Stanford researchers have built Biomni, a general-purpose biomedical AI agent that can execute a wide range of research tasks, from analyzing genomic data to designing lab protocols. The open-source platform, now used by more than 15,000 scientists, was spun out into a startup called Phylo in September 2025 with venture funding.
The team, led by former Stanford PhD candidate Kexin Huang and computer science professor Jure Leskovec, developed Biomni with funding from a Stanford HAI Hoffman-Yee Research Grant. It draws on expertise from Stanford's Computer Science, Genetics, Pathology, Medicine, and Pediatrics departments, along with contributions from Genentech, Arc Institute, the University of Washington, and UC San Francisco.
How Biomni works as a lab partner
Biomni consists of a cloud-based virtual work environment and an agentic architecture that lets the AI handle unfamiliar tasks without extra training. It combines large language models with more than 150 specialized bioinformatics tools, 59 curated databases covering protein structures and genomic variants, and over 100 software packages for molecular modeling and single-cell analysis. A scientist types a query into a chat window-for example, "Analyze the attached Perturb-seq data and generate a meaningful hypothesis"-and the agent selects the right tools, formulates a plan using its biomedical knowledge, and executes it step by step. The researcher can review the plan and intervene at any point.
"Today, we have abundant biomedical data, but we don't have enough human researchers to analyze it all," Huang said. "While human biologists are limited by specialized expertise, AI can integrate across disciplines and manage thousands of concurrent tasks." Biomni's ability to switch between domains and handle many workflows at once makes it a practical tool in AI for Science & Research.
Leskovec described the experience as a documented conversation. "You give the agent a task, and it writes Python code to use the advanced models and tools. From there, the assignment becomes a digital conversation that's fully documented and auditable. Scientists no longer have to worry about losing the history of their work in notebooks and Excel spreadsheets."
Early results across complex tasks
In its first nine months as an open-source project, over 15,000 scientists used Biomni to automate 100,000 different scientific workflows-formulating testable hypotheses, running bioinformatics analyses, and designing experimental protocols. The agent performed well on established biomedical Q&A benchmarks and on eight realistic scenarios it had never seen during development, indicating it can generalize without task-specific training.
One case study involved a researcher who gave Biomni 458 Excel files with continuous glucose monitor data from 30 participants. The open-ended prompt-"Can we uncover biologically meaningful thermogenic patterns?"-led the agent to autonomously generate and run a 10-step analysis plan. It inferred meal events from glucose spikes, extracted pre- and post-meal body temperature readings, and delivered a structured report with individual and population-level trends. In other tests, Biomni rapidly analyzed raw genomic sequence datasets and designed wet-lab protocols.
Limitations and the path to commercialization
Biomni approaches human-level performance on tasks like database querying, sequence analysis, and molecular cloning, but it still struggles where nuanced clinical judgment or deep biological synthesis is required. The team also notes it does not cover every biomedical field. The codebase remains fully open source, now under the Phylo startup's Biomni Lab, which offers an Academic Lab Program for universities. Huang leads the company, with Leskovec as scientific co-founder, continuing to support Research in both academic and commercial settings.
Why this matters for Science and Research
Biomni shows that AI agents can take over repetitive, fragmented analysis tasks that slow down biomedical discovery. By handling data wrangling, tool orchestration, and hypothesis generation, these systems free scientists to focus on experimental design and interpretation. The key takeaway: while the agent reduces manual work and speeds up preliminary analysis, human judgment remains essential for clinical decisions and novel reasoning. For research teams, tools like Biomni could shift how lab work gets done-less time spent on routine computation, more time on the questions that require a scientist's intuition.
Your membership also unlocks:
