KGTC-DDA:Drug-disease association prediction based on parallel graph isomorphism and transformer networks with cross-attention

Abstract

Drug development is a complex, time-consuming, and costly process, making drug repurposing an important strategy for identifying new therapeutic applications for existing drugs. Computational drug-disease association (DDA) prediction has therefore attracted increasing attention in recent years. Although many existing methods incorporate similarity information and graph neural networks for DDA prediction, effectively capturing both local structural patterns and global dependencies, as well as modeling complex interactions between drugs and diseases, remains challenging. To address these limitations, we propose Drug-disease association prediction model (KNN, GIN, Graph Transformer and Cross-attention, KGTC-DDA), a novel framework that combines Graph Isomorphism Networks (GIN), Graph Transformers (GT), and cross-attention mechanisms for drug-disease association prediction. Specifically, multiple biological similarities, including drug chemical structure similarity, disease phenotype similarity, and Gaussian interaction profile (GIP) kernel similarity, are integrated to construct comprehensive drug and disease similarity graphs. GIN and Graph Transformer modules are then employed in parallel to simultaneously learn local topological features and global contextual representations from the similarity networks. Furthermore, a cross-attention mechanism is introduced to capture high-order interdependencies between drug and disease embeddings, thereby enhancing heterogeneous feature interaction and representation learning. Extensive experiments on three benchmark datasets under 10-times 10-fold cross-validation demonstrate that KGTC-DDA achieves superior predictive performance compared with several state-of-the-art methods in terms of AUROC and AUPR. In addition, experiments on unseen disease prediction further validate the robustness and generalization capability of the proposed framework. These results indicate that KGTC-DDA provides an effective and promising approach for computational drug repositioning and drug-disease association prediction.