Property-Dependent Scoring Bias in Virtual Screening Validation: Evidence from Dual Scoring Function Cross-Validation of BCL-2 G-Quadruplex Ensemble Docking

Physics-based docking scores correlate with molecular size, yet virtual screening validation rarely controls for physicochemical property differences between actives and controls, risking inflated discrimination metrics. Here we address this problem through ensemble docking of the BCL-2 promoter G-quadruplex (PDB: 2F8U, 10 NMR models), cross-validated with AutoDock Vina (7,469 compounds, 114 ensemble-validated) and rDock (7,296 compounds, 113 validated), classifying each finding as scoring-function-agnostic or scoring-function-dependent. Agnostic results include the analytical framework, discrimination direction, the Kendall’s W category hierarchy, and G4-specific cross-scoring correlation (rho = 0.632). Dependent results include discrimination magnitude (Vina AUROC = 0.748 vs. rDock 0.602), compound rankings (rho = 0.263, n = 38), and model concordance patterns. Retrospective validation against 111 experimentally confirmed BCL-2 G4 binders (ROBIN dataset) yielded AUROC near 0.5 for both scoring functions (Vina: 0.466; rDock: 0.493), establishing that discrimination observed with size-diverse controls reflects property-dependent scoring bias, not target-specific recognition. Hydrogen bond donor count (rho = 0.647) and topological polar surface area (rho = 0.525) predict rank divergence between scoring functions, whereas rotatable bond count does not, implicating polar interaction parametrization—not conformational sampling—as the source of disagreement. Single-model screening recovered only 45–85% of ensemble top-20 compounds (Jaccard overlap 0.29–0.74). The ensemble docking methodology generalizes across scoring functions; compound prioritization, however, requires multi-scoring-function consensus.