A multi-modal computational strategy integrating ensemble learning, molecular docking, and drug-likeness profiling for the discovery of novel small-molecule TNF-α inhibitors.

BACKGROUND

Tumour necrosis factor-alpha (TNF-α) is a validated target in autoimmune disease, yet the clinic relies almost entirely on injectable biologics. Small-molecule inhibitors remain scarce: the functional surface is a shallow, symmetric trimer interface, and public resources hold few high-confidence actives.

METHODS

We assembled a five-stage pipeline joining ligand-based statistics to structure-based geometry and a physicochemical filter. A curated ChEMBL set of 1559 compounds (1344 active, 215 inactive) was encoded as a 2,215-bit fused fingerprint (MACCS 167, ECFP4 2048). Eight learners were benchmarked on a scaffold-disjoint partition, with three imbalance strategies compared. The top 500 of a 999,881-compound ZINC library were scaffold de-duplicated, docked with AutoDock Vina 1.2.5 against PDB 2AZ5 (SPD304 as positive control), then profiled with a composite drug-likeness score and tier-A predicted ADMET endpoints.

RESULTS

Random Forest gave the best held-out discrimination (ROC-AUC 0.9727, 95% CI 0.949-0.992); the soft-voting Ensemble finished 0.0051 AUC below it (paired bootstrap p = 0.028). Resampling shifted the Random Forest AUC by at most 0.001, and label permutation over 500 runs gave 0.550 ± 0.081 (ΔAUC +0.423, p < 0.002). The hits sat outside the model's applicability domain (median maximum Tanimoto 0.226 vs 0.842 within training). The SPD304 control failed to reproduce its crystallographic pose (best-of-nine RMSD 5.20 Å), so all Vina scores are unvalidated.

CONCLUSION

The pipeline returns a shortlist of tractable compounds, three negative results, and one qualification that bounds what it means. Binding, inhibition, and cytotoxicity assays are required before any compound can be called a TNF-α inhibitor.