Using multiple computer-predicted structures as molecular replacement models: application to the antiviral...

Area of Science:

• Structural Biology
• Computational Biology
• Protein Engineering

Background:

• Accurate protein structure determination is crucial for understanding biological function.
• Computational methods for solving protein structures via molecular replacement (MR) are continuously evolving.
• The engineered protein LCB2, a 58-residue three-helix bundle, serves as a test case for assessing novel structure prediction algorithms.

Purpose of the Study:

• To evaluate the performance of various contemporary computational tools in generating molecular replacement (MR) models for protein structure determination.
• To investigate the nature of structural differences observed between predicted models and experimentally determined structures.
• To explore the concept of benign model bias and its implications for interpreting protein conformational heterogeneity.

Main Methods:

• Utilized multiple computer-predicted molecular replacement (MR) models generated by AlphaFold3, AlphaFold2, MultiFOLD, Rosetta, RoseTTAFold, and trRosetta.
• Employed Coot and Phenix for multi-start structure determination, assessing convergence and structural accuracy (all-atom RMSD).
• Assigned B factors using predictor-generated confidence scores or accessible surface area (ASA) values.
• Analyzed side-chain conformations and electron density maps to identify sources of structural variation.
• Performed molecular dynamics (MD) simulations to support structural interpretations.

Main Results:

• All tested advanced predictors (AlphaFold3, AlphaFold2, MultiFOLD, Rosetta, RoseTTAFold, trRosetta) successfully generated MR models for the LCB2 protein.
• Six highly similar structures (within 0.25 Å all-atom RMSD) were determined, with differences largely attributed to a specific crystal contact.
• Observed variations in surface side-chain conformations across the six solutions, though individual electron densities suggested single rotameric states.
• Interpreted these variations as a benign model bias, where models matching actual rotamers enhance electron density.
• Grouping the six structures into a multiconformer ensemble significantly improved crystallographic refinement statistics (Rwork and Rfree).

Conclusions:

• Advanced computational tools are effective in generating starting models for protein structure determination via MR.
• Structural variations in the LCB2 dataset, initially appearing as model bias, represent a multiconformer ensemble reflecting protein dynamics.
• This multiconformer interpretation enhances the accuracy of structural models and provides insights into protein flexibility.
• The study highlights the utility of combining computational predictions with experimental data for a comprehensive understanding of protein structure and dynamics.