Biomni: A General-Purpose Biomedical AI Agent

Abstract

Biomedical research underpins progress in our understanding of human health and disease, drug discovery, and clinical care. However, with the growth of complex lab experiments, large datasets, many analytical tools, and expansive literature, biomedical research is increasingly constrained by repetitive and fragmented workflows that slow discovery and limit innovation, underscoring the need for a fundamentally new way to scale scientific expertise. Here, we introduce Biomni, a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields. To systematically map the biomedical action space, Biomni first employs an action discovery agent to create the first unified agentic environment - mining essential tools, databases, and protocols from tens of thousands of publications across 25 biomedical domains. Built on this foundation, Biomni features a generalist agentic architecture that integrates large language model (LLM) reasoning with retrieval-augmented planning and code-based execution, enabling it to dynamically compose and carry out complex biomedical workflows - entirely without relying on predefined templates or rigid task flows. Systematic benchmarking demonstrates that Biomni achieves strong generalization across heterogeneous biomedical tasks - including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning - without any task-specific prompt tuning. Real-world case studies further showcase Biomni's ability to interpret complex, multi-modal biomedical datasets and autonomously generate experimentally testable protocols. Biomni envisions a future where virtual AI biologists operate alongside and augment human scientists to dramatically enhance research productivity, clinical insight, and healthcare. Biomni is ready to use at https://biomni.stanford.edu, and we invite scientists to explore its capabilities, stress-test its limits, and co-create the next era of biomedical discoveries.

Figure 1:

Overview of the unified biomedical action space and agent environment in Biomni. (a) Workflow for systematically curating the unified biomedical action space. Actions necessary to conduct biomedical research were extracted from 2,500 recent bioRxiv publications across 25 biomedical subfields using an AI-driven discovery agent. Extracted actions were rigorously validated and curated by human experts, resulting in the integration of 105 biomedical software tools, 150 specialized biological tools (including wet-lab protocols, AI-driven predictive models, and domain-specific know-how), and 59 comprehensive biomedical databases. (b) Illustration of the unified biomedical action space spanning diverse biomedical subfields such as genetics, genomics, synthetic biology, cell biology, physiology, microbiology, pharmacology, bioengineering, biophysics, molecular biology, and pathology. Representative tools and databases integrated into Biomni’s environment are shown, highlighting its general-purpose capabilities. (c) Example workflow demonstrating Biomni’s reasoning and action composition process to autonomously answer a complex biological question. Biomni retrieves relevant tools based on the user’s query, formulates a structured reasoning plan, and composes executable code to perform comprehensive bioinformatics analyses, iteratively refining its reasoning based on observations until converging on a final, precise answer.

Figure 2:

Zero-shot generalization of Biomni across diverse realistic biomedical tasks. (a) Biomni is superior to 6 baselines in Q&A multiple choice benchmarks that broadly evaluate the model’s capability across biomedical fields. (b) Biomni demonstrates robust zero-shot performance across eight previously unseen, real-world biomedical scenarios spanning multiple biomedical subfields, without any task-specific fine-tuning or prompt engineering. Evaluated tasks include variant prioritization and GWAS causal gene detection (genetics and genomics), perturbation screen design (functional genomics, immunology), patient gene prioritization, rare disease diagnosis (clinical genomics), drug repurposing (pharmacology), microbiome disease-taxa bioinformatics analysis (microbiology), and single-cell RNA-seq cell annotation (single-cell biology). Across these diverse scenarios, Biomni consistently outperformed baseline models (Base LLM, ReAct+Code) and specialized environments (Biomni ReAct), highlighting its general-purpose biomedical capabilities and ability to autonomously adapt to new and complex biomedical tasks.

Figure 3:

Biomni autonomously executes complex multi-modal biomedical analyses to generate hypothesis. (a-d) Biomni rapidly analyzed CGM-derived thermogenic responses data and activity data from 30 individuals, comprising 458 raw Excel sheets. (b) Workflow demonstrating Biomni’s autonomous execution of data preprocessing, meal event detection, postprandial temperature analysis, and thermogenic response characterization. (c) Representative individual temperature-response plots and temperature increase distribution following meals, automatically generated by Biomni. (d) Summary of unique biological findings identified by Biomni, including significant increases in core body temperature post-meal intake (average 2.19C, median 1.10C), and notable inter-individual variability in thermogenic responses. (e-h) Biomni autonomously analyzed single-cell multiomics data from approximately 336,000 nucleus droplets, combining single-nucleus RNA (snRNA-seq) and single-nucleus ATAC sequencing (snATAC-seq) across human embryonic joint development (shoulder, hip, knee). (f) A detailed workflow diagram showing Biomni’s 10-step analysis pipeline for gene regulatory networks with multiomics. (g) Two key figures generated from Biomni: Left panel shows a heatmap of regulator activity by developmental stage, with color intensity indicating activity levels. Right panel displays a boxplot of RUNX2 regulon activity by cell type, showing variation in expression across different cell populations. (h) Key findings from the GRN analysis: 1) Novel transcription factors (AUTS2, ZFHX3, and PBX1) showing high regulatory activity across multiple skeletal lineages despite no previous association with skeletal development, and 2) Across the 566–589 regulons recovered, limb mesenchyme cells display the highest mean regulonactivity score, underscoring their prominent role in skeletal transcriptional control.

Figure 4:

Biomni designs wet-lab experimental protocol. (a) Open-ended cloning benchmark on 10 real cloning scenarios. We compared against base LLM, trainee-level human, and expert-level human scientists. We found that Biomni has similar accuracy as the expert level scientist, and significantly higher accuracy than trainee level, while using much less time. (b) Example of a user request to Biomni for cloning an sgRNA targeting the human B2M gene into the lentiCRISPR v2 Blast plasmid. (c) Biomni’s automated stepwise workflow, including plasmid analysis, sgRNA design, oligo synthesis, Golden Gate assembly, bacterial transformation, colony screening, and final plasmid mapping. (d) Biomni-generated detailed cloning protocol with step-by-step instructions and comprehensive plasmid map, enabling laboratory scientists to execute the experiment autonomously. (e) Validation of Biomni’s cloning protocol through successful colony growth on selection plates, followed by Sanger sequencing confirming perfect alignment of sgRNA insertion in picked colonies, demonstrating Biomni’s robust capability for precise and reliable experimental design.