Ligand-Dependent Affinity but State-Dominated Efficacy in Structure-Based GPCR Pharmacology Prediction

Abstract

Abstract High accuracy in structure-based prediction of G protein-coupled receptor (GPCR) efficacy may reflect the receptor state rather than ligand-specific information. We addressed these information sources using state rules, ligand-blind baselines, matched receptor-state contrasts, ligand permutation, and out-of-distribution (OOD) evaluation, with affinity prediction as a positive control. State rules and ligand-blind geometry achieved binary-efficacy-balanced accuracies of 0.844−0.910 and 0.777−0.870, respectively. Of 75 ligand-aware evaluations, 72 failed at least one prespecified ligand-sensitivity condition, and three lacked required evidence; none met all required criteria. Affinity pretraining gave no detectable efficacy benefit over matched random initialization. In contrast, ligand chemistry supported affinity prediction across 14,740 measured pairs, and all 12 primary affinity permutation comparisons deteriorated, although this advantage did not persist under scaffold, family, or combined OOD splits. Accuracy on endpoint structures therefore cannot alone establish ligand-specific efficacy learning; the findings do not set a general limit on such learning.