Inside Stanford's Virtual Biotech: 37,000 AI Agents and the Three Discovery Tasks They Performed

Stanford Medicine researchers have built a "virtual biotech" — an AI company with 37,000 agents and no human employees — and shown it can perform three concrete early-stage drug discovery tasks: prioritizing drug targets, validating a lung cancer target called B7-H3, and diagnosing why a failed ulcerative colitis trial went wrong. The platform, created by biomedical data science associate professor James Zou and graduate student Harrison Zhang, mirrors the organizational chart of a real biotech firm, with a chief science officer agent delegating work to specialized AI teams. Here is how it works, what it actually did, and where its limits lie.

Introduction to the Virtual Biotech Platform

What does it mean for a biotech company to have no laboratory space, no lunch breaks and no payroll? According to Stanford Medicine's announcement of the project, that is precisely the setup: tens of thousands of AI agents, each trained on some aspect of drug development, organized into a company-like structure. The project grew out of a virtual laboratory Zou launched in 2025, where AI scientists emulated researchers working across an academic lab. The Virtual Biotech extends that idea by giving the agents the structure of an established biotechnology firm rather than a single lab.

The stakes are real. As reporting on the project notes, around nine in every ten drug candidates entering clinical trials never reach regulatory approval, usually because of safety or efficacy problems. The Stanford team's ambition, as Zou framed it, was to ask: "Could we create a biotech company that takes on everything from looking for drug targets all the way to designing clinical trials?" and "Could we have a fully agentic company that tackles the extremely complex challenges of drug discovery?"

Organizational Structure: The Agentic Company

The platform's design borrows from how human biotechs divide labor. A chief science officer agent sits at the top, receiving scientific queries and delegating them to domain-specialized scientist agents, then integrating their outputs through data-driven reasoning. Those specialist agents draw on complementary tools and knowledge spanning statistical genetics, functional genomics, pathways and interactions, chemoinformatics, disease biology and clinical data.

One advantage of an AI company, Zou observed, is that you can skip the startup phase. The 37,000-agent figure describes agents spawned to perform tasks — including annotating clinical trial outcomes — rather than a permanent standing workforce. The structure matters because it lets specialized divisions work in parallel on different parts of the pipeline, such as identifying molecular targets and designing clinical trials, instead of running one monolithic model end to end.

How the Virtual Biotech Tackles Drug Discovery Challenges

The clearest way to understand the platform is to walk through its three demonstrated applications, each chosen to correspond to a common decision point in early drug development. In every case the pattern is the same: a scientific query enters at the top, the CSO agent breaks it into sub-tasks for specialized agents, and the outputs are synthesized into a strategic recommendation.

The first case study tested whether the platform could reduce the attrition rate of drugs entering trials by drawing together evidence from diverse sources. The second and third are unpacked below, because they show the system moving from ranking possibilities to generating therapeutic strategy.

Case Study: Validating a Cancer Drug Target (B7-H3)

In its second application, the Virtual Biotech evaluated B7-H3 as a lung cancer target. The scientist agents integrated statistical genetics, single-cell, spatial and clinicogenomic evidence — the kind of multi-scale picture a human team would assemble from several departments. The output was not just a yes-or-no verdict: the platform proposed an antibody–drug conjugate strategy while identifying key liabilities and differentiation opportunities, meaning it flagged both the risks and the ways a B7-H3 therapy could stand apart from existing approaches.

Stanford's announcement goes further, reporting that the platform independently designed a lung cancer therapy that was later validated in trials. That validation, however, happened outside the AI system — a reminder that the platform's proposals are computational outputs that still require experimental confirmation.

Case Study: Analyzing a Failed Ulcerative Colitis Trial

The third application is perhaps the most instructive for understanding what multi-agent AI adds. The platform analyzed a terminated ulcerative colitis trial that had targeted OSMR, and inferred potential mechanisms of failure. It then proposed biomarker-guided enrollment strategies to address precision-medicine gaps — in other words, identifying patient subgroups that might have benefited from the drug and suggesting how future trials could select for them.

This is the "clinical translation strategy after trial failure" decision point: the moment when a company must decide whether to abandon a target or reframe how it tests the drug. The Virtual Biotech's analysis provides insights for that decision, but the insights remain computational hypotheses until tested in new trials.

Limitations and Future Outlook of AI in Drug Discovery

Where does this sit in the overall pipeline? The demonstrated scope is early-stage, computational discovery: target prioritization, target validation and modality selection, and post-failure translation strategy. The platform's creators state that it keeps human scientists in the loop, and it does not replace physical lab work or human clinical trials — the expensive stages where most drug candidates still fail.

The team's stated ambitions point toward expansion rather than replacement. In their own words, they envision "expanding the Virtual Biotech to encompass even more of the drug discovery pipeline, such as molecule design, virtual screening, toxicity prediction, and clinical decision-making agents, bringing the benefits of interdisciplinary collaboration, speed, and objectivity to the entire end-to-end process." That is a future goal, not a current capability.

Accelerating Early Drug Discovery: A Practical Outlook

The Virtual Biotech platform offers a powerful computational tool for accelerating early-stage drug discovery by automating complex data integration and analysis tasks, allowing human scientists to focus on experimental validation and later-stage development. Its value is demonstrated through specific case studies — target prioritization, B7-H3 validation and the OSMR trial analysis — rather than proven as a general claim for all drug discovery, and every insight it produces still requires experimental validation in physical labs.

The measurable indicator to watch is continued publication of research demonstrating the platform's successful application to new drug targets or clinical trial analyses, and its integration into broader drug development pipelines. If the Stanford team delivers on its stated expansion into molecule design, virtual screening and toxicity prediction, the virtual biotech will have moved from an organizational curiosity to a genuine pipeline tool.

Sources

• Virtual biotech company puts thousands of AI scientist agents to work on drug discovery — News Center
• Virtual biotech deploys 37,000 AI agents across drug discovery — BIOENGINEER.ORG
• AI Agents Collaborate to Streamline Drug Discovery — Inside Precision Medicine
• EWSC Talk: The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development — Broad Institute