A Graph Neural Network Framework with Equivariance and Heterogeneity for Drug Repositioning

Drug repositioning is a strategy that enhances traditional drug development by identifying new therapeutic uses for approved or clinically investigated drugs through scientific approaches. However, the majority of drug repositioning strategies often overlook the critical three-dimensional (3D) structural information of drug molecules, despite its fundamental role in determining biological function. To tackle this challenge, we propose DREHGNN, a novel method that integrates an E(n)-equivariant graph neural network (EGNN) and a heterogeneous graph neural network (HGNN) for drug repositioning. First, EGNN is used to extract drug features directly from their 3D molecular structures. Concurrently, disease features are derived from the disease-symptom associations. Subsequently, the framework integrates drug-drug synergies, drug-disease associations, and disease-disease similarities into a heterogeneous network. Then, topological features are learned from this network using HGNN. Finally, a multi-layer perceptron (MLP) is employed as the final predictor to identify candidate drugs for repositioning. The experimental evaluation reveals that DREHGNN achieves substantially superior performance compared to benchmark methods in comparative evaluations. Ablation studies confirm the essential contribution of 3D structural feature learning via EGNN, and case studies validate predicted drug candidates with existing evidence. These results demonstrate the feasibility of leveraging molecular 3D structures via EGNN, thereby establishing HGNN as a reliable framework for drug repositioning. The data and codes of this study can be obtained through https://github.com/333jia/DREHGNN.git .