QKDTI A quantum kernel based machine learning model for drug target interaction prediction

Drug-target interaction (DTI) prediction is a critical task in computational drug discovery, enabling drug repurposing, precise medicine, and large-scale virtual screening. Traditional in-silico methods, such as molecular docking, classical machine learning, and deep learning, have made significant progress in addressing this issue. However, existing approaches are hindered by computational inefficiencies, reliance on manual feature engineering, and struggles to generalize across diverse molecular structures, limiting their molecular capabilities. Recent advancements in Quantum Machine Learning (QML) are paving the way for its practical applications, unlocking unprecedented capabilities in predictive accuracy, scalability, and efficiency by leveraging the unique powers of quantum computing, namely superposition and entanglement. This study proposes QKDTI - Quantum Kernel Drug-Target Interaction, a novel quantum-enhanced framework for DTI prediction. It used Quantum Support Vector Regression (QSVR) with quantum feature mapping that takes into account a quantum feature space for molecular descriptors and allows encoding molecular and protein features, improved predictions of binding affinities. To enhance the model to be more computationally feasible, integration of the Nystrom approximation into the model allows providing an efficient kernel approximation while reducing overhead expenses. QKDTI was evaluated on benchmark datasets - Davis and KIBA, and validated independently on BindingDB. This model achieves 94.21% accuracy on DAVIS, 99.99% on KIBA, and 89.26% on BindingDB, significantly outperforming classical and other quantum models. Further, the statistical tests have been conducted on the compared models to provide the reliability of the results. This indicates that introducing quantum computing into DTI pipeline can revolutionize computational drug discovery by improving predictive accuracy and providing a better generalization over multiple datasets.