Science: Stanford’s ‘Virtual Biotech’ of 37,000 AI agents mines ~50,000 clinical trials and designs a lung-cancer ADC later matched by pharma
Result confirmed
The takeaway
On Sept 17, 2026 James Zou’s Stanford group (lead author Harrison Zhang) published the “Virtual Biotech” in Science (aeg6779): about 37,000 specialized AI agents organized like a biotech company under a chief-scientific-officer agent.
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Peer-reviewed
- Importance
- 3 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra109 days after its cutoff
- Claude Opus 5.579 days after its cutoff
- Gemini 4 Hydrogen170 days after its cutoff
- Grok 4.7109 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 79 days before it.
Key facts
- Paper: Science, Sept 17, 2026, doi 10.1126/science.aeg6779; senior author James Zou, lead author Harrison Zhang (Stanford Medicine)
- Structure: a CSO agent takes a human’s question, delegates to specialised divisions and combines results; about 37,000 agents in total
- Clinical-trial analysis: ~50,000 trials in under a week
- Finding: drugs whose targets show high specificity and bimodal (‘switch-like’) expression were 40% more likely to advance from phase 1 to 2, 48% more likely to reach market, and had 32% fewer adverse events; the pattern held across cancer, brain, heart, kidney and lung conditions
- Drug design: an ADC against B7-H3 for lung cancer, using data from before January 2025; an independent company developed the same strategy (August 2025), which received FDA breakthrough-therapy designation; VentureBeat names Merck
- Grew out of Zou’s 5–8-agent ‘Virtual Lab’ that designed nanobodies (Nature, 2025); the LLM used is not named in the press coverage
- Print issue: Science, 22 Oct 2026. The abstract gives exact figures: ‘more than 37,000 agents annotated outcomes from 55,984 trials’, and drugs targeting cell type–specific genes were ‘48% more likely to reach market with 32% fewer adverse events’; it also analysed a terminated ulcerative colitis trial and ‘inferred potential mechanisms of failure’
What happened
Zou’s group scaled its earlier Virtual Lab, a handful of agents mirroring a university lab, to a company-scale organisation of tens of thousands of agents. Its two headline outputs were a statistical rule for which drug targets succeed and a concrete therapeutic design.
Why it matters
It is one of the largest multi-agent scientific systems published in a top journal, and an early test of whether “AI companies of agents” can do drug-discovery work. The drug design was matched by an independent programme, which is a form of external validation, but it was not new to the field.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 2 primary, 2 press
Primary
- Stanford Medicine: Virtual biotech company puts thousands of AI scientist agents to work on drug discoverymed.stanford.edu, official
- Crossref record of the Science paper (abstract)api.crossref.org, docs
Press
- VentureBeat: Stanford is running 37,000 AI agents as a virtual biotechventurebeat.com, press
- Singularity Hub: Virtual biotech company puts 37,000 AI agents to work on drug discoverysingularityhub.com, press
Changes
- Added the print-issue date and the exact trial count from the abstract
- Filed