Hayes et al., J. Am. Chem. Soc. 143, 14814–14826
(2021).
34. P. A. Cox et al., J. Am. Chem. Soc. 139, 13156–13165 (2017).
35. S. Pan et al., ACS Med. Chem. Lett. 1, 130–134 (2010).
36. J. J. F. Caturla, G. Warrellow, 2,3′-Bipyridines derivatives as
selective COX-2 inhibitors. World Intellectual Property Organization
(WIPO) patent application no. WO 2004/072037 A1 (2004).
37. B. M. Trost, D. L. Van Vranken, Chem. Rev. 96, 395–422 (1996).
38. S. Graßl, J. Singer, P. Knochel, Angew. Chem. Int. Ed. 59,
335–338 (2020).
ACKN OWLED GMEN TS

We acknowledge N. Faialaga for helpful discussions. We thank
I. Leibler, N. Faialaga, S. Li, and N. Naito for advice on the

preparation of the manuscript and the supplementary materials.
Funding: This work was supported by the William F. Milton Fund
and the Corning Fund for Faculty Development at Harvard
University. Author contributions: R.Y.L. conceived and directed
the execution of the study. P.O. and K.R. performed all
experiments. All authors contributed to the preparation of this
manuscript. Competing interests: The authors declare no
competing interests. Data and materials availability: All
data are available in the main text or the supplementary
materials. License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journalarticle-reuse
SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adl5359
Materials and Methods
Figs. S1 to S9
Tables S1 to S4
References (39–65)
Submitted 24 October 2023; accepted 5 January 2024
10.1126/science.adl5359

p
g
y
y g
y
,

Onnuch et al., Science 383, 1019–1024 (2024)

1 March 2024

6 of 6

WORKING LIFE
By Tae Seok Moon

From anxiety to action

“T

science.org SCIENCE

,

1 MARCH 2024 • VOL 383 ISSUE 6686

y

1026

Tae Seok Moon is a professor at Washington University in St. Louis. Send
your career story to SciCareerEditor@aaas.org.

y g

“I told myself I would stop fearing
career repercussions and would
push ahead with my ambitions.”

y

search. But I had never felt ready
or that I was a “big” enough name
to pull it off. Now, I told myself
I would stop fearing career repercussions and would push ahead
with my ambitions, focusing on
the good I could do rather than
the ways it might go wrong.
It wasn’t always easy, and my
anxiety didn’t immediately disappear. My colleagues and university leadership did not show
much interest in the project, and I
worried that my planned format—
which gave the big-name speakers
short presentation slots while leaving more time for those earlier in
their careers—might be taken as
an insult. But I learned to spend
more time trying than worrying. I
realized there are two types of problems: ones that can be solved, and
others that cannot. If it is solvable, my job is to try hard to
solve it. If not, my task is to find alternative options. Either
way, worrying does not help.
I’ve been running my weekly seminar series for
2.5 years now. It requires a tremendous amount of time, commitment, and effort. Nonetheless, I have been energized by
the excitement and passion I have seen from the young people involved. That is all the reward I have been looking for.
My regrets over the incident with my student helped me
realize that my goal in life should be fulfillment, not just
career success. I am now the happiest I have ever been because I have reconnected with my purpose: nurturing future generations. j

g

By many measures, my career
was flourishing. I had secured
grant funding, published highimpact journal articles, and given
conference talks. I was popular
among students. Even so, I was
consumed with anxiety. During
faculty meetings, I never spoke
up unless I was asked. I focused
only on research and education,
not campus politics. I thought
this was the way to success.
This was my mindset 5 years into
my faculty career, when the incident
with my student occurred. I refused
to kick him out, but I did make it
quiet. I persuaded him to rescind
his complaint, put his head down,
and focus on his research and job
search. (Editor’s note: Washington
University in St. Louis declined to
comment on the events in this story.)
Within the next year, the situation seemed resolved: I
was awarded tenure and my student found a job at a different company. But I was overcome with guilt about my
role. I felt like a coward. My anxiety worsened. I lost an unhealthy amount of weight and went to the emergency room
multiple times with severe pain. I call these my dying years.
A turning point came when my student ultimately
landed a job at the company he had been interviewing
with in the hallway and contributed to developing the first
COVID-19 vaccine. His perseverance inspired me. My goal
when I became a professor was to educate future leaders,
and I resolved not to let my anxious personality hold me
back any longer.
For years, I had wanted to start a seminar series that
covered both cutting-edge science and the personal and
professional challenges scientists face as we pursue our re-

p

ae Seok, you know what to do: Make it quiet and kick him out.” “Him” was one of my best
graduate students, who published four papers in 5 years while sending money home from the
United States to his family in Africa. But he had run afoul of another professor when he took
a phone call for a job interview in the hallway, where his cell signal was strongest. Based on
his skin color, the professor assumed he was not a student and called the police, who escorted
him away. Beyond the insult, it cost my student a dream job at his dream company. He filed a
complaint with the university’s discrimination office—and now my institution’s leadership was telling
me to make it go away. I was outraged, but I felt powerless. My tenure package was about to go up for
evaluation. I didn’t feel I was in a position to fight back.