Machine Learning-Based Classification Models for Non-Covalent Bruton's Tyrosine Kinase Inhibitors: Predictive Ability and Interpretability

Abstract

In this study, we built classification models using machine learning techniques to predict the bioactivity of non-covalent inhibitors of Bruton's tyrosine kinase (BTK) and to provide interpretable and transparent explanations for these predictions. To achieve this, we gathered data on BTK inhibitors from the Reaxys and ChEMBL databases, removing compounds with covalent bonds and duplicates to obtain a dataset of 3895 inhibitors of non-covalent. These inhibitors were characterized using MACCS fingerprints and Morgan fingerprints, and four traditional machine learning algorithms (decision trees (DT), random forests (RF), support vector machines (SVM), and extreme gradient boosting (XGBoost)) were used to build 16 classification models. In addition, four deep learning models were developed using deep neural networks (DNN). The best model, Model D_4, which was built using XGBoost and MACCS fingerprints, achieved an accuracy of 94.1% and a Mathews correlation coefficient (MCC) of 0.75 on the test set. To provide interpretable explanations, we employed the SHAP method to decompose the predicted values into the contributions of each feature. We also used K-means dimensionality reduction and hierarchical clustering to visualize the clustering effects of molecular structures of the inhibitors. The results of this study were validated using crystal structures, and we found that the interaction between the BTK amino acid residue and the important features of clustered scaffold was consistent with the known properties of the complex crystal structures. Overall, our models demonstrated high predictive ability, and a qualitative model can be converted to a quantitative model to some extent by SHAP, making them valuable for guiding the design of new BTK inhibitors with desired activity.