CSDTI: an interpretable cross-attention network with GNN-based drug molecule aggregation for drug-target interaction prediction

Drug-target interaction (DTI) is a critical and complex process that plays a vital role in drug discovery and design. In deep learning-based DTI methods, graph neural networks (GNNs) are employed for drug molecule modeling, attention mechanisms are utilized to simulate the interaction between drugs and targets. However, existing methods still face two limitations in these aspects. First, GNN primarily focus on local neighboring nodes, making it difficult to capture the global 3D structure and edge information. Second, the current attention-based methods for modeling drug-target interactions lack interpretability and do not fully utilize the deep representations of drugs and targets. To address the aforementioned issues, we propose an interpretable network architecture called CSDTI. It utilizes a cross-attention mechanism to capture the interaction features between drugs and targets. Meanwhile, we design a drug molecule aggregator to capture high-order dependencies within the drug molecular graph. These features are then utilized simultaneously for downstream tasks. Through rigorous experiments, we have demonstrated that CSDTI outperforms state-of-the-art methods in terms of performance metrics such as AUC, precision, and recall in DTI prediction tasks. Furthermore, the visualization mapping of attention weights indicates that CSDTI can provide chemical insights even without external knowledge.