Have you ever wondered what a scientist looks like? Traditionally, we picture someone in a lab coat with a pair of goggles on. But nowadays, being a ‘scientist’ could just as easily mean a lineup of humming computer servers tucked away in an air-conditioned facility.
This is exactly what a talented group from the Chan Zuckerberg Biohub and Stanford University has done, creating a “Virtual Lab” filled with AI systems that embark on serious research — in this case, they’re tackling possible treatments for COVID-19.
James Zou, a professor at Stanford and one of the study’s leading minds, expressed his excitement: “Imagine having AI assistant researchers for each scientist. It’s versatile and I believe the Virtual Lab could speed up advancements in various scientific areas.”
The AI-driven probes held digital meetings, coded away, and produced biological models, ultimately suggesting molecules that could help fight recent COVID variants. Researchers later tested these suggestions, as detailed in their recent paper published in Nature. While there’s still a long journey ahead for these potential treatments, the scientists believe that their AI cohort can significantly speed up future discoveries.
Scientists usually depend on collaboration among experts, bouncing ideas around to craft solutions to complex problems. This kind of teamwork can lead to breakthroughs, like the ongoing work that earned the Nobel Prize in Chemistry readied for release in 2024, comprising dozens of specialists from various fields. However, Zou and his team argue access to such collaborative networks isn’t commonplace.
Curious about whether AI could replicate these conversational exchanges among scientists, Zou decided to try. While some AI systems perform admirably in solving certain scientific queries, few have been trained to interact with each other meaningfully.
Pushing boundaries, the team set up a Virtual Lab with AI scientists assigned to find antibody treatments for the latest COVID-19 strains. Antibodies can be effective, but as the virus morphs, it reduces their efficiency, foreshadowing the importance of rapid development of new antibodies.
The Virtual Lab operated under an AI Principal Investigator, who organized a team of digital experts to tackle this complex task. With the help of sophisticated software for modeling proteins, individual AI researchers participated in collective brainstorming sessions and focused meetings to generate potential antibody treatments.
Using nanobodies— more diminutive versions of traditional antibodies— they garnered 92 proposals for potential COVID treatments in just two days! This highlighted the operations of the Virtual Lab, showcasing the speed at which AI can churn out hypotheses.
James Zou pointed out, “The AI meetings are typically way more productive than ours. They wrap up in minutes, conducting several sessions simultaneously without feeling weary!”
The standout aspect of the AI Lab was that it documented all discussions, which was instrumental for human researchers in dissecting the thought processes behind the AI conclusions. John Pak, biochemist and co-leader of the study, found this transparent interaction quite motivating, saying, “As a researcher, you can hesitate to involve AI in daily tasks, but our Virtual Lab felt more intuitive to engage with.”
Samuel Rodrigues, an AI researcher and bystander of the study, called the findings “a thrilling leap forward.” As the CEO of FutureHouse, a firm centered around streamlining scientific research through AI, he heralded the multi-AI collaboration as not only groundbreaking but vital for introducing AI into science. While he suggested future adjustments for expansion, he viewed challenges as minor compared to the advantages.
The researchers depicted a need for continued development of AI modalities, potentially enhancing them with additional training and tools. Still, they’d already accommodated these systems for versatility.
Despite remarkable advancements, limitations persist— AI can occasionally concoct inaccurate ideas or facts from flawed or incomplete data— take the early iteration of Google’s AI that humorously recommended adding glue to pizza sauce. To counteract these types of mishaps, an AI scientific critic was part of the Virtual Lab, scrutinizing claims and ensuring multiple meetings were held for the same queries for consistency. Remaining reliant on human oversight was crucial to validate decisions and authenticate evidence.
Lastly, while the potential treatments generated by nanobodies look good on paper, exacting tests will be indispensable before knowing how effective they might be in the more complicated environment of the human body.
Even with these caveats, Zou and Pak are confident in the significance of the Virtual Lab for various fields. As Pak noted, “We’re concentrating on exploratory research that could be beneficial to others! Looking forward to taking this out for testing across diverse scientific questions!”
