Virtual Biotech, a research platform driven by artificial intelligence developed in Stanford’s Department of Biomedical Data Science, could potentially streamline the discoveries that pharmaceutical companies can achieve from months to days. 

The platform uses a team of AI agents to emulate a human drug discovery organization, said James Zou, associate professor of biomedical data science. “We have tens of thousands of AI agents that will look for good drug targets, develop molecules and also assess the safety and efficacy of those targets,” Zou said.

Zou got the idea from a virtual lab that launched in 2025, where he collaborated with John Pak, a scientist at Chan Zuckerberg Biohub, and Kyle Swanson, a final-year PhD candidate, as lead author. 

The virtual lab was capable of designing new proteins that had been experimentally synthesized and validated. “We thought, can we scale this up?” Zou said.

Zou and lead author Harrison Zhang, an MD-PhD candidate, published a paper outlining their findings on Sept. 17 in Science, putting Virtual Biotech on the map in one of the biggest research journals.

Virtual Biotech is not a company, but rather “a research project and an idea that we have built from the ground up here at Stanford,” Zhang said. The agents were trained in the organizational hierarchies of biotech companies and perform specialized research tasks across the drug-development pipeline. The team organized the specialized agents into a framework resembling a company, with a virtual chief scientific officer (CSO) coordinating agents across target discovery, safety, modality selection and clinical development. 

Zou and Zhang tested the system on over 50,000 clinical trials. The analysis, which would have taken human researchers years, took less than a week. 

“Instead of trying to emulate a single human lab, it emulates a much more complex human R&D [Research & Development] organization,” Zou said.

Not only were the agents tasked with analyzing trial results, but also with data collected during the trial. In late 2024, the agents devised a therapeutic strategy to target B7-H3, a protein that frequently overexpresses in various types of cancer.

The AI scientists created an antibody-drug conjugate (ADC), a protein-based class of highly targeted cancer medicine that targets cells harboring B7-H3 proteins and administers treatment. “It’s sort of like a missile. Once it gets to the target, it will release a payload, which is a chemotherapy or toxic agent, which will then be released to kill the malignant cells,” Zou said.

He stressed the ADC’s precision-guided attack, directly linked to one of Virtual Biotech’s major findings: drugs targeting cell-type-specific genes were 48% more likely to reach the market and had 32% fewer adverse events.

Virtual Biotech’s model raises questions about its accuracy, Zou acknowledged, but the group is careful to verify the agents’ results. “We picked here several of the key findings from the virtual biotech agents,” he said. For each of the selected findings, a PhD-level human expert successfully reproduced the agents’ results to measure their precision.

In August 2025, a private pharmaceutical company independently arrived at the same B7-H3-targeting ADC strategy, he told The Daily. 

Zou acknowledged that Virtual Biotech’s recent launch raises another immediate question: as AI agents take on increasingly sophisticated research tasks, how will demand for the humans who perform those tasks change? 

“We view the Virtual Biotech as a way to complement and to accelerate human researchers rather than to replace human researchers entirely,” he said. Human experts were still responsible for guiding the agents to prevalent disease areas and compelling disease targets for clinical application, he added. 

“[Think] of [the agents] as a collaborative teammate — a co-scientist,” Zhang said. AI scientists would allow human experts to delve into intellectually stimulating matters, leaving more scientifically mundane tasks to the agents.

Following NIH budget disruptions, implementing AI in research could potentially “alleviate some pressure with regards to funding,” Zhang said.

“The cost will be less than $1 per agent,” Zou predicted. Virtual Biotech’s entire case study took “less than a week,” since the agents were able to work in parallel. 

While speed is a priority among many, “it’s really important to make sure these systems that we build are safe,” Zou said. He noted technological guardrails, including the presence of a reviewer agent.

Virtual Biotech’s meticulously constructed model is open source: anyone has access to this powerful biomedical engineering tool. The platform intends to support researchers across the broader scientific community.

The team isn’t pursuing the B7-H3 target any further, acknowledging that private pharmaceuticals have already moved into commercialization territory. “We are academic researchers,” Zhang said. “It’s important for us that we made it open source, because we wanted this framework and also its technical implementation to be broadly available to the research community.” 

For Zhang and his team, the next step is expanding the scope of what AI can do in science, including “how we can incorporate AI scientists into wet lab workflows.”