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.”

“It also relates to LLM-agent work in which models use tools to support scientific experimentation, and simulated clinical reasoning (Huang et al., 2025; Schmidgall et al., 2024).”

“Domain-specific systems constrain the planning action space to a narrow set of tools and workflows (Zou et al., 2025b; Huang et al., 2025; Darvish et al., 2025; Shao et al., 2025b; Zhou et al., 2025b), with feedback granularity ranging from pre-execution plan critique (Darvish et al., 2025) to…”

“BioMedAgent obtains 10.3% and 3.1% under the two correctness metrics, BixBench obtains 0.8% under each, and BioMaster and Biomni obtain no correct answers.”

“Recent work has extended LLM reasoning and agentic systems across scientific problem solving, chemistry, biomedicine, hypothesis generation, and increasingly autonomous research workflows (Ma et al., 2025; Guo et al., 2025; Gottweis et al., 2025; Swanson et al., 2025; Huang et al., 2025).”

“[6]) suggest that curated tools and unified execution interfaces can potentially turn capable models into accountable agents.”

“General-purpose agent frameworks automate analysis and tool use, but they usually do not make single-cell-specific decision points auditable or explicitly optimize for stable target convergence across runs [22–24].”

“The bio server ( palmetto-bio ) exposes discovery tools and domain tools such as genomics.* , validates arguments, and emits a deterministic RunSpec describing the intended execution; several genomics interfaces and implementations were adapted from Biomni [8].”

“Recent AI systems have demonstrated impressive capabilities for literature search and synthesis, hypothesis generation, coding, experimentation, and scientific writing [7, 21, 23, 26, 32, 37, 53, 63].”

“To identify potential TR4-binding sites within the human and mouse TREM2 promoters, we used Biomni AI 49 .”

“The integration of LLMs into bioinformatics has largely focused on Code Generation (e.g., BioGPT [23], StarCoder [22]) or Task Execution (e.g., autonomous agents like Biomni [16] or OpenBioLink [6] that execute specific workflows).”

“…built on this curated text — gene-representation models that embed it as features (Chen & Zou, 2025; 2024) and autonomous bio-AI agents that call it as a tool (Huang et al., 2025; Mitchener et al., 2025) — a more continuously updated and faithful representation of gene function is greatly needed.”

“We instantiate PLEAS on two production substrates — Biomni [7] and OpenClaude [8] — and demonstrate a preliminary portability probe on LangGraph [9].”

“Biomni provides a wide biomedical action space and code-based execution [11], which is valuable for general biomedical assistance but does not isolate routing across public, credential-free databases and artifact-contract completion.”

“Fine-tuning requires substantially less data and is increasingly supported by agentic biomedical platforms such as Biomni, which integrates LLM reasoning, retrieval-augmented planning, and code-based execution to automate biomedical research workflows 67 .”

“Cell2Sentence (C2S) [6] and Biomni [7] focus on biological data representation and general biomedical automation, but do not directly support user-facing microbiome interpretation workflows.”

“AI co-scientist tools were used to propose candidate analyses, assist with statistical implementation and figure generation, and support retrieval and integration of published literature for this dataset: OpenScientist (Roberts et al., 2026) and Biomni (Huang et al., 2025).”

“In biomedicine, OmniCellAgent [12] and Biomni [15] integrate domain databases and specialized analysis tools, but their validation remains field-specific and their designs are less suited to open-ended, end-to-end scientific studies.”

“Existing systems have explored agentic support for single-cell analysis [23, 24, 25] and broader biomedical research workflows [26, 27, 28].”

“…Tree Search over executable workflow graphs from execution feedback [Zhang et al., 2024], AgentSquare defines a modular space over planning, reasoning, tool use, and memory [Shang et al., 2024], and MaAS introduces an agentic supernet that samples query-dependent architectures [Zhang et al., 2025].”

“We evaluate how five frontier deep research systems handle dataset-construction requests: BioMNI-Phylo (a sandboxed Python coding agent) [18], Claude Research Mode (Opus 4.6 Extended), GPT-5.”

“Broader biological agent platforms include Virtual Lab [20], Biomni [21], BioLab [22], OriGene [23], and Medea [24], alongside systems for bioinformatics workflow automation [25, 26], self-verified gene-set analysis [27], and knowledge-graph construction [28].”

“Single-system agents such as Biomni [18], MedAgentBench [21], and AgentClinic [38] each ship one agent or one environment; LAB-Bench, LAB-Bench 2, and BixBench [13, 23, 28] similarly each Figure 2: BioMedArena casts evaluation as a configuration: the harness runs an agent —a backbone driving a…”

“They can retrieve scientific knowledge 12 , produce encyclopedia-like summaries 13 , propose novel scientific hypotheses 14,15 , assist in experimental data evaluation 16 , and even perform end-to-end scientific workflows 17 .”

“Biomni was selected as a validation agent given its state-of-the-art performance across a wide range of biomedical tasks by associated and independent benchmarks 12,17 .”

“Trace-level evaluations [2, 3], open-answer benchmarks, and longer analytical projects are the natural next tests.”

“In histopathology, foundation models trained on hematoxylin and eosin (H&E) stained slides have enabled representation learning [1, 2], biomarker discovery [3–5], and prediction of patient prognosis, treatment outcomes, and clinical markers [6–8].”

“Biomni advances a more integrated biomedical agent paradigm that combines reasoning, code execution, and a biomedical knowledge layer [35].”

“Off-the-shelf NLP tools typically underperform on specialized biological tasks without extensive customization (Wang et al. 2025).”

“Systems such as ChemLint 6 and Biomni 7 have shown that large - language - model - based agents can decompose multistep scientific tasks and invoke external tools in chemistry - related settings 8,9 , while recent reviews have highlighted the importance of domain - specific, tool - grounded, and…”

“The Biomni condition used Biomni Lab online, whereas the public Biomni paper and repository were used here only to describe the system at a high level [10, 11].”

“In computational biology, early LLM-based systems have demonstrated tool use and workflow assistance in domains such as genome analysis and transcriptomics Zhou et al. (2024); Huang et al. (2025); Su et al. (2025).”

“The functional profile for each region was done through a combined analysis of manual review, homology searches, and Biomni (Huang et al., 2025).”

“Current applications of LLMs in biomedical research typically focus on generalized AI assistants that may struggle with the specificity and reproducibility required for rigorous scientific workflows (Huang, et al., 2025), (Su, et al., 2024).”

“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.”

“…with particular fields: AlphaFold embeds protein physics [10, 11], au-50 tonomous laboratories encode reaction rules [12, 13], and biomedical platforms such as Biomni [14] orches-51 trate hundreds of specialized tools across 25 subfields to execute tasks from drug repurposing to molecular cloning.”

“Early AI co-scientist platforms 10,17 emphasize literature synthesis and work with experimental data in small samples supplied to them.”

“Systems such as MMedAgent (Li et al., 2024), Biomni (Huang et al., 2025), AgentMD (Jin et al., 2024), and MedAgent-Pro (Wang et al., 2025) demonstrate the potential of LLM agents to support biomedical reasoning through explicit planning, tool use, and workflow execution across a wide range of biomedical tasks (Rose et…”

“01155v1 Paired with the growing paradigm of agentic systems [13]– [15], where software agents orchestrate and query LLMs, this opens the door to human-interpretable, automatized causal reasoning in complex industrial pipelines [16], [17].”

“Prior studies have explored multi-agent architectures for biomedical discovery [5, 18, 9], autonomous coding agents [22], and debate-driven scientific reasoning systems [20, 11], demonstrating that agent collaboration can support hypothesis generation, planning, and code-level refinement.”

“Retrieval in science offers compelling use cases for accelerating scientific discovery [11, 13] and aiding researchers in coping with an exploding literature [1, 7].”

“By orchestrating tasks such as literature mining, experimental design, hypothesis generation, and multi-modal data interpretation, LLM-driven agents hold the potential to augment researchers’ capacity to navigate the vast and heterogeneous landscape of biological information [1, 2, 3, 4].”

““AI Agents for Science” are often large, complex systems—featuring hierarchical planning and large tool spaces—designed for high-level, open-ended discovery [12, 17, 29] or specialized scientific tasks [25, 36, 37, 47], rather than the targeted adaptation of existing tools.”

“…(RAG) 14,15 , thus overcoming two scaling bottlenecks: first, RAG replaces explicit tool specifications for individual analysis packages 12,15 and, therefore, scales directly with the size of the literature on published analysis code; second, RAG reliefs the LLMs used from requiring…”

“Data quality dominates architectural tweaks; attention becomes competitive with cleaner/larger datasets or pretraining [10].”

“Biomni [12] targets individual biomedical tasks using a monolithic agent that lacks end-to-end autonomy across stages of scientific discovery and requires detailed instructions (user prompts) for multi-step workflows.”

“A growing number of multi-agentic systems and general-purpose agentic frameworks have recently been proposed (e.g., BiOMNI [ 142 ]), and it is plausible that some of these could in principle be marshalled toward achieving similar scientific tasks.”

“[24] Both focus on the demonstration of their respective approaches in boosting performance, rather than critical areas of control, central to end users.”

“…Hyb SA GG (Dajani et al., 2025) DeepSeq SA Hyb SA GG (Nouri et al., 2025) Agentic AI SA Hyb SA GG (Martinek et al., 2025) Agentomics-ML SA Hyb A GG (Huang et al., 2025) Biomni SA Hyb SA GG (Su et al., 2025) BioMaster H Seq SA GG (Raghavan et al., 2025) CellAgentChat D Hyb A GG (Mehandru et al.,…”