A Stanford University-led team including two Chinese researchers said they built the first general-purpose biomedical AI agent capable of working alongside human scientists, taking on complex tasks that once required groups of specialists.
Jure Leskovec, a Stanford computer science professor who supervised the work, said the agent had been released as an open-source system with a web interface so that biologists could use it without writing code.
“We have over 10,000 scientists all over the world using the system for their everyday tasks,” Leskovec said on Tuesday.
In the journal Science this week, the researcher said the system, called Biomni, could turn a plain-language request into an entire research workflow – from searching databases and writing analysis code to finding disease-causing genes and even creating step-by-step lab instructions that scientists successfully followed in real experiments.
In one test, the virtual biologist was given hundreds of raw files collected from wearable devices and asked to look for biological patterns. It cleaned the data, ran the analysis and generated new hypotheses.
In another test, the AI agent analysed genetic data from developing human embryos and uncovered previously overlooked factors that may help control bone formation.
Kexin Huang, a PhD student in computer science at Stanford at the time of the work and an architect behind Biomni, said the team tested the system on more than 400 real-world research tasks and compared its performance with existing AI tools and human experts.
In one case, Biomni needed only 35 minutes to complete a complex biological data analysis that took a human expert three weeks.
“It reached expert-level accuracy while cutting the time to just minutes,” said Huang, who later co-founded Phylo, a San Francisco-based start-up that aims to make the AI system available to more researchers.
Yuanhao Qu, a cancer biology PhD student at Stanford at the time of the work and another main developer of Biomni, said he was also a user of the system, after spending years doing laboratory research himself.
One example was molecular cloning – the process of creating and replicating DNA constructs for experiments – a routine task that took Qu years to master and still consumed hours each time, he said.
When tested on such a task, Biomni “did the whole thing from end to end”, designing the experiment and producing a protocol that matched expert-level work in a fraction of the time.
“For me, Biomni is really changing the way biologists work,” said Qu, who is also a Phylo co-founder and originally from Beijing.
“Work that usually takes me hours now takes just minutes, so I can really spend my time on the science that actually needs a human.”
But Leskovec said the system was designed as a collaborator rather than a replacement for scientists, with humans still responsible for asking questions, judging results and deciding what directions to pursue.
“Biomni is a powerful tool, not a decision maker,” Leskovec added.
The researchers noted that Biomni still had major limitations, having only been tested on part of the biomedical field while continuing to struggle with tasks requiring deeper scientific judgment, original experimental ideas and complex reasoning.
Future versions of Biomni could become broader, smarter and more self-improving by learning from experience, incorporating more types of scientific data and continuously adding new research tools and knowledge, they said. -- SOUTH CHINA MORNING POST
