Key points are not available for this paper at this time.
Artificial intelligence (AI) has transformed prediction of protein structure and biomolecular interactions, yet modeling of allosteric regulation remains a persistent and unresolved challenge. We develop a dual explainable AI framework that systematically interrogates AI Co-Folding models AlphaFold3, Protenix, Boltz-2, Chai-1, and DynamicBind on rigorously stratified datasets of orthosteric and allosteric ligand-protein complexes. While all AI models excel in accurate modeling of orthosteric ligand binding, a universal and architecture-independent collapse emerges in prediction of allosteric complexes. The biophysical logic for this dichotomy is unveiled through physics-based lens of the energy landscape theory and local frustration analysis. Orthosteric binding creates dominant energetic funnels via ligand-induced minimal frustration quenching, while allosteric sites preserve neutral frustration landscapes in both apo and holo protein states. The findings show that conformational heterogeneity and evolutionary plasticity encoded in allosteric binding landscapes may conceal the recurrent recognition patterns AI models are trained to detect. By linking prediction outcomes to frustration landscapes, this study recasts AI shortcomings in allosteric ligand binding as diagnostic indicators of fundamental biophysical constraints, establishing a physics-informed framework that turns the allosteric blind spot into mechanistic insight for next-generation landscape-aware predictive tools.
Parikh et al. (Thu,) studied this question.