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.”

“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 .”

“AI agents have dramatically re-37 shaped fields such as software engineering 9 , law 10 , material science 11 and healthcare 12 by au-38 tomating repetitive tasks, enhancing productivity, and enabling breakthroughs that were previ-39 ously unimaginable.”

“Dataset 3, developed by Pasolli et al. 37 , combines eight human metagenomic studies with species-level features processed using MetaPhlAn2 37 .”

“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 .”

“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.”

“The pertur-443 bation screen design benchmark was curated from Schmidt et al. 34 .”

“The drug repurposing benchmark used a dataset from Huang et al. 40 , for the task of identifying the most likely drug from a pre-defined list of drugs for repurposing in a given indication.”

“Previous approaches have largely relied on specialist agentic workflows tailored to nar-47 row biomedical tasks 13–19 , which restricts their capacity to move fluidly and generalize across the full spectrum of biomedical domains, as needed to answer key research questions.”

“Previous approaches have largely relied on specialist agentic workflows tailored to nar-47 row biomedical tasks 13–19 , which restricts their capacity to move fluidly and generalize across the full spectrum of biomedical domains, as needed to answer key research questions.”

“…regulatory network (GRN) analysis (Figure 3h), Biomni re-capitulated known regulatory relationships between key osteogenic transcription factors such as RUNX2 and HHIP, confirming how they are regulated by a shared set of anti-osteogenic transcription factors including TWIST1, LMX1B, and ALX4 27 .”

“Although PBX1 is a well-established skele-266 tal regulator 28 and ZFHX3/AUTS2 have only limited or indirect skeletal reports (in mouse 29 or zebrafish 30 ), their broad activity here suggests under-appreciated roles across diverse skele-268 tal lineages.”

“We evaluated Biomni on three challenging multiple-choice benchmarks of general biomedical knowledge and reasoning: Humanity’s Last Exam (HLE) 23 and LAB-Bench 24 , which includes two key subtasks – DbQA (Database Question Answering) and SeqQA (Sequence Question Answering) (Figure 2a).”

“A vast volume of valuable biomedical data sits underutilized 7 , many so-30 phisticated analyses are not conducted, and many connections for past knowledge and literature are not made, not for lack of significance, but because the demand for expert researchers far exceeds the supply.”

“We evaluated Biomni on three challenging multiple-choice benchmarks of general biomedical knowledge and reasoning: Humanity’s Last Exam (HLE) 23 and LAB-Bench 24 , which includes two key subtasks – DbQA (Database Question Answering) and SeqQA (Sequence Question Answering) (Figure 2a).”

“Development and testing sets were created by sampling the LAB-Bench Database Question-Answering and Sequence Question-Answering benchmarks 24 .”

“Although PBX1 is a well-established skele-266 tal regulator 28 and ZFHX3/AUTS2 have only limited or indirect skeletal reports (in mouse 29 or zebrafish 30 ), their broad activity here suggests under-appreciated roles across diverse skele-268 tal lineages.”

“The GWAS causal gene detection benchmark utilized a dataset curated from Shringarpure et al 33 , using the original prompt: ”Your task is to identify likely causal genes within a locus for a given GWAS phenotype.”

“Dataset 1 comes from the MGM 2.0 platform 35 and includes relative microbial abundance across samples and another with sample labels, ideal for classification tasks 35 .”

“…have achieved remarkable advances across diverse domains, including code generation (e.g., Claude Code [1]), scientific discovery (e.g., Biomni [2], STELLA [3], and SciMaster [4]), and in-depth research (e.g., Tongyi, OpenAI, and Gemini DeepResearch [5–7]), showcasing their immense potential…”

“Significant evidence shows interest in post-training for agents [6, 87–93], yet Section 5.1 shows fine-tuning and reinforcement learning remain uncommon in deployment despite potential benefits.”

“More general systems such as Biomni (Huang et al., 2025) aim to support a broader range of biomedical tasks through larger tool ecosystems.”

“experimental protocol design and bioinformatic analyses 146 .”

“Biomni is an AI agent for diverse biological tasks [12].”

“OriGene continued its strong performance on the TRQA-lit (short answer) benchmark, designed to assess the generation of concise and accurate literature-derived responses, scoring 82.76%, which again surpassed all other models, including DeepSeek-V3 and GPT-4o.”

“Frameworks like Biomni [295] and SciToolAgent [276] establish comprehensive environments to integrate specialized tools and databases, enabling agents to autonomously execute diverse tasks ranging from gene prioritization to protein engineering.”

“Community-developed agentic AI systems 101,144,145,153,163 are led by academic institutes, non-profit organizations or consortia and are characterized by a focus on transparency, open access, ethical alignment and strict regulatory compliance.”

“…in single-cell analysis spans three trajectories: (i) targeted automation of labor-intensive, subjective steps (e.g., cell type annotation 71,72 ), (ii) end-to-end pipelines assembled from automated modules 73 , and (iii) most recently, broad-scope, LLM-enabled code-generation work fl ows 24 .”

“In biological research domain, Biomni [187] advances beyond task-level assistance by supporting complex biomedical research workflows, enabling agents to automatically identify, compose, and execute multi-step experimental pipelines.”

“LLM agents have been applied to autonomous chemical research (Boiko et al., 2023), biological data analysis with protocol generation (Huang et al., 2025), and AI research (Kon et al., 2025).”

“Here, STELLA is highly competitive with Biomni, outperforming it on most metrics.”

“…analyses (dimensionality reduction, clustering, DE, enrichment), producing ready-to-use figures and reports with minimal user intervention. scBaseCount [Youngblut et al., 2025] uses an agentic, hierarchical workflow to continuously mine, standardize, and expand a single-cell compendium now…”

“Agents have wide applications, and have delivered promising early results in diverse fields including drug discovery [14, 52], cell biology and genomics [61, 19], and hypothesis verification [18].”

“We analyzed 32,014 user queries collected from the open-source Biomni platform [Huang et al., 2025].”

“In biology, agents have been applied to automate gene editing [18], omics analysis [19, 20], nanobody design [21], and biological data analysis [22].”

“Multi-agent architectures such as Biomni [37], CellAgent [10], and Co-Scientist [97] emphasize collaborative reasoning and distributed task decomposition, while neuro-symbolic models including AutoMAxO [52],Tokensome[99],andmergen[101]embedsymbolicconstraintsor ontology structures within…”

“Biomni [40] combines planning, retrieval, code execution, and multi-omics tool-use within a single modular agent capable of gene prioritization, variant interpretation, and enrichment analysis.”

“Experimental designs generated by these AI Scientist systems often lack scientific rigor, innovation, and practicality, making it difficult to meet the demands of high-level research (Huang et al., 2025a, Weng et al., 2025).”

“1-8B-Instruct ) (Wu et al., 2025a), Baichuan-M1-14B-Instruct (Wang et al., 2025a), and Baichuan-M2-32B (Dou et al., 2025).”

“— Incorporate Knowledge with AI Agents : A paradigm-shifting future direction is the deep integration of GDL with Biomedical AI agents like Stella [98], Biomni [83], BioMiner [208], and BioLab [97].”

“This intermediate step avoids overwhelming the planner with the extensive set of tools available in Biomni-E1.”

“However, alternate approaches include pre-installing a comprehensive range of tools and datasets 8 , or providing “skill” files to the agents that outline high-level analytical strategies 25,26 .”

“Traditional AI models often struggle with this heterogeneity, but agentic AI can navigate complexity by retrieving literature, linking datasets, and generating adaptive insights [1, 7].”

“Additional LLM-based efforts, such as TextGrad [48], STELLA [25], and BIOMNI [22], have also been used in biomedical and chemistry applications; however, none of these methods are equipped for retrosynthesis planning tasks.”

“Furthermore, SP-Mind outperforms the strongest specialized baseline, Biomni-SP (55.4%), by a substantial margin.”

“In contrast, we run A UTO S CIENTISTS , Autoresearch, and rerun Biomni [10] under a unified experimental-compute setting of 4 hours on 1 H100 GPU with 16 CPUs and 48 GB memory for all tasks except for biomedical imaging which had a wall-clock budget of 16 hours instead.”

“…in areas such as drug repurposing, causal gene detection, single-cell annotation and disease diagnosis, but, while it excels at tasks such as database querying and sequence analysis, it is less effective for more complicated processes that require deeper biological reasoning and thought 148 .”

“These were categorised (Miller et al., 2025) as ‘generalist’ ML coding agents (AIDE (Jiang et al., 2025), MLAgentBench (Huang et al., 2024)), or more specialised ‘biomedical’ ML coding agents (STELLA (Jin et al., 2025), Biomni (Huang et al., 2025)).”

“Recent systems like Biomni and OriGene exemplify this evolution 40,41 .”

“This indicates that current AI scientists have largely mastered the procedural aspects of standard single-cell analysis workflows.”

“These technologies offer a powerful solution to speed up the extraction and structuring of complex information from scientific texts, a capability that has been successfully explored in fields like materials science [ 20 ] and genomics [ 21 ].”

“Biomedical agents extend the approach across literature, databases, and analysis workflows (e.g., Biomni (Huang et al., 2025)).”

“More complex agentic AI systems like Biomni 106 have demonstrated the ability to perform biomedical research tasks ranging from generating wet-lab cloning protocols to analyzing heterogeneous datasets for rare disease diagnosis.”

“These augmented systems, referred to hereafter as “AI agents”, include OpenAI’s ChatGPT Agent, Biomni (Huang et al., 2025), CRISPR-GPT (Qu et al., 2026), STELLA (Jin et al., 2025), and BioDiscoveryAgent (Roohani et al., 2025).”

“Biomni [103] belongs here because it couples biomedical reasoning to external tools and scientific resources, and MM-Agent [104] is relevant because it emphasizes scientific tool integration across multimodal settings.”

“Biomni [27] couples an agentic planner with a curated execution environment.”

“In contrast, emerging AI-first platforms—such as Biomni [4], FutureHouse [5], Kepler AI [6], and Potato AI [7]— offer flexible, intelligent systems that broaden access to data-driven discovery.”

“…relies on automated pipelines and artificial intelligence (AI)-driven systems, including emerging autonomous scientific agents such as Biomni [6] and Edison Scientific’s Kosmos [7], and the underlying large language models (LLMs) that enable them, including GPT and Claude, to construct…”

“General-purpose biomedical agents such as Biomni [31] demonstrate autonomous problem-solving across domains.”

“As a result, an analysis can run successfully but still be wrong, for example, when differential expression uses mismatched covariates that are propagated into downstream pathway ranking [22, 23].”

“Medical agent systems [24]–[26] have emerged as a promising paradigm to overcome the limitations of monolithic models by orchestrating multiple specialized components.”

“We additionally evaluate the performance of C2S--Scale [Rizvi et al., 2026], a language model fine-tuned on single-cell tasks, and Biomni [Huang et al., 2025], a biomedical agent with access to biology-relevant tools.”

“Biomni’s performance pattern provides a striking corroboration: it achieved 74% on scriptable property filtering but fell to 24.3% on binding affinity (below chance) and produced zero valid docking outputs [29], failing precisely on tasks requiring structured workflow orchestration.”

“In biomedical research, the Biomni [17] agent goes further: it uses an LLM to scan thousands of research papers and build a unified action space of hundreds of tools and databases, enabling dynamic composition of complex workflows.”

“In this domain, Biomni [44] addresses this via a unified space of 150+ verified tools and code generation for complex logic.”

“Although significant progress has been made in developing such agents for both general research contexts [3, 4] and specialized fields such as mathematics [5], chemistry [6], and life sciences [7], the combustion research community notably still lacks a practical and accessible AI agent system.”

“Existing agents may learn to refine a specific tool’s usage (Qiu et al. 2025; Jin et al. 2025) or improve a reasoning template for a sub-problem (Zhang et al. 2025b), but they operate within a fixed cognitive architecture.”

“With external tools, an LLM can retrieve up-to-date information, perform calculations, and execute code (Huang et al., 2025a).”

“Since this study, newer biology-specific models like Biomni Lab leverage their specialized training on genomic and protocol data to significantly reduce errors in sequence analysis, protocol generation, and biomedical data analysis (Huang et al. 2025).”

“We then compared the performance of our system against Biomni [23], a prominent biology-focused agent, and Gemini#net, a state-of-the-art generalist model with web access based on Gemini 2.5 Pro [24].”

“These new paradigms have already achieved great successes in automating complex scientific workflows across a range of fields, such as biomedicine [29], chemistry [30], and materials science [31].”

“Large language model (LLM) coding agents can write analysis code, fit statistical models, interpret results, and draft reports, and are increasingly being incorporated into scientific work-flows [10, 16, 21, 33, 41].”

“Emerging directions include chat-based interfaces for natural-language querying 167,168 , automated analysis pipe-lines 169 and agentic AI systems for experiment design and execution 170 .”

“Three gaps remain: zero­shot prediction, reaching more complex phenotypes, and pace of training / prediction for integration of these models as grounding data in agentic reasoning [8], [9].”

“[3, 5, 8, 19, 21, 41] Another important feature is the tool-calling capability of the latest LLM models offered by OpenAI, Anthropic, Google Gemini, and Mistral.”

“To overcome these limitations, we explore the potential of large language models (LLMs), which have recently demonstrated strong reasoning, flexible communication, and generalization capabilities across diverse domains [16, 23, 31].”

“The second paradigm consists of agents that develop a full solution pipeline from scratch: two ML-automation agents , MLEvolve [16] and AIRA-Dojo [12]; one general-purpose coding agent , Claude Code [38]; three scientific-discovery agents , Biomni [2] 2 , STELLA [3], and Agentomics [36].”

“Alongside these efforts, several general-purpose biomedical agents have also been developed [16, 17].”

“Beyond theorem proving, many recent efforts use AI to assist scientific research, with most systems targeting applied areas such as biomedicine [15], the natural sciences [48], and machine learning [25].”

“Most recently, CRISPR-GPT [42] was proposed for automatic gene-editing experiments, and Biomni [43], a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields.”

“Biomni runs on a CPU and can take input files up to a limit, has a runtime execution of 1 hour per iteration.”

“In the biomedical domain, AI agents like Biomni (Huang et al. 2025b) offer generalist capabilities from literature curation to experimental design.”

“This limitation has been recognized [8], and specialized software that leverages retrieval-augmented generation (RAG) [9], custom-developed tool suites [10, 11], or LLMs fine-tuned on biomedical data [12] to integrate previous knowledge has been developed.”

“While this may suggest that a better domain understanding in the underlying LLM enhances the tool orchestrating capabilities in that domain, a counterexample is the general-purpose biomedical AI agent Biomini for which this was not observed [13].”

“…from automated paper writing (Lu et al., 2024; Schmidgall et al., 2025) to evolutionary program search (Romera-Paredes et al., 2024; Novikov et al., 2025; Lange et al., 2025) and autonomous experimentation (Boiko et al., 2023; Swanson et al., 2025; Gottweis et al., 2025; Huang et al., 2025).”

“In broader computational science, tool-oriented agents have been developed to select and execute specialized scientific tools across domains (Huang et al., 2025).”

“General-purpose agents now demonstrate end-to-end research loops that integrate reasoning with tool use, encompassing literature survey, hypothesis generation, verification, implementation, and reporting [46].”

“…aiming to automate scientific workflows (Boiko et al., 2023; Lu et al., 2024; Mitchener et al., 2025), and frameworks iteratively searching and refining scientific hypotheses (Wang et al., 2025b; Shojaee et al., 2025a; Wang et al., 2025a; Lu et al., 2025; Gan et al., 2025; Du et al., 2025).”

“By integrating natural language reasoning with dynamic tool invocation, these systems can orchestrate specialized bioinformatic pipelines, query curated biological databases, and synthesize rapidly evolving scientific literature within a unified computational framework 1–6 .”

“In the era of agentic AI, an intuitive solution to the above problems seems to be integrating the LLMs with retrieval tools (e.g., ToolUniverse [10] and Biomni [13]), which can retrieve additional relevant guideline information (e.g., from PubMed and PMC) to the context.”

“A side-by-side comparison with seven prior biomedical and spatial analysis agents [26], [27], [28], [29], [30], [31], [32] is given in Supplementary Tables 1 and 2.”

“Among these, Biomni is presented as a general-purpose biomedical AI agent, whereas K-Dense Analyst is presented as a hierarchical multi-agent system for autonomous bioinformatics analysis [1, 2].”

“Recent agentic biomedical research platforms such as Biomni 39 , DeepEvidence 40 , and Alvessa 41 have demonstrated the feasibility of coordinated tool-based reasoning, but their evaluations primarily emphasize black-box task-level accuracy ( Append Table 1 ).”

“Existing systems either overlook these requirements or rely on users to interact directly with low-level code and data schemas 65 .”

“…and machine learning communities have developed benchmarks (Young-blut et al., 2025; Zhang et al., 2025a; Wu et al., 2025b; Levine et al., 2024; Wu et al., 2025a) and methods (Roohani et al., 2024; Lu et al., 2025; Adduri et al., 2025; Wenkel Work conducted during HITS internship by H. Kim and Y.”

“BiOmics was compared against LLMs (GPT-4o)[68], biological agents (Biomni)[10], and specialized bioinformatic toolkits.”

“More and more attempts have been made to build AI scientist agents in multiple fields including biology[6, 7, 8], chemistry[9, 10] and more generic tasks[11, 12, 13, 14].”