AI-Driven Drug Repurposing for Rare Diseases

Abstract

Rare diseases collectively affect millions of individuals worldwide, yet therapeutic development remains limited because conventional drug discovery is expensive and time-intensive. Artificial intelligence has emerged as an effective approach for accelerating drug repurposing by identifying new therapeutic applications for approved compounds. Nevertheless, existing AI models frequently rely on homogeneous biomedical data and rarely quantify prediction uncertainty, reducing confidence in computational recommendations for clinically underrepresented diseases. This study introduces an Uncertainty-Aware Multi-Modal Graph Learning Framework that combines biomedical knowledge graphs, transcriptomic profiles, pathway representations, and protein interaction networks to predict novel drug–disease associations. Bayesian uncertainty estimation is incorporated to distinguish high-confidence predictions from uncertain recommendations, improving decision support for downstream biological validation. The proposed framework emphasizes interpretability through attention-based graph aggregation and biological pathway contribution analysis. The research demonstrates how integrating heterogeneous biomedical evidence with uncertainty-aware learning can improve reliable candidate prioritization for rare diseases. The findings contribute toward trustworthy AI-assisted drug repurposing systems capable of supporting precision medicine while reducing computational bias associated with limited rare disease data