Stanford's all-agent biotech put one AI agent on each of 37,075 clinical trials

The 37,000 figure tracks the reading list. Nature reports that the chief scientific officer agent handed out 37,075 agents, one to each later-stage trial ^[6], and 37,075 rounded to the nearest thousand gives the 37,000 "employees" in the Stanford account, none of them human and none on a payroll ^[2][25]. Staffing followed the size of the corpus. The divisions describe how the work was split.

Zou did not turn a swarm loose on the literature. Each agent took a single trial and retrieved that trial's safety and effectiveness data ^[9]. Every output stays attached to one named study. Zou said the agents analyzed and catalogued some 50,000 trials in less than a week, work he put at years for human staff ^[8]. Nature describes the corpus as more than 55,000 published trial results ^[7].

The predictive signal came from agents working only on trials that had single-cell gene activity data ^[10]. They built one score for how narrowly a drug hit a specific cell type and a second for bimodality, meaning whether the targeted gene's activity behaves like an on-off switch or a dimmer ^[10]. Drugs against switch-like targets advanced from phase 1 to phase 2 40% more often, reached market 48% more often, and had 32% fewer adverse events than broad-acting drugs ^[11]. The pattern held across cancers, brain diseases, heart diseases, kidney and lung conditions ^[13]. Nature gives the market figure as nearly 50% and ties it to cell-type specificity; the Stanford account ties it to switch-like gene behaviour ^[12]. Neither account gives the number of catalogued trials that carried the single-cell data the scores needed ^[27].

The comparison runs across trials that already reported results, so it ranks features that approved drugs happened to have. Zou's explanation is mechanistic: a target that acts like a switch and sits in one cell type may be easier, and therefore safer, to control with a drug ^[14]. "The science the agents discovered is really exciting, and it shows that these single-cell features can be used to make better drugs. It points to the importance of collecting this kind of data," he said ^[15].

The lung-cancer demonstration had a human hand at both ends. Zou's team directed the CSO agent to investigate CD276, a protein that earlier work had linked to dampened immune responses and high expression in lung tumours ^[21]. The system confirmed the candidate using previously collected data and developed a strategy, a CD276-recognizing antibody tethered to an anticancer drug ^[22]. Zou and his collaborators, helped by external reviewers, concluded this was a promising avenue ^[23]. Nature reports that other scientists note the Virtual Biotech has not been vetted in real-world drug discovery and that its predictions were not validated through experiments, let alone clinical trials ^[19]. The Stanford account describes the same output more warmly, as a cancer therapy that a major pharmaceutical company later independently built ^[24].

The agents ran on versions of Claude, from Anthropic, and Zou says any advanced large language model will do, including open-source models researchers can run on their own computers ^[20]. "We want to see how far these agent teams of AI scientists can help us to really accelerate drug discovery and development," Zou told Nature ^[18].