RepoLLM: A Multi-modal Foundation Model for Drug Repurposing via Alignment of Molecules, EHRs, and Knowledge Graphs

ICML 2025 Workshop FM4LS Submission49 Authors

Keywords: drug repurposing, multi-modal learning, knowledge graphs, electronic health records, attention mechanisms, translational medicine, foundation models

TL;DR: We present a multi-modal foundation model combining drug molecules, EHRs, and knowledge graphs for interpretable drug repurposing, achieving SOTA prediction accuracy while preserving biological relationships.

Abstract: We propose a novel multi-modal foundation model for drug repurposing that integrates drug molecules (SMILES sequences), electronic health records (EHRs), and knowledge graphs (KGs) through a cross-modal attention mechanism. Our framework achieves state-of-the-art performance on drug-disease prediction (0.824 AUROC) while maintaining knowledge graph consistency (0.642 Hit@10), demonstrating significant improvements over unimodal baselines. The model exhibits exceptional robustness to missing data, retaining 92.1\% of performance when two modalities are absent. Clinical validation shows 83.4\% agreement with physician decisions, with attention-guided knowledge graph paths providing interpretable biological explanations. This work establishes a new paradigm for therapeutic discovery by effectively bridging molecular, clinical, and relational biomedical data.

Submission Number: 49