Overview
- Mid‑September 2026, Stanford published a Science paper describing a Virtual Biotech made up of about 37,000 coordinated AI scientist agents that analyzed more than 50,000 clinical trials and multimodal molecular data.
- The platform assigned agents to individual trials and built two scoring systems—cell‑type specificity and bimodality (switch‑like on/off expression)—that correlated with higher advancement rates and fewer adverse events in historical trial data.
- Using data available before January 2025, the system proposed an antibody‑drug conjugate targeting B7‑H3 (CD276) for lung cancer, and an independent drugmaker later pursued a similar B7‑H3 ADC that received FDA breakthrough therapy designation in August 2025.
- Authors and outside scientists emphasize these outputs are computational hypotheses that have not been validated in laboratory experiments or clinical trials and that human oversight and experimental follow‑up are required.
- The study shows agentic AI can speed large‑scale evidence curation and target prioritization but also highlights the need for governance measures such as role specialization, audit trails, time‑bounded data sources, and independent validation before clinical use.