Skip to content

BioInsight: Multi-Agent Orchestration for Interactive Biomedical Knowledge Discovery

🕒 Published (v1): 2026-06-19 00:15 UTC · Source: Arxiv · link

Why this paper was selected

Multi-agent orchestration for biomedical discovery; interactive evidence inspection matches medical+agent interests

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

BioInsight is a multi-agent harness that converts disease-associated protein tables into citation-grounded reports and interactive dashboards through a pipeline of typed intermediate artifacts. Rather than producing a static text answer, it preserves full evidence provenance—linking proteins, pathways, publications, and PPI modules—so researchers can inspect, filter, and refine hypotheses interactively. It achieves best or tied-best on BioASQ Phase B and outperforms GPT-5.5, Gemini Deep Research, and domain-tuned models on expert-rated end-to-end biomedical evidence synthesis.

Problem

Existing search-augmented LLMs and deep-research agents produce static text or tables as terminal outputs. For disease-centered protein interpretation, researchers need to trace claims back to source evidence, compare alternative mechanistic explanations, check uncertainty, and refine hypotheses—operations that static reports cannot support. There is no prior system that converts multi-source biomedical evidence into a provenance-preserving, explorable decision-support interface.

Method

BioInsight is a four-agent harness with three explicit layers:

  1. Evidence layer — A Planning Agent maps disease-associated proteins to enriched pathways via g:Profiler, deduplicates pathway names using BioBERT embedding similarity, and scores pathways by \(S_\text{path}(t) = 0.4 \cdot P_\text{norm}(t) + 0.6 \cdot (1 - L_\text{norm}(t))\), where \(P_\text{norm}\) is normalized enrichment P-value and \(L_\text{norm}\) is literature relevance. A Search Agent retrieves publications from PubMed and Semantic Scholar, scored as \(S_\text{pub}(q,a) = \text{clip}(S_\text{raw}(q,a) \cdot J(a),\ 0,\ 2.2)\) with \(S_\text{raw} = 0.45K + 0.35E + 0.20C\) (keyword match, BioBERT semantic similarity, log-citation count).

  2. Synthesis layer — A Reasoning Agent receives fixed evidence packets and returns structured reasoning notes (pathway interpretation, uncertainty statements, citation links) rather than prose. PPI edges are modeled as an undirected weighted graph; connected components are ranked by internal edge weight.

  3. Interface layer — A Writing Agent assembles a citation-grounded report; a Visualization Agent converts the same typed artifacts (not open-ended code) into an interactive dashboard schema with a pathway ranking panel, evidence graph (pathways ↔ proteins ↔ publications ↔ PPI/drug edges), and detail panels.

A central harness schedules stages sequentially, validates required fields, and passes only typed artifact contracts between agents. All LLM calls use GPT-4o.

Key Contributions

  • Formulation of evidence-centered interactive interface generation as an output target distinct from DeepSearch (single answers), DeepResearch (narrative reports), and WideSearch (tables).
  • BioInsight harness: four-agent, three-layer architecture with typed intermediate artifacts (\(T, E, N, R, S, H\)) enforcing provenance continuity from evidence retrieval to rendered dashboard.
  • BioInsight-1k benchmark: protein-function reasoning dataset built from UniProt and STRING; a 100-question expert-curated subset (BioInsight-100) used for evaluation.
  • Demonstration that search decomposition is necessary: ablating explicit search consistently degrades coverage, validity, traceability, and ranking scores.

Results

  • BioASQ Phase B (Batch 1): 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 protein-function reasoning: Highest mean expert score of 8.62/10 across five compared systems (GPT-5.5, DR-Tulu-8B, Gemma-4-31B, Qwen3.5-9B).
  • End-to-end synthesis (5 diseases — AD, MDD, AF, CKD, RA): BioInsight scores highest or tied-highest on automatic evaluation dimensions (coverage, biomedical validity, traceability, ranking/depth) and receives strongest human expert ratings specifically for traceability and prioritization; biomedical validity scores are closer across frontier models.
  • Ablation (w/o search decomposition) consistently underperforms full BioInsight across all evaluation dimensions and diseases.

Limitations

  • All LLM agents use GPT-4o as a fixed backbone; no ablation over base model choice.
  • Evaluation covers only five diseases and one BioASQ batch, limiting generalizability.
  • The benchmark (BioInsight-100) was partially generated by GPT-5.5 before expert curation, introducing potential contamination or bias.
  • Dashboard generation and interactive interface evaluation rely on expert scores without inter-rater reliability statistics reported.
  • The system is sequential (not parallel), which may be slow for large protein tables or many candidate pathways.
  • External knowledge bases (PubMed, UniProt, STRING, OpenTargets) introduce dependency on third-party availability and freshness.

Relevance to Harnesses / Meta-Harnesses

BioInsight is a direct instantiation of the harness pattern: a central coordinator schedules heterogeneous agents through typed artifact contracts, enforces stage ordering, validates outputs, and propagates structured evidence across layers rather than relying on a single monolithic model. The key design principle—that downstream agents receive only typed artifacts from upstream stages, not raw context—is a concrete implementation of the artifact-contract discipline central to robust multi-agent harnesses. The paper's framing of "harness-centered multi-agent system" and its explicit ablation of the search decomposition stage make it a useful case study for understanding how harness structure (vs. flat prompting) drives measurable gains in grounding, traceability, and ranking quality. It also demonstrates that typed intermediate artifacts serve dual purposes: feeding downstream agents and exposing provenance to end users, an architectural idea transferable to any domain-specific research harness.