Deep Learning-Based Prediction of Drug-Target Binding Affinities by Incorporating Local Structure of Protein

Traditional drug discovery methods are both time-consuming and expensive. Utilizing artificial intelligence to predict drug-target binding affinity (DTA) has become an essential approach for accelerating new drug discovery. While many deep learning methods have been developed for DTA prediction, most of them only consider the primary sequence structure of proteins. However, drug-target interactions occur only in specific regions of the protein, and the primary structure can only represent the global protein features, which fails to fully disclose the relationship between the drug and its target. In this study, we used both the primary and secondary protein structures to represent the protein. The primary structure served as the global feature, and the secondary structure as the local feature. We use convolutional neural networks (CNNs) and graph neural networks (GNNs) to model proteins and drugs separately, which helped to better capture the interactions between drugs and targets. As a result, our method demonstrated improved performance in predicting DTA comparing to the latest methods on two benchmark datasets.