Graph Neural Networks for Predicting Drug-Target Interactions in Rare Disease

Abstract

Rare diseases affect millions of people worldwide, yet the discovery of effective treatments remains hindered by limited biological data, sparse drug–target interaction (DTI) information, and the high cost of experimental validation. Existing computational approaches often struggle to maintain predictive performance under such data-constrained conditions. This study aims to develop a robust Graph Neural Network (GNN)-based framework capable of accurately predicting DTIs for rare diseases by leveraging heterogeneous biomedical knowledge graphs and advanced representation learning techniques. The proposed methodology integrates a self-supervised graph contrastive learning module with a rare-disease-aware edge-reweighting mechanism to improve representation quality and preserve disease-specific biological signals during message passing. An experimental evaluation was conducted using benchmark datasets derived from DrugBank, Orphanet, and UniProt across varying levels of data sparsity. The proposed framework achieved AUROC scores of 0.932, 0.904, and 0.859 and AUPRC scores of 0.871, 0.835, and 0.748 under moderate, high, and extreme sparsity conditions, respectively, consistently outperforming state-of-the-art baseline methods. The results further demonstrate that graph contrastive learning and contextual edge reweighting substantially improve predictive robustness and model generalization. In conclusion, the proposed framework provides an accurate, scalable, and interpretable computational solution for rare disease drug discovery, supporting prioritization of drug repurposing candidates and therapeutic target identification in data-scarce biomedical environments, subject to future biological and clinical validation.