Supplementary material from "Computational Discovery of TGF-β R1 inhibitors from the EMNPD database via machine learning, molecular docking and steered molecular dynamics"

Machine learning (ML) is increasingly applied in cheminformatics to predict physicochemical and biological properties from structural descriptors. In this work, eight ML algorithms, Linear Regression, Random Forest, XGBoost, LightGBM, CatBoost, k-Nearest Neighbours, Support Vector Regression, and Gradient Boosting, were benchmarked for predicting binding free energies of compounds targeting TGF-β Receptor I (TGFβR1). Among them, KNN achieved the best predictive accuracy (RMSE = 1.023, R = 0.704, MAE = 0.773) and was employed for virtual screening of the EMNPD database. Fourteen candidates below −12 kcal.mol-1 were prioritized. Docking validation confirmed reliability (RMSD < 2 Å), and five predicted hits displayed favorable docking scores. Molecular dynamics (MD) simulations over 100 ns revealed stable protein–ligand complexes, with interactions involving LYS232, LEU340, VAL219, and ILE211. To enhance energy prediction, steered molecular dynamics (SMD) was applied, yielding stronger correlation with experimental data (R = −0.7201) compared to docking (R = 0.5404). Nigrosporolide showed the greatest mechanical stability (ΔGSMD = −8.125 kcal.mol-1, ⟨Fmax⟩ = 852.78 pN). Toxicity profiling predicted LD50 values from 340–9000 mg/kg with variable organ-specific risks. Overall, integrating ML, docking, SMD, and toxicity analysis enabled effective identification of TGFβR1 inhibitors, highlighting nigrosporolide and trichocadinin E as promising leads for further study.