Drug Repurposing Using Graph Neural Networks (GNN)

Abstract

The discovery of new drugs is a time-consuming, expensive, and uncertain task. Drug repurposing is an efficient solution to this problem, which aims to find new uses for existing drugs. This paper introduces an AI-powered framework for computational drug repurposing based on Graph Neural Networks (GNNs) for predicting disease-drug associations. The biomedical knowledge graph is represented as a heterogeneous graph of diseases, genes, and drugs, which are linked by validated biological interactions. The framework leverages multimodal node representations, such as molecular embeddings for drugs and protein sequence embeddings from a pre-trained protein language model. A two-layer heterogeneous GraphSAGE model is used to aggregate information from the graph to learn biologically informative representations. Drug-disease similarity computation is performed by cosine similarity, and the model is trained with margin ranking loss and negative sampling. Experimental evalua- tion on AUROC indicates excellent performance in distinguishing actual therapeutic associations, which indicates a scalable and biologically informed approach to accelerate drug repurposing.