Stanford University researchers have built an AI system called Biomni that can read scientific papers, analyze datasets, write code, form hypotheses, and design experiments in minutes. The system, described in the journal Science, is designed to handle the labor-intensive work of biomedical research while leaving interpretation and decision-making to human scientists.
Biomni uses 150 specialized tools, 105 software packages, and 59 databases across 25 areas of biology. It completes complex research tasks at speeds far beyond traditional human workflows - completing in minutes what might take researchers weeks or months.
"If you think of an agent as a carpenter, a carpenter without tools is just a carpenter who can talk," said Jure Leskovec, professor of computer science at Stanford. "With Biomni, we give the carpenter a set of tools, so it can build."
A systems approach to research bottlenecks
Modern biomedical research produces new studies daily and datasets grow larger each year. But progress often stalls under the weight of routine tasks. Scientists must read literature, organize data, write code, and test ideas before reaching conclusions. Each step takes time.
"The hurdle in biomedical science is not intelligence or ideas; it is mechanics," Leskovec said. "It's this laborious stuff that slows innovation. Biomni can do this work in minutes."
The system aims to close the gap between what researchers know and what they can act on. A scientist can ask a question in plain language and receive a complete research workflow. Biomni reads papers, selects datasets, chooses tools, writes and runs code, interprets results, and suggests next steps.
"Biomni is able to understand a simple question like, 'Why are these patients responding differently to the drug?'" said Kexin Huang, who helped lead the project. "Then it digs in, doing a lot of the scientific legwork."
Speed that changes the pace of research
In one real-world example, a researcher uploaded more than 450 files of health data including glucose levels, food intake, and physical activity. The user asked the system to analyze the data and find meaningful patterns. Within 40 minutes, Biomni cleaned the data, combined it, created visualizations, and identified links between diet and body temperature. Leskovec said the same work could take a human more than 60 hours.
In tests, Biomni matched the accuracy of experienced researchers across gene identification and disease analysis tasks. For rare disease diagnosis, it reached 60 percent accuracy - similar to experts scoring between 60 and 70 percent - but finished in three minutes compared to nearly two hours for humans. In gene detection, it achieved 80 percent accuracy and reduced analysis time from about 90 minutes to four minutes.
From data analysis to real experiments
The system also designs experiments and guides lab work. In one case, Biomni generated a complete protocol for DNA cloning including detailed steps, materials, and validation methods. Researchers followed the instructions in the lab and the experiment worked as expected - colonies grew and DNA sequencing confirmed results.
Every step of Biomni's work is recorded and linked to sources, providing full citations for the data and methods it uses. This helps researchers verify results and reproduce findings.
"This is not about machines taking over science, but more about machines becoming a powerful new partner to augment human researchers," Huang said.
More than 10,000 labs have begun using the system in academic and industry settings. Researchers are working to expand its capabilities with more data sources and improved reasoning methods.
Why this matters for science and research professionals
For research scientists, Biomni addresses a concrete problem: the gap between having ideas and being able to test them. The system automates the mechanics of research - data cleaning, code writing, literature review - that often consumes weeks of time. This changes what's possible in resource-constrained settings. Scientists who lack access to large teams or specialized computational expertise can now perform complex analyses. For those interested in building these skills, training resources like the AI for Science & Research tag and this AI Learning Path for Research Scientists offer practical ways to understand how AI tools like Biomni can augment laboratory work.
Your membership also unlocks:
