MemBack: An Equivariant Graph Neural Network for Backmapping Lipid Membranes

We present MemBack, an SE(3) equivariant graph neural network designed to convert coarse-grained lipid membrane simulations to atomistic resolution. This is a crucial step in multiscale molecular simulation, but backmapping complex lipid membranes has proven difficult. MemBack solves this by predicting heavy atoms from Martini 3 configurations in a single pass, then applying automated post-processing.

Across a wide range of chemically diverse systems, MemBack demonstrated impressive accuracy. The mean superposition-free per-lipid heavy-atom RMSD was just 0.65 Å. MemBack maintained this accuracy even when applied to membrane systems and lipid types not present in the training data.

Most importantly, MemBack preserves the structural organization of the membrane across different resolutions. In a challenging 16-component red-blood-cell membrane model excluded from training, MemBack successfully reproduced key membrane properties after only restrained minimization. This includes maintaining correct bilayer thickness, lipid area, and acyl-chain order.

In a more complex case, MemBack was applied to a native phase-separated Martini 3 membrane. It successfully retained the lateral organization of ordered and disordered domains, directly after backmapping, without any atomistic equilibration.

Finally, MemBack was tested on native Martini 3 systems containing up to 1.4 million reconstructed atoms. The results showed that these systems were consistent with their original coarse-grained configurations and could be seamlessly integrated into atomistic simulations. This makes MemBack a powerful tool for efficiently bridging the gap between Martini 3 and CHARMM36 membrane simulations.