XG-DTA: Drug-Target Affinity Prediction Based on Drug Molecular Graph and Protein Sequence combined with XLNet

Drug-target affinity (DTA) prediction is critical in drug development. Accurate prediction of drug-target interactions can accelerate the development of new drugs and improve drug safety. However, DTA involves complex biomolecular interaction, and DTA prediction requires the processing of large amounts of chemical and bioinformatics data. Traditional methods rely on expensive and time-consuming experimental analysis, resulting in a slow and expensive drug development process. Despite recent advances in drug-target relationships (DTRs) prediction in deep learning algorithms, most computational approaches still focus on determining whether there is a binding interaction between a drug and its target, while neglecting to correctly discriminate between primary and non-targets through unbiased binding affinity values. In this study, we propose a deep learning model called XG-DTA that uses microstructural features of drug molecules and protein sequence features as inputs to predict DTA. In this model, we utilize GAT to explore the complex representation of drug molecule graphs, and adopt XLNet as a word vector model to encode protein sequences, from which high-dimensional semantic features are extracted. And the two are combined to obtain the pharmacological context of DTA. The proposed model exhibits better performance in predicting binding affinity values compared to current state-of-the-art baselines from experimental results. The results show that the XG-DTA model achieved the best Concordance Index (CI) and Mean Square Error (MSE) performance on two bench mark datasets. The case study experiments on two important Human Immunodeficiency Virus (HIV) proteins confirm that the proposed DTA model can be used as an effective pre-screening tool for drug discovery.