How Well Can AI and Physics-Based Simulations Predict the Probability a Cryptic Pocket Is Open?

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Cryptic pockets are crucial targets in drug discovery for modulating challenging proteins.
• Current AI structure prediction methods lack the physics to fully characterize protein conformational ensembles, including cryptic pockets.
• Several new AI-based methods (AlphaFlow, BioEmu, PocketMiner, CryptoBank) aim to address this limitation.

Purpose of the Study:

• To benchmark AI models and molecular dynamics (MD) simulations in characterizing known cryptic pockets.
• To evaluate the ability of these methods to recapitulate experimental thermodynamics of cryptic pockets in Ebola VP35 and TEM β-lactamase.
• To assess performance in predicting pocket opening probability changes due to mutations.

Main Methods:

• Benchmarking AI models (AlphaFlow, BioEmu, PocketMiner, CryptoBank) and physics-based molecular dynamics (MD) simulations.
• Utilizing experimentally characterized cryptic pockets in Ebola VP35 and TEM β-lactamase.
• Comparing predicted pocket opening probabilities with experimental data for wild-type proteins and mutants.

Main Results:

• Multiple AI and MD methods accurately predict whether mutations increase or decrease cryptic pocket opening probability.
• No method reliably predicts the absolute probability of pocket opening, particularly for rare pockets (<1% opening).
• BioEmu and PocketMiner show trends for pockets with >1% opening but exhibit systematic errors and struggle with rare pockets.

Conclusions:

• AI and simulation-based strategies hold significant promise for cryptic pocket characterization in drug discovery.
• Further improvements are needed for robust prediction of absolute cryptic pocket opening probabilities, especially for low-probability events.
• Current methods show potential for guiding mutation strategies but require refinement for precise thermodynamic characterization.