Accelerating Protein Folding Molecular Dynamics Using Inter-Residue Distances from Machine Learning Servers.

Area of Science:

• Computational biology
• Structural bioinformatics
• Biophysics

Background:

• Predicting native protein structures computationally is crucial but challenging for larger proteins.
• Current physics-based methods (computational molecular physics) require significant computational resources and scale poorly with protein size.

Purpose of the Study:

• To develop a novel computational method that improves the efficiency and accuracy of protein structure prediction.
• To integrate machine learning-derived contact information into physics-based molecular dynamics simulations.

Main Methods:

• Developed ML x MELD x MD, a hybrid approach combining machine learning (ML) contact predictions with MELD (a Bayesian accelerator for molecular dynamics).
• Used residue contacts from trRosetta-predicted distograms as spatial restraints within MELD's atomistic molecular dynamics (MD) simulations.
• Validated the method in the CASP14 blind prediction experiment.

Main Results:

• ML x MELD x MD successfully predicted 13 native protein structures with errors below 4.5 Å in CASP14.
• The method achieved accurate predictions for proteins up to 250 amino acids long.
• Demonstrated significantly improved scaling of simulation time with protein length compared to traditional MD (e^0.023 vs e^0.168).

Conclusions:

• Integrating ML predictions into physics-based MD simulations offers a powerful strategy for accurate and efficient protein structure prediction.
• ML x MELD x MD represents a significant advancement in computational protein modeling, overcoming limitations of previous methods for larger proteins.