Toward a chemically transferable coarse-grained model of protein-based materials learned from tripeptide fragments

Proteins and their assemblies pose significant potential for materials design due to their chemical diversity, biocompatibility, and facile synthesis. Unfortunately, their design space is vast, with only a fraction concomitant with functional sequences that can self-assemble. Conventional all-atom molecular dynamics simulations (AAMD) can be used to expedite exploration of this space but are too costly to span the time and length scales relevant to self-assembly. We introduce a coarse-grained (CG) framework that aims to improve transferability by decoupling intra- and intermolecular physics and representing all nonbonded intermolecular interactions with a minimal two-body Lennard-Jones-plus-Gaussian (LJG) potential. Intermolecular parameters are learned via relative-entropy minimization to match pair-distance statistics from AAMD simulations of GxG tripeptides (guest residue X) for all 210 amino acid pair combinations. Application of these parameters yields good agreement with AA potentials of mean force (PMF) across three benchmarks: the QVVAG dimerization motif, the coiled-coil E3/K3, and the globular complex Barnase–Barstar. We further predict assembly outcomes for tripeptides previously characterized experimentally: fibrillar (VFF), micellar (VYV), and non-assembling (ECG). These results show that a simple LJG basis trained only on tripeptide statistics can generalize from molecular PMFs to mesoscale structure and energetics relevant to protein-materials design.