AI news ·
Stanford team builds AI agent system that designs COVID-19 nanobodies in days
Stanford researchers built an AI system that designed 92 COVID-19 drug candidates in days, work that typically takes human researchers months. Several nanobodies showed promising results in lab tests.
Stanford Team Builds AI System That Designs Drug Candidates in Days, Not Months
Stanford researchers have created the Virtual Lab, a system where multiple AI agents collaborate as a research team to accelerate drug discovery. The system designed 92 candidate nanobodies targeting COVID-19 variants in a few days - work that would take human researchers weeks or months.
James Zou, a professor of biomedical data science, and Kyle Swanson, a Ph.D. candidate in computer science, led the project. They worked with John Pak from Biohub, a nonprofit focused on AI research in biology.
How the System Works
The Virtual Lab assigns AI agents different roles - principal investigator, biologist, chemist, machine learning researcher - and has them debate and refine hypotheses through structured meetings. One agent organizes discussions while others contribute specialized perspectives.
When given the task of designing molecules for COVID-19 variants, the agents debated initial questions and diverged from conventional antibody design. They chose nanobodies instead, which are smaller and easier to computationally optimize. The agents then selected tools, wrote code, and built a complete design pipeline.
"They did most of the discussion, most of the scientific design, most of the code writing," Swanson said.
The 92 candidate nanobodies were generated entirely in silico and sent to Biohub for testing. Several showed promising results, including two that bound effectively to both newer variants and the original virus.
Real-World Testing Reveals Gaps
The agents completed their core planning in one to two hours. Yet the system has clear limitations.
The agents lack awareness of real-world constraints. They don't know what equipment labs have, what interests researchers, or what experiments are practical. Pak noted that wet-lab scientists had to interpret and filter the agents' recommendations.
Swanson identified additional weaknesses: agents can miss context, suggest impractical experiments, or fail to challenge each other's assumptions. "They were too agreeable with each other," he said.
Most critically, AI agents powered by large language models cannot run wet-lab experiments themselves. Researchers are exploring robotic labs that could execute experiments and feed results back into the system.
Scaling to Full Drug Discovery
Zou scaled the concept into Virtual Biotech, a larger system simulating an entire drug discovery organization. Instead of a handful of agents, it uses thousands organized under a Chief Scientist Officer agent.
The system proposed an antibody-drug conjugate design for the lung cancer target B7-H3. Months later, Merck independently arrived at the same discovery, suggesting the system converges on ideas aligned with real research directions.
The work shows that AI agents can contribute to early-stage research planning and molecule design. The bottleneck remains moving from computational proposals to validated experimental results.