Interpretable machine learning models for prediction of kinase inhibitors types using protein grid energy-based descriptors

藍靖涵 Ching-Han Lan1, 許彥超 Yen-Chao Hsu1, 楊進木 Jinn-Moon Yang1

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

Dysregulation and mutations of protein kinases play an important role in human diseases and known as therapeutic targets. In recent years, kinase inhibitors (KIs) could form different binding modes, including type I, type I½ and type II, according to their binding pockets such as ATP binding pocket or allosteric sites. Different types of KI could cause conformational changes in different protein kinases. Therefore, it’s crucial to understand binding mechanisms  for designing highly selective inhibitors. However, current machine learning methods to predict the binding mode of KIs only consider that compound structure as features, which is lack of protein binding environment and cannot consider that may cause different kinase inhibitor types when the same inhibitor binds to different kinases. Here, we developed an interpretable model that utilizes compound and protein kinase as features to predict kinase inhibitor types. Functional group structural moieties and the interaction energy between the kinases and the inhibitors by 3D grid-based approach are features represent compound and protein kinases, respectively. By using three models created by each types and 4,624 labeled kinase-inhibitor-types dataset, it can not only achieves over 95% accuracy but also be relevant to the kinase-inhibitor interactions and analyze the binding environment. We believe it could be a strategy for highly selective and low-side-effect kinase drugs design.