A virtual biotech company with tens of thousands of AI agents predicted a lung cancer treatment target in a blinded analysis. Months later, pharmaceutical companies independently confirmed the finding and received FDA breakthrough therapy designation

A virtual biotech company with tens of thousands of AI agents predicted a lung cancer treatment target in a blinded analysis. Months later, pharmaceutical companies independently confirmed the finding and received FDA breakthrough therapy designation

Stanford’s Virtual Biotech system, led by an agent acting as chief scientific officer, analyzed lung cancer data and proposed a strategy targeting the protein B7-H3, with access limited to data available through January 2025. In August of that year, independently of the system, an antibody developed by Daiichi Sankyo and Merck against the same protein received breakthrough therapy designation, an FDA pathway that accelerates the development of promising drugs. On September 17, 2026, the peer-reviewed study was published in Science.

The system was tasked with assessing whether to develop a drug targeting B7-H3, a protein that suppresses the immune response in some tumors. The agents found no inherited genetic variants linking B7-H3 to lung cancer. The chief scientific officer, an agent coordinating the others, turned to another line of evidence: abnormal protein activity in tumor tissue. It directed the analysis toward data from individual cells in the tumor.

Those data showed that B7-H3 accumulates mainly in cancer-associated fibroblasts, connective tissue cells surrounding the tumor. Spatial analysis showed that immune cells were excluded from tissue in areas with high B7-H3 activity. A survival analysis of 566 patients found that the risk of death was 62% higher among patients with high B7-H3 activity than among those with low activity. The agents found no suitable binding site on the protein for a small-molecule drug, so they proposed an antibody that delivers a toxic drug directly to cells carrying B7-H3.

An antibody targeting the same protein, ifinatamab deruxtecan from Daiichi Sankyo and Merck, was already in clinical trials and received breakthrough therapy designation in August 2025: 48% of patients with small-cell lung cancer had a tumor response to treatment. Project lead James Zou commented on the convergence in an article in Singularity Hub:

“This is truly encouraging. It provides independent external confirmation of the effect and strategy proposed by the virtual biotech company.”

Virtual Biotech grew out of a smaller system, Virtual Lab, in which several agents simulated Zou’s laboratory at Stanford. The nanobodies it designed against COVID bound to the virus better than those designed by humans. That success persuaded the team to scale up to a “company” with tens of thousands of agents. They read papers and databases directly through Paperclip, a shared file system built by the team, Zou explained at VB Transform 2026.

The second case concerned MOONGLOW, a trial of Genentech’s antibody vixarelimab for ulcerative colitis that was discontinued after 79 patients. Its target, OSMRβ, a receptor in the gut involved in inflammation, had strong genetic support, but overall gene expression in intestinal tissue did not differ between patients and healthy individuals. Comparing five independent trials, the agents found that intestinal expression of the gene was elevated before treatment in patients who did not respond. Single-cell data pointed to fibroblasts, where its expression increases during treatment and inflammation begins. MOONGLOW did not select patients by OSMR levels, so the agents proposed doing so in a new trial.

In both cases, the agents found the explanation in the same cell type: fibroblasts, supporting cells that usually receive less attention. In February, the same system showed that such targets are 48% more likely to reach the market than targets with widespread activity. The agents derived their hypothesis about the lung cancer mechanism in a blinded analysis. It was independently confirmed by the decision of pharmaceutical companies that had already developed a drug against the same protein, and the work has now undergone peer review in Science.