How Stanford’s Virtual Biotech Used AI Agents to Analyze 55,984 Clinical Trials

Virtual Biotech is best understood as a large-scale computational research system, not a self-running pharmaceutical laboratory. Developed by Stanford researchers including James Zou and Harrison Zhang, it organizes specialized AI agents under a virtual chief scientific officer, allowing many evidence-gathering and analysis tasks to run in parallel. 12

Its most notable demonstration was the curation and analysis of 55,984 clinical trials. The work produced potentially useful patterns for choosing drug targets, alongside a lung-cancer treatment hypothesis. But none of those outputs establishes that a new therapy is safe or effective: that still requires laboratory research, clinical trials, and regulatory review. 12

How thousands of agents processed clinical-trial evidence

Rather than relying on one general-purpose model to summarize a vast literature, Virtual Biotech divided work among agents with different scientific roles. A chief scientific officer agent delegated tasks to specialized agents spanning areas such as target biology, molecular design, and clinical-trial analysis, then integrated their findings. 12

For the large trial-analysis task, the system deployed more than 37,000 clinical-trialist agents in parallel. They structured published safety and efficacy outcomes from 55,984 trials and linked drug targets to genomic, molecular, and single-cell annotations. Stanford reported that roughly 50,000 trials were catalogued in less than a week. 2

The speed came from parallelization: many agents could retrieve, annotate, and check discrete trial records at once. It was an evidence-curation workflow at extraordinary scale, rather than evidence that the system had independently run experiments or discovered a clinically validated medicine. 12

The target features linked with better historical outcomes

The analysis highlighted two features of drug targets:

• Cell-type specificity: whether a gene or protein is concentrated in the cell type most relevant to a disease rather than broadly active across many cells.
• Bimodality: whether gene activity behaves more like a discrete on/off state than a gradual continuum.

The researchers reported that targets with strong cell-type specificity and bimodal behavior were associated with better development outcomes across the historical trial data. In their analysis, drugs aimed at cell-type-specific genes were 40% more likely to progress from Phase I to Phase II, 48% more likely to reach market, and associated with 32% fewer adverse events. 2

These figures describe associations in a curated dataset, not guaranteed effects for every target or treatment. A plausible interpretation is that a target confined to disease-relevant cells may be easier to affect selectively, potentially reducing unintended effects elsewhere in the body. That biological explanation still needs experimental testing. 12

The B7-H3/CD276 lung-cancer hypothesis

Virtual Biotech also examined B7-H3, also called CD276, in lung cancer. Drawing on public information available before January 2025, the system identified a rationale involving B7-H3 expression in tumour-adjacent fibroblasts and proposed that these cells could contribute to local immune suppression. 2

It proposed an antibody-drug conjugate (ADC) strategy: an antibody directed at B7-H3, coupled to a cytotoxic payload. In principle, an ADC seeks to bring a cancer-killing compound preferentially to cells bearing the target. The output was a preclinical design and mechanistic hypothesis—not a manufactured product or a demonstrated treatment. 12

B7-H3 is also an active industry target. GSK has described risvutatug rezetecan (GSK5764227) as an investigational B7-H3-targeting ADC for small-cell lung cancer, and its materials report an FDA Breakthrough Therapy Designation for the asset in relapsed or refractory extensive-stage small-cell lung cancer.

However, the supplied record does not support presenting that program as a later clinical validation of Virtual Biotech’s proposal. GSK’s own cited announcement is dated August 2024, before the January 2025 public-information cutoff used in Stanford’s analysis. Shared interest in B7-H3 therefore shows strategic convergence around a target, not proof that the AI system originated, replicated, or clinically validated GSK’s drug program. 2

What Virtual Biotech has not shown

The key limitation is the distance between computational reasoning and medical proof. The AI agents did not:

• run cell, animal, or human experiments;
• establish an ADC’s selectivity, dosing, pharmacokinetics, toxicity, or resistance profile;
• manufacture a therapy; or
• demonstrate patient benefit in a clinical trial. 12

Other scientists cautioned that the platform had not yet faced the practical drug-discovery test of experimental validation. Historical correlations can be useful for prioritizing work, but they can also reflect data quality issues, confounding variables, or patterns that do not hold in a prospective program. 1

The next test is experimental validation

Virtual Biotech may be most useful as a way to narrow large scientific search spaces and make evidence review more systematic. Human scientists still need to define the question, inspect the underlying evidence, assess biological plausibility, and decide which hypotheses are worth pursuing. 12

For the B7-H3 proposal, meaningful validation would require independent replication followed by laboratory studies of target biology, ADC binding and selectivity, efficacy, toxicity, pharmacokinetics, and resistance. Only appropriately designed clinical trials can determine whether a treatment improves outcomes for patients. 12

The broader result is promising but measured: multi-agent AI can rapidly organize biomedical evidence and generate testable hypotheses. It has not removed the need for careful experimental science—or for human judgment at every consequential step.