The Pulse
Stanford Deploys 37,000 AI Agents to Design Drug Candidates
Stanford researchers created a virtual biotech company staffed by thousands of AI agents that analyzed clinical trials, identified biological features associated with drug-development success and proposed a lung-cancer therapy later indepen
AI.info Team ·
Stanford researchers have assembled a virtual biotech with 37,000 AI agents, none of them human, to analyze clinical trials, assess drug targets and propose therapies. The system examined roughly 50,000 trials in less than a week and identified biological features associated with better drug-development outcomes.
The work, published September 17 in Science, also produced a proposed lung-cancer treatment aimed at the protein B7-H3. Months after the AI system made its proposal using information available before January 2025, a pharmaceutical company independently developed the same antibody-drug conjugate strategy. Stanford describes that result as an external validation, not proof that the virtual company has created a drug ready for patients.
“Our idea was to see how far we could push this. Could we create a biotech company that takes on everything from looking for drug targets all the way to designing clinical trials? Could we have a fully agentic company that tackles the extremely complex challenges of drug discovery?” —James Zou, associate professor of biomedical data science at Stanford Medicine
37,000 agents split the work of a biotech company
The Virtual Biotech mirrors the structure of a conventional drug-development organization. A chief science officer agent receives a research question, assigns work to specialized divisions and combines their findings. The divisions work in parallel on core elements of drug design, including target identification and clinical-trial planning.
That architecture is intended to address a familiar problem in pharmaceutical research: important evidence sits in separate databases and is often reviewed by teams working in isolation. Individual agents can focus on one narrow task while a coordinating agent passes findings between specialties and requests additional analysis when the evidence is incomplete.
The trial analysis found two biological features tied to success
For its first major test, the virtual biotech assigned agents to extract structured information from clinical-trial records, papers and related sources. The agents analyzed and catalogued roughly 50,000 trials in less than a week, a task Stanford says would have taken human researchers years.
Alongside the trial results, the agents investigated molecular data collected during the trials. They built two scoring systems: one evaluated how specifically a drug targeted a certain cell type, while the other measured bimodality — whether a targeted gene’s activity behaved more like a light switch than a dimmer.
Trials with high scores in both categories fared better. Drugs aimed at switch-like genes were 40% more likely to move from Phase 1 to Phase 2, according to the Stanford report. They were also 48% more likely to reach the market and showed 32% fewer adverse events than drugs with broader activity patterns. The associations appeared across cancer, brain, heart, kidney and lung diseases, although those findings describe historical patterns rather than a guarantee for any individual candidate.
B7-H3 became the system’s real-world test
Stanford then asked the virtual biotech to assess B7-H3, also known as CD276, as a possible lung-cancer target. The agents analyzed relevant data from studies and biomedical repositories and reported that B7-H3 was highly expressed in fibroblasts, connective-tissue cells often found near tumors.
The analysis suggested that fibroblasts expressing B7-H3 signal to nearby immune cells and suppress their activity, helping tumors avoid immune attack. Based on that evidence, the agents proposed an antibody-drug conjugate: an antibody designed to bind B7-H3-bearing cells and deliver a toxic chemotherapy payload.
The system made its proposal using data available before January 2025. In August 2025, a private, well-established pharmaceutical company independently arrived at the same antibody-drug conjugate strategy against B7-H3. That therapy later received the U.S. Food and Drug Administration’s breakthrough therapy designation after showing effectiveness in a human study.
That sequence gives the Stanford work a stronger test than a retrospective match with an old paper, but it does not establish that the AI discovered a commercially viable medicine. The outside company developed and tested its therapy separately, while Stanford’s agents produced a computational hypothesis that still requires laboratory and clinical validation.
The system still stops before the laboratory
Zou’s group presents the Virtual Biotech as an early-stage research system, not an autonomous pharmaceutical company. Its agents can search evidence, write analyses, compare targets and suggest therapeutic designs, but they do not manufacture compounds, run experiments or enroll patients.
That boundary matters because drug development fails for reasons that databases cannot fully capture. A candidate must survive experiments in cells and animals, manufacturing constraints, toxicity testing, regulatory review and controlled human studies. A convincing computational rationale is only one part of that sequence.
The researchers say the next step is to take additional targets identified by the virtual biotech into physical laboratories and measure how many withstand experimental testing. The B7-H3 result gives the system a specific benchmark: whether its other proposals can move from coordinated analysis to reproducible biological evidence.
For now, Stanford has demonstrated a system that can distribute a large biomedical research task across thousands of specialized agents and return a testable therapeutic hypothesis. The next answer will come from experiments on the candidates that remain outside the screen.