Decoding Allosteric Grammar with Explainable AI Integrating Protein Language Models and Energy Landscape Analysis:...

Area of Science:

• Biophysics
• Computational Biology
• Structural Biology

Background:

• Allosteric regulation is crucial for protein kinases to integrate cellular signals.
• Understanding the energetic basis of regulatory binding sites is incomplete.
• Existing methods struggle to consistently identify allosteric sites.

Purpose of the Study:

• To investigate the biophysical organization of regulatory binding sites in protein kinases.
• To determine why allosteric sites are difficult to detect computationally.
• To link binding site detectability to their energetic properties.

Main Methods:

• Developed an explainable artificial intelligence (AI) framework using protein language models (PLMs).
• Integrated PLM predictions with energy landscape frustration analysis.
• Analyzed a dataset of 453 human kinases and performed atomic resolution analysis on ABL kinase.

Main Results:

• Orthosteric ATP binding sites are detected with high confidence due to their minimally frustrated energetic regions.
• Allosteric sites are consistently missed by PLMs, residing in neutrally frustrated zones.
• The 'allosteric blind spot' is an intrinsic biophysical property, not an algorithmic limitation.
• ABL kinase allosteric pockets remain neutrally frustrated across various states and ligands.

Conclusions:

• The detectability of binding sites is determined by their energetic embedding within protein energy landscapes.
• Allosteric sites are encoded in persistent, neutrally frustrated regions for context-dependent modulation.
• Explainable AI can uncover how protein energy landscapes shape functional plasticity and site detectability.