Toward Development of “FXR PREDICTOR”: A Web‐Based Application for Estimating Biological Activity of Chemical Compounds Toward Farnesoid X Receptor

This study focuses on the development of FXR PREDICTOR, an application designed to guide researchers in determining the activity of chemical compounds against non‐alcoholic fatty liver disease (NAFLD) and non‐alcoholic steatohepatitis (NASH) through modulation of the farnesoid X receptor (FXR). From ChEMBL database 1217 unique compounds were retrieved which were analyzed by: i . conventional 2D‐QSAR (Quantitative Structure‐Activity Relationship) with Dragon descriptors and molecular fingerprints, ii . Transformer‐CNN, leveraging SMILES augmentation. Morgan fingerprint‐based model with random forest classifier algorithm produced the best predictive model with strong Matthews correlation coefficient (MCC). Although the descriptor‐based model had lower predictive quality, it provided important insights into the structural features of the ligands through permutation and SHAP analyses. Contour maps for fingerprint‐based model specify favourable regions for lead optimization. 85 new compounds from ChEMBL were deployed as validation set; retrieving more than 80% balanced accuracy, confirming model robustness. This model was incorporated in Streamlit‐based FXR PREDICTOR, a user‐friendly application for virtual screening.