How AI Scientist Agents Are Transforming Research—and Why Humans Still Matter
Biochemist Anna Pertl typed out her question, hit enter and left for the night.
The artificial-intelligence systems she promoted are designed to generate innovative scientific hypotheses. Unlike typical chatbots, however, the AI tool called Co-Scientist does not simply produce an answer in a single pass. Instead, it launches several autonomous AI systems, known as agents, to search and synthesize information from scientific papers, evaluate competing explanations, refine and critique hypotheses, and repeatedly test ideas against published evidence.
This process can require significant computing power as the agents pursue explicit inferences before converging on a viable solution. It also takes time. Pertl jokes that, since receiving her doctorate from the Whitehead Institute for Biomedical Research in Cambridge, Massachusetts, she has attempted to complete a long-distance triathlon eight times.
On a rainy Tuesday in July, Pertl put her AI system to work on one of cancer biology’s most complex problems. She asked her “co-scientists” to find an obscure but practical way to harness the biology of molecular droplets, known as condensates, to block MYC—a protein that is out of control in most cancers and has long resisted attempts to target it.
The request was based on years of research by Pertl’s boss, Whitehead biologist Richard Young. In 2018, Young and his colleagues discovered that cells activate important genes by collecting regulatory proteins into condensates clustered at key genomic control regions called super-enhancers.1 They also discovered how cancer cells hijack these super-enhancers and overdrive the gene encoding MYC. This suggested that condensates help promote high levels of protein expression.2
For Young, the next step is clear: directly target tumor-cell condensates. Dewpoint Therapeutics, a Boston-based company he co-founded, is considering its options. Previous attempts to suppress MYC have failed, but Pertl wanted to see whether AI could suggest new methods of attack.
It took some work to arrive at an idea worth pursuing. Pertl and Whitehead bioengineer Karon Overholt spent nearly an hour going back and forth before the AI collaborator understood the question. Initially, the model assumed that the protein clusters at the center of the study were drugs rather than targets. It also assumed that the clusters always activated gene expression, when some actually do the opposite.
After Pertl and Overholt corrected these misconceptions, they released the system. By the next day, the AI co-scientists had reviewed more than 700 scientific papers and generated 108 possible strategies. All but one approach was rejected as unfeasible. The final strategy changed the way the laboratory thought about the problem.
Rather than dissolving the clusters—the approach being pursued by Dewpoint—the AI tool suggested gluing them together. The proposed approach uses a molecular-bonding technique called click chemistry to clump proteins that activate or inhibit cells. By bundling the genes into one gooey mass and triggering a chain reaction, the DNA of MYC could become inaccessible and no longer be read.
Scientists already knew that DNA and the proteins surrounding it can shift from a loose liquid into something more solid, similar to gelatin setting in a refrigerator. Using this transition to block cancer genes, however, was uncharted territory.
“This is conceptually very compelling,” says Overholt. “We certainly hadn’t thought about anything like this.”
AI-assisted scientific discovery: Trust but verify
This type of AI-assisted brainstorming is becoming increasingly common. Researchers have used Co-Scientist to identify combinations of drugs that kill leukemia cells in laboratory dishes.3 The system has also been used to identify treatments that could regenerate liver tissue damaged by disease in the laboratory.4
Co-Scientist is not the only system of its kind. Frontier AI laboratories such as Anthropic and OpenAI, as well as startups including FutureHouse5 and Philo6, are developing systems that can tackle tasks once reserved for human scientists.
These tools could free researchers to focus on the questions and decisions that matter most.
“We imagine it to be like a collaborator, a partner with you,” says Vivek Natarajan, an AI researcher at Google who helped develop Co-Scientist.
If that happens, AI could change not only how scientists work but also how scientific achievement is evaluated. For generations, scientific progress has depended on researchers who can ask difficult questions, devise ways to answer them and understand the results. As machines take over more of that work, scarce resources may shift toward human scientific judgment: knowing which questions are worth asking and which areas of investigation are worth pursuing.
“The most valuable thing to do right now is to actually ask questions,” says Ajay Agrawal, an economist at the Rotman School of Management at the University of Toronto who studies the impact of AI on innovation and entrepreneurship.
The goal is not to replace researchers but to free them from the mundane aspects of discovery, says Le Cong, a molecular geneticist at Stanford University and scientific co-founder of Philo.
“We are moving humans up the value chain,” he says.
But the introduction of AI research tools also creates a troubling paradox. Even as students have fewer opportunities to do the work themselves and build expertise, researchers may need deeper knowledge than ever to determine whether AI-generated ideas are sound. A tool that promises to accelerate discovery could therefore make it more difficult to train the scientists needed to oversee it.
Hector Zenil, a biomedical-computing researcher at King’s College London, concludes that “researchers with the human skills to understand the big picture will be more valuable than ever.”
For Fyodor Urnov, who researched genome editing at the University of California, Berkeley, and was an early adopter of FutureHouse’s Cosmos platform, the promise of an AI research assistant comes with a simple rule: its suggestions should not be taken at face value.
Urnov grew up in Moscow and remembers the phrase popularized during nuclear-arms-control negotiations between the United States and the Soviet Union in the 1980s.
“My relationship with Cosmos is based heavily on the Reagan and Gorbachev adage,” he says. “Trust, but verify.”
Anna Pertl and her collaborators sought ways to target the cancer-associated protein MYC.Credit: Anna Pertl
AI research assistants put scientific theories to the test
Researchers at Imperial College London used the platform to investigate how bacteria exchange pieces of DNA across species in 2025, incorporating a validation step into tests conducted with their Google collaborators.7
The researchers had already collected experimental data to form a theory. Rather than feed the unpublished results into the system, they asked the AI tools to work solely from published literature and datasets. Only after the system generated its theory did they reveal the experimental results and compare the AI’s conclusions with their own.
The co-scientists’ theory was highly consistent with what microbiologists José Penadez and Tiago Costa had spent years investigating: bacterial DNA elements could borrow parts of viruses and spread between distantly related hosts. Because of its ability to reason across information from many sources, Co-Scientist reached nearly the same conclusion in about two days.7
Costa says that if the Imperial researchers had used the tool from the beginning, they might not have needed years of exploratory experiments before identifying the correct mechanism.
“If you think of scientific discovery as a 100-metre race, the race starts at, say, 30 metres instead of zero,” he says.
AI research assistants typically use multiple agents to approach the same question from different angles. A scientific question can be divided into smaller tasks, allowing agents to work on different parts in parallel. Google’s systems and competing platforms from FutureHouse and Huawei Technologies of Shenzhen, China,8 as well as Fish AI in Tokyo,9 can explore multiple possible paths at once. Useful connections can emerge from interactions between agents rather than from a single chain of inference.
“We have AI agents all with different backgrounds to provide different perspectives,” says Kyle Swanson, a computer scientist at Stanford University who helped develop Virtual Lab, an early academic AI research-support platform.10 “You will have instant access to multidisciplinary experts.”
Why the best AI science still depends on good questions
Even a team of AI specialists needs someone to direct the tools toward a valuable problem and frame the challenge in a way that allows the system to tackle it productively.
For narrowly defined problems, the goal may be obvious. That could be why the decades-old problem posed by the late Hungarian mathematician Paul Erdős, and solved by AI tools in May, proved fertile ground for machine logic. In open-ended science, however, deciding which questions to ask and what counts as a meaningful answer remains fundamentally human work.
“You have to ask really good questions,” Swanson says. “If you ask a very general question, you’ll probably get a very general answer.”
Even questions that appear specific can create problems if they are not framed correctly. FutureHouse and Stanford University molecular geneticist Philine Guckelberger provided Cosmos with a large dataset of molecular contacts that help control which genes are turned on. They asked the system to look for patterns that human researchers might miss.
Cosmos initially struggled to label the data and failed to make distinctions that seemed obvious to Guckelberger. In a dataset containing 30 million rows, for example, it ignored columns that differentiated two types of DNA regions. After Guckelberger modified the data, the system discovered an unexpected pattern in cancer cells that she was able to confirm in an independent dataset.
“It accelerated like crazy,” she says.
From the beginning to the end of the research process
Another appeal of the AI Co-Scientist platform is its ability to compress the most tedious parts of research and take on complex, multistep workflows that scientists would otherwise have to complete manually.
Source: www.nature.com