The agents linked cell-specific drug targets to better clinical trial outcomes. They also proposed a lung cancer treatment strategy matching one independently pursued by a drugmaker.
Stanford researchers have built a system of more than 37,000 AI agents to sift the evidence behind drug development. Graduate student Harrison Zhang and biomedical data scientist James Zou organized them into research teams under a chief scientific officer agent, with human researchers guiding the work. Their paper on the system, which they call the Virtual Biotech, appeared in Science on September 17.
Zhang and Zou gave each clinical trial to a single agent, which pulled out its safety and effectiveness data. The agents catalogued 55,984 trials in less than a week.
About 90% of drug candidates entering clinical trials never reach approval, most often on efficacy or safety. Drugs aimed at genes specific to particular cell types did better: they were 40% likelier to pass from phase 1 to phase 2 and 48% likelier to reach market, and they came with 32% fewer adverse events. The pattern held across cancers and conditions affecting the brain, heart, kidneys and lungs. The agents found it in trials that had already run; none of the drugs was one they designed.
In a second exercise, the agents proposed treating lung cancer with an antibody that delivers chemotherapy to cells carrying the B7-H3 protein. Working only from information available before January 2025, they reached an approach a drugmaker independently pursued that August. Stanford presents the match as support for the system's reasoning.
In a third test, the agents worked back from a terminated ulcerative colitis trial to the likely reasons it failed, then proposed selecting patients by biomarker so a new trial would not fail the same way.
The agents' conclusions depend on the evidence available, which leaves poorly studied diseases and targets out of reach for now. Zou says the next step is to take other proposed targets into physical labs and see which hold up.