The accuracy of electrostatic interactions captured by AI protein structure prediction models

Area of Science:

• Biophysics
• Computational Biology
• Protein Science

Background:

• Deep learning models (AlphaFold2, RoseTTAFold2, OmegaFold, ESMFold) excel at predicting protein structures.
• However, their ability to accurately model the placement of ionizable residues within protein cores is questionable.

Purpose of the Study:

• To investigate the accuracy of AI models in predicting protein structures, specifically concerning the placement of ionizable residues.
• To assess whether AI models adhere to established physico-chemical principles in structure prediction.

Main Methods:

• Generated a U1A protein variant with altered ionizable residues.
• Utilized multiple AI tools (AlphaFold2, RoseTTAFold2, OmegaFold, ESMFold) for structure prediction.
• Performed biophysical measurements and short molecular dynamics simulations (50 ns) using CHARMM/AMBER force fields.

Main Results:

• AI models predicted structures nearly identical to wild-type U1A, with buried ionizable residues contradicting physical chemistry.
• Thousands of generated sequences and additional protein folds (acylphosphatase, TOP7) showed similar AI prediction patterns.
• Physics-based simulations rapidly corrected AI-predicted structures by exposing ionizable residues.

Conclusions:

• AI tools, while powerful for general structure prediction, do not reliably encode physico-chemical rules for ionizable residue placement.
• Brief molecular dynamics simulations are essential for validating AI-generated protein structures.
• This highlights a critical limitation in current AI-driven structural biology workflows.