A Stanford-led research team has built the Virtual Lab, a system in which multiple AI agents work together as a team of scientists.
The Virtual Lab designed 92 candidate nanobodies targeting evolving COVID-19 variants. Several showed promising results in experimental testing, including two that bound effectively to both the newer variants and the original virus.
The project was led by James Zou, professor of biomedical data science, and Kyle Swanson, a fifth-year Ph.D. candidate in computer science. They also collaborated with John Pak, a senior platform leader at Biohub, a nonprofit focused on AI research in biology with funding from Mark Zuckerberg and Priscilla Chan.
Virtual Lab’s system completed its design process in just a few days — compared with several weeks or months for a human researcher — with the core planning compressed into only one to two hours of agent discussion.
Zou said the project stemmed from the “limited time” available to academic researchers.
“Wouldn’t it be great if we actually have a team of AI agents that can emulate my physical lab so that they can tackle some of these problems in a more autonomous way?” he said.
To evaluate the system, the team gave it a concrete biological task: design molecules that could bind to evolving COVID-19 variants.
“Initially, we asked very broad questions,” Swanson said. The agents debated these queries and arrived at a surprising insight: they diverged from the conventional approach of designing antibodies and opted instead for nanobodies, which are smaller and easier to computationally design and optimize, according to Zou. The agents selected computational tools and wrote code to integrate them into a complete nanobody design pipeline.
“They did most of the discussion, most of the scientific design, most of the code writing,” Swanson said.
The result was 92 candidate nanobodies generated entirely in silico. These were then handed off to Biohub for experimental testing, where several showed promising results.
“It was very shocking. No one had really thought that a team of LLM agents could provide any sort of actionable wet lab protocols or suggestions,” Pak said. “The plans were very well thought out. They synthesized a number of different AI and ML tools to create a protein engineering pipeline.”
Despite the system’s capabilities, it lacks awareness of real-world constraints. “[The agents] don’t really know what we’re capable of, what sort of equipment we have, what interests us,” Pak said. “So there is a lot of interpretation on the hands of the wet lab scientists to pick out what they find interesting amongst the agent’s recommendations.”
Swanson pointed to additional limitations. The agents can miss context, suggest impractical experiments or fail to challenge each other. “They were too agreeable with each other,” he said.
Critically, while AI can propose designs, it cannot yet carry out the wet-lab experiments required to test them. Researchers are exploring ways to create automated robotic labs that can run experiments and feed results back into the system.
The multi-agent system is structured to mimic a real research group. “I think the big difference is the Virtual Lab is not just a tool, but it’s really a team of scientists,” Zou said. One agent plays the role of a principal investigator (PI) and organizes discussions, while others take on more specialized, domain-specific roles like biologists, chemists and machine learning researchers. These agents debate, refine hypotheses and propose methods through structured meetings.
Swanson said this design was intentional. “When we’ve worked on problems in drug discovery, it is a very interdisciplinary process. We have biologists, chemists, computer scientists — we always made the most progress by having all those different perspectives in the room,” Swanson said. “You’re getting more and better information by forcing [the system] to think from different perspectives.”
After building the Virtual Lab, Zou scaled the idea into the Virtual Biotech, a larger system designed to simulate an entire drug discovery organization. Instead of a small handful of agents, this system uses thousands.
“It has a Chief Scientist Officer agent which then organizes teams of all these different agents that would do everything from looking for targets, for drugs, all the way to designing molecules, designing clinical studies and so on,” Zou said.
In one case, the system independently proposed a design for an antibody-drug conjugate for the lung cancer target B7-H3. Months later, Merck, a large pharmaceutical company, independently arrived at the same discovery, suggesting the system can converge on ideas aligned with real-world research directions.