The latest company to spin out of a Stanford Medicine lab has 37,000 employees, and not one of them is human. There is no lab space, no payroll, and no lunch breaks. In a study published September 17 in the journal Science, researchers describe a fully AI-staffed virtual biotech company that analyzed tens of thousands of clinical trials, uncovered a biological signal that predicts which drugs will succeed, and independently designed a lung cancer therapy that a major pharmaceutical company later arrived at on its own.
An Entire Drug Company Made of Agents
The project is the brainchild of James Zou, associate professor of biomedical data science at Stanford Medicine, and graduate student Harrison Zhang, who led the work. It builds on the "virtual lab" of AI scientist agents that Zou's group launched in 2025 — but this time the team went further, creating an entire company with tens of thousands of AI agents trained to support the full pipeline of drug development.
"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?" Zou said in the Stanford Medicine announcement. "Could we have a fully agentic company that tackles the extremely complex challenges of drug discovery?"
The virtual company mirrors the organizational chart of an established biotech: a chief science officer agent leads research teams divided into specialized divisions that work in parallel on core elements of drug design, from identifying molecular targets to designing clinical trials.
What 37,000 Agents Found in 50,000 Trials
One of the biggest challenges in drug discovery is predicting which molecules will survive clinical trials — a process Zou notes can cost tens or even hundreds of millions of dollars and take many years. The agents were assigned a meticulous task: rather than sweeping the scientific literature broadly, each agent analyzed a specific clinical trial and retrieved data on safety and effectiveness. In total, they cataloged roughly 50,000 trials in less than a week, a task Zou says would have taken human researchers years.
The agents then examined molecular data collected during those trials and built two scoring systems. The first measured how specifically a drug targets a particular cell type, rather than affecting many cell types broadly. The second measured "bimodality" — whether a targeted gene's activity behaves like a light switch, snapping on and off, or like a dimmer with a range of intermediate states.
The pattern that emerged was striking. Drugs that targeted switch-like genes in a cell-specific way were 40% more likely to advance from phase 1 to phase 2 trials, 48% more likely to reach the market, and had 32% fewer adverse events compared with drugs that had broad activity. The effect persisted across cancers, brain diseases, heart disease, and kidney and lung conditions.
"The science the agents discovered is really exciting, and it shows that these single-cell features can be used to make better drugs," Zou said. "It points to the importance of collecting this kind of data. This could help the entire drug discovery industry."
The B7-H3 Test: From Prediction to Validation
The decisive question was whether an all-AI company could design an actual drug. To find out, the team pointed the agents at B7-H3, a protein long of interest to lung cancer researchers. The agents analyzed studies and biomedical data repositories and found B7-H3 highly expressed in fibroblasts — connective-tissue cells that often live near tumor cells. Digging into spatial transcriptomic data, they discovered that these B7-H3-expressing fibroblasts were signaling to nearby immune cells and suppressing them, effectively cloaking the tumor from the body's natural defenses.
The agents then designed an antibody-drug conjugate — a targeted protein that homes in on cells carrying abundant B7-H3 and delivers a chemotherapy payload directly to them. Crucially, the design relied only on information available before January 2025. Months later, in August 2025, an established pharmaceutical company independently developed the same antibody-drug conjugate strategy against B7-H3. That therapy went on to receive an FDA breakthrough therapy designation after showing effectiveness in a human study — an outcome Zou describes as independent, third-party validation of the virtual biotech's design.
Zou is not pursuing B7-H3 further, since another company is already shepherding it toward clinics. But he says the virtual biotech has surfaced other candidate targets designed in the same way.
Humans Remain the Conduit
Despite the scale of the agent workforce, Zou is explicit about the limits. "Humans, physical experimentation and validation will always be the conduit through which AI makes an impact," he said. The team's next step is moving the virtual biotech's new findings into real laboratories to test how many hold up in the physical world.
The research was funded by a Knight-Hennessy Scholarship, the National Institutes of Health, the National Science Foundation, and Chan Zuckerberg Biohub, with additional support from Stanford's Department of Biomedical Data Science.
Why It Matters
The study arrives amid a broader shift in how AI is applied to science — from models that summarize research to agentic systems that plan experiments, analyze data, and propose novel therapeutics. Coverage of the work by The New York Times and genomics industry press underscores the attention the "virtual biotech" framework is drawing.
If the design-before-validation result on B7-H3 is any indication, the pharmaceutical industry has a new class of research instrument to evaluate: organizations of AI agents that can compress years of literature review and trial analysis into days. The open question — the one Zou's team is now testing — is how much of what the agents propose survives contact with a physical lab bench.
Stay Ahead of AI
AI is transforming research at every level, from drug discovery to fundamental physics. For more breaking AI news and analysis of the research shaping the field, read more AI news →.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
