Quick Facts
- Stanford deployed 37,000 AI agents in parallel to annotate 55,984 clinical trials and design novel drug candidates autonomously.
- The system designed an antibody-drug conjugate targeting CD276 for lung cancer; Merck independently developed the same compound, which received FDA Breakthrough Therapy Designation in August 2025.
- Drugs targeting cell-type-specific genes were 48% more likely to reach market and showed 32% lower adverse event rates, according to the system’s analysis.
A Stanford research team has built a virtual biotech powered by tens of thousands of AI agents, and one of its autonomous drug designs was independently confirmed by Merck, whose version received FDA Breakthrough Therapy Designation.
The project is led by James Zou, associate professor of biomedical data science at Stanford. His team deployed a multi-agent system with 11 specialized agents that spawned 37,000 sub-agents to process nearly 56,000 clinical trials. A preprint describing the work was posted on February 23, 2026.
The system mirrors the structure of a human pharmaceutical company. A Chief Scientific Officer agent receives queries and delegates to divisions covering target discovery, molecule design, and clinical trials. Individual agents specialize further within each division.
‘Working with the CSO agent are different divisions that mirror the divisions found in a human biotech or pharma company,’ Zou said at VB Transform 2026.
The headline finding involves an antibody-drug conjugate (ADC) targeting the CD276 protein for lung cancer. The agents completed this design autonomously, drawing only on data published before January 2025. Months later, Merck independently developed and validated the same therapeutic approach. That drug, ifinatamab deruxtecan, received FDA Breakthrough Therapy Designation on August 18, 2025, for treating extensive-stage small cell lung cancer.
Zou called it ‘a third-party external validation of the therapeutic design provided by the virtual biotech agents.’
The clinical trial analysis produced several concrete findings. Drugs targeting cell-type-specific genes were 40% more likely to advance from Phase I to Phase II, 48% more likely to reach market, and showed 32% lower adverse event rates.
The architecture relies on a mixture of AI models. Zou noted that Claude often serves as the backbone for coding and data analysis, while other models are fine-tuned for specialized domains. Stanford also built what Zou calls an ‘agent school,’ where agents undergo supervised fine-tuning to sharpen expertise before deployment.
‘In these scientific virtual labs, the agents actually get into debates and disagreements. They have to convince the other AI scientists of their ideas, and all of that elicits much more creative and robust reasoning compared to if you have a single model trying to do the problem by itself from scratch,’ Zou said.
This work builds on Zou’s earlier Virtual Lab, published in Nature in July 2025. That system used LLM agents led by an AI principal investigator to design 92 novel nanobody binders against SARS-CoV-2 variants. Two showed improved binding in experimental validation. Zou said the AI-designed nanobodies outperformed previous human-designed versions.
The commercial implications are drawing investor attention. A Stanford-affiliated AI startup called Human Intelligence is reportedly in talks to raise $100 million at a valuation near $1 billion. The company is developing what it describes as a physiology foundation model.
Zou’s broader argument is that the next frontier in AI is not a single more capable model. It is thousands of specialized agents working in parallel, debating, and integrating outputs across an entire research pipeline.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.