High-accuracy protein structures by combining machine-learning with physics-based refinement.

Area of Science:

• Computational biology
• Structural biology
• Biophysics

Background:

• Protein structure prediction is crucial for understanding biological function.
• Homology modeling and machine learning (ML) methods like AlphaFold have advanced prediction accuracy.
• Predicting structures for sequences lacking templates remains a challenge.

Purpose of the Study:

• To improve protein structure prediction accuracy by combining ML models with physics-based refinement.
• To evaluate the performance of the combined approach against existing methods in the CASP competition.
• To assess the suitability of the refined models for experimental structure determination.

Main Methods:

• Utilized AlphaFold, a state-of-the-art ML-based protein structure prediction model.
• Applied physics-based refinement using molecular dynamics (MD) simulations.
• Integrated ML-derived distance restraints with MD simulations for enhanced accuracy.
• Benchmarked the combined method against other prediction tools in the latest CASP (Critical Assessment of protein Structure Prediction) round.

Main Results:

• The combined ML and MD refinement approach significantly outperformed all other tested prediction methods in CASP.
• Achieved highly accurate global and local protein structures.
• Demonstrated high accuracy at functionally important interface residues.
• Generated models suitable for crystal structure determination via molecular replacement.

Conclusions:

• Combining AlphaFold predictions with physics-based molecular dynamics refinement offers superior protein structure prediction.
• This integrated approach enhances accuracy, particularly at critical functional sites.
• The resulting high-quality models serve as excellent starting points for experimental structure determination, accelerating biological discovery.