Uncertainty-Guided Agents for Rare-Disease Hypothesis Discovery on Knowledge Graphs

Agents4Science 2025 Conference Submission272 Authors

Keywords: Rare Diseases, Knowledge Graphs, Uncertainty, Multi-Agent Systems, Hypothesis Discovery, Reproducibility

TL;DR: An uncertainty-guided multi-agent system improves rare-disease hypothesis discovery on knowledge graphs under tight experimental budgets.

Abstract: Rare disease discovery is hampered by data sparsity, fragmented evidence, and expensive validation. We present an uncertainty-guided multi-agent system that closes the loop between hypothesis generation, experiment selection, and self-audit on a biomedical knowledge graph (KG). A lightweight link scorer with Monte Carlo-style uncertainty feeds a planner that prioritizes experiments under a fixed budget; an auditor reports calibration at high-confidence thresholds. On a synthetic rare-disease KG benchmark, our agent improves precision--recall and budgeted discovery over heuristic and static baselines (e.g., +0.10 AUPRC and +0.9 Hit@10 on average) while maintaining reasonable calibration. Ablations confirm that uncertainty-driven selection is critical to early-budget gains; robustness sweeps show graceful degradation under increased sparsity and noise. The framework is fully reproducible with code that regenerates all figures, providing a tractable template for evaluating AI agents for scientific discovery.

Submission Number: 272