Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RareAgent: Self-Evolving Reasoning for Drug Repurposing in Rare Diseases

Lang Qin, Zijian Gan, Xu Cao, Pengcheng Jiang, Yankai Jiang, Jiawei Han, Kaishun Wu, Jintai Chen

TL;DR

RareAgent tackles rare-disease drug repurposing under the sparse-zero bipartite setting by reframing the task as active, evidence-seeking reasoning with a self-evolving multi-agent system. It builds task-specific evidence graphs (T-EGraphs) through adversarial debates among four agents, guided by textual feedback that refines policies and distills reasoning into transferable heuristics, yielding auditable reasoning trails. Empirically, RareAgent achieves state-of-the-art performance on rare-disease indication (AUPRC 0.438, AUROC 0.662, improved to 0.463 and 0.750 after self-evolution) and strong results on BioHopR (1-hop 33.87%, 2-hop 17.68%), outperforming reasoning baselines and large LLMs under budget constraints. The approach enhances interpretability and reliability in high-stakes drug discovery by grounding predictions in verifiable evidence and a growing repertoire of domain-specific heuristics.

Abstract

Computational drug repurposing for rare diseases is especially challenging when no prior associations exist between drugs and target diseases. Therefore, knowledge graph completion and message-passing GNNs have little reliable signal to learn and propagate, resulting in poor performance. We present RareAgent, a self-evolving multi-agent system that reframes this task from passive pattern recognition to active evidence-seeking reasoning. RareAgent organizes task-specific adversarial debates in which agents dynamically construct evidence graphs from diverse perspectives to support, refute, or entail hypotheses. The reasoning strategies are analyzed post hoc in a self-evolutionary loop, producing textual feedback that refines agent policies, while successful reasoning paths are distilled into transferable heuristics to accelerate future investigations. Comprehensive evaluations reveal that RareAgent improves the indication AUPRC by 18.1% over reasoning baselines and provides a transparent reasoning chain consistent with clinical evidence.

RareAgent: Self-Evolving Reasoning for Drug Repurposing in Rare Diseases

TL;DR

RareAgent tackles rare-disease drug repurposing under the sparse-zero bipartite setting by reframing the task as active, evidence-seeking reasoning with a self-evolving multi-agent system. It builds task-specific evidence graphs (T-EGraphs) through adversarial debates among four agents, guided by textual feedback that refines policies and distills reasoning into transferable heuristics, yielding auditable reasoning trails. Empirically, RareAgent achieves state-of-the-art performance on rare-disease indication (AUPRC 0.438, AUROC 0.662, improved to 0.463 and 0.750 after self-evolution) and strong results on BioHopR (1-hop 33.87%, 2-hop 17.68%), outperforming reasoning baselines and large LLMs under budget constraints. The approach enhances interpretability and reliability in high-stakes drug discovery by grounding predictions in verifiable evidence and a growing repertoire of domain-specific heuristics.

Abstract

Computational drug repurposing for rare diseases is especially challenging when no prior associations exist between drugs and target diseases. Therefore, knowledge graph completion and message-passing GNNs have little reliable signal to learn and propagate, resulting in poor performance. We present RareAgent, a self-evolving multi-agent system that reframes this task from passive pattern recognition to active evidence-seeking reasoning. RareAgent organizes task-specific adversarial debates in which agents dynamically construct evidence graphs from diverse perspectives to support, refute, or entail hypotheses. The reasoning strategies are analyzed post hoc in a self-evolutionary loop, producing textual feedback that refines agent policies, while successful reasoning paths are distilled into transferable heuristics to accelerate future investigations. Comprehensive evaluations reveal that RareAgent improves the indication AUPRC by 18.1% over reasoning baselines and provides a transparent reasoning chain consistent with clinical evidence.

Paper Structure

This paper contains 42 sections, 6 equations, 4 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formulation
  5. RareAgent Framework
  6. Hypothesis Generation
  7. Iterative Refinement
  8. Agents Evolution
  9. Experiments
  10. Datasets and implementation details
  11. RQ1: Rare disease drug repurposing
  12. RQ2: General biomedical reasoning
  13. RQ3: Ablations of key components
  14. RQ4: Case Study: Auditable Evidence Trail for ARVC (Flecainide)
  15. Conclusions
  16. ...and 27 more sections

Figures (4)

  • Figure 1: RareAgent overview. Explorer seeds a task-specific evidence graph (T-EGraph) with initial candidates based on the user query. The Proponent and Skeptic freely add and challenge evidence on a shared whiteboard, while the PI monitors, interrupts, and revises. Iteration yields ranked candidates with auditable trails. Finally, PI analyzes reasoning paths to generate feedback and distill heuristics, continuously improving the system's policies.
  • Figure 2: The cumulative effect of multiple self-evolution rounds on agent performance.
  • Figure 3: BioHopR relations and answers splits.Single asks for one correct answer per hop; Multi asks for all correct answers.
  • Figure 4: Case study. RareAgent identified Flecainide as a candidate therapy for ARVC, whereas the baseline found no novel therapy and proposed drugs with severe safety risks. The sole flecainide-ARVC trial supports this conclusion, and the RareAgent did not access that trial record during its operation.