Development of Coarse-Grained Lipid Force Fields Based on a Graph Neural Network.

Area of Science:

• Computational chemistry
• Biophysics
• Materials science

Background:

• Coarse-grained (CG) models accelerate molecular simulations of large biological systems like cell membranes.
• Achieving both computational efficiency and atomic-level accuracy in CG lipid models is a persistent challenge.
• Graph neural networks (GNNs) have shown potential as accurate force fields for CG simulations, particularly for proteins.

Purpose of the Study:

• To develop the first GNN-based coarse-grained (CG) lipid models for simulating lipid bilayers.
• To evaluate the accuracy and efficiency of these GNN-based CG lipid models compared to all-atom (AA) simulations.
• To explore the transferability and performance enhancement of GNNs trained on different lipid systems.

Main Methods:

• Generated datasets from all-atom (AA) simulations of 1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC), 1,2-dioleoyl-sn-glycero-3-phospho-L-serine (DOPS), and mixed DOPC/DOPS lipid bilayers.
• Developed GNN-based CG lipid models using the TorchMD-GN architecture.
• Validated models by comparing structural correlations and dynamics with AA simulations; tested performance on lipid self-assembly and vesicle simulations.

Main Results:

• The developed GNN-based CG lipid models accurately reproduced structural correlations from AA simulations.
• Lipid dynamics were accelerated by 9.4 times using the GNN-based CG models.
• Models demonstrated temperature transferability and improved performance when trained on lipid bicelles for self-assembly and vesicle simulations.

Conclusions:

• GNN-based CG force fields represent a promising advancement for efficient and accurate large-scale membrane simulations.
• This approach offers a powerful tool for studying complex membrane dynamics and lipid behavior.
• Training strategies, such as using lipid bicelles, can further enhance the predictive power of GNN lipid models.