KinBloc: an interpretable machine learning model for prediction of kinase inhibitors in the kinome with kinase-inhibitor interaction block

Wei-Lin Lin(林韋霖)1, Yen‑Chao Hsu1 and Jinn-Moon Yang1,2*

1Institute of Bioinformatics and Systems Biology, National Yang Ming Chiao Tung University, Hsinchu 300, Taiwan

2Department of Biological Science and Technology, National Yang Ming Chiao Tung University, Hsinchu 300, Taiwan

Protein kinase dysregulation, mutation, and overexpression are closely related to the pathogenesis of many diseases, making protein kinases important drug targets. However, the highly conserved binding sites of approximately 500 human protein kinases make the development of selective inhibitors a challenging task. Currently, many models for predicting kinase inhibitors are black-box and lack interpretability with biological significance, while only considering compounds leads to a lack of correlation between different kinases. We propose an interpretable machine learning model, based on the kinase-inhibitors interaction block (KinBloc), for predicting kinase inhibitors. In the proposed method, residue encoding is based on physicochemical properties, sequence conservation、environment, and structural position information, and combined with molecular fingerprints to predict and observe the selectivity mechanism of multiple kinases. We evaluated the performance of KinBloc on well-known benchmark datasets and verified the compound interaction correlations at the ATP binding pocket through protein kinase crystal structures. We believe that the KinBloc model can be used to predict the activity of a single compound in multiple kinases and discover selective molecular structures between different kinases, which can contribute to the development and design of selective inhibitors.