Source-linked AI summary
BioInsight: Multi-Agent Orchestration for Interactive Biomedical Knowledge Discovery
Jieyi Wang, Bingxuan Li, Nanyi Jiang, Desong Meng, Zirui Fan, Yuxin Guo, Jiayu Liu, Kunlun Zhu, Eddie Yang, Xiusi Chen, Pan Lu, Bingxin Zhao
TL;DR
Biomedical deep-research systems often collapse heterogeneous biomedical evidence into static text, making provenance difficult to inspect and reuse. BioInsight preserves evidence as structured, provenance-linked artifacts across multi-agent retrieval, reasoning, reporting, and visualization, and improves exact answering, protein-function reasoning, and traceable report synthesis across its evaluations.
Problem
Existing biomedical AI systems provide citation-grounded answers but lack a structured evidence state linking protein signals to pathways, interactions, publications, claims, and uncertainty for inspection.
Method
BioInsight uses typed artifact contracts and an Independent Search Agent to preserve shared provenance-linked evidence across retrieval, mechanistic reasoning, report generation, and interactive visualization.
Results
Across BioASQ, BioInsight-100, and expert evaluation, BioInsight improves exact answering, protein-function reasoning, and traceable report synthesis.
Takeaways & Limitations
Biomedical evidence synthesis can benefit from structured, auditable intermediate artifacts rather than fluent generation alone.
Takeaways & Limitations
BioInsight’s retrieved evidence may be incomplete or noisy, so its outputs require expert review and independent biomedical validation before downstream decisions.
Abstract
from arXiv · showhide
Biomedical deep-research systems increasingly retrieve and synthesize scientific evidence, but their outputs typically collapse heterogeneous evidence into static text, making provenance difficult to inspect and reuse. We formulate evidence-centered biomedical knowledge discovery, where disease-associated protein signals are transformed into a structured evidence state connecting proteins, pathways, publications, interactions, claims, and uncertainty. We introduce BioInsight, a provenance-preserving multi-agent orchestration framework built around typed artifact contracts and an independent Search Agent that decouples evidence acquisition from downstream mechanistic reasoning, supporting both the citation-grounded report and an interactive evidence workspace, without independently regenerating evidence for visualization. We evaluate BioInsight on standardized biomedical QA, challenging protein-function reasoning, and end-to-end biomedical evidence synthesis. The results demonstrate that BioInsight achieves better traceability and ranking performance than standard search-augmented baselines, and suggest that biomedical AI systems should move toward provenance-preserving, interactive evidence artifacts.
1 Introduction
BioInsight frames disease-centered protein interpretation as evidence-centered biomedical knowledge discovery, preserving links among protein signals, heterogeneous evidence, mechanisms, claims, and sources. Its provenance-preserving multi-agent architecture uses shared typed artifacts and independent evidence acquisition to support citation-grounded reports and interactive inspection.
- Motivation: Disease-associated protein signals require contextualization with pathways, interaction networks, literature, and drug–target evidence, while researchers need to inspect supporting evidence and competing interpretations.The introduction identifies protein ranking alone as insufficient for mechanistic interpretation and emphasizes links between conclusions, source proteins, and papers.
- Provenance layer: The framework maintains a persistent evidence layer with explicit schemas for proteins, pathways, publications, and relations, linking downstream claims to entities, analytical results, and source documents.Reasoning, writing, and visualization agents operate over shared evidence objects rather than regenerating evidence independently for each output.
- Problem formulation: BioInsight formulates disease-centered protein interpretation as an evidence-centered task that preserves lineage from observed protein signals to mechanisms, claims, and supporting sources.The task takes a disease name, disease-associated protein table, and optional cohort metadata, then retrieves, organizes, and synthesizes heterogeneous evidence.
- System architecture: BioInsight combines an Independent Search Agent, typed artifact contracts, a Reasoning Agent, a Writing Agent, and a Visualization Agent to acquire, synthesize, and render biomedical evidence.The system supports a citation-grounded report and an interactive evidence interface connecting hypotheses to proteins, pathways, publications, and interaction evidence.
- Evaluation: Across standardized biomedical QA, protein-function reasoning, and end-to-end report and dashboard evaluation, BioInsight achieves best or tied-best QA performance and stronger expert ratings for traceability, ranking quality, and dashboard usability.The introduction also reports the highest expert score on BioInsight-100.
2 Related Work
Prior work advances multi-step search, research writing, biomedical tool-using agents, and interactive scientific interfaces. BioInsight distinguishes itself by organizing heterogeneous biomedical evidence into typed, provenance-linked artifacts reusable for reasoning and presentation.
- Language agents and deep research: Recent language agents support multi-step search and long-form research writing, with Search-R1 and ASearcher interleaving reasoning with search.These systems connect search and reasoning within research workflows.
- Language agents and deep research: WebThinker, WebExplorer, Tongyi DeepResearch, and DR-Tulu emphasize well-attributed answers or reports over retrieved evidence.RAG and self-reflective retrieval evaluations likewise emphasize grounding and evidence use beyond final text quality.
- Biomedical agents: Biomedical LLM agents connect language models with literature, databases, tools, and code execution, while Biomni demonstrates broad task automation through planning.Biomedical retrieval systems improve access to PubMed and domain knowledge bases before generation.
- Biomedical agents: Unlike broader tool coverage alone, BioInsight adds an explicit evidence layer that keeps heterogeneous sources typed, provenance-linked, and reusable across reasoning and presentation.This design is summarized as the distinguishing architectural feature in the related-work comparison.
- Interactive scientific interfaces: Visual analytics, PaperVoyager, and generative UI systems enable scientific information inspection beyond static text, but biomedical visuals must remain tied to heterogeneous evidence.The cited biomedical evidence includes proteins and pathways.
3 Method
BioInsight orchestrates typed multi-agent stages that transform disease-associated protein signals into provenance-preserving evidence objects, reports, and interactive dashboards. The dashboard and report reuse the same structured evidence, while pathway planning, publication retrieval, reasoning, and visualization remain separately coordinated.
- Evidence-centered orchestration: BioInsight builds structured evidence objects from protein associations, pathway enrichment, literature retrieval, and interaction data, then generates both a citation-grounded narrative and dashboard.The report is one artifact in the process rather than the endpoint.
- Evidence-centered orchestration: Typed artifact contracts let the harness schedule stages, validate fields, and preserve provenance as agents reuse, revise, or expose upstream evidence.Search and planning construct ranked pathways and evidence packets; reasoning and writing refine them into notes and reports, while visualization renders the same evidence base.
- Pathway and publication evidence: Pathway planning maps proteins to enriched terms with g:Profiler, removes near-duplicate names using BioBERT similarity, and ranks candidates by enrichment strength and disease-specific literature support.Lower pathway scores are prioritized, favoring statistical support from the protein set alongside disease-relevant literature grounding.
- Pathway and publication evidence: The Search Agent retrieves pathway publications from PubMed and Semantic Scholar, scores them using lexical, semantic, citation, and journal signals, and retains validated evidence with Spub ≥0.25.Retained packets store the query, metadata, relevance score, PMID when available, and associated pathway.
- Reasoning and interactive interface: Reasoning agents produce structured pathway, protein, interaction, uncertainty, and citation notes, while visualization converts these artifacts into an explorable dashboard linking pathways, proteins, publications, and PPI edges.The dashboard supports inspection of association statistics, PMID-linked evidence, and graph filtering by edge type.
4 Experiments and Results
BioInsight is evaluated through grounding, cross-source reasoning, orchestration, and provenance questions using task-specific model and search controls. Across these evaluations, it improves exact biomedical answering, protein-centered evidence ranking, and traceable end-to-end synthesis.
- Evaluation design: The experiments test grounding, cross-source reasoning, orchestration, and provenance using task-specific comparisons with general-purpose, biomedical, open-weight, same-model, and search-enabled systems.RQ3 includes GPT-4o+Search versus BioInsight (GPT-4o), GPT-5.5 versus BioInsight (GPT-5.5), and BioInsight without the Independent Search Agent.
- Grounding: 51.10 vs. 40.49 for GPT-5.5 on List F-measure is the most consistent BioASQ signal, aligning with multi-entity retrieval and evidence aggregation.BioInsight maintains competitive exact-answer performance, although wide confidence intervals make point-estimate gaps uncertain.
- Cross-source reasoning: 8.62 is BioInsight’s highest mean expert score on BioInsight-100, with answers concentrated in the high-score range for protein-centered cross-evidence reasoning.The benchmark requires integrating functional annotations, PPIs, pathway context, disease mechanisms, and ranked explanations across heterogeneous sources.
- Orchestration and provenance: 0.11 vs. 0.27 for pathway omission shows BioInsight outperforming a strong GPT-5.5 typed-input control, while the control already makes unsupported entities and relations approximately zero.This indicates hallucination reduction alone is not BioInsight’s distinctive contribution.
- End-to-end synthesis: BioInsight receives the strongest expert ratings for traceable synthesis, ranking, and usability, while experts also find clear Markdown reports useful.The reported results across three evaluations combine stronger exact answering, cross-evidence reasoning, and more traceable and usable synthesis reports.
5 Case Study
The Alzheimer’s disease case demonstrates BioInsight as an evidence-centered, interactive interface that connects protein signals to pathways, literature, mechanisms, and hypotheses. By exposing enrichment and disease-specific publication support together, it makes interpretations auditable while preserving weaker evidence as exploratory.
- Case motivation: Alzheimer’s disease requires interpreting protein signals across synaptic dysfunction, lipid metabolism, glial activation, axonal injury, and neurodegeneration rather than one dominant pathway.This heterogeneity motivates an evidence-centered interface instead of a single-step report generator.
- Evidence orchestration: BioInsight orchestrates planning, disease-specific literature search, multimodal evidence reasoning, and visualization into an inspectable artifact chain.The chain connects protein statistics to pathways, publications, reasoning notes, and the dashboard.
- Evidence workspace: APOE links lipid receptor biology, vesicle organization, synaptic processes, and axon-related hypotheses, while GFAP, NEFL, SNAP25, and SYT1 support glial, axonal-injury, and presynaptic modules.SNAP25 and SYT1 specifically form a presynaptic vesicle and chemical synapse module.
- Uncertainty handling: Proteins with weaker pathway or literature support remain visible but are treated as exploratory rather than established interpretations.The workspace preserves these weaker signals without presenting them as definitive evidence.
- Auditability: Combining enrichment evidence with publication support distinguishes statistically enriched pathways from those with clearer Alzheimer’s disease relevance and creates an auditable path from proteins to follow-up hypotheses.The interactive artifact makes both evidence types visible for inspection.
6 Conclusion
BioInsight is a harness-centered multi-agent system for disease-centered protein interpretation that preserves evidence structure across biomedical discovery. Artifact contracts expose links among proteins, pathways, publications, and citations while supporting improvements in answering, protein-function reasoning, and traceable report synthesis.
- BioInsight is a harness-centered multi-agent system for disease-centered protein interpretation.
- Artifact contracts connect retrieval, reasoning, writing, and dashboard construction while exposing protein, pathway, publication, and citation links.
- Across BioASQ, BioInsight-100, and report-level expert evaluation, BioInsight improves exact answering, protein-function reasoning, and traceable report synthesis.
Limitations … A.2 Intermediate Artifacts and Reasoning Notes
BioInsight is intended for expert-supported biomedical interpretation and hypothesis generation rather than clinical decision-making. Its provenance-preserving artifact chain separates external resources, intermediate reasoning outputs, and inspectable evidence representations while explicitly exposing uncertainty and failure cases.
- Limitations: BioInsight does not support clinical diagnosis, treatment selection, or other clinical decisions, and generated mechanistic explanations may be overinterpreted as validated biology or therapeutic conclusions.Its outputs are grounded in retrieved publications and exposed through structured artifacts, but this does not validate the underlying mechanisms.
- Limitations: Citation links, protein-level statistics, intermediate artifacts, uncertainty notes, and failure handling enable inspection, but expert review and independent biomedical validation remain required.Outputs should be reviewed by domain experts before guiding experimental, translational, or clinical decisions.
- A Implementation and Artifact Details: The implementation appendix follows BioInsight’s artifact chain across external resources, typed intermediate artifacts, implementation parameters, and failure handling.This organization expands the harness details behind the method section.
- A.1 External Biomedical Knowledge Resources: Resource-specific modules assign defined roles to external biomedical knowledge, preventing heterogeneous evidence from being merged into an opaque retrieval result.The modular design preserves distinctions among evidence sources during ingestion.
- A.2 Intermediate Artifacts and Reasoning Notes: BioInsight stores pathway rankings, evidence packets, reasoning notes, citation-linked drafts, network views, dashboard schemas, and rendered dashboards as separate artifacts.Researchers can inspect enriched pathways, trace claims to proteins and publications, and verify that dashboards represent the report’s evidence.
- A.2 Intermediate Artifacts and Reasoning Notes: Reasoning notes record pathway interpretation, disease relevance, key proteins, PPI module explanations, uncertainty, and citations.The schema provides structured fields for connecting mechanistic interpretations to their supporting evidence.
- A.2 Intermediate Artifacts and Reasoning Notes: The Reasoning Agent uses a narrow, inspectable output channel that requires pathway function, disease relevance, driving proteins, supporting interaction modules, and evidence weakness.When relevant publications are unavailable, citation fields remain empty and the explanation is marked as exploratory.
A.3 Implementation Parameters
BioInsight exposes agent model choices as configuration parameters while keeping prompt templates fixed across diseases. Its implementation stores disease-specific artifacts and caches external biomedical fetches for reuse.
- Agent configuration: Model choices are configurable for the Planning, Search, Reasoning, Writing, and Visualization Agents, while prompt templates remain fixed across diseases.Disease specificity enters through the disease name, protein table, cohort metadata, pathway terms, and retrieved evidence packets.
- Artifact storage: Disease-specific result directories store ranked pathway tables, selected pathways, pathway overview drafts, report drafts, citation-formatted reports, dashboard schemas, and rendered dashboard pages.The cache organizes intermediate and final artifacts by disease.
- External data caching: External fetches are cached as JSON files so repeated runs can reuse STRING and DGIdb responses, subject to each resource’s licensing terms.The caching mechanism targets repeated access to external biomedical data.
A.4 Quality Control and Failure Handling … B.3 BioInsight-100
BioInsight is designed to expose evidence insufficiency through graceful degradation and artifact-preserving failure handling. Its experiments use matched inputs and fixed procedures to assess standardized biomedical answering and challenging protein-function reasoning.
- A.4 Quality Control and Failure Handling: The harness exposes evidence insufficiency through documented failure modes and graceful degradation rather than hiding deficiencies behind fluent text.Table 6 summarizes these failure modes and corresponding behaviors.
- A.4 Quality Control and Failure Handling: Absent significant pathway enrichment, the system reports that the proteins do not support pathway-level interpretation instead of hallucinating mechanisms.Enriched pathways with little or no PubMed support remain ranked but receive literature-weak or exploratory narrative labels.
- A.4 Quality Control and Failure Handling: Missing annotations, interaction edges, or translational records suppress unsupported protein, cluster, or therapeutic interpretations.The system treats proteins individually without interaction edges and does not infer therapeutic relevance from absent Open Targets or DGIdb records.
- A.4 Quality Control and Failure Handling: Failures at artifact boundaries preserve malformed reasoning outputs and pre-revision drafts so they can be inspected, retried, excluded, or parsed for damage.This applies to malformed JSON and coherence revisions that damage tables, links, or image syntax.
- B Experimental Setup Details: The appendix documents reproducibility details for the three experimental settings and surrounding experiments.It provides the setup context for the evaluation procedures reported in the paper.
- B.1 Baseline and Input Matching: All systems receive matched disease inputs and task instructions, while BioInsight holds prompts, retrieval, ranking, citation, and dashboard rules fixed across disease cases.Automatic metrics are computed from raw model outputs without manual correction.
- B.2 BioASQ Phase B: BioASQ Phase B Batch 1 evaluates yes/no accuracy, yes/no macro F1, factoid strict accuracy, factoid mean reciprocal rank, and list F-measure.The setting tests concise biomedical entity selection and evidence-supported answers before longer reports and dashboards.
- B.3 BioInsight-100: BioInsight-100 contains 100 expert-selected challenging protein-function questions requiring cross-evidence reasoning from UniProt function records and STRING interactions.Each answer is scored on a 0–10 scale for biomedical correctness, evidence use, specificity, pathway or interaction reasoning, and clarity; an example tests weak interaction context.
B.4 Disease Cases for Report Evaluation
End-to-end report and dashboard evaluation used five disease cases spanning neurodegenerative, psychiatric, cardiovascular, renal/metabolic, and autoimmune/inflammatory contexts.
- B.4 Disease Cases for Report Evaluation: Five disease cases were used for end-to-end report and dashboard evaluation, covering neurodegenerative, psychiatric, cardiovascular, renal/metabolic, and autoimmune/inflammatory contexts.These cases are summarized in Table 7.
C Human Evaluation Protocol · D The Use of Large Language Models (LLMs)
The human evaluation protocol compares anonymized biomedical reports and available evidence views across five disease–protein cases using four systems, four domain experts, and five scored dimensions. The appendix also states that large language models were used only to correct typos and streamline selected paragraphs.
- C Human Evaluation Protocol: Evaluators compare anonymized outputs from GPT-5.5 + Search, Gemini Deep Research, DR-Tulu-8B, and BioInsight across five disease–protein interpretation cases.Outputs include the generated report and, when available, associated dashboards or evidence views; presentation is randomized to reduce ordering and model-identity bias.
- C Human Evaluation Protocol: Four domain experts score each output from 1 to 5 and provide a brief justification for every evaluation dimension.The evaluation uses randomized output order, hidden system identities, identical forms, and a shared rubric to reduce bias.
- C Human Evaluation Protocol: The rubric covers comprehensiveness, biomedical validity, evidence grounding and traceability, prioritization and research depth, and readability/dashboard usability.All dimensions use a five-point Likert scale with dimension-specific criteria.
- C Human Evaluation Protocol: Comprehensiveness assesses end-to-end coverage linking proteins, pathways, mechanisms, disease relevance, interactions, therapeutic associations, citations, and dashboard evidence views.Reports are penalized when they cover only isolated proteins or pathways or omit important elements.
- C Human Evaluation Protocol: Biomedical validity and evidence traceability require scientifically plausible, appropriately qualified claims connected through an auditable chain to visible intermediate or source evidence.Evaluators examine biological statements, mechanisms, drug associations, network interpretations, protein statistics, pathways, literature, identifiers, and citation-linked claims.
- C Human Evaluation Protocol: Qualitative analysis categorizes evaluator justifications into recurring strengths and failure modes, including incomplete coverage, unsupported causal language, weak protein–pathway links, citation misalignment, and poor traceability.These findings are used to interpret the quantitative evaluation.
- C Human Evaluation Protocol: The protocol acknowledges limits from the number of disease cases and experts, judgment-dependent biomedical interpretation, and potential differences in weighing specificity, plausibility, and evidence strength.Blinding, common cases, a shared rubric, randomized order, and identical forms are used as bias controls.
- D The Use of Large Language Models (LLMs): Large language models are used to correct typos and streamline selected paragraphs, with prompts aimed at improving grammatical accuracy, logical structure, clarity, and readability.The appendix describes this use as writing refinement rather than evidence generation or analysis.