Biomedical Technology · us
From Finding Genes to Writing Experimental Protocols: Biomni Connects Fragmented Biomedical Research Tools into an AI Agent
The system can break down research questions, retrieve information, and execute analytical code, and has been tested across 25 biomedical domains. But its ability to accelerate workflows does not mean it can replace clinical judgment, experimental validation, or researchers’ biological insight.
Bottlenecks in biomedical research often arise not only from a lack of data, but also because data are scattered across different formats, databases, and analytical tools. A study published in *Science* introduces Biomni, a general-purpose biomedical AI agent that seeks to connect literature searches, tool selection, code writing, results analysis, and experimental protocol design. Researchers can pose questions in natural language, then inspect and revise the system’s work step by step.
Biomni combines language-model reasoning, information retrieval, and executable tools. According to information from the Stanford University team, its operating environment includes more than 150 specialized tools, 59 curated databases, and approximately 100 software packages; the official project page lists 105 software packages. When presented with tasks for which it has not been specifically trained, the system first selects resources and develops a plan, then writes and executes code, while researchers can review the process and change direction along the way.
The paper’s benchmark tests and case studies span 25 domains, including disease-associated gene prioritization, drug repurposing, rare-disease diagnosis, protein stability optimization, multimodal data analysis, and experimental protocol generation. Specific applications also include variant interpretation, Perturb-seq and single-cell RNA sequencing analysis, ADMET property prediction, molecular docking, molecular cloning, and CRISPR screen design. What these tasks collectively test is not one-off question answering, but whether multiple analytical steps can be connected into an executable workflow.
One case closer to a real-world research setting involved continuous glucose and body-temperature data accumulated over several months from 30 participants and distributed across 458 spreadsheets. Biomni created and executed a ten-step analysis, inferring eating events from rises in blood glucose and then organizing pre-meal and post-meal temperature changes into a report for researchers to review. This suggests that it may shorten the path from data cleaning to generating testable hypotheses, but patterns identified through the analysis still require confirmation with independent data and experiments.
The scale of real-world adoption also offers signals distinct from benchmark scores. The Stanford Institute for Human-Centered Artificial Intelligence said that more than 15,000 scientists used Biomni during the first nine months after it was open-sourced, collectively automating 100,000 workflows. These figures come from the development team and cannot be treated directly as evidence of research quality or replicated discoveries. The publicly available code is licensed under Apache 2.0, an approximately 11 GB data lake can be deployed locally, and multiple language-model providers are supported.
The system’s limitations are equally clear. The research team acknowledges that Biomni still has weaknesses in nuanced clinical judgment, genuinely novel experimental reasoning, and deep biological integration, and that some important domains have not yet been evaluated. The paper’s broad range of cases also cannot substitute for prospective clinical validation. If the system is used with patient data or for medical decisions, data governance, reproducibility, accountability, and regulatory requirements will all need to be addressed separately.
Deployment security presents a more immediate risk. The project repository warns that Biomni may execute model-generated code with full system privileges, so it should be placed in an isolated or sandboxed environment when sensitive data, credentials, or internal systems are involved. The repository also states that the public version frozen on April 15, 2025, is not entirely identical to the current web platform. When assessing performance and risk, users must first clarify which model, tools, and data versions are actually being used.