Computational & AI Journal of Chemical Theory and Computation (ACS)

Equivariant graph network converts Martini 3 lipid membranes to atomistic detail

  • MemBack, an SE(3)-equivariant graph network, predicts heavy-atom CHARMM36 lipid coordinates from Martini 3 coarse-grained configurations in one pass, then cleans up automatically.
  • Per-lipid error stayed low and similar for lipid species and membranes held out of training.
  • In a held-out red-blood-cell membrane, thickness, area per lipid and chain order matched atomistic references after restrained minimization alone, with phase-separated domains preserved; very large systems also stayed consistent.
Points Based on the abstract

Thickness, area per lipid and chain order after only restrained minimization is a decent test, but those are bulk averages that a model can hit while local packing is off. I'd want headgroup hydration, lipid-protein contacts and cholesterol orientation checked, plus a comparison against CG2AT-style fragment backmapping on cost and quality, before trusting it with a protein-embedded membrane.