Predicting drug-target interactions and binding affinity using an optimized deep learning approach

The accurate prediction of Drug-Target Interactions (DTIs) and Drug-Target Affinity (DTA) is crucial for reducing experimental costs and time, thereby accelerating drug discovery and repurposing efforts. The local biochemical contexts and the global structural dependencies between drugs and their target proteins are often insufficiently captured by conventional deep learning models, which restrict their predictive performance. In this work, we introduce a CNN based Dual Attention (nCNN-DA), which combines 1D convolutional feature extraction with channel and spatial attention mechanisms to enhance the representational power of features for drug SMILES and protein sequences. The model was trained and tested on three benchmark datasets: KIBA, Davis, and BindingDB, using AUPR, AUROC, MSE, Pearson correlation, and accuracy. Experimental results show that nCNN-DA significantly improves performance compared to well-established models (FusionNet, GraphormerDTI, DeepDTAGen, and DTBA-net). In particular, nCNN-DA achieved the best accuracy of 98.5%, 95.5%, and 97.5%, as well as the lowest MSE of 0.1559, 0.3189, and 0.2957 on KIBA, Davis, and BindingDB, respectively, and better scores for AUPR and Pearson Correlation. These findings further demonstrate that nCNN-DA issued for identifying putative DTI pairs and predicting binding affinities with high quality, making it a versatile and general method for drug discovery, virtual screening, and drug repurposing.