Encoding Drug-Target-Pathway-Disease profiles for Drug-Cancer Association Prediction using Graph Transformer-Convolution Networks

Abstract

New drug development is costly, time-consuming, and has a low success rate, leading to a decline in drug discovery efficiency over time. Drug repurposing has emerged as an effective alternative, applying existing, safe drugs to new diseases, thereby reducing development time and costs by bypassing preclinical toxicology testing. With the increasing availability of large-scale interaction data (e.g., drug-protein, protein-protein, and drug-disease networks) and advancements in generative AI, new opportunities have arisen for drug discovery. However, AI-based methods still face challenges: (1) ineffective integration of heterogeneous biological data across drugs, proteins, pathways, and diseases, and (2) lack of interpretability, limiting insights into drug mechanisms of action. To address these challenges, we propose a Graph Transformer-Convolution Network (GTCN) that integrates Graph Transformer Networks (GTNs) and Graph Convolution Networks (GCNs). By leveraging dynamic heterogeneous graph learning and attention mechanisms, our model optimizes relational structures within biological networks (drug-target-pathway-disease) and extracts more discriminative node features. Unlike traditional models that only encode direct drug-disease relationships, our approach captures how drugs act on proteins and regulate pathways to treat diseases. Furthermore, we design an interpretability framework that identifies critical elements for drug-cancer predictions, offering insights into disease mechanisms and drug mechanisms of action (MoA). This facilitates the discovery of new therapeutic strategies with biologically interpretable visualizations. The proposed dataset and code are available at https://github.com/zhengyutong99/GTCN