Virtual biotech company puts thousands of AI scientist agents to work

Researchers at Stanford built a fully virtual biotech company composed of roughly 37,000 autonomous AI scientist agents that together emulate the organizational structure and workflows of a real drug-development firm. Each agent was trained for specific tasks across the discovery-to-clinic pipeline; for example, individual agents were assigned to extract and catalogue data from about 50,000 clinical trials in under a week. The agents mined single-cell molecular data and developed two predictive scores - cell-type specificity and gene-expression bimodality (on/off vs. graded activity) - that correlated strongly with clinical success. Drugs whose targets scored high on both measures were 40% more likely to advance from phase 1 to phase 2, 48% more likely to reach market, and produced 32% fewer adverse events; these patterns held across cancers and multiple other disease areas.

The virtual company also independently designed a therapeutic strategy against B7-H3 in lung cancer by identifying B7-H3-expressing fibroblasts that suppress nearby immune cells and proposing an antibody-drug conjugate to selectively kill those cells. A commercial pharma later arrived at the same ADC approach and secured FDA breakthrough designation after human studies; the AI design used only data available before January 2025. The work, led by James Zou and Harrison Zhang, is published in Science (Sept. 17, 2026) and the team plans to move promising AI-generated targets into physical labs for experimental validation.