Drug discovery is one of humanity’s most expensive and least efficient processes. It takes over a decade, costs hundreds of millions of dollars, and still fails 90% of the time — most drug candidates that enter clinical trials never reach patients. The evidence that could help catch failures earlier exists, but it’s scattered across scientific disciplines, data formats, and decades of published literature that no human team could ever fully absorb.
Stanford researchers think they’ve found a better way. They’ve built a virtual biotech company staffed entirely by AI agents — 37,000 of them — and published the results in Science. The system didn’t just outperform existing drug selection methods. It independently proposed a lung cancer treatment that a major pharmaceutical company later arrived at on its own.
A Company With No Human Employees
The architecture mirrors a real drug development organisation. A virtual Chief Scientific Officer receives a query from a human user and delegates tasks downward to an army of specialised agent divisions: one focused on finding and validating drug targets, one assessing safety risks, one evaluating how drugs should be delivered, and one reviewing clinical trial data.
Each division has its own tools and databases. The entire system has built-in access to Open Targets, a massive public repository of clinical trial data — exactly the kind of broad, cross-disciplinary resource that is practically impossible for human researchers to fully exploit.
The First Test: Cleaning Up the Data
When the researchers gave the system an existing study on how genetic evidence predicts drug success, the virtual CSO’s first move was not to generate hypotheses. It was to improve the quality of the underlying data.
Many trials in the Open Targets database don’t clearly record whether the drug actually worked. So the CSO dispatched one AI agent to each of 37,075 individual Phase II and III clinical trials, tasking them with digging through trial registries, published papers, and press releases to determine outcomes. The entire job took around six hours. A team of human researchers doing the same work would take months or years.
With cleaner data in hand, another agent analysed which genes — potential drug targets — were most likely to lead to successful treatments. It developed a two-part scoring system: first measuring whether a gene was active in just one cell type or across many, and second assessing whether its activity behaved more like an on-off switch or a dimmer dial.
The pattern that emerged was striking. Drugs targeting switch-like genes active in only a small number of cell types were 48% more likely to reach the market, 40% more likely to advance from Phase 1 to Phase 2 trials, and caused 32% fewer adverse events than drugs aimed at more broadly active targets. It’s a finding that, if it holds up, could meaningfully reshape how drug companies choose which targets to pursue in the first place.
The Cancer Treatment Nobody Told It About
The researchers then pushed the system further, asking it to evaluate a protein called B7-H3 associated with lung cancer. The agents discovered the protein was particularly abundant in connective-tissue cells found close to tumours — and then found evidence that these cells were actively suppressing nearby immune cells, preventing the body from detecting the cancer.
The proposed solution: tag cells expressing B7-H3 with an antibody, then use it to direct a toxic chemotherapy drug precisely to those cells, sparing healthy tissue. It was an elegant strategy built entirely from data available before January 2025.
In August 2025, a major pharmaceutical company independently arrived at the same approach. Its B7-H3-targeted therapy — ifinatamab deruxtecan — received FDA breakthrough therapy designation. The virtual biotech had reached the same conclusion, from the same publicly available data, with no knowledge of what the drug company was working on.
“This was really exciting as an independent, third-party validation,” said senior author James Zou.
What This Can and Can’t Do
The researchers are clear about the limits. Identifying better drug targets is only one step in a process that still requires extensive laboratory testing and rigorous clinical trials before any treatment reaches a patient. The virtual biotech can’t compress or replace those steps — it can only help ensure that the candidates entering them are better chosen to begin with.
But given that 90% of drugs currently fail that gauntlet, a system that can meaningfully improve the odds at the selection stage is genuinely valuable. The bottleneck in drug discovery has never been a shortage of scientific ideas. It’s been the inability to rapidly and comprehensively evaluate which ideas are worth pursuing. An army of 37,000 AI scientists, working in parallel, crunching through decades of trial data in hours, represents a fundamentally different kind of tool for that problem.
The virtual biotech is not replacing pharmaceutical companies. But it may be the closest thing yet to giving them a research team that never sleeps, never misses a paper, and never runs out of bandwidth.