Biomni: A General-Purpose Biomedical AI Agent

“Biomedical research is a key pillar of modern science and medicine, driving discoveries in disease mechanisms, diagnostics, and therapeutics 1–4 .”

“Dataset 2 curated from a well-known Nature study, offers microbial abundance data in mice alongside metadata such as diet and sex, making it valuable for modeling host-microbiome interactions 36 .”

“Biomedical research is a key pillar of modern science and medicine, driving discoveries in disease mechanisms, diagnostics, and therapeutics 1–4 .”

“Finally, Dataset 5 is an in-house resource derived from the Human Microbiome Project 39 .”

“Enabling an AI agent to handle a broad range of biomedical tasks introduces substantial technical chal-50 lenges – most notably, the need to tightly couple advanced reasoning 20 with the ability to exe-51 cute highly specialized biomedical actions 21 .”

“Yet, with the growth in large-scale experiments, data, tools, and literature, progress is increasingly slowed by fragmented, complex workflows that require specialized tools, exhaustive literature reviews, intricate experimental design, and careful statistical modeling 5,6 .”

“Although LLM-based reasoning has seen significant advancements 22 , such LLMs need access to an environment that explicitly defines the biomedical action space, which is inherently diverse, domain-specific, and complex.”

“One line emphasizes automating experimental workflows, e.g., chemical synthesis or biomedical pipelines (M. Bran et al., 2024; Ruan et al., 2024; Huang et al., 2025b; Qu et al., 2025).”

“This tightly coupled language–code–3D loop reduces the time from question to evidence relative to agent-only or text-only systems.”

“…difficulty is not surprising, it is also reported in other broad-scope benchmarks and suggests that some reasoning skills may be particularly reliant on specific knowledge absent from pre-training or may exceed the reasoning capacity of the base models (Narayanan et al., 2025; Wang et al., 2025).”

“…reasoning, treatment planning (Ferber et al., 2025; Chen et al., 2025; Kuerbanjiang et al., 2025; Liu et al., 2026), and biomedical research applications such as drug discovery and biomarker reasoning (Inoue et al., 2025; Zuo et al., 2025; Jin et al., 2025; Huang et al., 2025; Bazgir et al., 2025).”

“Biomni [48] runs Claude Sonnet 5 with one replicate.”

“Five systems represent distinct agent paradigms: ChemCrow [3] for chemistry-specific tool use, Biomni [17] for biomedical research, AgentD [22] for modular drug discovery, and Codex [23] and Hermes [21] for general-purpose tool interaction.”

“Biomni, 24 released as this perspective was being prepared, presents a structured action space across biomedical subfields and satisfies the structural conditions of the paradigm at the broader biomedical layer rather than genomics specifically; we expect general-purpose biomedical agents of this…”

“In practice, experienced analytical scientists frequently employ customized optimal designs rather than rigid predefined templates [ 171, 172].”

“Looking forward, pipeline building and execution can be accelerated with use of artificial intelligence (AI) agent platforms or tools [195, 196] powered with large language models (LLM).”

“And recently, AI agents have been highlighted as partners to scientists in virtual labs that resulted in the identi fi cation of new nanobody binders for SARS-CoV-2 variants, and as independent biomedical AI agents that can perform tasks across different sub fi elds 25,26 .”

“IO Chatbot [9], and general biomedical agents such as Biomni [10].”

“Equipped with retrieval, database access, and other tool use, such systems have shown strong promise for biological problem solving [21–23].”

“…systems such as The AI Scientist [13] aim to automate substantial parts of the research cycle, and domain-specific agents such as ChemCrow [14] and Biomni [15] demonstrate that coupling LLMs with expert tools, databases, and executable environments can substantially expand scientific task coverage.”

“Biomni [8] is a general-purpose biomedical AI agent that unifies reasoning, retrieval, and code execution across diverse biological domains (genomics, immunology, microbiology, neuroscience), emphasizing breadth and autonomous task execution through free-form planning.”

“Regarding task-specific planning , we can integrate information proposed by known biomedical databases and divide the main task into several steps, based on Biomni (35).”

“Biomni 12 extends this direction as a general-purpose biomedical agent for tasks such as hypothesis generation, protocol design, clinical reasoning, gene prioritization, drug repurposing, rare disease diagnosis, and microbiome analysis.”

“Recent systems demonstrate the potential of general-purpose biomedical AI agents across these diverse tasks (Huang et al., 2025).”

“…of these agents lies in their ability to write, debug, and execute code [21, 39, 89], which makes them well-suited for aiding the automation of scientific discovery workflows that rely on code-based interfaces, spanning chemistry [7, 15], mathematics [92], physics [5, 63], and biology [33, 46, 83].”

“Biomni [12] : Biomni is a LLM–driven framework designed for general-purpose biological reasoning across diverse omics modalities.”

“The Reasoning agent uses GPT-OSS-120B [2], an open-source model hosted and served on ALCF Sophia, for inference with a sampling temperature of 0.7 and a maximum output length of 32768 tokens to generate structured JSON outputs.”

“Meanwhile, the number of available models and datasets in machine learning has grown dramatically, making it increasingly important and challenging to recommend appropriate baselines and datasets for a new research idea, see fig 1.”

“From a technical perspective, future efforts can enhance Traj-CoA’s performance, including access to external knowledge [59, 60], using more powerful base models, and fine-tuning via multi-agent training approaches [61, 62].”

“From a broader perspective, LLMs agents [72] and Retrieval Augmented Generation (RAG) [73] could become increasingly valuable for identifying and collecting omics datasets from vast public repositories that best match a given phenotype of interest.”

“Subsequent research extended these capabilities to cross-stage planning [22], automated hypothesis generation [11, 23, 24], multimodal knowledge integration [25, 26], and automated experimentation across scientific domains.”

“…systems can help overcome this limitation by surfacing relevant knowledge from across vast datasets and literature corpora, organizing connections, and enabling hypothesis generation that is informed by a broader and more comprehensive information space than a human could synthesize alone [21,22].”

“The AI Scientist automates an ML research loop end to end [16], and Biomni is a general-purpose biomedical agent [13].”

“Such systems in biology including Virtual Lab, ToolUniverse, and Biomni demonstrate the use of tools within agentic scientific workflows [10, 13, 14, 35].”

“While recent agentic systems have demonstrated that large language models (LLMs) can automatically decompose and execute complex scientific workflows 11,12 , translating this success to the highly specialized domain of S2F modeling reveals critical limitations.”

“Such approaches can enhance research and also aid in improving healthcare by generating clinical insights in complex cases [ 13].”

“Agentic AI systems compound this problem [17]: hallucinated pathways, i.e., chains of reasoning that traverse plausible-sounding molecular steps with no physical substrate, propagate silently across retrieval and multi-step inference, making them particularly dangerous in a therapeutic context…”

“…single-cell foundation models began treating biological sequences and cellular states as learnable representations, while biomedical agents such as Biomni and tool ecosystems such as ToolUniverse began connecting LLMs to databases, software packages, and executable analysis workflows [ 256, 260].”

“Current multi-agent biomedical AI systems, such as OriGene [ 31 ], Robin [ 30 ], Biomni [ 32 ], MAC-doctor [ 33 ], PharmaSwarm [ 34 ], and AI Co-Scientist [ 35 ], are generally optimized for specific stages of the scientific workflow or confined to particular biomedical subdomains.”

“We do not extensively evaluate domains such as mathematics, physics, materials science, or engineering.”

“Unlike Biomni [33] and BioDiscovery Agent [34], which center tool-rich biomedical workflow execution, AIDO.”

“Emerging medical agent frameworks [14, 43, 44] suggest a direction in which multiple agents with complementary expertise jointly deliberate over complex cases.”

“Biomni [ 6] constructs a unified action space from biomedical tools and databases, enabling dynamic task orchestration including gene prioritization.”

“Additional features ( Table 1 ) include continuous debugging, free usage, parameter explanations, and example datasets, distinguishing PlotGDP from similar tools such as Biomni (18), OLAF (25), HiOmics (11), Hiplot (26), ImageGP 2.0 (27), SangerBox (10), SRplot (9), BioLadder 2.0 (12), and Chiplot (https://www.chiplot…”

“We evaluate S1-NexusAgent on the Biomni-Eval1 benchmark (Huang et al., 2025), a comprehensive suite designed to assess expert-level reasoning and execution capabilities of biomedical agents under realistic, end-to-end scientific workflows.”

“Recent controlled evaluations [30, 35, 32, 28, 36, 37] further demonstrate that appropriately designed MAS can achieve human-level performance in complex medical tasks, underscoring the viability of this approach for pharmaceutical research and development [38].”

“Conjoint analysis was performed to integrate the DNA methylation and RNA sequencing datasets (with matched columns) and identify overlapping genes with a high correlation across different comparison groups using the Biomni tool, as recently described [22].”

“While all these strategies have pushed the frontier of scientific AI [18], it remains unclear how much of the success in Sci-LLMs comes from genuine reasoning over raw sequences.”

“In the case of bioinformatics related tasks, a version of Biomni (A1) with access to 150 specialized biomedical tools, 105 software packages and 59 databases was able to analyze single-cell RNA-seq and ATAQ-seq datasets based on instructions provided in natural language [11].”

“%) lags behind Biomni (10.7%), but surpasses it once retrieval is applied (12.1% at pass@2, 14.1% at pass@3).”

“Language-model agents can assist by retrieving evidence, querying databases, and executing computational workflows [9, 10, 18].”

“…and similar agentic frameworks have primarily been developed and evaluated for action-driven tasks such as digital assistance and data analysis (Wang et al., 2025; Huang et al., 2025), and while prior work has noted code’s potential for enhancing reasoning (e.g., Chen et al., 2023; Gao et al.,…”

“The agent may then settle on the most salient, easily editable symptom while missing a deeper methodological problem (Trehan and Chopra, 2026; Li et al., 2025; Qiao et al., 2025; Zhang et al., 2025).”

“General-purpose biomedical agents now exist, and Biomni [44] lists GWAS causal-gene detection among the workflows on which it is evaluated.”

“All hypotheses were evaluated by integrating co-fractionation, crosslinking, AF prediction, and literature data surveyed using large language models 45,46 .”

“Across all scientific domains, language-model agents now formulate hypotheses, build pipelines, write code, and draft manuscripts [1–6].”

“This design spans molecular discovery, genomics, single-cell analysis, and laboratory tasks [3, 4, 5, 9, 6, 8, 7, 10, 13, 19, 11, 12, 14].”

“The organization of implementation, tool, benchmark and validation disclosures was informed by published biomedical-agent reporting [3, 4, 6, 9].”

“Among domain-specific agents, Co-Scientist [15] applies a multi-agent generate-debate-evolve loop to biomedical hypothesis generation, Biomni [16] equips agents with curated biomedical tool libraries for generalist scientific tasks, and CodeSci-entist [17] chains code-based experimentation into…”

“Such agents have been used to plan chemical syntheses [9], operate laboratory systems [8] and answer biomedical questions using external tools and databases [31].”

“…Slone et al. [2025] and the peptide-disjoint cluster splits of (4) CDR3 β positional stability and hotspots derived from (3), adapted from Biomni [Huang et al., 2025]; and (5) structure-based CDR3 β –pMHC interface descriptors (buried surface area computed with FreeSASA [Mitternacht, 2016],…”

“Co-Scientist develops hypotheses through multi-agent refinement [8], AI Scientist-v2 connects research ideation with agentic experimental search [33], and Biomni integrates biomedical tools and data [12].”

“Robin and the Virtual Lab coordinate scientific reasoning and validation[11, 12]; CellVoyager proposes and implements single-cell analyses[13]; Biomni connects biomedical tasks to tools and databases[14]; and BiOmics supports the integrated interpretation of multiple types of molecular data[15].”

“Discovery platforms such as Biomni and Kosmos coordinate analyses across many tools and have led to experimentally validated discoveries [Huang et al., 2025, Mitchener et al., 2025].”

“LLMs are increasingly used to navigate and reason over biomedical data (Huang et al. 2025; Jin et al. 2025; Luo et al. 2025; Aamer et al. 2025; Zhou et al. 2025; Lan et al. 2025; Ifargan et al. 2025; Bi et al. 2024; Lu et al. 2024; Wysocki et al. 2024).”

“Within bioinformatics and biomedical research, agents increasingly support workflow planning, tool selection, and multi-stage analysis [26]–[28].”

“…from interpreting experimental data and designing candidate molecules to planning and executing experiments, and benchmarks have been built to measure these capabilities [Boiko et al., 2023, Bran et al., 2024, Laurent et al., 2024, Huang et al., 2025, Swanson et al., 2025, Chen et al., 2025].”

“For example, TxGemma was obtained by fine-tuning the generalist Gemma model on therapeutic tasks (Wang et al. 2025) while Google AI co-scientist leverages a multi-agent system built on Gemini 2.0 to generate hypotheses (Gottweis et al. 2025).”

“Emerging work has demonstrated the potential of AI agents for biomedical data analysis [10], though these approaches have not yet been combined with comprehensive literature mining to utilize the value of existing biomedical omics publications.”

“Other recent frameworks [29, 16] investigate multi-agent architectures for automating bioinformatics workflows, tool orchestration, error recovery, and computational analysis pipelines.”

“Biomni unifies biomedical resources through a code interface (Huang et al., 2025), while STELLA adds evolving reasoning templates and autonomous tool creation (Jin et al., 2025).”

“Agents We additionally evaluated scientific agent frameworks Biomni [9] and our AMixKit toolkit.”

“Large language models can actuate scientific workflows via tool calling, as in Biomni and Claude for Life Sciences [10, 1].”

“Discovery agents, such as Google’s AI co-scientist, reason from the existing literature toward new hypotheses, and can propose ideas that later hold up at the bench(12–20).”

“General biomedical agents integrate scientific reasoning, tool use, code execution, and access to databases [8] in broad research domains [9, 1].”

“Data processing was facilitated using the Biomni Lab platform https://biomni.phylo.bio/, accessed on 1 July 2026 [32]. https://doi.org/10.3390/cells15171535”

“On EHRFlowBench, HealthFlow reached 4.01, compared with 3.76 for the strongest baseline, Biomni 22 .”

“Multi-agent architectures such as BioMaster [27], BioAgents [18] and Biomni [10] enable inter-agent coordination, planning and task execution to nominate hypotheses for validation.”

“To determine whether this operon structure is conserved more broadly, the BT32 and PV4 operons were analyzed through a BLAST-base synteny analysis using the biological agent Biomni 76 .”

“…recovering active compounds for acute myeloid leukemia [37]; FutureHouse’s Robin closed a literature-to-experiment loop over successive rounds to nominate ripasudil for dry age-related macular degeneration [38]; and Biomni [39] executed tasks autonomously across many biomedical domains.”