Structure-based machine learning model for discovering pregnane X receptor (PXR) agonists and biological activity...

Pregnane X receptor (PXR), a nuclear receptor superfamily member, maintains bile acid homeostasis by regulating metabolic enzymes [e.g., cytochrome P450 3A4 (CYP3A4), uridine diphosphate glucuronosyltransferase 1A1 (UGT1A1)] and multidrug resistance protein 1 transporter, and alleviates liver/intestinal inflammation via inhibiting the nuclear factor kappa-B pathway, serving as a critical therapeutic target for cholestatic liver diseases and inflammatory bowel disease. In this study, we established a novel structure-based machine learning strategy to identify PXR agonists from a natural product database. With generated bioactive conformations binding with PXR, we integrated ligand-based and structure-based features to construct a comprehensive machine learning model using light gradient boosting machine. This model illustrated an R² of 0.874 for the internal validation set and an R² of 0.845 for the external test set, superior to the performance of other machine learning models, e.g. random forest regression, support vector regression, gradient boosting regression, K-nearest neighbors, and extreme gradient boosting. The model was used to predict the PXR agonistic activity of the candidate molecules screened out by the pharmacophore model. Promising candidates were selected out for further assay with HepG2 cell culture combined with a dual-luciferase reporter. Ultimately, natural products like schisantherin A, rhynchophylline, and irigenin were identified as potent PXR agonists, with half-maximal effective concentrations (EC₅₀) of 1.58 μM, 2.57 μM, and 20.67 μM, respectively. These PXR agonists act as potential candidates for targeted therapies against PXR-related diseases. We anticipate that this work will provide support for the design and discovery of PXR modulators.