AI biotech without humans? Stanford deploys 37,000 agents to hunt new therapies

What if an entire biotechnology company could operate without human employees, laboratories or a physical office? Researchers at Stanford Medicine have built a virtual biotech company that uses as many as 37,000 AI agents to work across the drug-development pipeline, from identifying potential drug targets to analysing clinical trials and designing therapies.

The system, developed by a team led by Stanford associate professor of biomedical data science James Zou and graduate student Harrison Zhang, is designed to mimic the structure of a conventional biotech company. A virtual chief science officer coordinates specialised AI agents working in different areas of drug discovery.

The research was published in Science on September 17.

AI agents analyse 50,000 clinical trials 

One of the biggest challenges in drug development is predicting which experimental medicines are likely to succeed in humans. Clinical trials can take years and cost tens or even hundreds of millions of dollars, while many drug candidates fail somewhere along the development process.

Stanford's virtual biotech was tasked with finding biological characteristics that could help predict successful drugs.

Rather than asking one AI system to process the entire scientific literature, the researchers assigned individual agents to specific clinical trials. The agents retrieved information on safety and effectiveness and analysed molecular data associated with the trials.

Together, they analysed and catalogued roughly 50,000 clinical trials in less than a week, a task Stanford researchers said could take human teams years.

The AI agents identified two potentially important characteristics of drug targets: cell-type specificity — whether a target is concentrated in a particular type of cell — and bimodality, which measures whether gene activity behaves more like an on-off switch or a dimmer.

The analysis found that drugs targeting genes with high cell-type specificity and switch-like activity performed better in historical clinical data. Such drugs were 40% more likely to advance from Phase 1 to Phase 2, 48% more likely to reach the market and associated with 32% fewer adverse events than drugs aimed at broader targets. These patterns appeared across diseases including cancer, brain, heart, kidney and lung conditions.

From data analysis to a potential cancer therapy 

The researchers then tested whether the AI company could go beyond analysing existing evidence and actually design a new therapy.

The system focused on B7-H3, a protein associated with lung cancer. Its analysis indicated that B7-H3 was highly expressed in fibroblasts, connective-tissue cells found near tumours.

The AI agents examined how these cells interact with nearby immune cells and concluded that B7-H3-expressing fibroblasts appeared to suppress immune activity around tumours.

The virtual biotech then proposed an antibody-drug conjugate targeting B7-H3. Such therapies use an antibody to locate cells carrying a particular protein and deliver a toxic drug payload to them.

The proposal was based only on information available before January 2025. Several months later, in August 2025, an established pharmaceutical company independently developed the same broad B7-H3 antibody-drug conjugate strategy. That therapy subsequently received US Food and Drug Administration breakthrough therapy designation, providing what the Stanford team described as independent validation of its AI-generated target and drug concept.

What this means for drug discovery 

The Stanford experiment points to a shift from AI being used simply as a research assistant to AI agents being organised into autonomous research teams capable of dividing complex scientific problems among themselves.

The potential advantage is speed. Instead of having individual researchers spend months searching databases, reviewing papers and compiling trial results, thousands of specialised agents can work simultaneously.

But the system is not a replacement for laboratories or clinical researchers. Stanford researchers stress that the AI-generated findings still need to be tested through physical experiments and, ultimately, human clinical trials.

The team is now planning to take other targets identified by the virtual biotech into real laboratories to determine how many of its predictions hold up experimentally.