Machine learning coarse-grained potentials of protein thermodynamics

Area of Science:

• Computational biology
• Biophysics
• Machine learning in structural biology

Background:

• Understanding protein dynamics is crucial for interpreting structure-function relationships in biological processes.
• Simulating protein dynamics accurately and efficiently remains a significant scientific challenge.

Purpose of the Study:

• To develop a novel approach for simulating protein dynamics using machine learning-based coarse-grained potentials.
• To accelerate molecular dynamics simulations while maintaining thermodynamic accuracy.

Main Methods:

• Constructed coarse-grained molecular potentials using artificial neural networks grounded in statistical mechanics.
• Trained models on a dataset of approximately 9 milliseconds of unbiased all-atom molecular dynamics simulations across twelve diverse proteins.
• Validated the coarse-grained models against all-atom simulations and experimental data.

Main Results:

• Coarse-grained models achieved over three orders of magnitude acceleration in dynamics simulation speed.
• The models preserved the thermodynamics of the simulated systems.
• Identified relevant structural states and their energetics, comparable to all-atom simulations.
• A single coarse-grained potential successfully integrated all twelve proteins and predicted experimental features of mutated proteins.

Conclusions:

• Machine learning-based coarse-grained potentials offer a feasible and efficient method for simulating protein dynamics.
• This approach can aid in understanding protein structure-function relationships and essential biological processes.
• The developed potentials show promise for integrating diverse protein systems and predicting effects of mutations.