When John Pak and his collaborators tasked a team of A.I. bots with redesigning part of a Covid-19 vaccine in 2024, he doubted that they would develop something successful or even coherent. At least they’d learn something about the deficiencies of these models, he thought.
The lab of A.I. agents had hundreds of conversations to brainstorm, unburdened by the distractions that typically bog human researchers down, like eating or sleeping. They often duplicated themselves and discussed the same topic in five different conversations simultaneously, each of which lasted only a few seconds.
As he watched the A.I. agents debate with one another, Dr. Pak, a human scientist who specializes in the kind of research the bots were doing, fought the urge to give them pointers.
“Take a step back,” his research partner, James Zou, told him. “See what they can do.”
Within a few days, the results were in: dozens of novel proteins that would theoretically bind to new Covid variants. Testing showed two of them actually worked — created at a fraction of the time and cost it would take a lab of humans.
“It was shocking,” Dr. Pak said.
In recent years, artificial intelligence has emerged as a powerful new tool in drug development, allowing scientists to predict the structure of proteins and how molecules will interact with them. But Dr. Zou’s team at Stanford University is working to develop a future where the technology is doing more than just carry out a researcher’s bidding. He envisions fleets of A.I. agents that work as “co-scientists,” rapidly accelerating the drug discovery process by independently generating hypotheses and running experiments on the computer with the oversight of a human researcher.
In a new paper published Thursday in the journal Science, the lab reported on its latest progress: Tens of thousands of A.I. agents predicted which approach to attacking a protein involved in lung cancer would be most effective at treating the disease in clinical trials. Weeks after the experiment started, a large pharmaceutical company announced promising results from a drug that used the same approach — a coincidence which Dr. Zou said was “a very nice, independent corroboration” of the agents’ work.
As fears over A.I.’s potential to create bioweapons or hack into critical infrastructure have ratcheted up, the scientists see this early research as a reason for optimism about some of the technology’s advancements.
“It really does feel like a new era where one person plus a bunch of GPT or Claude models can really make significant, fast progress in drug discovery,” said Kyle Swanson, a former graduate student in Dr. Zou’s lab who was involved in the Covid vaccine research.
Dr. Zou first devised the idea of the “virtual lab” as a way to deal with the mountain of ideas he had — more than his team could ever possibly tackle.
It was led by a “professor A.I.,” who compiled a team of other agents with various specialties, depending on the task. In the case of the vaccine problem, it picked immunology, computational biology and machine learning. After a few “meetings” together, the bots identified a research direction and gave Dr. Zou a list of required research tools.
“I’m very jealous that it doesn’t get tired,” Dr. Zou said. “It just goes 24/7, and its meetings are so much more efficient than mine.”
Dr. Zou and his colleagues are still tinkering with the system. For example, the team realized early on that the A.I. agents were too polite to one another during scientific debates, which Dr. Zou said was preventing them from “getting to the truth directly.”
They made the personas more critical and installed “professional devil’s advocate” bots whose task was to poke holes in the other agents’ ideas, and make sure they were not dangerous. Dr. Zou added that a human expert always reviews the agents’ work before any proteins are synthesized in real life, to minimize the risk of developing something harmful to humans.
Dr. Zou and his team are also experimenting with ways to incentivize the bots . (They don’t respond to mortal perks, like raises or promotions.) They created a leaderboard for the agents; top performers were rewarded with more computational resources.
The human scientists also have had to periodically jump in to ensure the bots stayed in budget.
“They were sometimes trying to plan out like a multiyear, multimillion-dollar set of experiments,” Dr. Swanson said. “That would be cool. But I was trying to finish my Ph.D.”
Over time, Dr. Zou added thousands more agents to the team and organized them into different departments of a virtual biotechnology company. It was that “company” that tackled the lung cancer research.
After extensively analyzing scientific studies on a protein called B7-H3 — a well-known and promising lung cancer drug target — the agents conducted their own analyses to study how the protein worked in cancer at a cellular level and then recommended the approach most likely to be successful in targeting it.
Dr. Zou said the A.I. agents could not have known about the successful trial results from the drug using a similar approach, which was developed by Merck. The bots were not connected to the internet, and the underlying language model had a knowledge cutoff of January 2025. (Merck did not respond to a request for comment.)
Dr. Charles Rudin, who studies this class of drugs at Memorial Sloan Kettering Cancer Center, noted that we wouldn’t necessarily know whether the approach the agents selected was the best, since no other drugs going after this target have been tested in a clinical trial.
The research is simply a proof of concept, and there are many parts of the drug discovery process the A.I. bots still can’t do like run experiments with cells in a lab. There’s also an open question of whether the technology can be as successful with another target that hasn’t been as well studied by humans.
Still, Dr. Rudin said he thought the study was a “nice example” of how it can accelerate drug development.
All in all, the work took the team of agents less than a day and about $46 in computing power.
The post A.I. Agents as ‘Co-Scientists’? This Lab Says They Could Speed Drug Research. appeared first on New York Times.
