Stanford builds AI drug-discovery system with 37,000 agents

A Stanford team has designed a computer-based drug company in which as many as 37,000 artificial intelligence agents divide up the work of drug discovery. The system can investigate possible drug targets, examine safety concerns, compare clinical-trial results and suggest treatments, according to reporting by Singularity Hub.

The researchers describe the system in a paper published in Science. In one test, it identified characteristics linked with drugs that are more likely to succeed in development. In another, it proposed a treatment strategy aimed at a lung-cancer target that a major pharmaceutical company later pursued independently.

The project addresses a persistent weakness in pharmaceutical research. About nine out of every 10 drugs entering clinical trials do not ultimately become approved products. Some fail because findings from laboratory experiments do not hold up in patients. Others produce harmful effects that were not identified early enough.

A company made of specialised AI teams

Drug-development evidence is spread across scientific fields, databases, trial registries, papers and company announcements. That fragmentation makes it difficult for a conventional research group to evaluate every relevant piece of information before deciding whether a target is worth pursuing.

Stanford’s virtual biotech is organised to imitate the structure of a drug company. A virtual chief scientific officer receives a question from a human researcher and assigns portions of the investigation to specialised agents. The agents are equipped with separate data sources and analytical tools, then operate in four broad areas: identifying and testing drug targets, examining potential safety problems, determining how a treatment might be delivered, and analysing previous clinical trials.

The system can also use Open Targets, a large public resource containing information related to drug targets and clinical research. To test whether the virtual company could improve on existing work, the researchers supplied it with a study indicating that human genetic evidence may help predict which medicines will succeed in trials.

The virtual chief scientific officer first focused on the quality of the available evidence. Many records did not make it clear whether a trial treatment had actually worked. The system therefore assigned individual agents to review 37,075 Phase II and Phase III trials. They searched registries, scientific publications and press releases for outcome information, completing the task in roughly six hours.

What the system found about drug targets

The next stage examined genes in a public database of human tissues. This resource shows which genes are active in particular cell types. The AI system assessed each candidate using two measures. One considered whether a gene was concentrated in a limited set of cell types or active across many. The other looked at whether its activity behaved more like a binary switch or could vary gradually.

When the resulting scores were compared with the improved trial-outcome dataset, a pattern emerged. Medicines aimed at genes with switch-like activity and limited distribution among cell types were 48 percent more likely to reach the market. They were also 40 percent more likely to progress from Phase I to Phase II, and were associated with 32 percent fewer adverse events than drugs aimed at targets active across wider ranges of tissue.

Those findings do not remove the uncertainty involved in drug development, but they could help researchers decide which biological targets deserve further investment. Avoiding weak candidates earlier may reduce the resources spent on programmes that are unlikely to succeed.

A lung-cancer test case

The researchers then asked the virtual biotech to study B7-H3, a protein linked to lung cancer. Its agents found that the protein was especially common in fibroblasts, connective-tissue cells often located near tumours. They also found evidence that these cells could inhibit nearby immune cells, making it harder for the immune system to recognise and attack cancer.

The proposed treatment would use an antibody to identify cells carrying B7-H3 and guide a toxic chemotherapy compound towards them. The system based its recommendation only on information available before January 2025.

In August 2025, ifinatamab deruxtecan, a B7-H3-directed therapy developed by a major pharmaceutical company, received breakthrough therapy status from the US Food and Drug Administration. The company had arrived at a similar strategy independently. Stanford senior author James Zou described that development as outside confirmation consistent with the virtual biotech’s proposed design.

The result is not evidence that the AI system can take a medicine from idea to approval on its own. Selecting a promising target is only an early step. Candidates still require laboratory experiments, safety testing and lengthy clinical trials, all of which remain expensive and difficult to accelerate.

Even so, the system demonstrates how large numbers of specialised AI tools could examine scattered evidence faster than a conventional team. If that capability helps researchers eliminate poor choices and identify stronger ones earlier, it could improve one of the drug industry’s most failure-prone stages.