Drug Target Affinity Prediction Based on Graph Structural Enhancement and Multi-scale Topological Feature Fusion

Drug target affinity prediction is a critical step in virtual screening, and the accuracy of its prediction can have a direct impact on the course of drug development. Various sequence-based and graph-based deep learning models have achieved good performance in drug target affinity (DTA) prediction, but most of them extract features at a single scale, and this approach is deficient in global topological feature extraction. In this paper, we propose a new DTA prediction model GSEM-DTA. In order to accurately extract the features of drugs and proteins, we constructed two graphs, drug molecule graph and motif graph, in the drug feature extraction stage of the model, and designed a graph structure enhancement module (GSEM) to obtain the features of drug molecule graph. In the protein feature extraction step we designed a multiscale convolutional coding module with a multi-head linear attention mechanism for extracting protein features from multiple scales and deep levels, and later used the attention mechanism for Multi-scale topological feature fusion. We conducted a series of experiments on two benchmark datasets, Davis and Kiba, and the experimental results show that our model achieves optimal performance.