Advanced drug-target interaction prediction using convolutional graph attention networks in expert systems

Predicting drug-target interaction (DTI) is crucial in drug discovery and repurposing, as it significantly cuts the time and costs associated with traditional experimental methods. To address these challenges, this study introduces an advanced deep learning framework that integrates graph-based neural networks with novel feature selection mechanisms to improve DTI prediction accuracy. A Convolutional Multilayer Extreme Adversarial Graph Attention-based Neural Network (CMEAG-ANN), combined with a Fast Correlation-Based Gradient Naïve Bayes and Binary Pattern Selection (FC-GNBBPS) algorithm, is proposed for the robust and biologically meaningful feature extraction from DNA molecule-derived data. Using graph attention algorithms that capture complex relationships within molecular graphs, CMEAG-ANN effectively integrates structural and evolutionary aspects of drugs and target proteins. It uses both molecular fingerprints and PSSM-based annotations, ensuring a rich representation of chemical and biological information. Experimental evaluations on benchmark datasets, including approved_drug_target dataset, ImDrug dataset, DrugProt dataset, and Drug Combination Extraction Dataset, are compared with the CMEAG-ANN and the baseline models. The CMEAG-ANN model achieves an accuracy of 99.17%, precision of 99.11%, recall of 98.83%, F1-score of 98.96%, and specificity of 98.74%. This study highlights the model's effectiveness in improving the reliability and efficiency of DTI systems through biologically grounded feature selection.