The topic at a glance
- A Science paper from Stanford describes a Virtual Biotech: specialist AI agents that retrieve, source-track, and synthesize public biomedical evidence for early drug-development decisions.
- The useful result is not agent headcount. It is that the agents could query structured evidence and leave a trail a human can inspect.
- Most internal R&D still lacks that record. OVAITY is building a scientific memory layer so project context can be used by researchers and, later, by AI agents.
Last week, Science published a paper that was easy to misread.
Harrison Zhang, James Zou and colleagues at Stanford describe the Virtual Biotech: a multi-agent system modeled on a drug-development company. Coverage led with scale. More than 37,000 AI agents annotated clinical-trial outcomes. None of them were human. [1,2]
The number is real. It is also a poor way to understand the work.
The standing organization was small. A virtual chief scientific officer coordinated specialist scientist agents across target identification, safety, modality selection and clinical development. Those agents queried public biomedical resources through Model Context Protocol tools, produced code and reports, and passed claims to a reviewer agent before a final synthesis. The 37,000 figure comes from one job: sending a dedicated clinical-trialist agent to each Phase II and Phase III record that needed outcome curation. [1,3]
What made the system useful was not headcount. It was access to evidence that agents could query, source-track and leave in a form a scientist could inspect.
What the agents actually produced
The authors tested the Virtual Biotech at three decision points.
First, they used it to enrich 55,984 trials from Open Targets with structured outcome annotations. Each trialist agent followed a documented cascade: ClinicalTrials.gov, then PubMed, then press releases and regulatory notices when fields were still missing. Human review of 100 trials found 89.7% agreement on primary endpoints. Against that curated set, drugs acting on cell-type-specific genes were 40% more likely to progress from Phase I to Phase II, 48% more likely to reach Phase IV, and associated with 32% lower adverse-event rates. The associations remained after adjustment for genetic evidence. [1,3]
Those analyses are observational. The authors say so. Trial indications, endpoints and reporting quality can still confound the result, and the work is limited to registered trials with mapped molecular targets. [1,4]
Second, they asked the system to evaluate B7-H3 in lung cancer, with web search disabled and a January 2025 model cutoff. The agents integrated genetics, single-cell and spatial transcriptomics, survival analysis and modality assessment, then argued for an antibody–drug conjugate. The run cost $46 in API credits. In August 2025, the FDA granted Breakthrough Therapy Designation to ifinatamab deruxtecan, a B7-H3-directed ADC, based on a Phase 2 trial in extensive-stage small-cell lung cancer. [1,3,5]
That later clinical progress does not mean the agents discovered the target. B7-H3 was already in the literature. What the authors claim is narrower: the system assembled a coherent, inspectable case from primary data without seeing the later readout.
Third, they pointed the system at a terminated Phase II ulcerative colitis trial of vixarelimab, an antibody that blocks OSMRβ. The agents inferred a precision-medicine gap — an unselected population, no OSMR enrichment — and proposed biomarker-guided enrollment. That is a hypothesis. It still has to be tested. [1,3]
The authors are explicit about the limits. Conclusions depend on the quality and breadth of available data. Poorly studied diseases and targets are a weak fit. Experimental validation remains necessary. Zou’s next step, as he put it, is to take findings into real laboratories and see how many hold up. [2,4]
The interesting behavior is the handoff
This is no longer a demonstration that an agent can draft a methods paragraph or write analysis code.
In the B7-H3 case, a reviewer agent flagged that ligand–receptor inferences from dissociated single-cell data had lost tissue architecture. The orchestrator then requested spatial analysis. That is the kind of scientific behavior organizations hope for from cross-functional teams, and often fail to get because the relevant dataset sits in another group’s folder. [3]
The same week, Novo Nordisk said it would test Anthropic’s Claude Science on specific R&D workflows, with data governance and human oversight written into the collaboration. The paper and the partnership are different things. One is a research system over public resources. The other is a company putting an AI workbench into live discovery. Both point to the same operational question. [6]
Agents will participate in target decisions. The open issue is what they will be allowed to see, and whether anyone can reconstruct the path from evidence to recommendation.
Public databases are not a lab’s memory
The Virtual Biotech could do this work because the evidence already existed in queryable form.
Open Targets, CELLxGENE, ClinicalTrials.gov, Tabula Sapiens, Tahoe-100M and cBioPortal are public scientific infrastructure. MCP gave the agents a uniform way to call them. Source-tracking made the annotations auditable. The interface exposed reasoning, tool use, code and reports. [3,7]
Most internal R&D does not look like that.
The sequencing run from last quarter may be on a share drive. The reason a sample was excluded may sit in a meeting note. The analysis that failed, and should not be repeated, may exist only as a local notebook. The decision to drop a target may live in a slide deck that never names the dataset version behind it.
An agent with access to public cell atlases can tell you what Tabula Sapiens says about a gene. It cannot tell you that your team already ran the experiment, got a contradictory result, and moved on — unless that history is stored with its context intact.
This is the same fragmentation problem we described in What is bioinformatics?: results survive, while the path that produced them does not. It is also why AI workbenches still need a memory layer. A workbench helps a scientist execute a task. A virtual biotech helps a swarm of agents execute many tasks. Neither replaces a durable record of the project.
Give agents only public data, and they will reason over public data. That can still be valuable. It will also miss the evidence that never left the organization: negative results, protocol deviations, the rationale for a failed series, the version of the assay that was retired.
Scale makes the gap more expensive. One agent working from an incomplete folder produces a weak answer. Thousands of agents working from the same incomplete folder can produce a confident case for the wrong conclusion.
What has to exist before agent teams are usable
If agentic R&D is going to be more than a literature-and-database overlay, the scientific environment underneath it has to provide a few ordinary capabilities.
- Data that people and machines can find, including failed experiments and the analyses that never became a figure.
- Provenance that travels with the result: dataset version, workflow, parameters, reviewer, and the later decision that depended on it.
- A trail complete enough that a scientist can inspect the agent’s inputs, code and rejected alternatives.
- Scoped permissions. Public MCP servers are not a model for an ELN, a LIMS, or identifiable clinical data.
The FAIR principles were written for this problem long before MCP existed. Machine-actionability is now a practical requirement. When a Virtual Biotech trialist agent labelled a trial outcome, it recorded the source. An internal equivalent would have to do the same for the organization’s own work. [3,8]
This is also one of the challenges we are exploring at OVAITY. We are building a Scientific Intelligence Platform that connects projects, experiments, datasets, analyses and decisions in one governed environment, so scientific context remains usable over time. Context-aware AI over that project record is being validated with researchers. Broader research agents and autonomous agent workflows remain prototypes. We are not building a 37,000-agent virtual company. We are trying to make the underlying scientific record coherent enough that humans, and later agents, can work from the same evidence.
A better question than headcount
The Virtual Biotech is a strong existence proof. Specialist agents, given structured public evidence and an inspectable workflow, can assemble target cases that look like the work of a cross-functional team.
It is not a blueprint for replacing a research organization.
The constraint that remains is local. Can your agents see the experiment that failed? Can a scientist see why the agent trusted one dataset over another? Can next year’s hire reconstruct the decision without finding the person who remembers it?
Until those questions have answers, adding agents will mainly accelerate access to the same public record everyone else can already query.
The organizations that get value from this moment will treat agent teams as a reason to fix the scientific record, not as a way to step around it.
If your team is exploring AI-native research workflows but struggling with context, traceability, or project memory, we’d love to learn from you.
References