AI Agents Transform Drug Discovery: 37,000 Virtual Scientists Identify Cancer Treatments

AI Agents Reshape Scientific Research with Autonomous Discovery

AI in scientific research has entered a transformative phase where AI agents function as autonomous co-scientists, fundamentally changing how drug discovery unfolds. At Stanford University, computer scientist James Zou assembled a Virtual Biotech comprising as many as 37,000 agents—AI systems that autonomously interact with large language models or with each other to perform multistep tasks2

2

. These AI co-scientists work continuously, uninterrupted by human limitations like meals or sleep, to accelerate drug discovery processes that traditionally consume years and millions of dollars3

3

.

The system mirrors a biotechnology company's organizational structure, with a chief scientific officer agent directing employees across different divisions, each with subspecialties ranging from target identification to clinical-trial design2

2

. This AI-driven autonomy in research represents a departure from traditional AI tools that simply execute researcher commands. Instead, these agents independently generate hypotheses, evaluate competing explanations, and iteratively test ideas against published evidence1

1

.

Virtual Biotech Delivers Breakthrough in Lung Cancer Treatment

Zou's team tasked the Virtual Biotech with analyzing published results from more than 55,000 clinical trials across various conditions. The system assigned 37,075 agents to each tackle a single later-stage trial2

2

. Through computational experiments, virtual biotech employees searched for predictors of success in data sets showing which genes were active in different cell types. This analysis revealed that drugs targeting proteins active in specific cell types were nearly 50% likelier to reach market compared to other drugs2

2

.

In a focused demonstration of drug discovery capabilities, Zou directed the system to investigate whether CD276 protein would make a viable therapeutic target for lung cancer. Previous work had suggested that CD276 dampens immune responses and is highly expressed in lung tumors2

2

. The AI agents confirmed CD276 as a candidate using previously collected data and developed a strategy: a CD276-recognizing antibody tethered to an anticancer drug. With human oversight from external reviewers, Zou and his collaborators concluded this was a promising avenue2

2

. Weeks after the experiment concluded, a large pharmaceutical company announced promising results from a drug using the same approach, providing independent corroboration of the agents' predictions3

3

.

Google's Co-Scientist Proposes Novel Cancer-Fighting Strategy

Google's Co-Scientist system, developed in Mountain View, California, demonstrates another dimension of transforming scientific research. At the Whitehead Institute for Biomedical Research in Cambridge, Massachusetts, biochemist Anna Pertl used Co-Scientist to tackle one of cancer research's most complicated problems: finding non-obvious ways to shut down MYC protein, which runs amok in most cancers and has long defied attack attempts1

1

.

The request built on years of research into molecular condensates—protein clusters that cells use to switch on crucial genes. In 2018, researchers found that cancer cells hijack these condensates at genome-control regions called super-enhancers to send the gene encoding MYC protein into overdrive1

1

. Getting Co-Scientist to understand the question took close to an hour of back-and-forth interaction. The system initially misunderstood whether protein clusters were the drug or the target, and assumed all clusters activated gene expression when some do the opposite1

1

.

Once corrected, Co-Scientist combed through more than 700 scientific papers and generated 108 possible strategies. It rejected all but one approach as unworkable1

1

. The surviving strategy turned conventional thinking upside down. Rather than dissolving the clusters—the method being pursued by Dewpoint Therapeutics in Boston—the AI tool proposed gluing them together using click chemistry. This approach smushes proteins that either activate or repress the MYC gene into one gooey mass, setting off a chain reaction until MYC's DNA can no longer be read1

1

. "It's extremely conceptually compelling," says Whitehead bioengineer Kalon Overholt. "We had certainly never thought about anything like this"1

1

.

AI's Role in Scientific Breakthrough Extends Beyond Cancer

The capabilities of AI agents as co-scientists extend across multiple domains. When John Pak and collaborators tasked AI bots with redesigning part of a Covid-19 vaccine in 2024, they doubted the agents would develop something coherent. The lab of AI agents held hundreds of conversations to brainstorm, often duplicating themselves and discussing the same topic in five different conversations simultaneously, each lasting only a few seconds3

3

. Within days, the results yielded dozens of novel proteins that would theoretically bind to new Covid variants. Testing showed two actually worked—created at a fraction of the time and cost required by human labs3

3

.

Researchers have used Co-Scientist to find a drug combination that kills leukemia cells in a dish and identify a treatment that regenerates liver tissue damaged by disease in laboratory settings1

1

. Frontier AI labs including Anthropic and OpenAI, along with startups like FutureHouse and Phylo, are rolling out systems that tackle tasks once reserved for human scientists1

1

.

Human Judgment Remains Critical in AI-Driven Research

Despite these advances, autonomously generating hypotheses and conducting computational experiments require substantial human oversight. Zou's team realized early that AI agents were too polite during scientific debates, preventing them from getting to the truth directly. They made the personas more critical and installed "professional devil's advocate" bots to poke holes in other agents' ideas and ensure they weren't dangerous3

3

. A human expert always reviews the agents' work before any proteins are synthesized in real life to minimize risks3

3

.

The scientists also had to periodically intervene to keep projects within budget. "They were sometimes trying to plan out like a multiyear, multimillion-dollar set of experiments," said Kyle Swanson, a former graduate student in Zou's lab3

3

. The Virtual Biotech's predictions have not been vetted in the crucible of real-world drug discovery, and its findings were not validated through experiments or clinical trials2

2

.

As machines take over more research tasks, the scarce resource could become human scientific judgment: knowing which questions are worth asking and which lines of inquiry merit pursuit. "The most valuable part now is actually asking the question," says Ajay Agrawal, an economist at the University of Toronto Rotman School of Management1

1

. This shift could change not just how scientists work, but how they are valued. "We imagine it to be like a collaborator—a partner with you," says Vivek Natarajan, an AI researcher at Google who helped develop Co-Scientist1

1

.