Transferable Coarse Graining via Contrastive Learning of Graph Neural Networks

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Coarse-grained (CG) force fields are crucial for simulating large biomolecules efficiently.
• Developing accurate and transferable CG force fields is challenging due to complex interactions and parameterization difficulties.
• Current methods struggle to capture the nuances required for precise biomolecular simulations.

Approach:

• A novel machine learning approach using graph neural networks (GNNs) to represent CG force fields.
• Parameterization of GNN models is achieved using atomistic simulation data.
• A transferable GNN implicit solvent model was developed using extensive atomistic configurations of proteins.

Key Points:

• The GNN model significantly improves solvation free energy estimations compared to existing methods.
• The model accurately reproduces configurational distributions observed in explicit solvent simulations.
• Demonstrated reasonable transferability of the GNN model beyond the training dataset.

Conclusions:

• This machine learning-based strategy offers a powerful bottom-up approach for constructing accurate CG force fields.
• The GNN model advances the capability to simulate complex biomolecular systems with enhanced realism and efficiency.
• Provides a foundation for developing more sophisticated and reliable CG models in computational biophysics.