# Run a virtual biotech staffed by AI agents

Stanford's Virtual Biotech is staffed by AI agents: more than 37,000 of them annotated outcomes from 55,984 clinical trials, the team reports in Science.

Organisation: Stanford University (Zou group) and collaborators
Status: Research (published in Science)
Place: United States
Categories: Science & discovery, Health & care
Built with: Claude
Try it with: Claude, ChatGPT / OpenAI models, Gemini, Grok
Time to try: 1 hour

## Why it matters

Most drug candidates fail before they reach patients. A company of AI agents that reads trial outcomes, with people setting the questions, helps researchers see which kinds of targets are more likely to succeed.

## How it works

A chief scientific officer agent plans each question and hands parts of it to specialist scientist agents for target discovery, safety, choice of drug type and clinical development. Each specialist queries its own databases, such as Open Targets, ClinicalTrials.gov and CELLxGENE, through MCP tools, a reviewer agent checks the evidence and the chief scientific officer agent pulls the findings together, while humans set the questions and steer the analysis. Across those trials, the agents found that drugs aimed at genes active in specific cell types were 48% more likely to reach market, with 32% fewer adverse events (Science). Asked to assess a known lung cancer target, the agents proposed an antibody-drug conjugate using information from before January 2025; Stanford Medicine says a pharmaceutical company independently arrived at the same strategy in August 2025. The code is open source under the MIT License and runs on Anthropic's Claude models. It builds on the Zou group's Virtual Lab, whose AI agents designed nanobodies reported in Nature in 2025.

## Who it's for

Research groups, biotech teams and funders choosing which drug targets to study.

## The prompt

```
Plan a virtual biotech review of this question: [scientific question, disease area and what a useful answer would change].

This is a research planning exercise. It is not medical advice and it must not recommend a treatment for a real patient.

Agents:
- Chief scientific officer: splits the question into parts and pulls the findings together.
- Specialists for target discovery, safety, choice of drug type and clinical development. Each one names the database it would query, such as Open Targets, ClinicalTrials.gov or CELLxGENE.
- Reviewer: checks the methods and whether the evidence answers the question.

Work in this order:
1. Restate the question and list the parts the chief scientific officer would hand out.
2. For each part, name the database to query and what result would count as evidence.
3. Have the reviewer name the weakest evidence and what is missing.
4. End with a brief synthesis and the questions a human should decide before anyone acts.

Use only sources you have actually retrieved. If you cannot retrieve a source, mark the claim [CITATION NEEDED]. Do not invent trial results, response rates or patient outcomes.
```

## Caveats

These are computational results. They have not been tested in the lab or in trials. The work comes from the same Stanford lab as the Virtual Lab pick on the site.

## Sources

The Science paper, the Stanford Medicine article and the code repository are the Stanford team's own work. No independent evaluation was found.

1. Science: More than 37,000 agents annotated outcomes from 55,984 trials; drugs aimed at genes active in specific cell types were 48% more likely to reach market, with 32% fewer adverse events. The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development (https://www.science.org/doi/10.1126/science.aeg6779)
2. Stanford Medicine: The agents proposed an antibody-drug conjugate for a known lung cancer target using information from before January 2025; a pharmaceutical company independently arrived at the same strategy in August 2025. Virtual biotech company puts thousands of AI scientist agents to work on drug discovery (https://med.stanford.edu/news/all-news/2026/09/virtual-biotech-company.html)
3. GitHub (Stanford, Zou group): A chief scientific officer agent coordinates scientist agents through MCP tools, including Open Targets, ClinicalTrials.gov and CELLxGENE; the code is open source under the MIT License and runs on Claude. harrisongzhang/TheVirtualBiotech (https://github.com/harrisongzhang/TheVirtualBiotech)

Fact-checked: 17 Oct 2026

Page: https://escapevelocitysociety.com/picks/stanford-virtual-biotech-agents

See Selected Projects: https://escapevelocitysociety.com/picks.md
