Biology

An international scientific competition widely credited with spurring the development of artificial intelligence (AI) for biology appears to be on its deathbed. Known as the Critical Assessment of protein Structure Prediction (CASP), the 3-decade-old competition has run out of funding from the U.S. National Institutes of Health (NIH) and will exhaust emergency support from the University of California (UC) Davis, which oversees the grant, on 8 August. UC Davis has told the two researchers who run the program that their jobs will end in weeks.

NIH officials have offered no reassurance about the program’s future, despite repeated inquiries from CASP organizers, who submitted a request to renew their $800,000 grant last year. The agency did not respond to multiple requests for comment. John Moult, a CASP co-founder at the University of Maryland, says contest organizers are “scrambling” to find alternative funding from foundations and other countries.

If CASP disappears, “it would be a tremendous loss for the community,” says Jianlin Cheng, a protein modeler at the University of Missouri. “CASP is the most successful and prestigious scientific competition in the world.” Not only did it fuel the creation of AI programs such as RosettaFold and Google DeepMind’s AlphaFold, whose developers shared in last year’s Nobel Prize in Chemistry, but CASP has evolved to host new competitions, such as determining how drug molecules bind with protein targets, a key to discovering new medicines. “CASP continues to have a major impact on a wide range of scientific fields,” Cheng says.

Started in 1994, the biennial CASP competition began as a way to independently evaluate computer models aimed at solving the so-called protein folding problem: predicting how long linear chains of amino acids assemble into the 3D shapes required for their function. At the time, research teams developing the models were regularly claiming breakthroughs. “Everybody thinks their method is the greatest. That’s just the way people are,” says Nick Grishin, a biochemist at the University of Texas, Southwestern. CASP’s goal was to compare the methods head to head, he adds. “When the methods are stacked against one another they see where the failures are and that’s what drives science forward.”

Moult and his colleagues began by asking experimental groups which protein structures they expected to solve soon, using x-ray crystallography or other direct measures. They then invited computer modelers to predict the same structures, and had independent experts compare the models with the finished experimental data. In CASP1, modelers’ predictions for the most challenging class of proteins were “near random,” Moult says.

But in CASP14 in 2020, the year RosettaFold and AlphaFold burst on the scene, bringing a dramatic jump in accuracy in structure prediction. In last year’s CASP16, modelers achieved a nearly 95% success rate with the most difficult proteins. “The problem of folding single proteins is nearly solved,” concludes a 2 June preprint by Grishin and his colleagues summarizing the most recent CASP16 results.

“That’s a tremendous achievement,” Grishin says. “But what lies ahead is even more important.” Biochemistry, he notes, works through molecular interactions: proteins with other proteins, and proteins with DNA, RNA, and small molecules. Because those pairings are hard to study in the lab, there is less experimental data available for training AIs to predict the structures that result. As a result, the models struggle. In CASP16, for example, AlphaFold3—an AI designed to predict structures of protein complexes—hit the mark in only 52.5% of test cases, and even that was considered an impressive jump over previous efforts.

As protein models have improved, CASP organizers have come up with new ways to challenge the modelers, says its director, Krzysztof Fidelis, who got his 45-day notice from UC Davis last week. In last year’s CASP 16, 209 modeling groups worldwide tried to solve not only the structure of individual proteins, but also the 3D structures of RNA molecules, ensembles of proteins, molecules such as drugs bound to proteins, and the variety of shapes that so-called disordered proteins can adopt. “Every CASP is essentially a new set of competitions,” says Andriy Kryshtafovych, a mathematician who helps run the UC Davis program and has also been given a termination notice.

In addition to paying the salaries of Kryshtafovych and Fidelis, the UC Davis grant funds the coordination of legions of scientific volunteers, including hundreds of experimental groups that come up with targets, panels that decide the foci of each competition, modeling groups, assessors, and a conference after every competition to announce the winners and share lessons on what worked. “There’s a lot of leverage of the investment,” Moult says.

Organizers still hold out hope CASP’s grant might be renewed. NIH has been slow to spend its 2025 budget under President Donald Trump’s administration, stoking fears that much of the agency’s $48 billion in funding for this fiscal year will go unused. CASP might simply be caught in the slowdown.

In case NIH or an alternative funder doesn’t come through, Moult and others say UC Davis is currently exploring whether it can afford to archive CASP’s computer servers. That way the competition’s decades of resources won’t be lost. But even if the money does come through, there is no guarantee CASP’s success will endure after those running the program have moved on and the network of scientific volunteers has disbanded. Says Kryshtafovych: “It will not be easy to reconstruct this.”

By training a large language model (LLM) on a vast data set of human behavior, researchers say they have built an artificial intelligence (AI) system that can mimic a human mind. In a paper published today in Nature, they report that their model, Centaur, can “predict and simulate” human behavior in any experiment that can be written out in natural language.

But other scientists raise their eyebrows at the claim. “I think there’s going to be a big portion of the scientific community that will view this paper very skeptically and be very harsh on it,” says Blake Richards, a computational neuroscientist at McGill University and Mila – Quebec Artificial Intelligence Institute. He and others say the model doesn’t meaningfully mimic human cognitive processes, and that it can’t be trusted to produce results that would match human behavior.

Cognitive scientists often build models to help them understand the systems underlying abilities such as vision and memory. Each of these models captures only a very small, isolated part of human cognition, says Marcel Binz, a cognitive scientist at the Institute for Human-Centered AI at Helmholtz Munich. But with recent advances in LLMs, “we suddenly got this new exciting set of tools,” that might be used to understand the mind as a whole, he says.

To develop such a model, Binz and his colleagues created a data set called Psych-101, which contained data from 160 previously published psychology experiments, covering more than 60,000 participants who made more than 10 million choices in total. For example, in two “two-armed bandit” experiments, participants had to repeatedly choose between two virtual slot machines rigged to have unknown or changing probabilities of paying out.

The researchers then trained Llama, an LLM produced by Meta, by feeding it the information about the decisions participants faced in each experiment, and the choices they made. They called the resulting model “Centaur”—the closest mythical beast they could find to something half-llama, half-human, Binz says.

For each experiment, they used 90% of the human data to train the model and then tested whether its output matched the remaining 10%. Across experiments, they found Centaur aligned with the human data more closely than did more task-specific cognitive models. When it came to the two-armed bandit decisions, for example, the model produced data that looked more like the slot machine choices made by participants than a model specifically designed to capture how people make decisions in this task.

Centaur also produced humanlike outputs on modified tasks that weren’t in its training data, such as a version of the two-armed bandit experiment that adds a third slot machine. That means researchers could use Centaur to develop experiments “in silico” before taking them to human participants, Binz says, or to develop new theories of human behavior.

But Jeffrey Bowers, a cognitive scientist at the University of Bristol, thinks the model is “absurd.” He and his colleagues tested Centaur—which Binz’s team had made public when it published a first draft of the paper as a preprint—and found decidedly un-humanlike behavior . In tests of short-term memory, it could recall up to 256 digits, whereas humans can commonly remember approximately seven. In a test of reaction time, the model could be prompted to respond in “superhuman” times of 1 millisecond, Bowers says. This means the model can’t be trusted to generalize beyond its training data, he concludes.

More important, Bowers says, is that Centaur can’t explain anything about human cognition. Much like an analog and digital clock can agree on the time but have vastly different internal processes, Centaur can give humanlike outputs but relies on mechanisms that are nothing like those of a human mind, he says.

Federico Adolfi, a computational cognitive scientist at the Max Planck Society’s Ernst Strüngmann Institute for Neuroscience, agrees. Further stringent tests are likely to show that the model is “very easy to break,” he says. And he points out that although the Psych-101 data set is impressively large, 160 experiments is “a grain of sand in the infinite pool of cognition.”

But others see some value in the paper. Rachel Heaton, a vision scientist at the University of Illinois Urbana-Champaign, says the model doesn’t offer useful tools for understanding human cognition, but thinks the Psych-101 data set is a useful contribution in its own right because other researchers can use it to test the success of their models. Richards says future studies to understand what’s going on under the hood of Centaur could also be valuable.

Many computational neuroscientists are “cautiously excited” about new tools like Centaur, adds Katherine Storrs, a computational visual neuroscientist at the University of Auckland. The paper makes some unjustified sweeping claims, she says, but a lot of time and effort has gone into the data set and model, and the work “may end up paying off scientifically in the long run.”

When Azzurra Ruggeri of the Technical University of Munich saw the grant solicitation for bold new ideas in cultural studies, she was eager to submit her proposal to use her expertise as a developmental psychologist to study whether artificial intelligence (AI) could be designed to behave ethically. Only after Ruggeri applied did she notice an unusual condition in the grant competition’s rules: All applicants were required to review others who responded to the same solicitation, letting the funder take advantage of a ready-made pool of relevant experts.

The unconventional approach, called distributed peer review (DPR), has been touted for its potential to alleviate long-standing ailments with traditional reviews of grant applications and journal-article manuscripts, in which a panel of invited outside experts score and debate submissions. The problems include delays in decisions, the growing workload from an ever-rising number of submissions, the risk of bias or groupthink among the panels, and a lack of consistency across reviews. DPR, however, has its own potential shortcomings, including the risk that reviewers might score other applications poorly in hopes of making theirs stand out—a form of gaming that grant funders using the model have sought to deter by checking reviews and other mechanisms.

Now, results from the solicitation Ruggeri participated in—run by the Volkswagen Foundation—and another this year by the funder UK Research and Innovation (UKRI), to choose fellows who will study how AI is shaping science, offer some perspective on what works about DPR, as well as areas for improvement. The UKRI trial, for example, found the strategy cut the time from submission to funding decision by as much as two-thirds, to 2 months, compared with normal panel reviews. And for both grant solicitations, substantial numbers of applicants surveyed were open to participating again in another competition using DPR. Still, many worried whether they and their fellow reviewers possessed the right expertise to do the job well.

The findings on DPR, presented this week at the Metascience 2025 Conference in London, suggest the approach is unlikely to completely overtake traditional panel review. But it could improve reviews for certain kinds of solicitations, analysts say, while also providing data that could help improve grant-funding reviews of all types. “We need to build the evidence base to allow funders to make better choices about how to innovate in the funding system,” says Tom Stafford, a cognitive scientist at the University of Sheffield who helped lead an outside evaluation of the Volkswagen trial, which the funder commissioned.

The Volkswagen Foundation tested DPR using a selective, annual grants solicitation in 2024 for novel, interdisciplinary research ideas in cultural studies and the humanities. Applications were limited to a page and a half, phrased in terms understandable to academics outside the applicants’ discipline, which made the review burden less onerous. Grants were awarded based on the average score from other applicants, with each application receiving nine or 10 reviews. The “wisdom of the crowd” approach was well suited for this type of solicitation, Ruggeri says, “because it’s all about the idea.”

Acknowledging that the unorthodox peer-review experiment might deter some applications, the foundation ran all applications through a parallel traditional panel review and made 10 awards—up to €400,000 for 18 months—through each track. Of the 140 applications, three qualified for funding under both peer-review methods. (Each successful applicant was only given a single award.) Although that overlap may appear small, the scores from DPR and the traditional panel approach were similar enough to suggest that the differently constituted groups of reviewers applied the solicitation’s criteria in comparable ways, says Stafford, who joined colleagues at the Research on Research Institute (RoRI) to evaluate the two methods. For a competition so selective, and given the known subjectivity of peer reviews, greater overlap wouldn’t be realistically expected, he tells Science .

DPR shouldn’t be judged only on how well it replicates the traditional panel method but also on its own merits, Stafford says. Among its advantages is democratizing reviews by enabling more early-career scientists to participate. In the end, Volkswagen’s DPR track had 323 reviewers, including co-applicants; the traditional one, only eight. Spreading the reviewer burden more broadly is important, Ruggeri says. “There are all those free riders in the world submitting tons of papers and doing zero reviews, not really contributing to the community effort.”

The Volkswagen evaluators also performed a statistical analysis indicating that increasing the number of reviewers per application raises the consistency of the applications’ scores across reviewers, although no number emerged as an ideal cap. “DPR is the only way to get this kind of data, because most funders wouldn’t usually be getting 10 reviews per application,” says Ben Steyn, co-head of the UKRI’s Metascience Unit, which conducted that agency’s trial, describing his conference presentation.

To date, few other funders have taken up DPR, and only on a limited basis. They include the Dutch Research Council and the European Southern Observatory, which uses the strategy to vet the many applications it receives for telescope observing time. More than a decade ago, a trial of DPR by the U.S. National Science Foundation found it saved time and money while increasing the quality of reviews. But proposals to expand the experiment never gained traction among the agency’s leaders, says George Hazelrigg, a former NSF program officer who oversaw the effort.

Still, many researchers who participated in the recent grant solicitations were open to trying DPR again. For the Volkswagen call, 83% of applicants who won funding, and 60% who didn’t, were game—though, as Ruggeri notes, keeping the burden manageable is key. If the applications had been much longer, she says, “I’m not sure [doing the reviews] would have been worth the opportunity” for funding. (As it turned out, her application was funded.) For the UKRI trial, 36% said they would participate again. The U.K. funder surveyed applicants after they completed reviews but hasn’t shared reviews with applicants or announced grant decisions; without having seen the reviews for their own work, participants may not yet know whether to trust the method, Steyn says.

For its part, the Volkswagen Foundation plans to continue assessing DPR for a new round of the same funding program in the coming year, says Hanna Denecke, a program officer there. But she says it will be difficult for DPR to displace panels except for certain funding calls. The foundation considered but decided against using DPR for a call this year for proposals to study neurodegenerative diseases , in part because of concerns over preserving confidentiality. “Some people might not be willing to apply if they know that 10 people will read their proposal,” she says.

And many scientists adhere to long-standing conventions that favor panel review, Denecke says. “It’s still the gold standard.”

Less than 15 months into a scheduled 5-year mission, a pioneering satellite built to track rogue emissions of planet-warming methane has been lost.

The demise of MethaneSAT was announced today by the Environmental Defense Fund (EDF), the nonprofit behind the $88 million satellite. “Not one of my better days or weeks,” says Steven Hamburg, EDF’s chief scientist and leader of the mission.

Scientists have been unable to communicate with the satellite since 20 June, when it unexpectedly went silent. There were no previous indications of a problem, Hamburg says, and the cause is still under investigation.

The satellite’s failure leaves a gap in the ability to monitor emissions of a potent greenhouse gas, which over 20 years has 80 times the warming power of the same amount of carbon dioxide. Methane is emitted by natural sources, such as wetlands, but also by leaky oil and gas infrastructure. Stanching those leaks is an efficient way to slow global warming , many researchers argue, and MethaneSAT was developed specifically to identify them.

Some existing satellites, such as the European Space Agency’s Sentinel-5, can map methane on broader scales across hundreds of kilometers. Others can pinpoint large individual polluters such as a refinery. But MethaneSAT, funded with the help of a $100 million grant from Jeff Bezos’s Earth Fund, was unique in its ability to detect smaller emissions across entire oil and gas fields while also zeroing in on hot spots with a resolution of roughly four soccer fields. It could detect methane concentrations as low as 2 parts per billion, Hamburg says.

The satellite was also unusual because it was developed and operated by a nonprofit, rather than a for-profit company or government, says Paul Wennberg, an atmospheric scientist at the California Institute of Technology who served on a scientific advisory board during the development of the satellite. “It really was the demonstration of a pathway both from the perspective of who could guide this [work] but also how you would do it,” adds Wennberg, who recently became a paid senior contributing scientist for EDF.

Although the satellite’s lifespan proved short, Wennberg predicts its legacy will be longer. Data from its first year are still being processed, potentially revealing unknown leaks. The algorithms built to convert its observations into emissions estimates for particular sources could be adapted for use by other satellites, such as Japan’s GOSAT-GW, which was launched on 28 June, Wennberg says. And it heralds a future where more groups launch satellites to monitor greenhouse gas releases. The nonprofit Carbon Mapper, for example, is part of a group that last year launched the first of four planned satellites capable of tracking carbon emissions at a very fine resolution.

But Riley Duren, a remote sensing scientist and CEO of Carbon Mapper, says those satellites won’t match MethaneSAT’s wider angle view. “The loss of MethaneSAT, make no mistake, represents a gap in our community’s ability to monitor and quantify methane emissions. And it’s a pretty big gap,” Duren says.

Is there a possibility that EDF would launch a second satellite? Hamburg can’t say. “We’re going to take a pause. Obviously, we’ve suffered a loss,” he says. “I have a large team of people who have put their heart and soul into what many people said was impossible.”

More than 2 billion people worldwide lack access to clean drinking water, with global warming and competing demands from farms and industry expected to worsen shortages. But the skies may soon provide relief, not in the form of rain but humidity, sucked out of the air by “atmospheric water harvesters.” The devices have existed for decades but typically are too expensive, energy-hungry, or unproductive to be practical.

Now, however, two classes of materials called hydrogels and metal-organic frameworks have touched off what Evelyn Wang, a mechanical engineer at the Massachusetts Institute of Technology (MIT), calls “an explosion of efforts related to atmospheric water harvesting.”

So far, none of the devices can compete with established approaches to augment water supplies, such as desalinating seawater. But some applications—cooling data centers and slaking the thirst of soldiers on the move—could support higher costs until the technology scales up, says Samer Taha, CEO of Atoco, a California-based startup. “There are many applications where atmospheric water harvesting can help.”

Water capture technology may date back to the Inca who, living on the desert coast of South America, are thought to have collected morning dew on mesh nets, feeding it into cisterns. More recently, companies have deployed devices that use air-conditioners to cool air below the dew point, causing water vapor to condense, or water-absorbing desiccant materials such as salts, which are then heated to release the liquid. But both approaches require lots of energy, raising costs and limiting their reach.

The trick is to find a material that captures lots of water but readily frees it, too. Hydrogels—soft, porous networks of polymer fibers often impregnated with salts—seem to fit the bill. In a June report in Nature Water , Xuanhe Zhao, a mechanical engineer at MIT, and his colleagues describe a water harvester that, thanks to a novel hydrogel, requires no external energy input at all.

The team sandwiched the hydrogel, which contains lithium-chloride salt, between two glass sheets. At night, water vapor enters the gel and is trapped by the salts. During the day, sunlight heats the gel, evaporating the water. The vapor condenses on the glass panels, forming droplets that trickle down and are captured. So far, the results are modest: Prototypes can produce up to 1.2 liters of water per kilogram of hydrogel per day in the dry desert air of Death Valley, California.

Other researchers are using modest amounts of energy and dirt-cheap materials to harvest much more water. For example, in a February report in Advanced Materials , Guihua Yu, a chemist at the University of Texas at Austin, and his colleagues describe a promising hydrogel made by altering cellulose, chitosan, and starch—complex carbohydrates common in agricultural and food waste. The biomaterials have a dense structure that limits the amount of water they can store, and they tend to hold onto much of what they snag, even when heated.

So Yu’s team modified its hydrogel with chemical compounds known as zwitterionic groups that repel one another, stretching open the carbohydrates and making them more porous. Then the researchers added other compounds that cause the hydrogel to shrink when heat is applied, helping to squeeze trapped water out. Together, these changes enabled a prototype device to harvest up to 14 liters of water per kilogram of hydrogel daily. For now, getting the water out still requires raising temperatures to 60°C with an electric heater. However, Yaxuan Zhao, a graduate student in Yu’s lab, says the low cost of the biomaterials mean the device could be deployed along with solar panels in off-grid communities and emergency relief efforts.

Jeremy Cho, a mechanical engineer at the University of Nevada, Las Vegas, and his colleagues believe dividing a water capture device into two layers can help keep energy costs down. They use a hydrogel membrane containing salts that attract water vapor. Once concentrated, the water is pulled farther into a salty liquid desiccant layer for storage. The process empties the pores in the hydrogel, freeing it up to capture more water. Releasing the water from the desiccant takes only a modest amount of heat. “It’s a lot easier to heat a liquid than a solid,” Cho says, which raises efficiency. According to a October 2024 report in the Proceedings of the National Academy of Sciences , the setup could collect nearly 17 liters of water per day for each kilogram of absorbing material in humid environments, and a still respectable 5.5 liters in a Las Vegas-type arid environment.

Some of the most productive devices rely on a different kind of material: metal organic frameworks (MOFs). These porous atomic scaffolds have channels and pockets that can be designed to attract and store specific molecules—in this case, water. Although MOFs tend to hold less water than hydrogels, they can capture and release it more quickly, allowing them to go through dozens of such cycles in the time it takes hydrogels to go through one.

The key is to tailor them with alternating chemical groups that attract and repel water, says University of California, Berkeley chemist Omar Yaghi. In 2023 he and his colleagues reported an aluminum-based MOF that was cheap to make in bulk and that could wring water from desert air. In preliminary, unpublished tests, Yaghi says, prototype devices using a tweaked version of his team’s MOF can produce 200 liters of water per kilogram per day with only small amounts of added heat.

Yaghi has licensed the technology to Atoco, which is exploring using it to generate water to cool data centers, harnessing their waste heat to speed the cycling. Atoco plans to open pilot scale facilities in Texas and Arizona next year to test scaled-up versions, Taha says.

Despite all of these emerging solutions, “there is still room for improvement,” says Cody Friesen, an atmospheric water harvesting pioneer at Arizona State University. Today, he notes, desalination plants can convert large amounts of seawater to drinking water at a cost of less than 1 cent per liter. Water harvesting devices are orders of magnitude more expensive and nowhere near as prolific. With the community still sorting through various materials options and device designs, “We are not all singing out of the same hymn book,” he says.

But Friesen’s own company, Source Global, is an encouraging example. It has installed water-producing “hydropanels,” which use a proprietary desiccant, at more than 450 sites worldwide, mostly in remote, off-grid locations. And Friesen believes costs will drop as manufacturing is scaled up, much as has happened with solar panels and batteries.

“Atmospheric water harvesting will eventually be the lowest cost delivered potable water on the planet,” he predicts.