Stanford Medicine researchers built a simulated biotechnology company with about 37,000 artificial intelligence(AI) agents. The researchers arranged a chief science officer program to oversee groups that hunt drug targets and plan studies in people. This work is published in Science.
Guessing which medicines will pass clinical trials is a central problem. One agent was assigned to each study. The programs catalogued about 50,000 trials in less than a week and inspected molecular measurements collected during those studies.
Where single-cell data existed - records of which genes are active in individual cells - the agents scored how tightly a drug hit one cell type and whether a target gene showed bimodality, acting more like an on-off switch than a dimmer. High scores on both lined up with better results. Drugs aimed at switch-like genes were 40 percent more likely to move from early safety tests (phase 1) to larger tests (phase 2), 48 percent more likely to reach the market, and had 32 percent fewer harmful side effects than broadly acting drugs. The pattern appeared in cancers and in brain, heart, kidney, and lung disease.
A later match in the clinic
The programs then studied B7-H3, a protein watched in lung cancer, and said it is abundant on fibroblasts, connective-tissue cells near tumors. Location maps of gene activity suggested those cells quiet nearby immune cells and help hide the tumor. Using only information available before January 2025, the agents proposed an antibody-drug conjugate, an antibody that delivers chemotherapy to cells rich in B7-H3. In August 2025 a private company independently chose the same strategy. That medicine later received a Food and Drug Administration breakthrough therapy designation, a status that can speed review after early benefit in people. The Stanford Madicine researchers are not pursuing B7-H3. They say other targets have appeared and that laboratory tests by people must still check which findings hold.