Machine Learning Modeling for ABC Transporter Efflux and Inhibition: Data Curation, Model Development, and New Compound Interaction Predictions.

In recent years, multiple computational studies have used machine learning models to predict substrate binding and inhibition of ATP-binding cassette (ABC) transporters. However, many of these studies relied on relatively small training sets with limited applicability. In this study, we manually curated over 24,000 bioactivity records (i.e., inhibition, binding affinity, permeability) for the ABC transporters P-gp, BCRP, MRP1, and MRP2 from more than 900 literature sources in ChEMBL, with additional data from PubChem and Metrabase. This effort yielded eight data sets, comprising around 8800 unique chemicals with one or more substrate binding or inhibition activities for these four efflux transporters. Quantitative structure-activity relationship (QSAR) models were developed for each of the eight data sets using combinations of four machine learning algorithms and three sets of chemical descriptors. The resulting models demonstrated excellent performance by 5-fold cross-validation, achieving an average correct classification rate (CCR) of 0.764 for the substrate binding models and 0.839 for the inhibition models. Models were validated with additional compounds from DrugBank that were known substrates or inhibitors. We further analyzed how model predictions for efflux transporter activity could estimate exposure of the brain to xenobiotics. Notably, compounds predicted as P-gp and BCRP substrates were twice or more likely to have low brain exposure compared to compounds with high brain exposure. This study provides a large and curated drug transporter binding and inhibition database for computational modeling. Applicable models based on this large database for predicting transporter substrate binding and inhibition can be used to evaluate more complex drug bioactivities, such as exposure of protected tissues to chemicals.