Prediction of Farnesoid X Receptor Disruptors with Machine Learning Methods.

The farnesoid X receptor (FXR) emerges as a promising drug target involved in regulating various metabolic pathways, yet some xenobiotic compounds binding to FXR would be an important determinant to induce the receptor dysfunctions that lead to undesirable side effects. Thus, it is critical to identify potential xenobiotics that disrupt normal FXR functions. In this work, five machine learning methods coupled with eight molecular fingerprints and 20 molecular descriptors were used to develop classification models for prediction of FXR binders. The built models were evaluated using the test set and two external validation sets. The best model was obtained using a combination of molecular descriptors and fingerprints, which exhibited the AUC values of 0.83 and 0.92 for the test set and the first external validation set, respectively. The overall prediction accuracy for the second external validation set with the best model was over 85%. Furthermore, several representative privileged substructures that are essential for FXR binders, such as benzimidazole, indole, and stilbene moiety, were detected using information gain and substructure frequency analysis. The applicability domain analysis via the Euclidean distance-based approach demonstrated a marked impact on the improvement of prediction accuracy. Overall, our built models could be helpful to rapidly identify potential chemicals binding to FXR.