Deep Drug–Target Binding Affinity Prediction Base on Multiple Feature Extraction and Fusion

Accurate drug–target binding affinity (DTA) prediction is crucial in drug discovery. Recently, deep learning methods for DTA prediction have made significant progress. However, there are still two challenges: (1) recent models always ignore the correlations in drug and target data in the drug/target representation process and (2) the interaction learning of drug–target pairs always is by simple concatenation, which is insufficient to explore their fusion. To overcome these challenges, we propose an end-to-end sequence-based model called BTDHDTA. In the feature extraction process, the bidirectional gated recurrent unit (GRU), transformer encoder, and dilated convolution are employed to extract global, local, and their correlation patterns of drug and target input. Additionally, a module combining convolutional neural networks with a Highway connection is introduced to fuse drug and protein deep features. We evaluate the performance of BTDHDTA on three benchmark data sets (Davis, KIBA, and Metz), demonstrating its superiority over several current state-of-the-art methods in key metrics such as Mean Squared Error (MSE), Concordance Index (CI), and Regression toward the mean (Rm2). The results indicate that our method achieves a better performance in DTA prediction. In the case study, we use the BTDHDTA model to predict the binding affinities between 3137 FDA-approved drugs and severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) replication-related proteins, validating the model's effectiveness in practical scenarios.