HighRelax: Physics-Based Refinement of Deep Learning Protein Predictions with Noncanonical Amino Acids.

Area of Science:

• Biochemistry
• Structural Biology
• Computational Chemistry

Background:

• Noncanonical amino acids (NCAAs) offer advantages in protein engineering and drug development, such as improved stability and permeability.
• Accurate 3D structures of NCAA-containing proteins are vital for understanding function and rational design.
• Current protein structure prediction models struggle with NCAAs, leading to errors like chirality violations and steric clashes.

Purpose of the Study:

• To develop a computational method for refining protein structures containing noncanonical amino acids.
• To address limitations in existing protein structure prediction frameworks when integrating NCAAs.
• To enhance the accuracy and reliability of NCAA-containing protein models for downstream applications.

Main Methods:

• Expanded the AMBER force field to include parameters for 139 NCAAs.
• Developed an enhanced Amber-relax protocol named HighRelax.
• Integrated HighRelax with state-of-the-art structure prediction models like AlphaFold3 for complex systems.

Main Results:

• HighRelax effectively reduces steric clashes and restores residue chirality in NCAA-containing structures.
• The protocol improves the overall structural quality of proteins with noncanonical amino acids.
• HighRelax is compatible with complex systems, including cyclic peptides and structures from advanced prediction models.

Conclusions:

• HighRelax provides a general and effective postprocessing strategy for refining NCAA-containing protein structures.
• This method facilitates the application of NCAA-containing proteins in molecular simulation, peptide drug design, and protein engineering.
• The enhanced force field and protocol enable more accurate modeling of proteins with nonstandard amino acids.