Developing machine learning models to predict chemical-breast cancer resistance protein (BCRP) interaction

Abstract

The breast cancer resistance protein (BCRP; ABCG2 ) is an efflux transporter that affects drug pharmacokinetics, contributes to multidrug resistance, and can cause drug-drug interactions (DDIs). Predicting whether a compound interacts with BCRP can aid in the development of adjuvant anti-cancer drugs and help anticipate potential DDIs. In this study, we built and systematically compared 50 classification models spanning seven molecular representations and 10 machine learning algorithms to predict BCRP inhibition. We used a Butina cluster-based split to create structurally distinct training and test sets and evaluated all models with five-repeat × five-fold stratified cross-validation and statistical testing. The top-ranked model configuration, Mordred descriptors with TabPFN, achieved a cross-validation MCC of 0.79 and AUROC of 0.96, with several leading configurations showing statistically comparable MCC values. On the cluster-split test set, this model achieved an MCC of 0.70 and AUROC of 0.94, with specificity (0.95) substantially exceeding sensitivity (0.71), indicating that BCRP inhibitors are more difficult to classify than non-inhibitors. We also found that classical fingerprints and descriptors generally outperformed frozen transformer-based embeddings on this dataset. Misclassification analysis showed that false negatives occupy sparse regions in the chemical space. This work provides a transporter-specific case study using a systematic model benchmarking approach. The newly developed BCRP inhibition model will be incorporated into the multi-transporter screening platform MONSTROUS, replacing its existing BCRP model.