Domain-Specific Biomedical Agents

This page describes the Domain-Specific Biomedical Agents row of the Papers.md matrix: LLM agents purpose-built for one specific biomedical task or domain, such as RNA-seq analysis, spatial biology, gene editing, or a single named assay or modality. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The distinguishing feature is narrow, dedicated focus, as against General-Purpose Biomedical Agents, which claim breadth and say so. The test is the paper’s own framing rather than the architecture, since both rows build from the same components: a planner, a tool set, a memory, and an execution loop. What differs is whether the tools were chosen for one job.

This is the largest of the agent rows, which is itself informative: it is easier to build an agent that does one biological thing well than one that does everything, and the field has largely gone that way. It is also the row where cell-ag-adjacent work concentrates, because a domain-specific agent for metabolic engineering, cell-image segmentation or single-cell data curation is doing a task a cultivated-meat programme actually has.

Cellular Engineering

#125 BioDiscoveryAgent (Roohani et al. 2025, arXiv, Leskovec lab): designs genetic perturbation experiments, finding the small subset of genes whose perturbation produces a target phenotype. Its distinguishing claim is that it uses the LLM’s biological background knowledge in place of a trained acquisition function, so no surrogate model needs fitting first. The evaluation is deliberately constrained to be realistic: 128 genes perturbed per round and only five rounds, because each round costs money and introduces experimental variation, scored on hit ratio against six datasets covering over 18,000 single-gene knockdowns across different cell types, sites and publication dates, nearly all released after 2021 to limit what the model could have memorized. Code at snap-stanford/BioDiscoveryAgent; catalogued in Software.md.
#97 PerTurboAgent (Hao et al. 2025, bioRxiv, Regev lab): a self-planning agent for sequential Perturb-seq design with an action memory that lets a multi-round plan adapt mid-round. Described in Active Learning.
#93 CellForge (Tang et al. 2026, arXiv, Gerstein lab): designs the neural architecture for a perturbation-prediction task rather than running the experiment. Described in Agent Infrastructure, where its framework contribution is placed.

AI Tooling / Methodology

Bioinformatics and omics analysis

#277 AutoBA (Zhou et al. 2024, Advanced Science): a fully autonomous agent for multi-omic analysis that reduces the user’s job to three YAML fields, the data path, a plain-language data description, and the analysis goal. Everything downstream, selecting tools, configuring the environment, installing software, writing and debugging code, is the agent’s. It covers WGS, WES, RNA-seq, single-cell RNA-seq, ChIP-seq, ATAC-seq and spatial transcriptomics, and self-designs the analysis process from the input data rather than filling a fixed template. Code at JoshuaChou2018/AutoBA.
#43 (Lee et al. 2025, arXiv): an NGS downstream-analysis agent aimed explicitly at researchers without a biological or computational background, which is the gap it names in prior RNA-seq agents: they automate the workflow but not the interpretation. A Streamlit interface takes an expression matrix and clinical table, normalizes, relates clinical variables to the user’s stated hypothesis by correlation or t-test, runs differential expression, then supplies literature-backed interpretation and recommends further analyses. The interpretation-and-recommendation half is the contribution; the statistics are ordinary.
#51 Medea (Sui et al. 2026, bioRxiv, Zitnik lab): an omics agent built around transparency as an engineering requirement rather than a virtue. Four modules: research planning with context and integrity verification, code execution with pre- and post-run checks, literature reasoning that assesses evidence strength, and a consensus stage reconciling evidence across datasets, tools and literature. Its tool space is 20 tools from 17 sources, and the design principle is that intermediate decisions must be explicit and each one validated against data and tool constraints as the analysis unfolds, which is what a long-horizon analysis needs if anyone is to trust its conclusion. Code at mims-harvard/Medea; dataset on Hugging Face.
#56 SpatialAgent (Wang et al. 2025, Genentech): an autonomous agent for spatial biology spanning experimental design, multimodal analysis and hypothesis generation. Three components, a memory holding both semantic and episodic information, a planner doing chain-of-thought task decomposition, and an action module over a curated toolbox, run in a self-governing loop that records outcomes and revises on execution failure, which is what lets it handle new tissue types without a predefined workflow. Tested on two million cells across human brain, heart and a mouse colitis model. It also offers a co-pilot mode where the agent’s memory is updated from user input each turn. Code at Genentech/SpatialAgent.
#126 scBaseCount (Youngblut et al. 2025, Arc Institute): an agent whose product is a database. SRAgent mines all 10x Genomics datasets from the Sequence Read Archive, extracts metadata, and standardizes processing, yielding over 502 million cells across 27 organisms, self-updating as new data arrives. The engineering detail is instructive for anyone building a curation agent: it runs on Cloud Run at 3 to 5 datasets per job triggered every 1 to 5 minutes to respect NCBI rate limits, roughly 300 datasets an hour at 80 seconds each, and resolves tissue and disease terms by semantic search over a vector database of the Uberon and MONDO ontologies, then walks neighbouring ontology terms and queries the Ontology Lookup Service to validate. Unstandardized metadata is what makes the SRA underused, and the agent is aimed at exactly that. Code at ArcInstitute/scRecounter.
#90 GenCellAgent (Yu et al. 2026, arXiv): training-free cellular image segmentation, orchestrating specialist segmenters and generalist vision-language models through a planner-executor-evaluator loop with long-term memory. It routes an image to the best available tool, adapts on the fly from a few reference images when imaging conditions differ from what a tool expects, and supports text-guided segmentation of organelles no existing model covers. That last capability is the useful one for cell-ag phenotyping, where the structure of interest is often not what the off-the-shelf model was trained to find. Segmentation is otherwise a CNN problem; what puts this here is that the agent selects and adapts rather than segments.

Metabolic and systems modelling

#50 Talk2Biomodels (Wehling et al. 2025, BMC Bioinformatics): natural-language interaction with SBML kinetic models, built on LangGraph as a stateful computational graph where nodes are steps and a global state propagates through execution. A ReAct agent selects among tools annotated with their capabilities and input/output schemas, and reaches BioModels, the Ontology Lookup Service and UniProt. The problem it solves is access: kinetic models of regulatory processes are valuable and simulating them requires specialist expertise. Part of the AIAgents4Pharma family; code at VirtualPatientEngine/AIAgents4Pharma.
#167 (Singh et al. 2025, ICLR MLGenX workshop): the companion paper pairing Talk2Biomodels with Talk2KnowledgeGraphs under a supervisor agent that routes user input to whichever specialist applies. T2B covers model search, simulation, plotting, parameter scanning, steady-state stability analysis, annotation extraction and retrieval over uploaded papers; T2KG does subgraph extraction, summarization and GraphRAG reasoning. The stated aim, replacing a rigid GUI workflow with something adaptive, is the recurring argument of this whole row.
#69 KinModGPT (Maeda & Kurata 2023, IJMS): generates SBML kinetic models directly from natural-language descriptions of biochemical reactions, using GPT as the language interpreter and Tellurium as the SBML generator via the human-readable Antimony format. Temperature is set to 0 for reproducibility, a small detail worth copying in any agent whose output is a model someone will simulate. Three GPT variants were compared. Code at kmaeda16/KinModGPT.
#68 D2Cell (Li et al. 2024, bioRxiv): mines metabolic-engineering strategies from the literature into a 29,006-entry database feeding a hybrid target predictor. Described in GNN.

Therapeutics, protein and gene editing

#40 TxAgent (Gao et al. 2025, arXiv, Zitnik lab): therapeutic reasoning across 211 tools, analysing drug interactions at molecular, pharmacokinetic and clinical levels, identifying contraindications from comorbidities and concurrent medications, and tailoring recommendations to patient characteristics. The design commitment is white-box reasoning: every answer carries the thought processes, function calls and tool-derived evidence behind it, so a user can verify the basis rather than accept the conclusion. It runs on the ToolUniverse described in the infrastructure row. Code at mims-harvard/TxAgent.
#276 CRISPR-GPT (Qu et al. 2025, Nature Biomedical Engineering): automates CRISPR gene-editing design and data analysis through four components, an LLM planner, tool providers, task executors and a user-proxy agent. The executors are state machines, with 22 implemented tasks each decomposed into subgoals whose states issue instructions and guide the user through decisions over several rounds of text. That is a meaningful architectural choice: where most agents in this row let the LLM control flow, this one constrains it to a state machine and uses the LLM within each state, which is what you want when a wrong step costs a cloning run.
#275 ProtAgents (Ghafarollahi & Buehler 2024, Digital Discovery, MIT): de novo protein design by multi-agent collaboration in AutoGen, with user-proxy, Assistant, Planner and Critic agents under a group-chat manager, each given its role by a profile in its system message. Its argument against the prevailing approach is worth noting: surrogate models linking protein structure to material properties are end-to-end and therefore inflexible when out-of-domain knowledge or broader analysis is needed. Tools are plain Python functions registered with the assistant, including a purpose-built ProteinForceGPT predicting unfolding force-extension behaviour from sequence. Relevant to cell-ag for scaffold proteins and recombinant growth factors, where mechanical properties are the design target. Code at lamm-mit/ProtAgents.
#54 ClockBase Agent (Ying et al. 2025): mines decades of public molecular data for aging-intervention effects that were never analysed for them, reanalysing millions of human and mouse methylation and RNA-seq samples against over 40 aging clocks. The data engineering is the substance: all public GEO methylation data deposited before January 2025, RNA-seq from ARCHS4, soft filtering to genes with at least 10 reads in at least 20% of samples, RLE normalization, and per-dataset relative expression computed by subtracting the median control profile within the same dataset and tissue to suppress batch and tissue baseline effects. Specialized agents then generate aging-focused hypotheses and evaluate interventions on top. The pattern, agentic reanalysis of an existing public corpus for a question it was not collected to answer, transfers directly to cell-ag.

Adjacent methods

General-Purpose Biomedical Agents: the breadth-claiming counterpart, and where several of these agents are benchmarked against Biomni.
Agent Infrastructure (Frameworks, KGs, Protocols): ToolUniverse, LangGraph and the knowledge graphs these agents are assembled on.
Scientific Literature & Discovery Agents: the retrieval components several of these embed.
Active Learning and Bayesian Optimization: the classical experiment-selection machinery that #125 and #97 replace with an LLM’s prior.
CNN: the segmentation models GenCellAgent routes between rather than replaces.