Skip to content
CAAIL

General-Purpose Biomedical Agents

This page describes the General-Purpose Biomedical Agents row of the Papers.md matrix: LLM agents built to be broadly applicable across biomedical tasks without task-specific tuning, and that say so. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The distinction from Domain-Specific Biomedical Agents is breadth of intended use, and the taxonomy adds a condition that is easy to miss: the paper has to claim that breadth. This makes the row unusually checkable, because generality is not inferred from the architecture but read off the authors’ own framing, and most of these papers state it in the title or abstract. An agent aimed at one modelling problem, however flexibly built, belongs in the domain-specific row; one whose tools are chemistry tools belongs in Chemistry / Synthesis Agents; one whose substrate is the literature belongs in Scientific Literature & Discovery Agents.

Every reference sits in AI Tooling / Methodology, which follows from the definition: an agent that claims general applicability has by construction not committed to a cell-ag application. What a cell-ag team is choosing between here is a way of organizing tools, not a solution to a cell-ag problem.

AI Tooling / Methodology

  • #49 Biomni (Huang et al. 2025, bioRxiv, Stanford SNAP): maps the biomedical action space before building the agent, which is the part worth studying. An action-discovery agent parsed publications to extract three categories of actionable item, tasks, software and databases; roughly 1,900 recurring tasks were filtered to a curated set of 150 specialized tools, deliberately excluding anything an LLM could implement in a few lines, so a tool exists only where real domain expertise is encoded. Every tool had to satisfy a five-point checklist including a passing test case and outputs formatted as research logs an LLM can read. Databases with web APIs are reached through one unified natural-language query function that parses schemas at runtime; those without are downloaded and preprocessed into DataFrames. The agent itself is built on CodeAct, combining reasoning, retrieval-augmented planning and code execution with no predefined templates. Benchmarked across causal gene prioritization, drug repurposing, rare-disease diagnosis, microbiome analysis and molecular cloning without task-specific prompt tuning. Code at snap-stanford/Biomni; catalogued in Software.md.
  • #94 BRAD (Pickard et al. 2025, Bioinformatics): an agent whose stated motivation is reproducibility rather than capability. The paper’s objection to commercial AI systems is that they obscure data provenance and cannot be audited, so BRAD’s Agent class writes a JSON log entry for every interaction recording the prompt, the output, the exact LLM and settings, the chain of thought describing how the tool was used, and the sources referenced. Built on LangChain, runnable locally or against hosted models, with all intermediate steps written to an output directory. Demonstrated on a biomarker-discovery pipeline that runs gene enrichment over RNA-seq data and generates a report with verifiable references back into the user’s own documents. The provenance discipline is the transferable part for any lab that has to defend an analysis later. Code at Jpickard1/BRAD; catalogued in Software.md.
  • #95 OLAF (Riffle et al. 2025, arXiv): targets the accessibility problem directly, letting a non-programmer run single-cell RNA-seq workflows, gene annotation and visualization on real data (including .h5ad) through a web interface. Its contribution is an architectural pattern rather than a model: an agent-pipe-router split, where agents encapsulate one capability each, routers decide which agent handles a request and can re-invoke themselves to chain operations, and pipes are reusable asynchronous transformations for pre- and post-processing. An Angular front end sits over a Python and Firebase backend, containerized for local reproduction. The clean separation of routing from agent logic is what makes it testable, which most conversational-bioinformatics systems are not.
  • #96 STELLA (Jin et al. 2025, arXiv): argues that a manually curated toolset is the binding limitation on agents like the ones above, and answers it with two self-extending components: a Template Library of reasoning strategies that grows with experience, and a Tool Ocean expanded by a dedicated Tool Creation Agent that discovers and integrates new bioinformatics tools. Reported at roughly 26% on Humanity’s Last Exam: Biomedicine, 54% on LAB-Bench DBQA and 63% on LAB-Bench LitQA, ahead of leading models by up to 6 points, and evaluated against Biomni as the domain baseline. The claim the paper puts most weight on is not the headline score but its trajectory: accuracy on Humanity’s Last Exam almost doubles with more trials. Different LLMs are assigned to different internal roles, Claude 4 Sonnet for the Dev and Tool Creation agents, Gemini 2.5 Pro for the Manager and Critic. Note when reading the numbers that LAB-Bench evaluation used a 12.5% sample of the sub-benchmarks and HLE a 50-question subset, following Biomni’s protocol.
  • #98 BioMANIA (Dong et al. 2024, bioRxiv): turns any well-documented open-source Python tool into a conversational interface, rather than hand-writing tool wrappers. An abstract-syntax-tree parser extracts candidate APIs from source, each with its arguments, description and return type; APIs are then filtered by removing undocumented, non-executable and deprecated ones, cross-comparing the source-derived list against the documentation-derived list and keeping only the intersection, then testing what survives. It also includes a chit-chat detection stage, trained against the Topical-Chat and QnAmaker corpora, so casual conversation is not misread as a request to execute an analysis. The authors’ argument is that grounding in real tool APIs reduces hallucination relative to asking a general LLM to write the analysis from scratch. Demonstrated from single-cell omics to electronic health records.
  • #223 BioScientist Agent (Zhang et al. 2025, bioRxiv): three modules over a very large knowledge graph, RTX-KG2c v2.9.2, reduced after filtering to 6.38 million nodes across 37 categories and 40.99 million edges across 67 types, drawn from 70 public biomedical sources. A variational graph autoencoder with a relational GCN encoder learns node embeddings and predicts drug-disease links; an adversarial actor-critic reinforcement-learning module traverses the graph to recover mechanistic paths, bootstrapped from 1.9 million demonstration paths built from DrugBank and Molecular Data Provider drug-target interactions filtered by a PubMed normalized-Google-distance score below 0.6; and an LLM multi-agent layer orchestrates the two and writes causal reports. The gene-to-pathway-to-phenotype shape retargets naturally to cell-ag hypothesis generation, though the paper does not attempt it.
  • #237 BioHarness (Xiao et al. 2026, arXiv): addresses a failure that affects every retrieval-based agent above, which is that a fixed RAG workflow has no way to notice that the kind of evidence it retrieved is wrong. Its answer is substrate-aware cascade control: answer from reranked literature first, and escalate to curated knowledge bases or to REPL-style executable assembly over atlas-derived measurements only when the current evidence is uncertain, weakly grounded, or mismatched to the question’s real need (a gene alias to resolve, an identifier to normalize, a measurement to look up). Literature retrieval itself runs two views, the literal question and an LLM-written hypothetical PubMed-style abstract, merged, de-duplicated and reranked against the original question to undo hypothesis drift. Across 19,302 QA items spanning seven answer formats the pooled score rises from 65.9 to 71.0 over the strongest non-oracle baseline, and the ablations attribute the gain to repairing evidence-substrate mismatch rather than to more reasoning steps, more retrieved literature, or a bigger answer model. Code at coco11563/BioHarness.
  • #272 The Virtual Lab (Swanson et al. 2025, Nature, Zou group): organizes agents as a research group rather than a toolchain. A Principal Investigator agent, given a short project description from a human, writes the team into existence by specifying each scientist agent’s title, expertise, goal and role; a Scientific Critic agent can be added to any meeting to catch errors and push back. Work happens in structured team and individual meetings with an agenda, optional agenda questions and rules, summaries of prior meetings, and a set number of discussion rounds (typically three), with the human providing high-level feedback between them. Applied to SARS-CoV-2 nanobody design, it assembled a pipeline over ESM, AlphaFold-Multimer and Rosetta and designed 92 nanobodies, of which experimental validation found a range of functional binders and two with improved binding to the JN.1 or KP.3 variants while retaining binding to the ancestral spike. It is one of the few agent papers in the matrix with wet-lab-confirmed output. Code at zou-group/virtual-lab; catalogued in Software.md.

Reading the row as a set, the interesting axis is where each system puts its trust. Biomni and STELLA trust a curated or self-growing toolset; BRAD trusts a log; BioMANIA trusts the intersection of source and documentation; BioHarness trusts an explicit sufficiency check; the Virtual Lab trusts a critic agent and a human in the loop. None of them has been run on a cell-ag problem, so the choice of which discipline matters most is still open.

Adjacent methods

Further reading

Cite · Community · Privacy · Content licensed MIT · canonical source on GitHub · Get help

Linked external resources are independent of TUCCA and Tufts University and remain under their own licenses.