AQuaRef: machine learning accelerated quantum refinement of protein structures

Area of Science:

• Structural Biology
• Computational Chemistry
• Artificial Intelligence

Background:

• Macromolecular models from cryo-electron microscopy (cryo-EM) and X-ray crystallography require refinement.
• Current refinement methods use limited stereochemical data and miss noncovalent interactions.
• Quantum mechanical (QM) calculations offer better accuracy but are computationally expensive for large biomolecules.

Purpose of the Study:

• To introduce a novel AI-enabled Quantum Refinement (AQuaRef) method.
• To mimic QM calculations at reduced computational cost for biomolecular structure refinement.
• To improve the geometric quality and experimental fit of atomic models.

Main Methods:

• Developed AQuaRef utilizing the AIMNet2 machine learned interatomic potential (MLIP).
• Applied AQuaRef to refine 41 cryo-EM and 30 X-ray structures.
• Compared AQuaRef results against standard refinement techniques.

Main Results:

• AQuaRef produced atomic models with superior geometric quality.
• The method maintained or improved the fit to experimental data compared to standard techniques.
• AQuaRef successfully determined proton positions, including challenging short hydrogen bonds in DJ-1 and YajL proteins.

Conclusions:

• AI-enabled QM refinement offers a computationally efficient alternative for improving biomolecular models.
• AQuaRef enhances structural accuracy and provides insights into hydrogen bonding and protonation states.
• This method holds promise for advancing structural biology and understanding disease-associated proteins.