In Silico Regression Modeling and Improved Interpretability To Predict the Transport Inhibitory Activity of Breast Cancer Resistance Protein

Breast cancer-resistance protein (BCRP) functions as an efflux transporter, and its inhibition is an important pharmacokinetic parameter in drug–drug interactions (DDIs). In silico approaches that predict compound profiles from chemical structures are widely used in early drug discovery, and binary classification models have been developed for BCRP inhibition. While classification models distinguish inhibitors from noninhibitors, regression models provide numerical predictions that enable ranking. Additionally, there has been a growing demand for models with better interpretability. In this study, we aimed to construct regression models that can predict the IC50 values of BCRP inhibitory activity and to interpret the predicted results. Prediction models were built using a high-quality data set comprising 870 compounds with IC50 values obtained from publicly available sources, and the best model achieved an R2 value of 0.736 on the test set. Although the model demonstrated satisfactory performance, the use of descriptors that were difficult to interpret was found to limit the explainability of its predictions. To overcome this problem, we applied a new set of machine learning models using only substructure-based fingerprints and a tree-based algorithm. The best model showed an R2 value of 0.703 on the test set. Based on SHapley Additive exPlanations (SHAP) values, several substructures were identified as key features by the finalized model, and the prediction results revealed seven potential BCRP inhibitors among approved drugs. These models contribute to accelerating compound screening and reducing drug development costs, while also aiding in the exclusion of potential dropout candidates.