Navigating protein landscapes with a machine-learned transferable coarse-grained model.

Area of Science:

• Computational Biology
• Protein Dynamics
• Machine Learning in Biochemistry

Background:

• All-atom molecular dynamics simulations are highly predictive but computationally expensive.
• Developing a computationally efficient coarse-grained (CG) model with universal predictive power for proteins remains a significant challenge.

Purpose of the Study:

• To create a universal, computationally efficient coarse-grained (CG) force field for protein simulations.
• To achieve prediction performance comparable to all-atom models but with significantly reduced computational cost.

Main Methods:

• Combined deep learning techniques with a large dataset of all-atom protein simulations.
• Developed a bottom-up CG force field characterized by chemical transferability.
• Enabled extrapolative molecular dynamics on novel protein sequences.

Main Results:

• The developed CG model accurately predicts metastable states (folded, unfolded, intermediate structures).
• Successfully models fluctuations in intrinsically disordered proteins.
• Predicts relative folding free energies for protein mutants with high efficiency, orders of magnitude faster than all-atom methods.

Conclusions:

• Demonstrates the feasibility of a universal, machine-learned CG model for protein simulations.
• Highlights the potential of deep learning to accelerate molecular dynamics and protein structure prediction.