PATITO: Transferable Generative Molecular Dynamics for Peptide Assembly

Supramolecular systems emerge through slow collective interactions among many molecules, making their dynamics costly to resolve with molecular dynamics (MD). We introduce PATITO, a transferable generative molecular dynamics framework that learns transferable implicit transfer operators (TITO) for peptide assembly. Conditioned on molecular identity, topology, and periodic geometry, PATITO recursively generates stochastic long-time trajectories. On held-out peptide sequences, it reproduces structural and dynamical assembly behaviour, maintains molecular fidelity, and predicts late structures more accurately than direct structure generation under comparable data budgets. The learned dynamics also transfer across concentrations when the relevant local environments are represented during training. Although demonstrated for peptide assembly, PATITO supports variable molecular graphs in periodic systems and provides a general route to GenMD of supramolecular and other collective molecular processes.