TL;DR
Stanford researchers built a virtual biotech staffed by 37,000 AI agents that combed 55,984 clinical trials, discovered a predictive marker for drug success, and independently proposed a lung cancer therapy the FDA later fast-tracked.
What happened
- Stanford's James Zou lab published findings in Science showing a virtual biotech system can run end-to-end drug discovery, from target identification to clinical trial design.
- 37,075 agents deployed simultaneously processed Phase II and III trial records from ClinicalTrials.gov in roughly six hours, a task Zou said would have taken years by hand.
- Agents labeled outcomes for 55,984 trials total, with human reviewers agreeing with agent assessments 88% of the time on a 100-trial spot check.
- The system identified that drugs targeting narrowly active genes (active in few cell types, switch-like behavior) were 40% more likely to advance from Phase 1 to Phase 2, 48% more likely to reach Phase 4, and reported 32% fewer adverse events.
- Running the agents cost Anthropic API fees of a median 23 cents per trial, making the economics radically cheaper than traditional trial analysis.
Why it matters
- Roughly 90% of drugs entering clinical trials never reach market, so a reliable early filter for target selection could redirect billions in R&D spending toward higher-probability candidates.
- The system independently proposed an antibody-drug conjugate targeting B7-H3 in lung cancer, using only data before January 2025. In August 2025, the FDA granted breakthrough therapy designation to ifinatamab deruxtecan, a drug using the identical strategy.
- The new cell-type specificity marker adds a second predictive signal alongside the existing genetic-evidence standard, and it held up even after controlling for that advantage.
- Four independent human reviewers repeated eight analyses and reached the same conclusions as the agents every time, with no code errors that changed results, strengthening reproducibility claims.
- The colitis case study suggests the system can diagnose why failed trials failed, pointing to multi-receptor inflammation pathways that a single-target drug would miss.
What to watch next
- Whether lab validation of the lung cancer and colitis findings confirms the agents' hypotheses. Zou named this the explicit next step, and it is the critical gap between correlation and proof.
- Adoption by commercial pharma and biotech of the cell-type specificity scoring system as a standard target-prioritization filter, which would signal the finding has moved from academic paper to industry practice.
- Expansion of the model to rare and understudied diseases, where the authors themselves flag weaker expected performance due to sparse training data and limited trial registries.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.