Exploring the World of Lipids with Machine-Learned Coarse-Grained Models
Lipid bilayers are central to cellular organization and function, yet their atomistic modelling remains constrained by its high computational cost. While coarse-grained (CG) models can overcome this bottleneck, they are limited in accuracy due to the use of simplified functional forms based on molecular mechanics. Using a hybrid functional form combining physics-based terms with a graph neural network, we present implicit-solvent, machine-learned CG models for POPC. When trained on single-lipid data, the model recovers the atomistic conformational landscape. By adding short atomistic simulations of a small flat bilayer to the training data and combining force matching with a denoising objective, we train a model that accurately reproduces bilayer structure and dynamics, in closer agreement with the atomistic reference than existing models, like Martini3 and DryMartini. Beyond its training data, the model stabilizes curved geometries, including vesicles, and captures complex cooperative phenomena such as bilayer and vesicle self-assembly and pore sealing. While limitations remain, particularly regarding chemical transferability, our results establish machine-learned coarse-grained potentials as a promising path towards robust and accurate CG models of lipids and membranes.