The Virtual Biotech: When 37,000 AI Agents Go Drug Hunting

Posted on 19 September, 2026 by Mehrdad Fathi

Drug discovery has always been a numbers game. The more hypotheses you can generate and test, the better your odds of finding something that works. What happens when you remove the practical ceiling on how many scientists can work a problem simultaneously?

A team at Stanford University, led by computer scientist James Zou, has spent the past year building an answer. They call it the Virtual Biotech — an artificial-intelligence system that replicates the organizational structure of a drug development company, staffed not by humans but by up to 37,000 AI agents, each capable of performing multi-step tasks by interacting with large language models or with one another. The work was published today in Science.

The Org Chart That Scales Infinitely

The Virtual Biotech is structured like a real biotechnology company, with one agent acting as chief scientific officer and directing a hierarchy of specialized departments — target identification, clinical-trial analysis, drug design, and others. For the study reported in Science, Zou’s team tasked the system with one of the most demanding problems in the field: extracting signal from the published record of human clinical trials.

The CSO dispatched 37,075 agents, each assigned to a single late-stage trial from a corpus of more than 55,000 published studies. Simultaneously, other agents cross-referenced those results against gene-expression datasets to look for biological features that could predict whether a drug would succeed or fail in humans.

The analysis surfaced a meaningful finding: drugs that target proteins expressed in specific, relevant cell types are nearly 50% more likely to reach market than those targeting broadly expressed proteins. This kind of insight — obvious in retrospect, hard to extract manually from tens of thousands of documents — is precisely the domain where large-scale AI coordination has an advantage.

A Lung Cancer Candidate

The more operationally relevant demonstration involved a protein called CD276, which had been flagged in prior research as a suppressor of immune responses that is highly expressed in lung tumours. With that starting point, the Virtual Biotech confirmed CD276 as a credible therapeutic target using existing datasets and proposed a specific targeting strategy: an antibody designed to recognize CD276, tethered to an anticancer payload.

This is a well-established drug format — antibody-drug conjugates have been one of oncology’s more productive structural frameworks in recent years — but the system arrived at it independently, through its own analysis of the literature and genomic data.

The result gained an unexpected external reference point months later. In August 2025, Merck announced that a drug working on a similar mechanism — an ADC targeting a comparable checkpoint — had been fast-tracked by the FDA. Merck developed it independently and through a different path, but the convergence of a 37,000-agent AI system and one of the world’s largest pharmaceutical companies on the same general strategy is at minimum a useful data point.

What the System Can and Cannot Do

The Virtual Biotech had access to more than 100 tools and external databases. Ola Spjuth at Uppsala University, whose own group has published work on agentic AI for drug repurposing, notes that the breadth of that toolset is what enabled the system to contribute meaningfully rather than merely summarize literature.

But the criticisms are pointed and worth taking seriously. Andreas Bender at Khalifa University draws a clear line between retrospective analysis — the Virtual Biotech’s forte — and the real decision-making environment of drug development. When a company is betting its existence on a single clinical trial, the tolerance for uncertainty is categorically different from an academic benchmarking study. The Virtual Biotech has not been asked to make that kind of call, and no one yet knows whether it could.

More concretely, nothing the system proposed has been validated in the laboratory. The CD276 strategy remains a computationally derived hypothesis. The next step Zou is planning — integrating the Virtual Biotech with autonomous, self-driving laboratories that could run cell-based assays — would be a significant upgrade, and would move the system from analysis toward something closer to the experimental loop that actually separates good ideas from drug candidates.

The Coordination Problem

What the Virtual Biotech represents, at a structural level, is a solution to a coordination problem that has always constrained drug discovery: the volume of existing scientific literature and clinical data vastly exceeds any human team’s capacity to synthesize it quickly. Individual agents in the system are not doing anything a skilled analyst could not do. Thirty-seven thousand of them, working in parallel, are doing something qualitatively different.

Whether that difference translates into better drugs reaching patients faster is still an open question. Zou plans to pursue partnerships with pharmaceutical companies to test the system in real development pipelines. That is where the proof will come — not in the retrospective analysis of trials already run, but in the prospective identification of candidates that would otherwise have been missed, and the ultimate fate of those candidates in the clinic.

The pharmaceutical industry has been promised AI-accelerated drug discovery for over a decade. The Virtual Biotech does not settle that promise. It does, however, represent one of the more structurally serious attempts to operationalize it.

Source: Nature (News & Views). doi: https://doi.org/10.1038/d41586-026-02954-y

References:

1. Zhang, H. G., Eckmann, P., Miao, J., Mahon, A. B. & Zou, J. A virtual biotech powered by large language model agents. Science https://doi.org/10.1126/science.aeg6779 (2026).

2. Huynh, D. L. et al. Agentic AI for drug repurposing. Preprint at bioRxiv https://doi.org/10.64898/2026.04.20.719538 (2026).