MARS-guided PubChem fingerprint selection and random forest classification of hepatitis C virus NS5B inhibitors with scaffold-independent prioritization of active chemotypes

Rational drug design accelerates early anti-hepatitis C discovery by learning the substructures that govern activity against the NS5B RNA-dependent RNA polymerase. This study proposes a two-stage model that couples multivariate adaptive regression splines with random forest classification to identify the PubChem fingerprints driving NS5B inhibition and to classify compounds as active or inactive. Multivariate adaptive regression splines capture non-linear main effects and pairwise interactions among binary fingerprints without distributional assumptions. After removal of training-set-constant fingerprints, 522 informative bits remained, from which the method selected 66 features. The analysis uses 578 non-redundant NS5B inhibitors curated from ChEMBL, comprising 412 active and 166 inactive compounds described by 881-bit PubChem substructure fingerprints. On a Bemis-Murcko scaffold-based external partition whose scaffolds are absent from training, random forest on the 66 selected bits reached a balanced accuracy of 84.4%, a Matthews correlation coefficient of 0.69, an area under the ROC curve of 0.91 and an average precision of 0.94, and it ranked all 20 of its top-scoring novel-scaffold compounds as truly active. Under stratified ten-fold cross-validation, with feature selection performed within each training fold, the same model reached a balanced accuracy of 89.0% and a Matthews correlation coefficient of 0.79, achieving the highest or near-highest performance among the classifiers tested with nearly unchanged accuracy despite the eightfold reduction in dimensionality. Structure-activity analysis identifies extended and branched aliphatic chains together with halogenated aromatic motifs as the substructures most enriched in NS5B inhibitory activity, and converts them into explicit rules for prioritizing active chemotypes.