Researchers at Stanford University have created Biomni, an AI-powered research agent that can independently design experiments, analyze data and generate scientific hypotheses—work that previously took human scientists weeks or months to complete.
Unlike conventional chatbots, Biomni functions as a full biomedical "co-scientist" capable of reading scientific literature, forming hypotheses, selecting datasets and tools, writing code, interpreting results and suggesting next-stage experiments as part of a complete research workflow.
For the study, published in Science, the researchers trained Biomni specifically on biomedical science by incorporating the full text of publicly available papers, code and data stored on bioRxiv. The system layers in 150 specialized biomedical tools, 105 software packages and 59 databases spanning all 25 biomedical subdomains defined by bioRxiv, from genetics to neurology. That foundation allows Biomni to identify the common software, tools and databases already used across biomedical research and apply them to new questions.
In Biomni’s first nine months as an open-source Stanford project, more than 15,000 scientists asked the AI research assistant to automate 100,000 different scientific workflows, such as formulating testable hypotheses, performing complex bioinformatics analyses, and designing rigorous experimental protocols.
In its initial testing, Biomni excelled on established Q&A benchmarks for biomedical knowledge and reasoning. It also performed well on eight challenging, realistic scenarios never encountered during development, indicating it can generalize across domains without task-specific training.
“Across each of these use cases, Biomni accelerated the path from messy real-world data to testable hypotheses, and supported applications in domains as diverse as metabolic research and precision health,” said Kexin Huang, Biomni co-creator.
Although Biomni approaches human-level performance in some tasks, such as database querying, sequence analysis and molecular cloning, the researchers note it still struggles in areas that require nuanced clinical judgment, novel experimental reasons or deep biological thinking and synthesis. It also doesn’t cover every field.
Overcoming hurdles
The speed of Biomni—reducing some work from 60 hours to mere minutes—is especially critical, says Jure Leskovec, senior author and professor of computer science in Stanford's School of Engineering. Leskovec believes there is an inverse relationship between scientific information and the pace of discovery. As the volume of knowledge, data and tools has grown, innovation has slowed.
“The hurdle in biomedical science is not intelligence or ideas; it is mechanics,” said Leskovec. “It’s this laborious stuff that slows innovation. Biomni can do this work in minutes.”
Huang said Biomni is not designed to replace humans but rather free them to concentrate on the value of the scientist: ideation and judgment.
“This is not about machines taking over science, but more about machines becoming a powerful new partner to augment human researchers,” Huang said. “With Biomni, scientists have a fast and tireless collaborator that empowers them to focus on the important work of science.”
In September 2025, Huang spun Biomni out of the Stanford AI Lab with a seed round of venture funding. He now leads the startup, called Phylo, while Leskovec remains involved as scientific co-founder. The original public platform has migrated to the new entity as Biomni Lab, with an Academic Lab Program available to universities. The codebase remains fully open source.
A prototype Biomni is already in use by more than 10,000 labs in academia and industry, making it the most widely used AI co-scientist system in biomedicine.