Researchers create virtual scientists to solve biological problems

Stanford Medicine researchers have built what sounds like science fiction: a team of virtual scientists, powered entirely by artificial intelligence, working side by side with human researchers. These aren't your average chatbots spitting out quick answers. They're modeled after an actual Stanford lab, complete with an AI principal investigator, virtual colleagues, and even a critic whose job is to poke holes in the team's ideas.

As James Zou, PhD, associate professor of biomedical data science and the project's lead, put it: "Good science happens when we have deep, interdisciplinary collaborations... often that's one of the main bottlenecks and challenging parts of research." The virtual lab aims to shortcut those bottlenecks, delivering days' worth of brainstorming before a human has even finished their first cup of coffee.

How does it work?

The virtual lab kicks off just like any human lab — with a problem. A researcher presents the AI principal investigator (PI) with a scientific challenge, and the AI PI assembles a dream team of agents to tackle it. These agents aren't generic assistants; they're specialists in immunology, computational biology, machine learning, and more. One always plays devil's advocate, acting as critic-in-chief.

The AI scientists are equipped with tools like AlphaFold for protein modeling, and they're not shy about asking for more. "They would ask for access to certain tools, and we'd build it into the model to let them use it," said Zou. Meetings happen continuously — sometimes hundreds in the span of a morning — with back-and-forth exchanges, critiques, and brainstorming all logged for human oversight. Unlike human collaborators, they don't need snacks or breaks, and they're perfectly happy running meetings in parallel.

Why does it matter?

Because science is slow. Breakthroughs often take years of trial and error. But when Zou's team tasked their virtual lab with developing a better vaccine for SARS-CoV-2, the virus behind COVID-19, the AI team delivered a promising candidate in just a few days.

Instead of antibodies — the classic route — the AI scientists suggested nanobodies, smaller and simpler molecules that are easier to model and design computationally. Pak's real-world lab took those designs, tested them, and found they were not only stable and feasible but also bound tightly to the virus, including recent variants. Even more compelling, the nanobodies clung to the original Wuhan strain too, hinting at the potential for a broadly protective vaccine.

As Zou explained, "By the time I've had my morning coffee, they've already had hundreds of research discussions." In other words, the speed and scale of these AI-powered collaborations could dramatically accelerate how we respond to global health crises — or any scientific challenge that demands quick, creative solutions.

The context

The project sits at the crossroads of two big shifts in science: the explosion of biological data and the rapid evolution of AI. Biology and medicine generate mountains of complex datasets, yet human researchers are only scratching the surface in analyzing them. Zou believes AI agents can help re-examine published papers and uncover insights we've missed.

"Often the AI agents are able to come up with new findings beyond what the previous human researchers published on. I think that's really exciting," he said. And it's not just about COVID. The Stanford team is already eyeing other thorny biological problems, training AI collaborators to act as sophisticated analysts and idea generators.

The study, published in Nature and backed by Stanford's Human-Centered AI Institute, shows a glimpse of what science could look like when humans and machines co-create knowledge. If traditional labs are like sailboats navigating a vast ocean, these AI-powered virtual labs are motorboats — fast, relentless, and able to go further on the same tide.