China’s cities are playing a key role in the development of specialist technologies such as solar energy.Credit: Yaorusheng/Getty
Many of the patterns evident in the data for this year’s Nature Index Science Cities supplement will be familiar to watchers of global science trends over the past decade. China’s research output in the journals tracked by the Nature Index continues to grow strongly, demonstrated by Beijing extending its lead at the summit of the science cities ranking to almost double the Share of the second-placed city. The fact that this second place is now taken by Shanghai, pushing New York into third, only reinforces this trajectory.
Perhaps a more interesting development in the science cities data this year is the clear emergence of China’s provincial capitals. From Chengdu in the west, to Hefei in the east, these cities — lesser known in the West — are now rubbing shoulders in the top 30 with long-established scientific centres in Europe and North America.
Nature Index 2024 Science cities
The rise of these locations as globally competitive centres for research is as much due to economics and politics as it is science, as China seeks to spread the impact of its knowledge and innovation far and wide. Many of these rapidly developing cities are playing a specialist role in key technology areas such as electric vehicles and solar energy. And their local goals are very much aligned to national strategies to bolster the country’s economic self-sufficiency, such as ‘Made in China 2025’, a policy designed to shift the country towards knowledge-driven high-tech industries.
Even in the health sciences, an area where Chinese cities still lag behind their Western counterparts, there is evident progress. Within a decade, there is every chance that the leading Science Cities in this field — currently dominated by the dense academic–health-care–industry networks built up over many years in areas such as Boston or London — might be in China, too.
We are pleased to acknowledge the financial support of the Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park in producing this supplement. As always, Nature retains sole responsibility for all editorial content.
This article is part of Nature Index 2024 Science cities, a supplement produced with financial support from the Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park. Nature maintains full independence in all editorial decisions related to the content. For more information about Nature Index, see the homepage.
With its continued focus on sustainability and ecological protection, China is cementing itself as a leader in Earth and environmental sciences. Beijing, Nanjing and Guangzhou, the leading three cities in the subject, respectively, have all recorded increases in adjusted Share between 2022 and 2023. Wuhan’s momentum is particularly strong — the city’s adjusted Share in the subject rose 16.24%, propelling it from 8th place in 2022 to 4th in 2023.
The four US cities in the top 10 have lost ground, each recording a decrease in their adjusted Share between 2022 and 2023. The Los Angeles metropolitan area, in 9th place, had an almost 20% decline in adjusted Share in 2023.
Mirroring a broader trend across all subjects in the Nature Index, the movement of smaller Chinese cities stands out this year. Nanjing, China’s 13th largest city by population size, contributes around 11.6% of the country’s Earth and environmental science research in the index. The city has retained 2nd place in the subject for the past four years.
The regional centres of Guangzhou and Qingdao, known for their water purification and renewable-energy initiatives, have also been improving their outputs, with increases of 106.3% and 179.8%, respectively, in adjusted Share between 2019 and 2023.
In 2019, Beijing’s Share in Earth and environmental sciences was more than double that of Baltimore–Washington, in 2nd place. In 2023, its Share has grown to almost five times that of Baltimore–Washington, which has fallen to 6th place in the subject. Several institutional partnerships have helped to drive a 60.3% increase in Beijing’s adjusted Share over the past five years, positioning it well ahead of its competitors. Among these are collaborations between the University of Chinese Academy of Sciences (UCAS), the Research Center for Eco-Environmental Sciences, Tsinghua University and Peking University.
Source: Nature Index; Data analysis: Aayush Kagathra; Data visualization: Tanner Maxwell and Simon Baker
With a bilateral collaboration score (BCS) of 78.71, the partnership between UCAS and the Research Centre for Eco-Environmental Sciences significantly outperforms other leading academic pairings in the top five cities for the subject. The collaborative efforts of the Shanghai Municipal Bureau of Ecology and Environment and Fudan University, in Shanghai, form the next strongest partnership (BCS 18.75). This, paired with Shanghai’s 94.9% increase in adjusted Share from 2019 to 2023, suggests that the city will continue to develop its capacity in the subject.
Source: Nature Index; Data analysis: Aayush Kagathra; Data visualization: Tanner Maxwell and Simon Baker
This article is part of Nature Index 2024 Science cities, a supplement produced with financial support from the Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park. Nature maintains full independence in all editorial decisions related to the content. For more information about Nature Index, see the homepage.
An analytical chemist works in the lab of biotech company Arcaea, which is based in Boston, Massachusetts.Credit: Boston Globe/Getty
When a delegation of scientists from Japan recently visited Harvard University in Cambridge, Massachusetts, they asked their hosts a familiar question: what are the secret factors that make the Boston area, which includes Cambridge, such a hotbed for health-sciences research and innovation? In response, George Daley, dean of the Faculty of Medicine at Harvard Medical School, gave the half-joking answer he normally uses when asked similar questions: “Just incubate two of the most important educational institutions on the planet, support them for 200 years, and watch the magic happen.”
The Boston area is home to a critical mass of leading universities, hospitals, biotechnology and pharmaceutical companies, and independent research institutions that all interact synergistically, says Dan Barouch, an immunologist at Harvard Medical School and director of the Center for Virology and Vaccine Research at Beth Israel Deaconess Medical Center in Boston. “The quality, depth and sheer breadth and scope of research in Boston is just astounding.”
Nature Index 2024 Science cities
It’s no surprise, then, that the Boston metropolitan area leads the Nature Index Science Cities rankings in health sciences, based on 2023 research output in journals tracked by the database. According to the findings, the New York City metropolitan area ranks second after Boston, followed by the urban area formed by Baltimore and Washington DC; London; the San Francisco Bay Area; Beijing and Shanghai.
Science cities is tracking health sciences for the first time this year after data from journals in the subject were added to the Nature Index in 2022, but already, the data reveal some new trends. For one, US cities and London take the top five positions, whereas for most other tracked disciplines — including chemistry, physical sciences and Earth and environmental sciences — China now dominates the top positions.
Science cities rankings are not adjusted for population size, which means large cities such as Beijing and Shanghai — with populations of 21.5 and 26.3 million, respectively — have strong advantages for research output. But this also highlights the oversized contribution to health-sciences research by smaller leading cities such as Boston, whose greater metropolitan population is just 4.9 million. Boston is clearly “still very dominant in this area”, says Yiming Dong, a Chinese studies researcher at King’s College London. But this could change soon, with Dong emphasizing that China is moving quite quickly in the subject.
A tremendous anthill
Lots of cities around the world have good universities, smart people and some industry and capital for research, but few possess “this alchemy that creates, effectively, gold out of these regular materials”, says Paul Sagan, a senior adviser at General Catalyst, a venture-capital firm founded in Cambridge, Masachusetts. Scale, in terms of a concentration of elite scientific research institutions, and repetition, in terms of spinning out a continuous stream of new ideas — some of which succeed and spawn new biotech companies, are key to transforming a city into a true hub of excellence for science and innovation, Sagan continues. Among such hubs for health sciences and biotechnology, he adds, it’s clear “that Boston has sped ahead of everyone”. There are several probable reasons for this, he continues, including the presence of elite research institutions, start-ups and international companies with headquarters there, and a number of government initiatives over the years that have promoted and supported biotech research.
The Boston metropolitan area contains a familiar list of the leading institutions in the health sciences. Harvard University ranks first in the world in the Nature Index for the subject by a large distance and the leading two health-care facilities — Brigham and Women’s Hospital and Massachusetts General Hospital — are located nearby. Boston’s biotechnology sector is also growing quickly, Daley says, and most of the top pharmaceutical companies have established major research centres there.
A large and growing pot of venture capital also fuels health-sciences innovation in Boston. “Because drug development is so expensive, public research funding will never carry all the costs,” says Andrea Braun Střelcová, who studies science policy and research collaboration, with an emphasis on China, at the Max Planck Institute for the History of Science in Berlin. “So, the role of the market is really important.”
Although California has a strong venture-capital presence, too, “the big difference” for the Boston area is the presence of leading pharmaceutical companies — many of which are just a walk from the Massachusetts Institute of Technology (MIT) and Harvard, says Nobel laureate Phillip Sharp, who holds an emeritus position at MIT’s Koch Institute for Integrative Cancer Research.
The size of Boston’s talent pool is also notable, Daley says. Harvard’s full-time medical faculty alone numbers 10,000-plus — more than three times the size of other large medical schools in the United States. Considering all the other Boston-area health-sciences institutions, “you’ve got tens of thousands of clinicians and scientists working towards common goals in confronting disease and solving fundamental biomedical questions”, Daley says. “That’s just a tremendous anthill of activity all within a very small radius.”
Entrepreneurial and cool
Other top cities for health-sciences research possess the same features that make Boston stand out —only on a smaller scale. The New York City metropolitan area, for example, has Memorial Sloan Kettering Cancer Center, ranked sixth in the world among health-care institutions in the Nature Index for health sciences, and the Mount Sinai Health System, ranked eighth. Experts at many top-ranked institutions collaborate, too, which amplifies their impact and output. In the health sciences, collaborations between Harvard, MIT, Johns Hopkins University in Baltimore and the University of California, San Francisco, are among the most productive in the world, according to Nature Index data.
Like its US counterparts, London also has top-notch universities and strong biotechnology and pharmaceutical industries, says Rebecca Shipley, director of the Academic Health Science Centre at UCLPartners in London — an organization that brings together universities and health-care providers to accelerate the translation of research into improved outcomes. Unlike in the United States, researchers in London can benefit from the United Kingdom’s National Health Service, which operates across the country and makes it easier to obtain patient data and run clinical trials. Shipley predicts that London will continue to hold its spot among the leading five science cities in health sciences and has the potential to rise even higher. For example, the UK National Institute for Health and Care Research, which is the major funder of research to improve the population’s health, has awarded nearly £800 million (US$1 billion) in funding over 5 years to 20 university-hospital research centres around the United Kingdom — seven of which are in London — to translate basic discoveries into real-world patient care. There is also an increasing investment in London and nationally to build infrastructure to make patient data better available for research and innovation, Shipley says. This includes secure access for researchers to NHS patient data on a national level through a specialized platform, as well as a London-specific information-sharing hub called OneLondon that connects health and care staff to patient records, among other things. “There’s a real appetite in London to be innovative and build on this momentum,” Shipley says.
Visitors view a medicine and health exhibition at the 2024 Beijing Science and Technology Week, held in Beijing, China.Credit: NurPhoto/Getty
Indeed, for any sort of innovation hub to take off, there has to be a culture of entrepreneurialism and a mindset of “not being afraid to fail”, Sagan says. To attract and retain talent, the hub itself also must be somewhere that people want to live. “There are great research universities that might have some innovation, like the University of Illinois Urbana-Champaign, but by and large, that’s not a place where people aspire to live because it’s a small town, and small towns are limited, by definition,” Sagan says. “Not to demean small towns, but most ambitious entrepreneurs and researchers want to go to top-tier cities like New York City, Boston, or Silicon Valley because they are places where their partners can also get good jobs, their kids can go to great schools and their community offers great cultural diversity — and it’s just cool to be there.”
Surprising slips
The United States is showing some puzzling trends for health-sciences research output, however. Unlike the Boston metropolitan area, which increased its adjusted Share in the Nature Index by 6.6% from 2022 to 2023, the other leading four US cities lost ground. The San Francisco Bay Area experienced the steepest decline of 13.2%.
One explanation is likely to be that the Nature Index represents a relatively fixed set of research articles. If cities in one part of the world, such as China, are rapidly increasing their Share then others must fall to compensate. This makes Boston’s performance even more remarkable.
Stacie Bloom, the vice-provost for research and chief research officer at New York University, says she is surprised by New York’s results and that “all the messaging we get indicates that things are going in a more positive direction”. Daley also says that his perspective is that the US cities experiencing a drop in adjusted Share remain strong. “New York City has been on fire, and the Baltimore–Washington DC corridor is a hotbed of innovation,” he says. The San Francisco Bay Area also remains Boston’s “main competition” for cutting-edge biotechnology.
Daley adds that another explanation is health-sciences research from 2022 to 2023 was probably still affected by problems linked to the COVID-19 pandemic. The pandemic caused significant supply-chain issues for biomedical materials, he says, and across many industries, including in science, some people changed careers or took a while to return to work. Boston was probably more insulated from these impacts than other US cities, he adds, because of its higher density of people and institutions.
Daley expects that any decline in top US cities’ health-sciences research output will be “a momentary blip”, and that those hubs of innovation will “return to productivity and growth very soon”.
China rising
For now, cities in the United States, alongside London, still lead in health sciences, but experts predict that China will continue to gain ground. Logistically, this makes sense, says Yu-Xuan Lyu, a scientist at the Southern University of Science and Technology in Shenzhen, who studies ageing. It’s only in the past 10 to 15 years that China rapidly expanded its international research presence and rose to the top in natural-science subjects such as chemistry, which don’t require a close collaboration between universities, hospitals and industry. It has taken a bit more time for China to lay the structural groundwork to conduct world-class health-sciences research, but now that that is beginning to take shape, “the conditions are really good for China to start performing even better”, Střelcová says.
Beijing increased its health-sciences research output in the Nature Index by 17.6% between 2022 and 2023, while Shanghai’s contribution rose by nearly 4%. The southern city of Guangzhou, which is currently ranked 12th in the world for health-sciences research, is also growing quickly, with a 32.4% increase in the year to 2023. This growth is largely because health care and health-sciences research are priorities for the Chinese government, Dong says. “They’re spending a huge amount of money on this.” Health-sciences research accounted for 36%, or 97.6 billion yuan (US$13.8 billion), of the 2024 budget for the National Health Commission — an executive department under the State Council that’s responsible for health policies and health-related emergency management in mainland China.
Scientific advancement in health research is a key pillar of the Healthy China 2030 plan, a set of strategic public health goals first published in 2016. The country’s 14th five-year plan — which outlines overall objectives for long-term domestic economic development and innovation — also includes health-sciences goals, including specific plans to address China’s ageing population and improve health care. China’s National Health Commission’s science strategy also highlights similar goals, and the government is additionally investing in studying and developing traditional Chinese medicine. Some of the largest research grants in the health-sciences field in China are currently being given by the Ministry of Science and Technology and other public funders to university–hospital collaborations for translational research in service of these goals, Lyu says.
In 2022, construction also began in Shanghai on the first of a nationwide network of hospitals that are intended to act as comprehensive national medical centres. Some of the people who work there are likely to be expat Chinese scientists who are being attracted back from the United States or other Western countries, Dong says, through more than 100 talent-recruitment programmes operating at the national, provincial and city level, and also by high salaries offered by Chinese universities and research institutions. Many of these experts have left corporate positions abroad or vacated tenured roles at top-tier American universities, Dong says, including Harvard and MIT.
China’s provinces and cities can also introduce their own targeted priorities, and in both Beijing and Shanghai, that includes biosciences, says Glen Noble, the founder and director of Noble Endeavours, a London-based consultancy focused on research and academia in the United Kingdom, European Union and China. Both cities have “huge amounts of leeway and resources” to implement things such as tax breaks, subsidies, talent-recruiting programmes, science parks and research funding, Noble says. This allows health-science researchers to tap into support from multiple initiatives and levels of government.
Collaborations in China between academia and industry have also started “booming” over the past year or so, Lyu says, and grants are specifically set up to encourage and enable these partnerships. China still has issues regarding intellectual property (IP) protections that draw criticism from the United States and the West, Střelcová adds, including concerns over IP theft and economic espionage. On the other hand, she continues, over the past decade or so, China has “improved and professionalized” the IP protection landscape compared with the past, especially through its regulatory framework and enforcement. “The caveat is that the intent is not confined to intellectual-property rights protection itself, but rather to the overall desire to strengthen national security and increase the country’s competitiveness,” Střelcová says. Regardless of the intent, though, this is a boon for innovators, Dong says, because of the size of China’s market.
Regardless of whether Chinese cities do overtake locations in the United States and other Western cities such as London in health-sciences research, Noble hopes that researchers around the world will be able to maintain strong international collaborations despite political tensions. Currently, however, policies around research security in the West “are primarily calibrated around preventing China accessing Western technology — as if China wasn’t already a scientific power in its own right, across many disciplines”, he says. “Increasingly, we need the science happening in China to be disseminated back to us in the West.”
Researcher Yan Zeng looks over the machinery at the A-Lab, a fully automated laboratory at the Lawrence Berkeley National Laboratory in California.Credit: Marilyn Sargent/Berkeley Lab
Stephan Noack’s official title is bioprocess engineer. In simple terms, he is a problem-solver. His colleagues at the Jülich Research Centre in Germany knock on his office door armed with some of their thorniest questions about the process of coaxing bacteria, algae and other microbes into mass-producing valuable chemicals, such as ethanol and amino acids. Optimizing such processes requires making tiny adjustments to several variables, including the microbes’ food source and growing temperature. It’s trial and error — mostly error. “During the set-up of these workflows, a lot of failure happens,” Noack says.
Even the most efficient laboratories, with an ample amount of students to conduct the trials, can fail to complete the lengthy, laborious process. “It was a huge bottleneck,” he says. Noack and his engineers have therefore been turning to robotics and automation to speed up the process of growing microorganisms on plates of gelatinous agar. By combining a range of equipment, from robotic arms to liquid handlers, researchers have been able to swap out large single plates of agar for ones containing 96 or 384 tiny wells. This has increased throughput nearly 100-fold according to Noack.
Nature Outlook: Robotics and artificial intelligence
Although they are common in large industrial research facilities, robotics and automation have only begun to trickle down into smaller academic labs in the past five years, says Ian Holland, a postdoctoral researcher at the University of Edinburgh, UK. Historically, he says, academia has relied on large populations of students and postdocs to do the time-consuming work. But with scientific advances requiring ever-increasing amounts of data generation and analysis, lab workers can’t work quickly enough. But robots can.
The advances include robotic arms that can pipette more accurately than can human scientists1 and fully automated ‘cloud’ labs that experimenters can access online and command a robot workforce to perform their instructions from anywhere in the world2. Researchers who are leaning towards automation hope that the shift will decrease cost, save time and generate fewer errors while improving reproducibility.
But these changes don’t come without challenges. Scientists need a deep understanding of their experiments to program machinery and to prevent the propagation of errors. The equipment can be expensive and require hours of labour to fix and maintain. If done correctly, however, laboratory automation can transform science, according to Dennis Knobbe, a roboticist at the Technical University of Munich in Germany. “It’s not about excluding the human from these processes,” Knobbe says. “It’s instead about using robotics to enhance researchers’ capabilities.”
Rise of the machines
In 2012, Matheus Carvalho, a research technician and fisheries biologist at the Southern Cross University in Lismore, Australia, encountered AutoIt, a programming language originally created for automating Microsoft Windows tasks. Around the same time, he came across a toy robotic arm that could be controlled through a computer. Carvalho reasoned that if he could combine the toy robot with AutoIt, he could automate some of his tedious sampling tasks in the lab. Although the first robotic arm broke almost immediately, Carvalho convinced his supervisor to purchase a higher-quality, second-hand arm, which was built into an automated sampling machine that continues to operate more than a decade later. Carvalho was quickly sold on the idea of laboratory automation, which is the topic of a book he published in 2017.
Matheus Carvalho at the Southern Cross University in Lismore, Australia, created an automated sampling machine using a reprogrammed toy robotic arm.Credit: Matheus Carvalho de Carvalho
He aimed to automate more lab procedures without precluding human involvement. His lab used non-radioactive isotopes to understand organic material in water samples — a process that requires weighing and measuring tiny amounts of powders, often to a fraction of a milligram. Every powder they tested had a different grain size and texture, which made it impossible to program a robot to measure out all the samples. Instead, Carvalho devised a protocol that allowed people and machines to each do what they were best at: a human lab technician weighed out the powder samples, and a small, mobile robot was programmed to retrieve containers and calibrate the scales. “It’s better to automate what is easy but leave the hard parts for us humans,” Carvalho says.
In the 2010s, Dina Zielinski, who was then a technician at the Whitehead Institute in Cambridge, Massachusetts, faced similar challenges with automation while working on a different type of test. She wanted to sequence tissue samples from people with Parkinson’s disease to understand the genes contributing to the condition. The job required pipetting — a lot of pipetting. Zielinski saw the task in front of her as a fast track to repetitive strain injury.
“Molecular biology essentially entails combining minuscule clear volumes with other miniscule clear volumes,” Zielinski says. “If you didn’t combine the right tiny volumes, you would have wasted a ton of money on sequencing.”
Even worse, she says, these samples were rare and hard to obtain. Yaniv Erlich, who was then a principal investigator, and his late collaborator Susan Lindquist, a biomedical researcher at the Whitehead Institute, began investigating various robotics, including automated liquid handlers, to speed up the process and to save Zielinski’s hands from injury. But none of the robots they investigated could provide both the precision and flexibility that the lab needed. So, Zielinski, Lindquist and Erlich, who is now chief executive of Eleven Therapeutics in Cambridge, UK, decided to build something different.
The iPipet app can be used to illuminate sections of a 96-well plate and help researchers to ensure they combine the correct samples.Credit: Dina Zielinski
The idea they came up with didn’t handle the pipetting itself. Instead, the team built an iPad app that users could program to help them pipette the correct samples into the correct position. The iPipet app illuminates sections of 96- or 384-well plates to enable a scientist to ensure they combine the correct samples3. When Zielinski pitted a researcher using iPipet against a top-of-the-line robot, the app-assisted human was the clear winner. “The error was much lower with human pipetting than with the liquid-handling robot,” she says.
Mind the gap
What makes efforts such as these so trailblazing isn’t their complexity but rather their simplicity. The goal is to find a middle ground between the expensive instruments that can perform every aspect of an experiment and the labour of a single student performing all their tasks manually, Holland says. Ideally, such technology would make it possible for researchers to spend time planning experiments and analysing results instead of pipetting samples.
“If automation can take some of the load off you, you can do more things and be a better researcher,” Holland says. And the academic environment is well-suited for this melding of human and machine. “You’ve got engineering students looking for projects and we’ve got biologists who have problems that need solving.”
However, the changes come at a cost says Holland. Since the dawn of the industrial revolution, people have invested time, resources and money into developing machinery to make products more quickly and cheaply. In commercial settings the benefits were clear, says Holland — investment in automation paid off because it allowed production of more commodities with low labour costs.
Academia is different. Industry focuses on profit, whereas academic labs place a greater emphasis on training the next generation of scientists and producing knowledge. A steady flow of students who are willing to work long hours — some of whom have their own grants and stipends as salary — means that labour costs aren’t as important. What’s more, the focus on teaching and training means that many scientists have conventionally seen automation as anathema to their mission as educators.
Postdoc researcher Julia Tenhaef and bioprocess engineer Stephan Noack at the Jülich Research Centre in Germany use an automated laboratory system called the AutoBioTech platform. Credit: Stephan NoackCredit: Stephan Noack
“In academia, you could spend US$100,000 on this machine, but it’s only going to make your output a bit faster,” says Holland. “That’s a lot harder to justify.” As a result, many academic labs have much less robotic equipment than do commercial and industrial labs — something Holland refers to as the automation gap4.
Joshua Pearce faced down these technological costs when he founded his lab at Michigan Technological University in Houghton in the mid-2000s. Now an engineer at Western University in London, Canada, Pearce was developing methods to build better photovoltaic systems to generate electricity from sunlight. He wanted to improve solar cells’ ability to absorb different wavelengths of light, but the automated filter wheel changer, which adjusted the wavelengths on his custom-built machine, broke. The replacement was $2,500 (an exorbitant price for a simple part) and had a five-month lead time.
Pearce realized that he was at a university filled with budding engineers, so he hired some students to help him 3D print the necessary components. What resulted was a bespoke device crafted entirely from open-source hardware and software that cost $50 and did exactly what Pearce needed it to. “It was something that wasn’t available on the market,” Pearce says. “You can make really high-end equipment, exactly what you want, and do it fairly easily for extremely low cost.”
With his equipment that could automatically adjust light wavelengths for his tests, Pearce began campaigning about the potential of open-source design as a cost-effective way to reap the benefits of lab automation5. He is now editor-in-chief of the journal HardwareX, a publication that allows researchers to share their code and blueprints — while also helping to bolster their CVs and tenure qualifications.
Pearce’s experiences challenge the idea that investing in automation hampers a scientist’s ability to train students, along with the opinion that robotics are prohibitively expensive.
Plain and simple
When it comes to the future of lab robotics, Knobbe thinks that inventions such as those created by Pearce, Carvalho and Zielinski will be key: modular, multipurpose and budget-friendly. “We don’t want to just build a huge machine, like an encapsulated system,” Knobbe says. “We want to integrate these robotic systems into everyday laboratories.”
He also imagines fully fledged robotic lab assistants that can perform basic experimental tasks with minimal supervision. Although this technology is nowhere near ready, Knobbe says, he thinks researchers will be able to deploy modular automated systems that can interact with each other and be controlled by a robotic assistant in the next ten years. One of the biggest challenges will be balancing robustness, flexibility, the ability to detect errors and asking for help.
A robotic lab assistant in Dennis Knobbe’s lab at the Technical University of Munich in Germany uses finger-like appendages to pipette autonomously. Credit: Dennis Knobbe/TUM, 2024
Building or buying a top-of-the-line machine that only does pipetting would force lab technicians to work around the machine. Knobbe wanted a robot that would work with his team, follow basic commands and scan the environment for obstacles. He is therefore building a robotic pipette with finger-like appendages. Early testing shows that this machine has met industry standards, he says.
Although reducing variability and mistakes has long been one of the selling points of robotics and automation, Knobbe says that robots can also propagate errors1,4. Knobbe also speculates that robots might create types of catastrophic failure.
A cautionary tale emerged in November last year, when a team of scientists from Google DeepMind in London, the University of California, Berkeley, and the Lawrence Berkeley National Laboratory in California teamed up to predict nearly 400,000 new compounds using artificial intelligence (AI) and then to synthesize these compounds in a fully automated laboratory, called A-Lab. The project was an endeavour to identify new high-performance, low-cost materials by automating both the physical synthesis of compounds and their subsequent analysis. A resulting Nature paper6 seemed to showcase the benefits of automation.
“It was a high-risk, high-reward project,” says co-author Yan Zeng, a former researcher at the Lawrence Berkeley National Laboratory who started her own lab at Florida State University in Tallahassee this year. “It was a little bit crazy, to be fully automated.”
Several weeks later, however, some scientists began raising questions about the AI’s ability to predict truly new materials. What seemed to be new in the computer’s modelling might have been different versions of known compounds. “This paper did not at all live up to its claims,” says Leslie Schoop, a chemist at Princeton University in New Jersey.
To Zeng, however, the study was as much about the process — demonstrating how such a system could be built, operated and used by materials scientists — as it was about the results. In fact, Zeng says, the robotic synthesis aspects of the study performed exactly as expected. She concedes that the initial programming steps took months and required a team of technicians to troubleshoot the process. But they quickly recouped the lost time as the robots required minimal human contact.
Zeng is now working to automate parts of her lab in Florida. Her first target is hydrothermal synthesis — a process that requires high temperatures and pressurized tubes. It’s a complex project, but her time at Berkeley gave her valuable experience in breaking down complex robotics into more manageable steps, and she hopes to begin automating this process as she scales up her lab.
Despite the scepticism over A-Lab, she remains optimistic about automation. Robotics could provide the key to future breakthroughs, she says, equipping researchers with the freedom and flexibility to think up the experiments of tomorrow. “This is a rising field, and it’s rising up pretty fast,” says Zeng.
When Sam Rodriques was a neurobiology graduate student, he was struck by a fundamental limitation of science. Even if researchers had already produced all the information needed to understand a human cell or a brain, “I’m not sure we would know it”, he says, “because no human has the ability to understand or read all the literature and get a comprehensive view.”
Five years later, Rodriques says he is closer to solving that problem using artificial intelligence (AI). In September, he and his team at the US start-up FutureHouse announced that an AI-based system they had built could, within minutes, produce syntheses of scientific knowledge that were more accurate than Wikipedia pages1. The team promptly generated Wikipedia-style entries on around 17,000 human genes, most of which previously lacked a detailed page.
How AI-powered science search engines can speed up your research
Rodriques is not the only one turning to AI to help synthesize science. For decades, scholars have been trying to accelerate the onerous task of compiling bodies of research into reviews. “They’re too long, they’re incredibly intensive and they’re often out of date by the time they’re written,” says Iain Marshall, who studies research synthesis at King’s College London. The explosion of interest in large language models (LLMs), the generative-AI programs that underlie tools such as ChatGPT, is prompting fresh excitement about automating the task.
Some of the newer AI-powered science search engines can already help people to produce narrative literature reviews — a written tour of studies — by finding, sorting and summarizing publications. But they can’t yet produce a high-quality review by themselves. The toughest challenge of all is the ‘gold-standard’ systematic review, which involves stringent procedures to search and assess papers, and often a meta-analysis to synthesize the results. Most researchers agree that these are a long way from being fully automated. “I’m sure we’ll eventually get there,” says Paul Glasziou, a specialist in evidence and systematic reviews at Bond University in Gold Coast, Australia. “I just can’t tell you whether that’s 10 years away or 100 years away.”
At the same time, however, researchers fear that AI tools could lead to more sloppy, inaccurate or misleading reviews polluting the literature. “The worry is that all the decades of research on how to do good evidence synthesis starts to be undermined,” says James Thomas, who studies evidence synthesis at University College London.
Computer-assisted reviews
Computer software has been helping researchers to search and parse the research literature for decades. Well before LLMs emerged, scientists were using machine-learning and other algorithms to help to identify particular studies or to quickly pull findings out of papers. But the advent of systems such as ChatGPT has triggered a frenzy of interest in speeding up this process by combining LLMs with other software.
AI science search engines are exploding in number — are they any good?
It would be terribly naive to ask ChatGPT — or any other AI chatbot — to simply write an academic literature review from scratch, researchers say. These LLMs generate text by training on enormous amounts of writing, but most commercial AI firms do not reveal what data the models were trained on. If asked to review research on a topic, an LLM such as ChatGPT is likely to draw on credible academic research, inaccurate blogs and who knows what other information, says Marshall. “There’ll be no weighing up of what the most pertinent, high-quality literature is,” he says. And because LLMs work by repeatedly generating statistically plausible words in response to a query, they produce different answers to the same question and ‘hallucinate’ errors — including, notoriously, non-existent academic references. “None of the processes which are regarded as good practice in research synthesis take place,” Marshall says.
A more sophisticated process involves uploading a corpus of pre-selected papers to an LLM, and asking it to extract insights from them, basing its answer only on those studies. This ‘retrieval-augmented generation’ seems to cut down on hallucinations, although it does not prevent them. The process can also be set up so that the LLM will reference the sources it drew its information from.
This is the basis for specialized, AI-powered science search engines such as Consensus and Elicit. Most companies do not reveal exact details of how their systems work. But they typically turn a user’s question into a computerized search across academic databases such as Semantic Scholar and PubMed, returning the most relevant results.
An LLM then summarizes each of these studies and synthesizes them into an answer that cites its sources; the user is given various options to filter the work they want to include. “They are search engines first and foremost,” says Aaron Tay, who heads data services at Singapore Management University and blogs about AI tools. “At the very least, what they cite is definitely real.”
These tools “can certainly make your review and writing processes efficient”, says Mushtaq Bilal, a postdoctoral researcher at the University of Southern Denmark in Odense, who trains academics in AI tools and has designed his own, called Research Kick. Another AI system called Scite, for example, can quickly generate a detailed breakdown of papers that support or refute a claim. Elicit and other systems can also extract insights from different sections of papers — the methods, conclusions and so on. There’s “a huge amount of labour that you can outsource”, Bilal says.
Elicit, like several AI-powered tools, aims to help with academic literature reviews by summarising papers and extracting data.Credit: Nature
But most AI science search engines cannot produce an accurate literature review autonomously, Bilal says. Their output is more “at the level of an undergraduate student who pulls an all-nighter and comes up with the main points of a few papers”. It is better for researchers to use the tools to optimize parts of the review process, he says. James Brady, head of engineering at Elicit, says that its users are augmenting steps of reviewing “to great effect”.
Another limitation of some tools, including Elicit, is that they can only search open-access papers and abstracts, rather than the full text of articles. (Elicit, in Oakland, California, searches about 125 million papers; Consensus, in Boston, Massachusetts, looks at more than 200 million.) Bilal notes that much of the research literature is paywalled and it’s computationally intensive to search a lot of full text. “Running an AI app through the whole text of millions of articles will take a lot of time, and it will become prohibitively expensive,” he says.
Full-text search
For Rodriques, money was in plentiful supply, because FutureHouse, a non-profit organization in San Francisco, California, is backed by former Google chief executive Eric Schmidt and other funders. Founded in 2023, FutureHouse aims to automate research tasks using AI.
Could AI help you to write your next paper?
This September, Rodriques and his team revealed PaperQA2, FutureHouse’s open-source, prototype AI system1. When it is given a query, PaperQA2 searches several academic databases for relevant papers and tries to access the full text of both open-access and paywalled content. (Rodriques says the team has access to many paywalled papers through its members’ academic affiliations.) The system then identifies and summarizes the most relevant elements. In part because PaperQA2 digests the full text of papers, running it is expensive, he says.
The FutureHouse team tested the system by using it to generate Wikipedia-style articles on individual human genes. They then gave several hundred AI-written statements from these articles, along with statements from real (human-written) Wikipedia articles on the same topic, to a blinded panel of PhD and postdoctoral biologists. The panel found that human-authored articles contained twice as many ‘reasoning errors’ — in which a written claim is not properly supported by the citation — than did ones written by the AI tool. Because the tool outperforms people in this way, the team titled its paper ‘Language agents achieve superhuman synthesis of scientific knowledge’.
The team at US start-up FutureHouse, which has launched AI systems to summarize scientific literature. Sam Rodriques, their director and co-founder, is on the chair, third from right.Credit: FutureHouse
Tay says that PaperQA2 and another tool called Undermind take longer than conventional search engines to return results — minutes rather than seconds — because they conduct more-sophisticated searches, using the results of the initial search to track down other citations and key phrases, for example. “That all adds up to being very computationally expensive and slow, but gives a substantially higher quality search,” he says.
Systematic challenge
Narrative summaries of the literature are hard enough to produce, but systematic reviews are even worse. They can take people many months or even years to complete2.
A systematic review involves at least 25 careful steps, according to a breakdown from Glasziou’s team. After combing through the literature, a researcher must filter their longlist to find the most pertinent papers, then extract data, screen studies for potential bias and synthesize the results. (Many of these steps are done in duplicate by another researcher to check for inconsistencies.) This laborious method — which is supposed to be rigorous, transparent and reproducible — is considered worthwhile in medicine, for instance, because clinicians use the results to guide important decisions about treating patients.
ChatGPT-like AIs are coming to major science search engines
In 2019, before ChatGPT came along, Glasziou and his colleagues set out to achieve a world record in science: a systematic review in two weeks. He and others, including Marshall and Thomas, had already developed computer tools to reduce the time involved. The menu of software available by that time included RobotSearch, a machine-learning model trained to quickly identify randomized trials from a collection of studies. RobotReviewer, another AI system, helps to assess whether a study is at risk of bias because it was not adequately blinded, for instance. “All of those are important little tools in shaving down the time of doing a systematic review,” Glasziou says.
The clock started at 9:30 a.m. on Monday 21 January 2019. The team cruised across the line at lunchtime on Friday 1 February, after a total of nine working days3. “I was excited,” says epidemiologist Anna Mae Scott at the University of Oxford, UK, who led the study while at Bond University; everyone celebrated with cake. Since then, the team has pared its record down to five days.
Could the process get faster? Other researchers have been working to automate aspects of systematic reviews, too. In 2015, Glasziou founded the International Collaboration for the Automation of Systematic Reviews, a niche community that, fittingly, has produced several systematic reviews about tools for automating systematic reviews4. But even so, “not very many [tools] have seen widespread acceptance”, says Marshall. “It’s just a question of how mature the technology is.”
Elicit is one company that says its tool helps researchers with systematic reviews, not just narrative ones. The firm does not offer systematic reviews at the push of a button, says Brady, but its system does automate some of the steps — including screening papers and extracting data and insights. Brady says that most researchers who use it for systematic reviews have uploaded relevant papers they find using other search techniques.
Systematic-review aficionados worry that AI tools are at risk of failing to meet two essential criteria of the studies: transparency and reproducibility. “If I can’t see the methods used, then it is not a systematic review, it is simply a review article,” says Justin Clark, who builds review automation tools as part of Glasziou’s team. Brady says that the papers that reviewers upload to Elicit “are an excellent, transparent record” of their starting literature. As for reproducibility: “We don’t guarantee that our results are always going to be identical across repeats of the same steps, but we aim to make it so — within reason,” he says, adding that transparency and reproducibility will be important as the firm improves its system.
Specialists in reviewing say they would like to see more published evaluations of the accuracy and reproducibility of AI systems that have been designed to help produce literature reviews. “Building cool tools and trying stuff out is really good fun,” says Clark. “Doing a hardcore evaluative study is a lot of hard work.”
Audit AI search tools now, before they skew research
Earlier this year, Clark led a systematic review of studies that had used generative AI tools to help with systematic reviewing. He and his team found only 15 published studies in which the AI’s performance had been adequately compared with that of a person. The results, which have not yet been published or peer reviewed, suggest that these AI systems can extract some data from uploaded studies and assess the risk of bias of clinical trials. “It seems to do OK with reading and assessing papers,” Clark says, “but it did very badly at all these other tasks”, including designing and conducting a thorough literature search. (Existing computer software can already do the final step of synthesizing data using a meta-analysis.)
Glasziou and his team are still trying to shave time off their reviewing record through improved tools, which are available on a website they call the Evidence Review Accelerator. “It won’t be one big thing. It’s that every year you’ll get faster and faster,” Glasziou predicts. In 2022, for instance, the group released a computerized tool called Methods Wizard, which asks users a series of questions about their methods and then writes a protocol for them without using AI.
Rushed reviews?
Automating the synthesis of information also comes with risks. Researchers have known for years that many systematic reviews are redundant or of poor quality5, and AI could make these problems worse. Authors might knowingly or unknowingly use AI tools to race through a review that does not follow rigorous procedures, or which includes poor-quality work, and get a misleading result.
By contrast, says Glasziou, AI could also encourage researchers to do a quick check of previously published literature when they wouldn’t have bothered before. “AI may raise their game,” he says. And Brady says that, in future, AI tools could help to flag and filter out poor-quality papers by looking for telltale signs such as P-hacking, a form of data manipulation.
Glasziou sees the situation as a balance of two forces: AI tools could help scientists to produce high-quality reviews, but might also fuel the rapid generation of substandard ones. “I don’t know what the net impact is going to be on the published literature,” he says.
Some people argue that the ability to synthesize and make sense of the world’s knowledge should not lie solely in the hands of opaque, profit-making companies. Clark wants to see non-profit groups build and carefully test AI tools. He and other researchers welcomed the announcement from two UK funders last month that they are investing more than US$70 million in evidence-synthesis systems. “We just want to be cautious and careful,” Clark says. “We want to make sure that the answers that [technology] is helping to provide to us are correct.”
Carl Laflamme knew what protein he wanted to study, but not where to find it. It is encoded by a gene called C9ORF72, which is mutated in some people with the devastating neurological condition motor neuron disease, also known as amyotrophic lateral sclerosis. And Laflamme wanted to understand its role in the disease.
When he started his postdoctoral fellowship at the Montreal Neurological Institute-Hospital in Canada, Laflamme scoured the literature, searching for information on the protein. The problem was that none of the papers seemed to agree where in the cell this mysterious molecule operates. “There was so much confusion in the field,” Laflamme says.
He wondered whether a reagent was to blame, in particular the antibodies that scientists used to measure the amount of the protein and track its position in the cell. So, he and his colleagues decided to test the antibodies that were available. They identified 16 commercial antibodies that were advertised as able to bind to the protein encoded by C9ORF72. When the researchers put them through their paces, only three performed well — meaning that the antibodies bound to the protein of interest without binding to other molecules. But not one published study had used these antibodies. About 15 papers described experiments using an antibody that didn’t even bind the key protein in Laflamme’s testing. And those papers had been collectively cited more than 3,000 times1.
Serious errors plague DNA tool that’s a workhorse of biology
Laflamme’s experience isn’t unusual. Scientists have long known that many commercial antibodies don’t work as they should — they often fail to recognize a specific protein or non-selectively bind to several other targets. The result is a waste of time and resources that some say has contributed to a ‘reproducibility crisis’ in the biological sciences, potentially slowing the pace of discovery and drug development.
Laflamme is part of a growing community that wants to solve the problem of unreliable antibodies in research. He teamed up with molecular geneticist Aled Edwards at the University of Toronto, Canada, to set up Antibody Characterization through Open Science (YCharOS, pronounced ‘Icarus’), an initiative that aims to characterize commercially available research antibodies for every human protein.
There are also efforts under way to produce better-performing antibodies, to make it easier for researchers to find them and to encourage the research community to adopt best practices when it comes to choosing and working with these molecules. Antibody vendors, funding agencies and scientific publishers are all getting in on the action, says Harvinder Virk, a physician–scientist at the University of Leicester, UK. “It’s hard to imagine that a problem that has been going on so long will suddenly change — but I’m hopeful.”
Putting antibodies to the test
The immune system produces antibodies in response to foreign substances, such as viruses and bacteria, flagging them for destruction. This makes antibodies useful in laboratory experiments. Scientists co-opt this ability by using them to mark or quantify specific biological molecules, such as a segment of a protein. To be effective, these molecular tags need to have both specificity — a strong affinity for the target — and selectivity — the ability to leave other proteins unmarked.
For decades, scientists created these antibodies themselves. They injected proteins into animals, such as rabbits, whose immune systems would generate antibodies against the foreign molecules. To create a longer-term, more consistent supply of antibodies, researchers extracted immune cells from animals and combined them with immortalized cancer cells. When reagent companies began the mass production of antibodies in the 1990s, most researchers shifted to purchasing antibodies from a catalogue. Today, there are around 7.7 million research antibody products on the market, sold by almost 350 antibody suppliers around the world.
In the late 2000s, scientists began reporting problems with both the specificity and selectivity of many commercially available antibodies, leading researchers to call for an independent body to certify that the molecules work as advertised. Over the years, a handful of groups have launched efforts to evaluate antibodies.
What sets YCharOS apart is the level of cooperation that it has obtained from companies that sell antibodies. When Laflamme and Edwards set out to start YCharOS, they called every single vendor they could find; more than a dozen were interested in collaborating. YCharOS’s industry partners provide the antibodies for testing, free of charge. The partners, along with the funders of the initiative (which include various non-profit organizations and funding agencies), are given the chance to review characterization reports and provide feedback before they are published.
YCharOS tests antibodies by comparing their specificity in a cell line that expresses the target protein at normal biological levels against their performance in what’s called a knock-out cell line that lacks the protein (see ‘Ways to validate’).
In an analysis published in eLife last year, the YCharOS team used this method to assess 614 commercial antibodies, targeting a total of 65 neuroscience-related proteins2. Two-thirds of them did not work as recommended by manufacturers.
“It never fails to amaze me how much of a hit or miss antibodies are,” says Riham Ayoubi, director of operations at YCharOS. “It shows you how important it is to include that negative control in the work.”
Antibody manufacturers reassessed more than half of the underperforming antibodies that YCharOS flagged in 2023. They issued updated recommendations for 153 of them and removed 73 from the market. The YCharOS team has now tested more than 1,000 antibodies that are meant to bind to more than 100 human proteins.
“There’s still a lot of work ahead,” Laflamme says. He estimates that, of the 1.6 million commercially available antibodies to human proteins, roughly 200,000 are unique (many suppliers sell the same antibodies under different names).
“I think the YCharOS initiative can really make a difference,” says Cecilia Williams, a cancer researcher at the KTH Royal Institute of Technology in Stockholm. “But it’s not everything, because researchers will use these antibodies in other protocols, and in other tissues and cells that may express the protein differently,” she says. The context in which antibodies are used can change how they perform.
Other characterization efforts are trying to tackle this challenge. Andrea Radtke and her collaborators were part of a cell-mapping consortium called the Human BioMolecular Atlas Program when they set up the Organ Mapping Antibody Panels (OMAPs). OMAPs are collections of community-validated antibodies used in multiplex imaging — a technique that involves visualizing several proteins in a single specimen. Unlike YCharOS, which focuses on conducting rigorous characterizations of antibodies for various applications in one specific context, OMAPs is looking at a single application for the antibodies, but in several contexts, such as in different human tissues and imaging methods. To do so, OMAPs recruits scientists from both academia and industry to conduct validations in their own labs.
“Vendors cannot test all possible applications of their antibodies, but as a community we can say ‘let’s try this’,” says Radtke, who now works as a principal scientist at the instrumentation company Leica Microsystems in Bethesda, Maryland. “People are testing things that you would never think you could test.”
Expanding the toolbox
Even if good antibodies are available, they are not always easy to find. In 2009, Anita Bandrowski, founder and chief executive of the data-sharing platform SciCrunch in San Diego, California, and her colleagues were examining how difficult it was to identify antibodies in journal articles. After sifting through papers in the Journal of Neuroscience, they found that 90% of the antibodies cited lacked a catalogue number (codes used by vendors to label specific products) — making them almost impossible to track down. To replicate an experiment, it’s important to have the right reagents — and proper labelling is crucial to finding them, Bandrowski says.
After seeing that a similar problem plagued other journals, Bandrowski and her colleagues decided to create unique, persistent identifiers for antibodies and other scientific resources, such as model organisms, which they called research resource identifiers, or RRIDs. Catalogue numbers can disappear if a company discontinues a product — and because companies create them independently, two different products might end up with the same one. RRIDs solve this.
‘A landmark moment’: scientists use AI to design antibodies from scratch
In 2014, Bandrowski and her team started a pilot project3 with 25 journals, in which they asked authors to include RRIDs in their manuscripts. In the years since, more than 1,000 journals have adopted policies that request these identifiers. “We currently have nearly one million citations to RRIDs from papers,” says Bandrowski.
Ultimately, the hope is that authors of every journal article will clearly label the resources they used, such as antibodies, with RRIDs, Bandrowski says. “That won’t change reproducibility by itself, but it is the first step.”
In addition to being able to track down antibodies, researchers need a way to choose which ones to use. In 2012, Andrew Chalmers, who was then a researcher at the University of Bath, UK, co-founded CiteAb, a search engine to help researchers find the most highly cited antibodies. Over the years, the platform has grown to include more than seven million antibodies — and now also includes, when available, information regarding validations. In May, CiteAb began integrating YCharOS’s characterization data onto its site.
“The big challenge is that antibodies are just used in so many different ways, for so many different species that you can’t tick off that an antibody is good or bad,” Chalmers says. Many say that knock-out validation is key, but less than 5% of antibodies on CiteAb have been validated in this way, either by suppliers or through other independent initiatives, such as YCharOS. “There’s a long way to go,” Chalmers says.
Stakeholders get involved
Like many others, Virk developed an interest in antibody reliability after a personal experience with bad antibodies. In 2016, Virk received a big grant to study the role of a protein called TRPA1 in airway inflammation. But one of his colleagues mentioned that, on the basis of his own experience, the antibodies he was working with might not be reliable.
When Virk put TRPA1 antibodies to the test, he discovered that his colleague was right: of the three most-cited antibodies used to study TRPA1, two didn’t detect the human protein at all, and the other detected several other proteins at the same time. “That was a shock,” Virk says. “At that point, I wanted to leave science — because if things are really this unreliable, what’s the point?”
Instead of leaving academia, Virk co-founded the Only Good Antibodies (OGA) community last year, with the aim of bringing together stakeholders — such as researchers, antibody manufacturers, funding agencies and publishers — to tackle the problem of poorly performing antibodies. In February, the OGA community hosted its first workshop, which included individuals from these various groups to discuss how to improve the reproducibility of research conducted with antibodies. They were joined by NC3Rs, a scientific organization and funder, based in London that focuses on reducing the use of animals in research. Better antibodies means fewer animals are used in the process of producing these molecules and conducting experiments with them.
Blame it on the antibodies
Currently, the OGA community is working on a project to help researchers choose the right antibodies for their work and to make it easier for them to identify, use and share data about antibody quality. It is also piloting an YCharOS site at the University of Leicester — the first outside Canada — which will focus on antibodies used in respiratory sciences. The OGA community is also working with funders and publishers to find ways to reward researchers for adopting antibody-related best practices. Examples of such rewards include grants for scientists taking part in antibody-validation initiatives.
Manufacturers have also been taking steps to improve antibody performance. In addition to increasingly conducting their own knock-out validations, a number of suppliers are also altering the way some of their products are made.
The need to modify antibody-production practices was brought to the fore in 2015, when a group of more than 100 scientists penned a commentary in Nature calling for the community to shift from antibodies generated by immune cells or immune–cancer-cell hybrids, to what are known as recombinant antibodies4. Recombinant antibodies are produced in genetically engineered cells programmed to make a specific antibody. Using these antibodies exclusively, the authors argued, would enable infinite production of antibodies that do not vary from batch to batch — a key problem with the older methods.
A few manufacturers are shifting towards making more recombinant antibodies. For example, Abcam, an antibody supplier in Cambridge, UK, has added more than 32,000 of them to their portfolio. “Facilitating the move towards recombinants across life-science research is a key part of improving reproducibility,” says Hannah Cable, the vice-president of new product development at Abcam. “That’s something that antibody suppliers should be doing.”
Rob Meijers, director of the antibody platform at the Institute for Protein Innovation in Boston, Massachusetts, a non-profit research organization that makes recombinant antibodies, says that this shift simply makes more business sense. “They’re much more reproducible, you can standardize the process for them, and the user feedback is very positive,” he says.
Standardize antibodies used in research
CiteAb’s data have revealed that scientists’ behaviour around antibody use has shifted drastically over the past decade. About 20% of papers from 2023 that involved antibodies used recombinants. “That’s a big change from where we were ten years ago,” says Chalmers, who is now CiteAb’s chief executive.
Although the ongoing efforts to improve antibody reliability are a move in the right direction, changing scientists’ behaviour remains one of the biggest challenges, say those leading the charge. There are cases in which researchers don’t want to hear that an antibody they’ve been using for their experiments isn’t actually doing what it’s meant to, Williams says. “If somebody is happy with the result of an antibody, it’s being used regardless, even if it’s certain that it doesn’t bind this protein,” Williams says. Ultimately, she adds, “you can never get around the fact that the researcher will have to do validations”.
Still, many scientists are hopeful that recent efforts will lead to much needed change. “I’m optimistic that things are getting better,” Radtke says. “What I’m so encouraged by is the young generation of scientists, who have more of a wolf-pack mentality, and are working together to solve this problem as a community.”
An unusually long and complex research paper has caught the attention of the scientific community, sparking questions about the ideal length of a paper. The study, by computational biologist Manolis Kellis and his colleagues, was published in Nature1 in July. Spanning 35 pages, it contains more than 20,000 words, and has 16 figures — or 61, if those in the Supplementary information are included. It describes changes in the genes, cellular pathways and cell types of people with Alzheimer’s disease across six regions of the brain, and provides a detailed atlas of gene expression.
When Kellis, who runs a computational-biology laboratory at the Massachusetts Institute of Technology (MIT) in Cambridge, shared the paper on X (formerly Twitter), the size of it seemed to divide his peers. Some were complimentary: “It must have taken a lot of effort and resources to get this done, so all in all, it is a great paper,” one response read. Others were concerned about its usefulness. “How can anyone read this article, let alone review and critique the work?” asked another user.
An analysis of paper characteristics across scientific fields2, published in 2023, suggests that this study is an outlier in the medical and health sciences, where papers typically hover around ten pages in length. However, it is not so unusual when compared with papers in subject areas such as mathematics, law or the humanities — all of which often exceed 20 pages.
Kellis’s work raises the question of how accessible research papers should be, and how readers in and beyond academia are expected to consume them. For example, is it better to publish large data sets alongside long and dense papers, to keep the information contained in one place? Or should researchers home in on specific topics and publish their results across several papers?
Alireza Haghighi, a geneticist at Harvard Medical School in Boston, Massachusetts, says that there is value in the former approach, particularly at a time when data sets are becoming increasingly large. “Although focus has traditionally been important in publications, we must acknowledge the complexity of new methods and the huge volumes of data generated today,” says Haghighi. “Not all papers can or should be understood in one hour.”
Does size matter?
Papers that provide broad, detailed overviews and extensive data sets — sometimes called ‘atlases’, in the omics fields of genomics, transcriptomics and proteomics — allow researchers to see the big picture, says Haghighi. They enable readers to “identify connections across different areas, and generate new hypotheses”, he explains, and adds that he sees them as drivers of innovation that can guide large-scale, integrative research initiatives better than a more focused paper might.
Responding to the discussion on X, Kellis said he understands that some people will be overwhelmed by his lab’s paper. He likened the work to “a good book with many chapters and many pages”, and said that “each paragraph, parenthesis, panel, supplementary figure, can hide potential hints and secrets that the authors themselves may have missed”. Kellis also suggested that for those who were overwhelmed by the results, tools such as the ChatGPT Consensus app, which is regarded as an academic search engine, could be useful for summarizing some of the paper’s findings.
2024 Research Leaders
Li-Huei Tsai, a neuroscientist at MIT and a corresponding author on the paper, told Nature Index that she is proud of the work, which has “produced important insights into genomic underpinnings of Alzheimer’s vulnerability and resilience”. Kellis did not respond to Nature Index’s request for comment.
Researchers who spoke to Nature Index flagged a number of issues with big, data-dense articles. Luke Dabin, an epigeneticist at the Indiana University School of Medicine in Indianapolis, is a “huge fan” of big data sets and atlas papers, because they have the potential to be a hotbed for generating hypotheses and can inform the design of future experiments. But Dabin says that such papers can sometimes be difficult to interpret — even by scientists working in the same field — and can have quality-control issues. “The Kellis paper has 475 figure panels and is difficult for me to digest, let alone someone with no training or experience in single-cell omics,” Dabin says. Haghighi agrees that accuracy can become a problem in large papers. “We should appreciate that atlas maps are more prone to inaccuracies due to their scope and complexity,” he says.
Such papers can also be resource-heavy for journal editors to publish. It took almost two years for Kellis’s paper to progress from acceptance to publication, although it might not have been under review the entire time. A spokesperson for Nature noted that “the length of the review process for papers submitted to Nature varies considerably from manuscript to manuscript”, and said that its primary focus is “to ensure that a rigorous peer-review process takes place”. (Nature Index’s news and supplement content is editorially independent of its publisher, Springer Nature.)
On X, Kellis noted that “it was a Herculean task by the reviewers and editors, as it was of course for the authors, to go through every figure, every panel, and every result” as part of the publishing process.
A case for brevity?
Some researchers argue that there is simply not enough time to read such long and dense papers. “The readership on most academic papers is low anyway, so writing a long paper is just inviting it not to be read even more,” says Daniel Price, an astrophysicist at Monash University in Melbourne, Australia, and former editor-in-chief of the journal Publications of the Astronomical Society of Australia, which publishes research on data-heavy topics such as modelling and computational astronomy.
Price says it’s unlikely that anyone has ever read the entirety of one of his monster astrophysics papers3, which clocks in at 82 pages and has 57 figures. “It’s definitely too long,” he says of the paper, admitting that it could easily have been 60 pages instead. The problem with going long, he adds, is that it’s “undisciplined” and compromises the ability to self-edit.
Haghighi says some improvements could be made to long, data-heavy papers. He suggests that publishers standardize the way such papers are formatted and published by introducing new editorial guidelines and implementing “a dynamic, continuous review process” that allows authors to update their work regularly over time, after publication. “I appreciate that this might not be easy,” says Haghighi, but “it would make the review process more effective and consistent and make it easier for the scientific community”.
Formatting guidelines at most major journals tend to favour shorter articles with fewer figures. Nature, for instance, suggests the typical length of biological, clinical and social sciences papers should not normally exceed 8 pages, or 4,300 words, and 5–6 figures. That said, it does not enforce specific limits, and instead leaves this up to the editor’s discretion.
In astrophysics, a field that is characterized by vast data sets that are often analysed by large, international teams, there are some examples of how a research finding can be broken down into more digestible parts. For instance, after the first image of a black hole was captured by the Event Horizon Telescope — a global network of radio telescopes run by a group of more than 300 scientists — the team published 6 papers in a special edition of The Astrophysical Journal Letters. Each paper presented an aspect of the research, looking at methodology, specific features of the black hole and the image itself.
Price thinks paper series such as this are “definitely a better idea” than one long paper, and adds that there is a lot to be said for concise papers. He points to a 2016 paper published by the LIGO Scientific Collaboration4 — a conglomerate of more than 100 institutions collaborating in the search for gravitational waves — after its seminal detection of gravitational waves using instruments in Washington and Louisiana. “It’s eight pages [ten, including references] and it revolutionized astrophysics,” he says.
All of these images were generated by AI.Credit: Proofig AI, 2024
From scientists manipulating figures to the mass production of fake papers by paper mills, problematic manuscripts have long plagued the scholarly literature. Science sleuths work tirelessly to uncover this misconduct to correct the scientific record. But their job is becoming harder, owing to the introduction of a powerful new tool for fraudsters: generative artificial intelligence (AI).
“Generative AI is evolving very fast,” says Jana Christopher, an image-integrity analyst at FEBS Press in Heidelberg, Germany. “The people that work in my field — image integrity and publication ethics — are getting increasingly worried about the possibilities that it offers.”
AI-generated images and video are here: how could they shape research?
The ease with which generative-AI tools can create text, images and data raises fears of an increasingly untrustworthy scientific literature awash with fake figures, manuscripts and conclusions that are difficult for humans to spot. Already, an arms race is emerging as integrity specialists, publishers and technology companies race to develop AI tools that can assist in rapidly detecting deceptive, AI-generated elements of papers.
“It’s a scary development,” Christopher says. “But there are also clever people and good structural changes that are being suggested.”
Research-integrity specialists say that, although AI-generated text is already permitted by many journals under some circumstances, the use of such tools for creating images or other data is less likely to be viewed as acceptable. “In the near future, we may be okay with AI-generated text,” says Elisabeth Bik, an image-forensics specialist and consultant in San Francisco, California. “But I draw the line at generating data.”
What ChatGPT and generative AI mean for science
Bik, Christopher and others suspect that data, including images, fabricated using generative AI are already widespread in the literature, and that paper mills are taking advantage of AI tools to produce manuscripts en masse (see ‘Quiz: can you spot AI fakes?’).
Under the radar
Pinpointing AI-produced images poses a huge challenge: they are often almost impossible to distinguish from real ones, at least with the naked eye. “We get the feeling that we encounter AI-generated images every day,” Christopher says. “But as long as you can’t prove it, there’s really very little you can do.”
There are some clear instances of generative-AI use in scientific images, such as the now-infamous figure of a rat with absurdly large genitalia and nonsensical labels, created using the image tool Midjourney. The graphic, published by a journal in February, sparked a social-media storm and was retracted days later.
Credit: Proofig (generated images)
Most cases aren’t so obvious. Figures fabricated with Adobe Photoshop or similar tools before the rise of generative-AI — especially in molecular and cell biology — often contain telltale signs that sleuths can spot, such as identical backgrounds or an unusual absence of smears or stains. AI-made figures often lack such signs. “I see tonnes of papers where I think, these Western blots do not look real — but there’s no smoking gun,” Bik says. “You can only say they just look weird, and that of course isn’t enough evidence to write to an editor.”
But signs suggest that AI-made figures are appearing in published manuscripts. Text written using tools such as ChatGPT is on the rise in papers, given away by standard chatbot phrases that authors forget to remove and telltale words that AI models tend to use. “So we have to assume that it’s also happening for data and for images,” says Bik.
Another clue that fraudsters are using sophisticated image tools is that most of the issues that sleuths are currently detecting are in papers that are several years old. “In the past couple of years, we’ve seen fewer and fewer image problems,” Bik says. “I think most folks who have gotten caught doing image manipulation have moved on to creating cleaner images.”
How to create images
Creating clean images using generative AI is not difficult. Kevin Patrick, a scientific-image sleuth known as Cheshire on social media, has demonstrated just how easy it can be and posted his results on X. Using Photoshop’s AI tool Generative Fill, Patrick created realistic images — that could feasibly appear in scientific papers — of tumours, cell cultures, Western blots and more. Most of the images took less than a minute to produce (see ‘Generating bogus science’).
“If I can do this, certainly the people who are getting paid to generate fake data are going to be doing this,” Patrick says. “There’s probably a whole bunch of other data that could be generated with tools like this.”
Some publishers say that they have found evidence of AI-generated content in published studies. These include PLoS, which has been alerted to suspicious content and found evidence of AI-generated text and data in papers and submissions through internal investigations, says Renée Hoch, managing editor of PLoS’s publication-ethics team in San Francisco, California. (Hoch notes that AI use is not forbidden in PLoS journals, and that its AI policy focuses on author accountability and transparent disclosures.)
Credit: Kevin Patrick
Other tools might also provide opportunities for people wishing to create fake content. Last month, researchers published1 a generative-AI model for creating high-resolution microscopy images — and some integrity specialists have raised concerned about the work. “This technology can easily be used by people with bad intentions to quickly generate hundreds or thousands of fake images,” Bik says.
Yoav Shechtman at the Technion–Israel Institute of Technology in Haifa, the tool’s creator, says that the tool is helpful for producing training data for models because high-resolution microscopy images are difficult to obtain. But, he adds, it isn’t useful for generating fake because users have little control over the output. Existing imaging software such as Photoshop is more useful for manipulating figures, he suggests.
Weeding out fakes
Human eyes might not be able to catch generative AI-made images, but AI might (see ‘AI images are hard to spot’).
The makers behind tools such as Imagetwin and Proofig, which use AI to detect integrity issues in scientific figures, are expanding their software to weed out images created by generative AI. Because such images are so difficult to detect, both companies are creating their own databases of generative-AI images to train their algorithms.
AI models fed AI-generated data quickly spew nonsense
Proofig has already released a feature in its tool for detecting AI-generated microscopy images. Company co-founder Dror Kolodkin-Gal in Rehovot, Israel, says that, when tested on thousands of AI-generated and real images from papers, the algorithm identified AI images 98% of the time and had a 0.02% false-positive rate. Dror adds that the team is now working on trying to understand what, exactly, their algorithm detects.
“I have great hopes for these tools,” Christopher says. But she notes that their outputs will always need to be assessed by an expert who can verify the issues they flag. Christopher hasn’t yet seen evidence that AI image-detection software are reliable (Proofig’s internal evaluation has not been published). These tools are “limited, but certainly very useful, as it means we can scale up our effort of screening submissions,” she adds.
Multiple publishers and research institutions already use Proofig and Imagetwin. The Science journals, for example, use Proofig to scan for image-integrity issues. According to Meagan Phelan, communications director for Science in Washington DC, the tool has not yet uncovered any AI-generated images.
Springer Nature, which publishes Nature, is developing its own detection tools for text and images, called Geppetto and SnapShot, which flag irregularities that are then assessed by humans. (The Nature news team is editorially independent of its publisher.)
Fraudsters, beware
Publishing groups are also taking steps to address AI-made images. A spokesperson for the International Association of Scientific, Technical and Medical (STM) Publishers in Oxford, UK, said that it is taking the problem “very seriously” and pointed to initiatives such as United2Act and the STM Integrity Hub, which are tackling paper mills and other scientific-integrity issues.
ChatGPT one year on: who is using it, how and why?
Christopher, who is chairing an STM working group on image alterations and duplications, says that there is a growing realization that developing ways to verify raw data — such as labelling images taken from microscopes with invisible watermarks akin to those being used in AI-generated text — might be the way forward. This will require new technologies and new standards for equipment manufacturers, she adds.
Patrick and others are worried that publishers will not act quickly enough to address the threat. “We’re concerned that this will just be another generation of problems in the literature that they don’t get to until it’s too late,” he says.
Still, some are optimistic that the AI-generated content that enters papers today will be discovered in the future.
“I have full confidence that technology will improve to the point that it can detect the stuff that’s getting done today — because at some point, it will be viewed as relatively crude,” Patrick says. “Fraudsters shouldn’t sleep well at night. They could fool today’s process, but I don’t think they’ll be able to fool the process forever.”
Unravelling how the billions of interacting neurons in the human brain conjure consciousness is one of the greatest challenges in twenty-first-century science. Over the past decade, large, well-funded initiatives, including in the United States, Europe and China, have been launched to unlock the mysteries of cognitive function — mental processes such as memory, language, perception and problem-solving — by coming at it from all angles.
For the millions of people around the world who will develop an incurable or treatment-resistant brain disorder this year, the need to better understand cognitive function and dysfunction is pressing, says Christopher Rozell, a computational neuroengineer at Georgia Institute of Technology in Atlanta. Rozell co-leads a multidisciplinary team that is developing technology-based therapies for depression, the leading cause of ill health and disability worldwide. “Globally, more than 300 million people will have a major depressive episode this year — and that’s just one neurological disorder subtype,” he says.
Nature Index 2024 Neuroscience
Rozell is exploring a therapy for treatment-resistant depression based on deep-brain stimulation, in which implanted electrodes electrically stimulate specific brain areas to provide long-term symptom relief. The work is funded by the US National Institutes of Health’s (NIH) Brain Research Through Advancing Innovative Neurotechnologies (BRAIN) Initiative, a major project launched in 2013, which to date has invested more than US$4 billion across neuroscience research. The BRAIN Initiative’s strategy is to develop tools, and then use these advances to gain a deeper understanding of brain function. According to Rozell, the decade-long investment is beginning to pay off.
In depression treatment, for example, doctors have always had to make subjective clinical judgements and trial-and-error therapy adjustments when trying to manage the condition. But, in 2023, Rozell and his collaborators used new brain-implant and big-data processing technologies to identify changes in brain activity that can indicate a patient’s current clinical state, enabling doctors to adjust treatment in response1. At the end of the six-month trial, 90% of patients showed significant improvement and 70% were in remission or no longer depressed. BRAIN Initiative funding was key. “We work with clinicians and engineers in teams with a breadth of expertise that would have been very difficult to imagine under conventional funding programmes,” Rozell says. “Every week now, you see large, interdisciplinary teams making incredible advances that would not be happening if it were not for a programme like the BRAIN Initiative.”
Likewise, proponents of the Human Brain Project (HBP), one of the largest research endeavours ever funded by the European Union (EU), which spent €600 million (US$668 million) over ten years before its completion in 2023, point to several advances. New brain-implant technologies that could restore partial vision in certain forms of blindness and brain-like ‘neuromorphic’ computer chips for more sophisticated artificial intelligence (AI) are important outcomes.
But concerns remain that core questions in neuroscience have not been addressed by big projects. It’s not clear how cognitive function emerges from patterns of brain activity, for instance, let alone how these processes go awry caused by disease.
And although big-neuroscience funding has increased in China over the past few years, it has been cut significantly in the EU and the United States, threatening the trajectory of brain science advancement.
Uncharted territory
If understanding human brain function is the ‘moonshot’ of neuroscience, we’ll never make it without the right maps, says Rozell. Creating brain atlases, each focused on different structural features, has been a key aim. In late 2023, the BRAIN Initiative’s Cell Census Network (BICCN), a multi-centre effort led by the Allen Institute for Brain Science in Seattle, Washington, produced the most detailed map yet of the cells that make up the human brain. Using single-cell genome sequencing — a technique that allows all or part of an individual cell’s genome to be sequenced — the team identified more than 3,000 different cell types in the human brain, many previously undescribed.
BICCN researchers also produced the first complete cellular atlas of a mammalian brain, pinpointing the location and identity of each of the more than 32 million cells in a mouse brain2. When the team launched the project 10 years ago, it was unclear whether this was even feasible, says Allen Institute director, Hongkui Zeng, who led the work. But the rapid development and scaling-up of single-cell genomic technology has revolutionized the field.
“Previously, the brain was just an unknown number of faceless cells,” says Zeng. “Now, we have the molecular identities for specific cells in specific brain regions, and we can start to label each cell type and see what they do.”
Immune cells in a mouse brain, intertwined with tiny blood vessels, captured for a BRAIN-funded project.Credit: Josephine Liwang, Yongsoo Kim lab/Penn State College of Medicine, PA
BICCN’s open-access brain-cell atlases are an indispensable resource, says Sebastian Seung, a computer scientist and neuroscientist at Princeton University in New Jersey. “To go from mapping the brain as a bunch of regions, to mapping cell types, is a huge jump in precision,” he says. Brain-cell atlases are foundational data supporting Seung’s own research, which focuses on the wiring between brain cells, known as the connectome. Together with cell mapping, new tools in connectomics — including those developed in Seung’s lab with BRAIN funding, which use AI to automate brain-scan image processing — allow scientists to study the brain in ways they’ve never done before.
A different approach was used to build the Human Brain Atlas, the most detailed 3D anatomical map of a human brain yet assembled3. A team led by Katrin Amunts, a neuroscientist at the Jülich Research Centre, a large-scale national facility in the Helmholtz Association of German Research Centres, took a postmortem brain and analysed it, slice by slice, to build the atlas not from the cells up, but from a whole brain down. The Human Brain Atlas forms a core part of EBRAINS, an open-access digital platform that combines tools, services and data generated by the HBP, which has been used by more 10,000 people worldwide.
The platform’s ‘virtual brain’ tool is being used to create personalized patient brain models to guide clinical decision-making in epilepsy, multiple sclerosis, depression and Parkinson’s, and its brain atlases and data are being accessed by researchers in neuroimaging, neurology, AI and basic science. In January, the EBRAINS project won a further €38 million from the European Commission to fund its continued development.
There is an argument that although BRAIN and the HBP did not specifically focus on conceptual questions in neuroscience, the foundational resources that they have provided can help to fill major knowledge gaps that will benefit those working in both basic and applied neuroscience areas. Seung says this is why the BRAIN Initiative’s strategy of prioritizing neuroscience tool development was the right approach. “So much of the study of neuroscience has been limited by the scarcity of data,” he says. “The NIH would normally not necessarily fund technology development, but sometimes to get to important science, we need a technological revolution.”
New model
Still in its early phases, China’s big neuroscience project can benefit from lessons learned by its international counterparts. Conceived in 2013 — closely following the launch of BRAIN and the HBP — the China Brain Project (CBP) began in 2021 with ten-year funding of 12 billion yuan (US$1.66 billion) to advance brain-disease studies and basic neuroscience, as well as brain-inspired technologies and brain–computer interfaces. The project involves more than 500 laboratories across the country, and aims to build on China’s research strengths, including in connectomics and non-human primate animal models, a valuable, but contentious, aspect of neuroscience. “You cannot do invasive experiments in the human brain to understand what’s going on, so animal models are very important,” says Zeng.
The protocols and standards for non-human primate research in China are based on those set by the NIH, but the work is easier to conduct because animal-rights groups don’t protest against animal use in research like they do in the United States, says Muming Poo, scientific director of the Institute of Neuroscience at the Chinese Academy of Sciences in Shanghai, who has led the CBP organizing committee since 2020. “There is a great need in the community for using non-human primate disease models because mouse models for brain disease, especially psychiatric disease, are just not working,” says Poo. He notes the slow global pace of drug development for brain disease, which is mostly based on rodent models, and says non-human primates, as our closest living relatives, should offer better models of the human brain.
Poo’s group is developing a toolbox of genetic-engineering techniques to produce non-human primate models of disease that they hope can be used in drug testing. In late 2023, they reported the first live-born monkey chimaera4, created by taking stem cells from one macaque embryo and adding them to another. The work is a key step towards creating transgenic non-human primate models of human brain diseases, akin to way that transgenic rodent models of disease are currently made.
Another strength that the CBP hopes to build on is China’s vast population, from which researchers can draw on extensive patient cohorts. According to Jialin Zheng, dean of the Tongji University School of Medicine in Shanghai, autism spectrum disorder in children, depression in adults and Alzheimer’s disease in ageing populations are the priority conditions addressed by CBP research.
In parts of the CBP that are related to brain-inspired technology, such as AI and brain–computer interfaces, there is strong competition between institutions in China and abroad, says Poo. But in basic neuroscience and brain medicine, the CBP was specifically designed to complement work conducted by other countries. “We made a strong point of taking the directions that are deficient in the United States and Europe,” such as non-human primate models and large-cohort studies, says Poo. Some of the first internationally collaborative research conducted within the project are now close to publication, he adds. “I think it’s like the global-warming problem — brain disease is an urgent problem shared by all of society, and we should solve it together.”
In many ways, the approaches and priorities of the big-brain projects in the United States, Europe and China complement each other to make the most of international resources and talent. In the United States, for instance, the BRAIN Initiative pooled resources to push technology development, whereas the HBP’s strategy focused on coordinating multidisciplinary research, such as bringing neuroscientists together with computer scientists to develop new treatments. China’s strategy is to use its unique strengths to fill important gaps and expand on them through international collaboration.
There are challenges ahead if researchers want to build on the outputs of the three initiatives. For example, Zheng says it’s going to require coordination between governments to decide how genetic information and biological samples can safely be shared between countries. “Different countries have different regulation in terms of data. How can it be shared more openly? We are dealing with the same diseases, so, how can we work together to address these challenges?”
In addition to restrictions on data sharing, coordination between different data centres is a major issue, says Poo. “It has been difficult to set up a generally agreed, smooth way of data-sharing among many big projects, because each big project has its own data centre,” he says. “We are in international discussions about the data problem, but there is no solution yet.”
There are also concerns about whether long-standing questions around cognitive function can be answered by the kinds of projects being funded by big brain programmes. On the one hand, finding answers will require parallel studies of brain activity at the molecular, anatomical and physiological levels — something that large-scale initiatives are designed to facilitate, says Zeng.
Source: Nature Index
But knowing how to piece this information together to explain cognitive function will require new ideas and hypotheses at a foundational level that none of the big neuroscience projects has yet produced, says Yves Frégnac, emeritus research director in cognitive science at the University of Paris-Saclay in France. “New concepts are not evolving at the same pace as technologies,” he says. “Reading out signs of cognitive activity is very different from understanding the brain.”
For China, the CBP has brought a much-needed injection of cash to a field that has struggled to find funding in the past. Poo says the initiative, which so far seems to be on track to meet its decade-long funding promise, will not only advance neuroscience in highly applied areas, but also in fundamental research. “In other countries, there are avenues of support for basic research in brain science, through organizations such as the US NIH or National Science Foundation — but not in China,” he says.
As the CBP builds momentum, researchers in Europe are trying to regain their footing, a year after the end of the HBP. Raising just over half of the expected €1 billion in funding from the EU and its member states, the HBP feels to many scientists like an opportunity not quite fulfilled, despite the progress made. “This money was needed in the field of brain sciences,” says Frégnac, who wrote an opinion piece on how the initiative could have been done better5. “People talk about €1 billion, US$4 billion, but if you compare it to initiatives in physics, this is peanuts.” NASA’s James Webb Space Telescope, for example, cost $10 billion, and the $1.5 billion annual budget of the European particle-physics laboratory, CERN, dwarfs the HBP’s entire ten-year funding. “If we want to be serious about the brain, we need to put more money in,” says Frégnac, who adds that the possibility of a well-funded follow-up to the HBP looks remote.
The future of BRAIN Initiative-supported research is also unclear. In 2024, as the ten-year pot of funds set aside in 2016 entered its ramp-down phase, a budget cap across all federal spending constrained the US Congress from making up the shortfall. The result was a 40% cut to BRAIN Initiative funding, compared with 2023. Researchers such as Rozell, whose work on treating depression is directly threatened by the cuts, are worried. “We’ve made enormous progress, but this work is not finished — it is not an approved therapy,” says Rozell. With the global economic cost of mental disorders estimated at US$5 trillion, the need for investment is clear, he adds. “To have spent a decade of money, time and expertise to reach a place where we’re starting to see the returns, and then have the threat of these programmes being taken away, it’s enormously concerning.”
Neuroscience has undergone remarkable progress. Researchers can now study specific areas of the brain with unprecedented detail thanks to cutting-edge imaging and genetic tools. Advanced modelling techniques, driven by artificial intelligence, have facilitated whole-brain mapping to track cognitive development over a lifetime. But fundamental questions about how the brain’s core functions emerge from cellular and molecular processes remain unanswered, limiting treatment options for neurological conditions.
Nature Index 2024 Neuroscience
Countries are pooling their resources and expertise to up the ante. Large-scale neuroscience projects are making use of unique strengths, including China’s vast population data and the United States’ med-tech industry. But researchers are calling for more funding, pointing out that budgets in other areas, such as the European particle-physics laboratory, CERN, and NASA’s James Webb Space Telescope, dwarf those of the biggest neuroscience initiatives.
Raising more money is not the only challenge. Over the past decade, tens of billions have been spent on finding effective treatments for Alzheimer’s disease, with limited patient benefit. A greater understanding of how brain conditions relate to other organs and biological systems, and vice versa, is needed. Studies investigating long COVID, for instance, could have major implications for autoimmune diseases.
In the coming years, technological advances such as implantable brain–computer interfaces (BCIs) are expected to fundamentally change how neuroscience is researched, and how neurological disorders are treated and diagnosed. The United States, the leading country in Nature Index neuroscience output by some margin, is setting the pace for BCI regulation, and other nations will need to find their footing fast.
