henhao
Sign in
BioinformaticsbioRxiv
Preprint

Biomni: A General-Purpose Biomedical AI Agent

Kexin HuangiD, Serena ZhangiD, Hanchen WangiD, Yuanhao QuiD, Yingzhou LuiD, Yusuf RoohaniiD, Ryan J. LiiD, Lin Qiu, G. L. LiiD, Junze ZhangiD, Di YiniD, Shruti MarwahaiD, Jennefer N. CarteriD, Xin ZhouiD, Matthew Thomas WheeleriD, Jonathan Adam BernsteiniD, Mengdi WangiD, Peng HeiD, Jingtian ZhouiD, Michael P. SnyderiD, Le CongiD, Aviv RegeviD, Jure LeskoveciD

PDFDOI
Post to X
Threads on XPost to BlueskyThreads on Bluesky
In-depth full-text digest available · Jump to digest ↓

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, micro-biome analysis, and molecular cloning - without any task-specific prompt tuning. Real-world case studies further showcase Biomnis 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.

02 /Specification & Classification

Click any tag to browse matching papers
Citations
125
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (2 Jun 2025) · CC BY 4.0

An LLM agent with a literature-mined environment of 150 tools, 105 packages and 59 databases plans in code, matching reported human accuracy on database and sequence question answering.

Earlier biomedical agents were built for narrow tasks, so they do not move across subfields, and general LLMs lack an explicit action space of specialised biomedical tools and data. Biomni pairs an environment mined by an LLM from recent literature across 25 bioRxiv categories with an agent that retrieves relevant resources, plans, and acts by writing and running code rather than following fixed workflows.

Key results

  1. ShownOn multiple-choice knowledge benchmarks, the agent outperformed all baselines on database question answering, reached the human scores quoted for both LAB-Bench subtasks, and on Humanity's Last Exam beat the base LLM, coding agent and literature agent.

    Database question answering: 74.4% accuracy versus 74.7% for human experts and 40.8% for a ReAct+Code agent. Sequence question answering: 81.9% versus 78.8% human. Humanity's Last Exam (52 questions, 14 subfields): 17.3% versus 6.0% base LLM, 12.8% coding agent, 12.2% literature agent. LAB-Bench used 315 held-out questions averaged over three runs. · Biomni excels on general biomedical knowledge…; Q&A Benchmarks

  2. ShownOn eight newly built realistic tasks (variant prioritization, GWAS causal gene, CRISPR screen design, rare disease diagnosis, drug repurposing, scRNA-seq annotation, microbiome association, patient gene prioritization), the full agent beat its baselines without task-specific prompt tuning.

    Average relative gains: 402.3% over a tool-free Claude Sonnet 3.7, 43.0% over a ReAct+Code agent, and 20.4% over Biomni-ReAct, an ablation with ReAct-style chaining instead of code-based planning. Per-task scores are in supplementary notes and figures that the extracted text does not include. · Biomni generalizes to new, real-world biomedical…; Curating real-world benchmarks

  3. SuggestedIn a case study on heterogeneous wearable data, the agent ran its own multi-step pipeline end to end and reported a postprandial temperature rise with wide between-person spread, plus sleep and multi-omics analyses.

    458 Excel files of continuous glucose monitoring and temperature data from 30 participants: mean rise 2.19°C, median 1.10°C, range -0.11°C to 15.56°C. Separately, 227 nights from 10 participants (sleep efficiency highest on Wednesdays, timing regularity associated with quality more than duration) and 652 lipid, 731 metabolite and 1,470 protein features analysed jointly with the glucose data. · Biomni jointly analyzes 458 wearable sensor…; Wearable analysis case study

  4. SuggestedGiven a published human embryonic skeletal multi-omics dataset, the agent planned and ran a ten-stage regulatory network pipeline, recovering known osteogenic regulators and flagging further transcription factors as broadly active; humans checked the traces afterwards.

    336,162 single-nucleus RNA-seq and ATAC-seq profiles from embryos 5-11 weeks post-conception; the run took just over five hours, handling execution errors by subsampling and debugging. It reproduced regulation of RUNX2 and HHIP by TWIST1, LMX1B and ALX4, and reported high activity for AUTS2, ZFHX3 and PBX1 in osteoblasts, preosteoblasts and chondrocytes. No new experiment tests these. · Biomni automates complex multi-omics analysis to…; Multiome analysis case study

  5. SuggestedOn an expert-designed molecular cloning benchmark scored blind for accuracy and completeness, the agent's protocols were rated comparable to a senior human expert and better than a trainee, and one agent-designed construct was built and sequence-verified in the lab.

    Tasks spanned four assembly strategies with single and pooled variants plus validation steps, scored 1-5 by a blinded senior reviewer; the agent finished far faster than the expert, whose task times were recorded. The text gives 10 tasks in the results and 20 in the methods, and reports no numerical scores. Wet-lab check: colonies grew and two sequenced clones aligned as designed. · Biomni designs wet-lab validated experimental…; Wetlab Benchmark Development and Evaluation; Wetlab Validation

Methods, data and reproducibility

An LLM action-discovery agent read 100 recent papers from each of 25 bioRxiv categories to extract tasks, software and databases; about 1,900 recurring tasks were filtered and human-verified into Biomni-E1 (150 tools with test cases, 105 packages, 59 databases). Biomni-A1, built on CodeAct, retrieves relevant resources, writes a numbered plan, and executes Python, R or Bash iteratively.

  • Action-discovery corpus: recent bioRxiv publications per subject category · n = 100 papers per category, 25 categories
  • LAB-Bench DbQA and SeqQA subsets (12.5% sample): development and held-out test sets · n = 45 development, 315 test questions
  • Humanity's Last Exam biology/medicine sample, up to five questions per subdiscipline · n = 52 questions
  • Eight curated real-world benchmarks drawn from Open Targets Genetics, Shringarpure et al., Schmidt et al., Huang et al., MyGene2 (Alsentzer et al.) and public scRNA-seq portals · multiple species · n = 5 microbiome datasets; scRNA-seq sets subsampled to 20k-50k cells
  • Wearable study: CGM and temperature files (485 meal-centred files in methods), sleep records, and blood omics · human · n = 458 files, 30 people; 227 nights, 10 people; omics on 147 and 20 samples
  • Published human embryonic skeletal multi-omics (snRNA-seq, snATAC-seq, spatial transcriptomics) across five anatomical regions, used unmodified · human · n = 336,162 nuclei, 5-11 weeks post-conception
  • Molecular cloning benchmark tasks curated with gene-editing researchers, plus one laboratory validation task · human (validation target B2M) · n = 10 tasks in results, 20 in methods
Methods & tools
Biomni (Biomni-E1, Biomni-A1), CodeAct, Claude Sonnet 3.7, ReAct+Code baseline, LAB-Bench, Humanity's Last Exam, pySCENIC, GRNBoost2, cisTarget, AUCell, scanpy, MetaPhlAn2
Reagents & environment
A containerized execution environment preinstalled with 105 biological software packages, supporting Python, R and command-line tools.
Experimental conditions
The skeletal multi-omics regulatory network run finished in just over five hours; no hardware is described, and benchmark sizes were limited by API cost.
Reproducibility gaps
The text says the code is open-sourced, a web interface exists and the data are on Harvard Dataverse, but every link was dropped from the extracted text; tool, database and score tables are in supplementary material. The public interface is stated not to be for protected health information.
Limitations (5)
  • The tasks evaluated cover only part of biomedicine, and key domains were not tested; the authors state the agent has not reached expert-level performance across all categories. · stated by the authors
  • Favouring the most recent literature in action discovery risks missing older but still useful concepts and techniques; future versions should cover more publications. · stated by the authors
  • The agent still struggles with nuanced clinical judgement, novel experimental reasoning and deeper biological synthesis. · stated by the authors
  • Benchmarks were subsampled (12.5% of LAB-Bench, for API cost; 52 HLE questions), and results on the eight realistic tasks are given only as averaged relative gains, without per-task scores or variability in the extracted text. · noted in this summary
  • The wearable and skeletal case studies are single demonstrations reported through the agent's own summaries, and the skeletal findings were checked by reading its traces rather than by new experiments; counts also differ between results and methods (458 vs 485 files; 10 vs 20 cloning tasks). · noted in this summary

Who should read:Read this if you are weighing whether an LLM agent with a curated tool-and-database environment can take on your analysis or protocol-design work, and want the benchmark numbers behind that claim.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISQu '26Miller '25Matsumoto '26Aamer '25Martinek '26Zhang '26
72% semantic matchBioinformatics
Yuanhao Qu, Yingzhou Lu, Xinming Tu+21 more·bioRxiv·May 18, 2026· 1 citations