Fig. 4

Excluding non-orthosteric residues markedly diminishes HOLIgraph performance across multiple classification metrics. Box plots show accuracy, precision, recall, F1, balanced accuracy, and AUC for the outward-facing HOLIgraph model trained with only orthosteric (orange), orthosteric and opportunistic (green), or all residues (magenta). Removing non-orthosteric/opportunistic residues significantly impairs performance, underscoring the importance of the complete binding interface for accurate inhibitor prediction. Statistical signficance was assessed via Wilcoxon tests followed by Holm-Bonferroni correction (*p < 0.05, ***p < 5E−4)