Skip to content

BioInsight: Multi-Agent Orchestration for Interactive Biomedical Knowledge Discovery

🕒 Published (v1): 2026-06-19 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

BioInsight is a four-agent orchestration system (Search, Reasoning, Writing, Visualization) that converts disease-associated protein tables into provenance-preserving, interactive biomedical dashboards rather than static text reports. It decomposes evidence retrieval from mechanistic reasoning via typed intermediate artifacts, keeping citations and uncertainty traceable through both the narrative report and the rendered interface. On BioASQ, BioInsight-1k, and five end-to-end disease cases it outperforms frontier models including GPT-5.5+Search and Gemini Deep Research on traceability, ranking quality, and dashboard usability.

Problem

Existing biomedical AI systems—search-augmented LLMs, deep-research agents—produce static text reports or structured tables that flatten the evidence chain. Biomedical researchers need to interrogate provenance (which proteins drive a pathway, which publications support a claim, where evidence is uncertain), compare alternative mechanisms, and refine hypotheses dynamically—tasks a single fixed narrative cannot support.

Method

BioInsight is organized as a harness-coordinated multi-agent pipeline over three layers:

  1. Evidence layer — A Planning Agent maps proteins to enriched biological terms via g:Profiler, de-duplicates near-duplicate pathway names with BioBERT embedding similarity, and ranks pathways by a composite score \(S_{\text{path}}(t) = 0.4\cdot P_{\text{norm}}(t) + 0.6\cdot(1 - L_{\text{norm}}(t))\) (lower = higher priority). A Search Agent retrieves publications from PubMed and Semantic Scholar, scoring each as \(S_{\text{pub}}(q,a) = \text{clip}(S_{\text{raw}}(q,a)\cdot J(a),\,0,\,2.2)\) where \(S_{\text{raw}} = 0.45K + 0.35E + 0.20C\) (keyword match, BioBERT semantic similarity, log-scaled citation count).

  2. Synthesis layer — A Reasoning Agent works over the fixed ranked evidence set, building PPI graphs from STRING, identifying connected components, and emitting structured reasoning notes (pathway interpretation, uncertainty statements, citation links). A Writing Agent assembles these notes into a citation-grounded report.

  3. Interface layer — A Visualization Agent consumes the same typed evidence artifacts (not free-form text) to generate a dashboard schema and rendered interactive interface with pathway ranking panels, protein–pathway–publication–PPI graph views, and PMID-linked detail panels.

All agents share GPT-4o as the base model. The harness enforces typed artifact contracts between stages so downstream agents cannot introduce unsupported claims.

Key Contributions

  • Formulates disease-centered protein interpretation as evidence-centered interactive interface generation, producing an explorable workspace rather than a text answer.
  • BioInsight multi-agent system with typed intermediate artifacts (ranked pathway table \(T\), evidence packets \(E\), reasoning notes \(N\), citation-grounded report \(R\), dashboard schema \(S\), rendered dashboard \(H\)) that preserve full provenance.
  • BioInsight-1k benchmark of protein-function reasoning questions derived from UniProt and STRING, with 100 expert-curated hard questions (BioInsight-100) requiring cross-evidence integration.
  • Empirical demonstration that separating retrieval from reasoning (explicit search decomposition) is necessary—ablation without search degrades coverage, validity, traceability, and ranking scores consistently.

Results

  • BioASQ Phase B (Batch 1): Best or tied-best across all five exact-answer metrics; +1.9 points factoid MRR and +4.4 points list F-measure over the strongest baseline.
  • BioInsight-100 (expert 0–10 scoring): BioInsight mean score 8.62, highest among five systems, with score distribution concentrated in the high range.
  • End-to-end disease synthesis (5 diseases: AD, MDD, AF, CKD, RA): BioInsight consistently leads or ties on automatic coverage, validity, traceability, and ranking dimensions; strongest expert human ratings for traceability, ranking quality, and dashboard usability vs. GPT-5.5+Search, Gemini Deep Research, and DR-Tulu-8B.
  • Ablation (BioInsight w/o search decomposition) shows lower scores across all dimensions, confirming that explicit staged retrieval is critical.

Limitations

  • All LLM agents use GPT-4o; no open-weight or smaller-model ablation is reported, so generalizability to other base models is unknown.
  • Evaluation scope is restricted to five diseases and one BioASQ batch; broader generalization across disease types or data modalities is not demonstrated.
  • The interactive dashboard is constrained to evidence objects present in upstream artifacts—novel visual hypotheses not in retrieved literature cannot be surfaced, which may limit exploratory use.
  • End-to-end human evaluation involves a small expert panel; inter-rater reliability statistics are not reported in the excerpt.
  • Retrieval relies on PubMed, Semantic Scholar, OpenTargets, UniProt, and STRING; coverage gaps in these databases propagate directly into reasoning quality.

Relevance to Agentic AI / LLM Agents

BioInsight is a concrete instantiation of harness-centered multi-agent orchestration where the central design challenge is enforcing typed artifact contracts between specialized agents rather than allowing unconstrained inter-agent communication—a pattern directly relevant to agentic AI system design. It demonstrates that decomposing retrieval from reasoning from report-writing from visualization into separate agents with fixed handoff schemas improves downstream accuracy and traceability compared to monolithic deep-research pipelines. The typed intermediate artifact framework addresses a key open problem in agentic AI: how to maintain provenance and prevent hallucination propagation across a multi-agent chain. The paper also contributes a domain-specific benchmark (BioInsight-1k) for evaluating wide-evidence reasoning agents, adding to the growing ecosystem of agentic evaluation resources.