Modeled after a well-established Stanford School of Medicine research group, the virtual lab is complete with an AI principal investigator and seasoned scientists.
“Good science happens when we have deep, interdisciplinary collaborations where people from different backgrounds work together, and often that’s one of the main bottlenecks and challenging parts of research,” said James Zou, PhD, associate professor of biomedical data science who led a study detailing the development of the virtual lab. “In parallel, we’ve seen this tremendous advance in AI agents, which, in a nutshell, are AI systems based on language models that are able to take more proactive actions.”
People often think of large language models, the type of AI harnessed in this study, as simple question-and-answer bots. “But these are systems that can retrieve data, use different tools, and communicate with each other and with us through human language,” Zou said. (The collaboration shown through these AI models is an example of agentic or agential AI, a structure of AI systems that work together to solve complex problems.)
The leap in capability gave Zou the idea to start training these models to mimic top-tier scientists in the same way that they think critically about a problem, research certain questions, pose different solutions based on a given area of expertise and bounce ideas off one another to develop a hypothesis worth testing. “There’s no shortage of challenges for the world’s scientists to solve,” said Zou. “The virtual lab could help expedite the development of solutions for a variety of problems.”
Already, Zou’s team has been able to demonstrate the AI lab’s potential after tasking the “team” to devise a better way to create a vaccine for SARS-CoV-2, the virus that causes COVID-19. And it took the AI lab only a few days.
A paper describing the findings of the study will be published July 29 in Nature. Zou and John Pak, PhD, a scientist at Chan Zuckerberg Biohub, are the senior authors of the paper. Kyle Swanson, a computer science graduate student at Stanford University, is the lead author.
Running a virtual lab
The virtual lab begins a research project just like any other human lab — with a problem to solve, presented by the lab’s leader. The human researcher gives the AI principal investigator, or AI PI, a scientific challenge, and the AI PI takes it from there.
“It’s the AI PI’s job to figure out the other agents and expertise needed to tackle the project,” Zou said. For the SARS-CoV-2 project, for instance, the PI agent created an immunology agent, a computation biology agent and a machine learning agent. And, in every project, no matter the topic, there’s one agent that assumes the role of critic. Its job is to poke holes, caution against common pitfalls and provide constructive criticism to other agents.
Zou and his team equipped the virtual scientists with tools and software systems, such as the protein modeling AI system AlphaFold, to better stimulate creative “thinking” skills. The agents even created their own wish list. “They would ask for access to certain tools, and we’d build it into the model to let them use it,” Zou said.
As research labs go, the virtual team runs a swift operation. Just like Zou’s research group, the virtual lab has regular meetings during which agents generate ideas and engage in a conversational back-and-forth. They also have one-on-one meetings, allowing lab members to meet with the PI agent individually to discuss ideas.
But unlike human meetings, these virtual gatherings take a few seconds or minutes. On top of that, AI scientists don’t get tired, and they don’t need snacks or bathroom breaks, so multiple meetings run in parallel.
“By the time I’ve had my morning coffee, they’ve already had hundreds of research discussions,” Zou said during the RAISE Health Symposium, during which he presented on this work.
Moreover, the virtual lab is an independent operation. Aside from the initial prompt, the main guideline consistently given to the AI lab members is budget-related, barring any extravagant or outlandish ideas that aren’t feasible to validate in the physical lab. Not one prone to micromanagement — in the real or virtual world — Zou estimates that he or his lab members intervene about 1% of the time.
“I don’t want to tell the AI scientists exactly how they should do their work. That really limits their creativity,” Zou said. “I want them to come up with new solutions and ideas that are beyond what I would think about.”
But that doesn’t mean they’re not keeping a close eye on what’s going on — each meeting, exchange and interaction in the virtual lab is captured via a transcript, allowing human researchers to track progress and redirect the project if needed.
SARS-CoV-2 and beyond
Zou’s team put the virtual lab to the test by asking it to devise a new basis for a vaccine against recent COVID-19 variants. Instead of opting for the tried-and-true antibody (a molecule that recognizes and attaches to a foreign substance in the body), the AI team opted for a more unorthodox approach: nanobodies, a fragment of an antibody that’s smaller and simpler.
“From the beginning of their meetings the AI scientists decided that nanobodies would be a more promising strategy than antibodies — and they provided explanations. They said nanobodies are typically much smaller than antibodies, so that makes the machine learning scientist’s job much easier, because when you computationally model proteins, working with smaller molecules means you can have more confidence in modeling and designing them,” Zou said.
So far, it seems like the AI team is onto something. Pak’s team took the nanobody structural designs from the AI researchers and created them in his real-world lab. Not only did they find that the nanobody was experimentally feasible and stable, they also tested its ability to bind to one of the new SARS-CoV-2 variants — a key factor in determining the effectiveness of a new vaccine — and saw that it clung tightly to the virus, more so than existing antibodies designed in the lab.
They also measured off-target effects, or whether the nanobody errantly binds to something other than the targeted virus, and found it didn’t stray from the COVID-19 spike protein. “The other thing that’s promising about these nanobodies is that, in addition to binding well to the recent COVID strain, they’re also good at binding to the original strain from Wuhan from five years ago,” Zou said, referring to the nanobody’s potential to ground a broadly effective vaccine. Now, Zou and his team are analyzing the nanobody’s ability to help create a new vaccine. And as they do, they’re feeding the experimental data back to the AI lab to further hone the molecular designs.
The research team is eager to apply the virtual lab to other scientific questions, and they’ve recently developed agents that act as sophisticated data analysts that can reassess previously published papers.
“The datasets that we collect in biology and medicine are very complex, and we’re just scratching the surface when we analyze those data,” Zou said. “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.”
This study was supported by the Knight-Hennessy Scholarship and the Stanford Bio-X Fellowship.
Stanford’s Human Centered AI Institute and Department of Biomedical Data Science also supported the work.
# # #
About Stanford Medicine
Stanford Medicine is an integrated academic health system comprising the Stanford School of Medicine and adult and pediatric health care delivery systems. Together, they harness the full potential of biomedicine through collaborative research, education and clinical care for patients. For more information, please visit med.stanford.edu.
END