A biotech company spun out of a Stanford Medicine lab has 37,000 employees, and none of them are human. Research published in Science and announced by Stanford Medicine on September 17, 2026, describes a virtual biotech company built from tens of thousands of AI agents spanning the entire drug-development pipeline, from hypothesis generation to trial-outcome prediction. The project is led by James Zou, an associate professor of biomedical data science, and graduate student Harrison Zhang, growing out of a virtual lab concept the pair launched in 2025.
An Entire Company, Simulated
Instead of a handful of AI tools bolted onto existing lab workflows, Zou’s team built specialized agents that mimic different roles inside a real pharmaceutical company: agents that analyze biological data, agents that propose therapeutic hypotheses, agents that predict clinical trial outcomes, and agents that critique each other’s work before a proposal moves forward. Coordinating 37,000 of these agents required infrastructure that treats each one less like a chatbot and more like an employee with a defined job function and reporting structure, which is part of what makes the project notable beyond its scientific results.
The Numbers Behind the Claim
The team’s retrospective analysis found that drugs targeting cell-type-specific genes, the kind of nuanced targets its agents were designed to flag, were 40% more likely to progress from Phase I to Phase II trials, 48% more likely to eventually reach the market, and associated with 32% fewer adverse events than a comparison group of less targeted drugs. Those figures came from analyzing patterns across large volumes of existing trial data rather than new experiments, giving the system a track record to validate its reasoning against before it was tasked with generating something new.
The Lung Cancer Case That Got Attention
The most concrete result involved lung cancer. The virtual biotech’s agents analyzed multiple layers of biological data and identified unusually high levels of a protein called B7-H3 in fibroblasts surrounding lung tumors, reasoning that these cells might be helping tumors hide from the immune system. Based on that finding, the system proposed an antibody-drug conjugate targeting B7-H3 as a possible treatment strategy. Separately, a pharmaceutical company working through conventional means arrived at a similar targeting strategy, which went on to receive FDA breakthrough therapy designation, giving Zou’s team an unusual real-world check on whether their AI’s reasoning matched what actual drug development independently converged on.
Why Some Researchers Are Withholding Judgment
Drug discovery specialists outside Stanford caution against over-reading a single convergence as proof the system generates novel science rather than pattern-matching against publicly available literature that human teams were also drawing from. It’s difficult to rule out that both the AI system and the pharmaceutical company were responding to the same published signals about B7-H3 biology that were already circulating before either effort began. There’s also a structural question hanging over any agent-based system this large: with 37,000 agents producing outputs, some fraction will look prescient in hindsight purely by chance, and the paper’s framing emphasizes its hits without the same visibility into how many proposed targets went nowhere.
What This Means for Drug Discovery Going Forward
Even skeptics generally agree the approach demonstrates something practically useful: a way to generate and triage far more therapeutic hypotheses than a human research team could evaluate manually, then rank them using outcome patterns learned from historical trial data. Zou’s team has framed the project as an open scientific instrument rather than a commercial venture, meaning other labs may be able to adapt pieces of the agent architecture rather than treating it as a black box. The next real test isn’t another retrospective analysis, it’s whether the virtual biotech, or a system built on its methods, can independently propose a therapeutic hypothesis that no human team has already converged on, and have that hypothesis hold up in a prospective trial designed specifically to test it.
Photo: Artem Podrez / PEXELS via Pexels
