Enhancing Drug-Target Interaction Prediction with CNN-Based Deep Learning and Systematic Encoding Strategies

Abstract

Drug-target interaction (DTI) prediction is a crucial task in drug discovery, aiming to identify novel therapeutic applications for existing drugs and reduce the time and cost associated with drug development. This study proposes a deep learning-based approach for DTI prediction, leveraging convolutional neural networks (CNNs) and advanced feature representation techniques. The methodology involves encoding drug-target pairs using a combination of drug molecular descriptors generated by PaDEL-Descriptor and protein sequence properties derived from the AAindex1 database. Moran autocorrelation is employed to capture the structural and functional characteristics of the proteins. The concatenated feature vectors are then projected onto a lower-dimensional space using random projection and reshaped into matrices to serve as inputs for the CNN model. The CNN architecture, based on LeNet-5, learns hierarchical feature representations from the input matrices. An ensemble of multiple CNN predictors is used to enhance the robustness and accuracy of the predictions. The model is evaluated on benchmark datasets, demonstrating superior performance compared to traditional machine learning algorithms and existing deep learning-based methods. The proposed framework exhibits high accuracy, sensitivity, precision, and area under the curve (AUC) values, highlighting its effectiveness in capturing complex drug-target relationships. Furthermore, the model shows strong generalization ability on external validation datasets and benefits from a systematic approach for generating reliable negative samples. The results suggest that the integration of deep learning with advanced feature representation techniques offers a promising tool for accelerating drug discovery and understanding drug mechanisms.