LGABAN: An Integrated Multi-Scale Approach Combining Graph and Sequence Features for Enhanced Prediction of Drug–Protein Interactions

The accurate identification of drug-target interactions is crucial for shortening the timeline and lowering the expenses of pharmaceutical research, as the discovery of novel drugs remains a highly complex, resource-intensive, and lengthy endeavor. Despite progress in the use of deep learning for drug-target interaction prediction, these methods still face substantial challenges in feature representation and model interpretability, especially when dealing with complex, multiscale interaction relationships. To address this, we propose a novel deep learning framework, LGABAN, which jointly models multilevel information from both drugs and proteins by parallelly extracting local and global features through a dual-branch structure. To explicitly model the complex multiscale interactions between drugs and proteins, LGABAN integrates four types of feature pairs─local-local, local-global, global-local, and global-global─using a bilinear attention network (BAN). Additionally, we introduce a multihead graph attention network (GAT) to further enhance the representational capacity of drug graph representations. Experimental results on four publicly available data sets reveal that LGABAN surpasses six state-of-the-art baseline models in overall performance. Furthermore, satisfactory interpretability results are provided for all aspects of drug-target interaction prediction.