SE-DTA: A Spatial Equivariant Network for Drug-Target Binding Affinity Prediction
Xinyi Tu, Zhe Li, Wenbin Lin · 2024
The accuracy of drug-target binding affinity (DTA) prediction directly affects the efficiency and success rate of drug discovery. While traditional drug screening methods are usually time-consuming and labor-intensive, the introduction of deep learning provides new possibilities for improving DTA prediction. However, most existing work primarily focuses on one-dimensional (1D) sequence features and 2D structural features of drugs and proteins. In this study, we propose a novel equivariant graph neural network (EGNN) model from the biological perspective, called SE-DTA, which comprehensively integrates the 3D spatial information of drugs and proteins. In SE-DTA, we convert drug molecules into 3D graph structures and perform special processing on proteins, that is, we extract the binding sites of each protein from UniProtKB and then construct the residue-level protein graph and atomic-level protein graph based on ATP binding sites. Finally, we apply EGNN to learn feature representations of drugs and proteins. EGNN can ensure the equivariance of molecular structures, thereby effectively capturing the spatial information and enhancing the predictability and interpretability of the model. We evaluate the performance of SE-DTA on two widespread datasets. Experimental results show that SE-DTA achieves exciting performance, with$CI$values of 0.913 and 0.903 on the Davis and KIBA datasets respectively, which are at least 0.03 higher than the state-of-the-art baseline model. Additionally, the high$r_{m}^{2}$index also verifies the strong generalization ability of our model.