Answers are grounded in this paper and cite their sources.
A rescoring benchmark compares scoring functions on a fixed decoy set; neither how many poses are kept nor their character is usually reported. We ask which pool properties decide whether such a comparison is well posed. Three pools were compared under one protocol: a perturbation pool built from the bound pose (all 22 rigid untrained complexes of Benchmark 5), the same perturbation on unbound structures (15 complexes), and the published ZDOCK 3.0.2 decoy set, a real Fourier output (14 complexes). The first two differ only in the structures, the second and third only in the pose generator. Only the bound pool supports the claim that a knowledge-based potential beats a shape-complementarity proxy: there it wins 21 of 22 complexes at K = 20 (p < 0.0001). The same perturbation on unbound structures removes most of it (8 of 15, p = 0.79), and on the search output no advantage remains (all p ≥ 0.09). The cause is measurable: the potential’s score is nearly monotonic in the clash count, and only the bound pool supplies clash-free near-native poses, so elsewhere it penalizes the poses it is asked to find. On the search output, the program’s own rescoring algorithm is at chance at K = 20 and helps only from K = 400; densifying the search fifteenfold leaves the pool unchanged. The comparison is well posed only when the size of the candidate set, its origin and composition, the relaxation of its near-native poses, the measure and the null are stated.
No takes yet. Share an insight, caveat, or question.
孙庭广 et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: