Quantifying conformational diversity in protein–ligand ensembles for structure-based virtual screening

Structure-based virtual screening (SBVS) requires accurate representations of protein and ligand conformational states, since both components can adopt diverse geometries that influence binding energetics. Using 79 HIV-1 protease-ligand complexes, we quantify how conformational heterogeneity and ensemble clustering shape the evaluation of interaction energies by comparing native complexes, nonnative pairings, protein ensembles derived from apo MD, and ligand ensembles sampled in solution. Native complexes consistently yield favorable interactions, whereas nonnative pairings are rarely favorable and often highly unfavorable, indicating that many failures in screening that uses a single structure arise from protein–ligand geometric mismatch rather than from the scoring function alone. We further show that structural reduction of protein and ligand ensembles decreases the recovery of favorable interaction states, and that simultaneous reduction of both ensembles constrains the availability of complementary structural pairs. Because SBVS is often limited by computational cost, receptor ensembles are typically restricted to a small number of conformations. Here, we quantify the sampling scale and the level of ensemble reduction required to retain binding-compatible geometries and stable interaction-energy trends, providing practical guidance for ensemble-based screening.