MegaTrans-machine learning models for drug transporters corresponding to the FDA guidance

Regulatory guidances (eg, FDA and European Medicines Agency) require an understanding of the interactions of novel drugs, natural products, and environmental toxicants with key transporters to avoid compounds with undesirable side effects. Computational approaches to predict such interactions using machine learning models trained on in vitro data could prevent compounds that are transporter inhibitors with potential for drug-drug interactions from reaching the more costly development stages. We now describe the curation and machine learning model building for the transporters covered in the FDA guidance (organic anion transporter 1, organic anion transporter 3, organic cation transporter 2, organic anion transporting polypeptide 1B1, organic anion transporting polypeptide 1B3, P-glycoprotein, breast cancer resistance protein, multidrug and toxin extruder 1, and multidrug and toxin extruder protein 2K) that enabled the creation of MegaTrans, a web-based software product that enables users to input molecules and predict the inhibition of transporters of interest. Each dataset was used to generate machine learning models using extended-connectivity fingerprint 6 with 5-fold cross-validation. These models were also applied to a dataset (N = 39) derived from recent FDA-approved drugs and chemical property overlap with our models was determined using t-distributed stochastic neighbor embedding. In aggregate, the model predictions had an accuracy of 66%, and a second round of transporter models displayed an accuracy of 76%. Our application of literature in vitro data on transporters for building these machine learning models can be used for compound profiling, representing an approach which could minimize the need for in vitro screening experiments and help to prioritize resources. Such applications of machine learning models could also be integrated at the drug design stages in artificial intelligence approaches, including generative artificial intelligence, further broadening the impact. SIGNIFICANCE STATEMENT: MegaTrans is a computational tool to predict the inhibition of small molecules across a range of clinically relevant human drug transporters, allowing visualization of model metrics, highlighting atom-bond contributions, and visualizing chemical space overlaps between queries and model datasets.