Stanford Medicine researchers have developed a “Virtual Biotech” comprising tens of thousands of coordinated AI agents organized to mirror the divisions of a drug development company, according to a study published September 17 in Science. The system was used to analyze more than 55,000 clinical trials, investigate potential therapeutic targets, and examine reasons for clinical development failure.
Led by Stanford associate professor of biomedical data science James Zou, the Virtual Biotech places specialized AI agents into groups covering areas including target discovery, safety, therapeutic modality selection, and clinical development, coordinated by a chief science officer agent. The agents integrate evidence from sources including single-cell and spatial transcriptomics, genetics, proteomics, and clinical trial databases.
In an analysis of 55,984 clinical trial records, the researchers identified two target characteristics associated with better development outcomes: expression restricted to relatively narrow cell populations and a bimodal, or on/off, expression pattern rather than continuous expression across cells. Drugs directed against genes displaying both characteristics were associated with a 48% higher likelihood of reaching the market and 32% fewer adverse events than drugs targeting genes without those features.
The researchers also used the Virtual Biotech to investigate therapeutic strategies for lung cancer. By integrating single-cell, spatial, genetic, and clinical evidence, the system prioritized B7-H3 (CD276) and proposed an antibody-drug conjugate (ADC) strategy targeting B7-H3-expressing cells in the tumor microenvironment. The finding converges with an approach already being pursued clinically by Daiichi Sankyo and Merck with ifinatamab deruxtecan, a B7-H3-directed ADC currently under US FDA Priority Review for previously treated extensive-stage small cell lung cancer. Its PDUFA date is October 10, 2026.
The study does not establish that the Virtual Biotech predicted that program prospectively, but the overlap provides a real-world example of the system arriving computationally at a target and modality combination that has independently advanced into late-stage drug development.