Archive · 20 Sep 2026
Stanford Deploys 37,000 AI Agents to Hunt New Therapies
Stanford researchers built a large team of computer programs that can work on science questions together.
The programs studied about 50,000 clinical trials in less than a week.
They looked for clues that might show whether a new medicine will work safely in people.
Two clues were whether a drug targets a specific kind of cell and whether a gene acts like an on-off switch.
Medicines aimed at targets with these clues did better in older trial data.
The computer team then studied a protein called B7-H3 linked to lung cancer.
It suggested a medicine that could carry a toxic drug directly to cells with that protein.
Scientists still need to test the idea in laboratories and human trials.
Stanford researchers organized 37,000 AI agents to analyze approximately 50,000 clinical trials in less than a week.
The agents identified cell-type specificity and switch-like gene activity as possible markers of drug-development success.
Drugs targeting genes with these characteristics were more likely to advance and reach the market, with fewer adverse events in historical data.
The system proposed a B7-H3 antibody-drug conjugate for lung cancer after analyzing tumor-associated fibroblasts and immune activity.
Researchers said the findings still require laboratory experiments and human clinical trials before the approach can be validated.
- Who
- Stanford researchers and their 37,000 AI agents conducted the analysis.
- What
- The AI system analyzed clinical trials, identified possible predictors of drug success, and proposed a B7-H3 antibody-drug conjugate for lung cancer.
- When
- The research was published in Science on September 17; the analysis used information available before January 2025, and an independent related development occurred in August 2025.
- Where
- The analysis covered clinical-trial and molecular data across diseases including cancer, brain, heart, kidney, and lung conditions; the proposed therapy focused on lung cancer.
- Why
- The project aimed to identify drug targets more likely to succeed and accelerate early-stage drug discovery.
Potential of Autonomous AI
Need for Human Validation
Speed and scale
Potential of Autonomous AI
Thousands of specialized agents can divide complex research tasks and analyze tens of thousands of trials in less than a week.
Need for Human Validation
The article does not establish that rapid analysis alone produces reliable treatments; the predictions still require physical testing.
Drug-discovery capability
Potential of Autonomous AI
The system identified target characteristics associated with better historical outcomes and proposed a B7-H3 cancer-therapy strategy later echoed by an established pharmaceutical company.
Need for Human Validation
The proposed findings are not yet proven therapies and must be tested in laboratories and, ultimately, human clinical trials.
- AI agents deployed
- 37,000
- Clinical trials analyzed
- Approximately 50,000
- Analysis time
- Less than one week
- Key target characteristics
- Cell-type specificity and bimodal, switch-like gene activity
- Historical success association
- 40% greater likelihood of advancing from Phase 1 to Phase 2
- Market association
- 48% greater likelihood of reaching the market
- Safety association
- 32% fewer adverse events
- Proposed therapy
- A B7-H3 antibody-drug conjugate for lung cancer