A Stanford Medicine team has built a fully virtual biotech company staffed by 37,000 AI agents - no lab space, no payroll, no humans. The agents handle the entire drug development pipeline, from identifying molecular targets to designing clinical trials, and they have already surfaced a biological signal that predicts which drug candidates are more likely to succeed.
The project, led by associate professor of biomedical data science James Zou, PhD, and graduate student Harrison Zhang, extends Zou's earlier virtual lab concept. "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. The work was published Sept. 17 in Science, with Zou as senior author and Zhang as lead author.
How the virtual company operates
The company mirrors the organizational chart of a brick-and-mortar biotech. A chief science officer agent leads research teams broken into specialized divisions that work in parallel. Each division tackles a core element of drug design, such as molecular target identification or clinical trial planning.
One advantage of an AI company is skipping the startup phase entirely. The agents are trained to support the full drug development pipeline, and their output, while intangible, has produced results that align with real-world pharmaceutical work.
A biological clue hidden in clinical data
Zou tasked the agents with hunting for characteristics that set successful drugs apart. Instead of flooding agents into scientific literature, he assigned a single agent to analyze one specific clinical trial at a time, retrieving safety and effectiveness data. The agents catalogued roughly 50,000 trials in under a week - work that would take human teams years.
For trials with single-cell gene activity data, the agents built two scoring systems. One measured how specifically a drug targeted a particular cell type. The other measured bimodality, which indicates whether a targeted gene behaves like an on-off switch or a dimmer. Drugs targeting switch-like genes were 40% more likely to advance from phase 1 to phase 2 trials, 48% more likely to reach market, and had 32% fewer adverse events compared with drugs that had a broad spectrum of activity. The pattern held across cancers, brain diseases, heart diseases, 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. His theory: a target that acts like an on-off switch and homes in on a specific cell type may be easier and safer to control with a drug.
An AI-designed therapy validated independently
To test whether the virtual company could design a viable drug, the agents focused on B7-H3, a protein lung cancer researchers have studied for years. The agents analyzed data from multiple studies and biomedical repositories and found that B7-H3 was highly expressed in fibroblasts, connective tissue cells often found near tumors.
Spatial-transcriptomic analyses and cell communication data revealed that fibroblasts expressing B7-H3 appeared to suppress nearby immune cells, cloaking the tumor from immune defenses. The agents designed an antibody-drug conjugate - a protein-based delivery system that targets cells carrying B7-H3 proteins and delivers a toxic chemotherapy payload directly to them. The design used only information available before January 2025. In August 2025, a private pharmaceutical company independently arrived at the same antibody-drug conjugate strategy against B7-H3, and the therapy later received an FDA breakthrough therapy designation.
"This was really exciting as an independent, third-party validation that's consistent with the effects and the design proposed by the virtual biotech," Zou said.
Why this matters for Science and Research
For research scientists, this work demonstrates that AI agents can independently surface actionable biological insights from existing clinical data - insights that later align with independent drug development efforts. The approach does not replace lab work, but it can compress years of literature review and trial analysis into days, flagging high-probability drug targets before expensive wet-lab experiments begin. Researchers interested in applying AI to their own workflows can explore structured training in this area through an AI Learning Path for Research Scientists or follow developments in AI for Science & Research.
Zou said the next step is bringing the virtual biotech's new candidate targets into real laboratories. "Humans, physical experimentation and validation will always be the conduit through which AI makes an impact." The study was funded by a Knight-Hennessy Scholarship, the National Institutes of Health, the National Science Foundation, and Chan Zuckerberg Biohub.
Your membership also unlocks:
