Identification of small-molecule TNF-α inhibitor candidates using machine learning-guided screening and multiscale molecular modelling.

Lumbar disc herniation (LDH) is a major cause of chronic low back pain, in which tumor necrosis factor-alpha (TNF-α) plays a central role in inflammation and pain signaling. While biologic TNF-α inhibitors have shown therapeutic benefit, their systemic administration, high cost, and limited penetration into the avascular disc environment restrict their clinical utility. In this study, we present a multiscale computational framework to identify small-molecule TNF-α inhibitor candidates. A curated dataset of experimentally validated TNF-α inhibitors was used to train supervised machine learning models, among which a Random Forest classifier achieved the best performance (ROC-AUC = 0.92). The optimized model was applied to screen 61,534 compounds from the ChemDiv database, yielding high-confidence candidates that were further evaluated through Glide XP docking, ADMET prediction, and molecular dynamics (MD) simulations. Docking protocol validation was performed via redocking of the co-crystallized ligand, achieving an RMSD < 2.0 Å. The top-ranked compound (8009-0259) exhibited favorable binding affinity, stable interaction patterns during 100 ns MD simulation, and consistent engagement with key residues (Tyr151, Gln61). Binding free energy analysis (MM/GBSA) suggested that hydrophobic interactions are the dominant contributors to ligand stabilization. Density functional theory (DFT) analysis indicated moderate electronic stability of the lead compound, supporting its potential for intermolecular interactions. Overall, this study provides a computational prioritization framework for identifying TNF-α inhibitor candidates and offers mechanistic insights into their binding behavior. The identified compounds warrant further experimental validation for their therapeutic potential in LDH.