Decoding the Allosteric Paradox: A Dual Framework Integrating AI Cofolding Models with Landscape-Guided Interpretable...

Area of Science:

• Computational biology
• Biophysics
• Artificial intelligence in drug discovery

Background:

• Artificial intelligence (AI) has revolutionized protein structure and biomolecular interaction prediction.
• Modeling allosteric regulation in ligand-protein complexes remains a significant challenge for current AI approaches.

Purpose of the Study:

• To develop and apply a dual explainable AI framework to analyze AI Co-Folding models' performance on orthosteric and allosteric ligand-protein complexes.
• To elucidate the biophysical underpinnings of AI model performance disparities between orthosteric and allosteric binding.

Main Methods:

• Systematic interrogation of multiple AI Co-Folding models (AlphaFold3, Protenix, Boltz-2, Chai-1, DynamicBind) using stratified datasets.
• Application of energy landscape theory and local frustration analysis to understand prediction outcomes.
• Comparative analysis of AI model performance on orthosteric versus allosteric ligand-protein complexes.

Main Results:

• All tested AI models demonstrated high accuracy for orthosteric ligand binding prediction.
• A universal, architecture-independent performance collapse was observed in the prediction of allosteric complexes.
• Orthosteric binding was linked to ligand-induced minimal frustration quenching, creating dominant energetic funnels, while allosteric sites exhibited neutral frustration landscapes.

Conclusions:

• The conformational heterogeneity and evolutionary plasticity of allosteric binding landscapes may obscure patterns detectable by current AI models.
• AI model limitations in allosteric ligand binding prediction highlight fundamental biophysical constraints.
• A physics-informed framework is established, integrating energy landscape insights to improve future AI-driven predictive tools for allosteric regulation.