The accuracy of electrostatic interactions captured by AI protein structure prediction models
TL;DR
AI protein structure prediction models like AlphaFold2 incorrectly predict that ionizable residues can be buried in nonpolar cores for mutated sequences, a physical impossibility that can be corrected using short molecular dynamics simulations.
Problem / question
The study investigates whether state-of-the-art AI protein structure prediction models accurately encode fundamental physico-chemical principles, specifically the placement of ionizable residues, when predicting structures for non-natural or mutated sequences.
Methods
The authors compared biophysical measurements of a U1A protein variant containing four ionizable substitutions against predictions from AlphaFold2, RoseTTAFold2, OmegaFold, and ESMFold. They then computationally generated thousands of sequences with up to 12 ionizable residues in the nonpolar cores of U1A, acylphosphatase, and TOP7, and tested whether 50-nanosecond molecular dynamics simulations using CHARMM and AMBER force fields could correct the AI-generated structures.
Key findings
Biophysical measurements revealed the U1A variant is trimeric with twice the helical structure of the wild-type, yet all tested AI models incorrectly predicted a monomeric structure nearly identical to the wild-type with a backbone RMSD under 1 Angstrom. Across thousands of sequences for U1A, acylphosphatase, and TOP7, the AI models consistently buried ionizable residues inside the nonpolar core while retaining the parent fold. Running 50-nanosecond molecular dynamics simulations rapidly relaxed these physically impossible structures and properly exposed the ionizable residues.
Why it matters
This reveals that while AI models excel at predicting natural protein structures, they fail to learn underlying physical rules like electrostatics, making physics-based validation steps essential for de novo protein design and variant effect prediction.
Limitations
The provided text does not detail extensive limitations, but the experiments were specifically constrained to testing the burial of ionizable residues in the nonpolar cores of three similarly sized proteins.
Takeaway
AI protein structure predictions for non-natural sequences should be validated with brief physics-based molecular dynamics simulations to ensure they do not violate basic physico-chemical principles like burying ionizable residues.