Stanford AI Team Designs 92 COVID Treatments in Days

What if a team of AI scientists could solve problems in days that take human researchers months to crack?

Stanford researchers just turned that vision into reality. Led by professor James Zou and Ph.D. candidate Kyle Swanson, the team built Virtual Lab, a system where multiple AI agents collaborate like a real research team to tackle scientific challenges.

The system proved its worth fast. Given the task of designing molecules to fight evolving COVID-19 variants, the AI team completed its design process in days, with core planning compressed into just one to two hours of discussion.

The results surprised even the researchers. Virtual Lab generated 92 candidate nanobodies targeting COVID variants, and several showed real promise in experimental testing.

Two molecules stood out. They bound effectively to both newer variants and the original virus, a significant achievement in the race to keep pace with evolving threats.

The AI team didn't just follow instructions. It made strategic decisions, choosing to design nanobodies instead of conventional antibodies because they're smaller and easier to optimize computationally.

Each AI agent takes on a specialized role, mimicking how real research teams work. One acts as the principal investigator organizing discussions, while others serve as biologists, chemists, and machine learning experts.

"When we've worked on problems in drug discovery, it is a very interdisciplinary process," Swanson explained. The different perspectives help the system generate better solutions.

John Pak from Biohub, who tested the AI's designs, was stunned. "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," he said.

The system has clear limits. The AI agents don't understand real-world lab constraints or equipment availability, and they sometimes suggest impractical experiments or agree too readily with each other.

The Ripple Effect

Zou expanded the concept into Virtual Biotech, using thousands of AI agents to simulate an entire drug discovery organization. In one test, the system independently proposed an antibody-drug conjugate for lung cancer that Merck later discovered through traditional research, suggesting the AI can arrive at genuinely valuable insights.

The breakthrough addresses a fundamental problem in academic research: limited time and resources. "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?" Zou asked.

Researchers are now exploring automated robotic labs that could run experiments and feed results back to the AI, closing the loop between computational design and real-world testing.

This technology could accelerate discoveries across medicine, giving human researchers a tireless partner in the race to develop life-saving treatments.