SupraTITO: Transferable Generative Molecular Dynamics for Supramolecular Systems

The periodic formulation is the paper’s clearest strength. By decomposing displacements into a common molecular translation and an internal bond-vector change under minimum-image conventions, the model keeps bonded sites together through wrapped generation and exact topology preservation. The systematic comparison with a direct ensemble prediction baseline trained on identical data is fair and rigorous: finite-lag learning yields more accurate late-time ensembles and is more data-efficient, corroborating prior observations in the single-molecule GenMD literature.

Several limitations temper the scope. The benchmark relies on one MD trajectory per sequence, so the model cannot capture polymorphic pathways or quantify pathwise variability; the authors concede that 'Peptide assembly can follow multiple pathways and produce distinct metastable or polymorphic morphologies.' Additionally, transfer to unseen system sizes is tested but markedly weaker than concentration transfer, and the architecture is not rotation equivariant and scales quadratically with system size.

The coarse-grained Martini 2.1 representation also restricts fidelity. Assembly accuracy is bounded by the reference force field, and specific interactions such as hydrogen bonding and detailed side-chain packing are absent. The result is a useful surrogate for collective morphology and encounter dynamics, but not yet for atomistically resolved assembly mechanisms.

The evidence broadly supports the core claims. Sequence-conditional transfer is demonstrated with Pearson correlations above 0.98 for six observables across 80 held-out sequences, and concentration-response curves follow MD trends for unseen dilute conditions. The comparison with direct ensemble prediction is fair because both models share the same training trajectories and architectural backbone.

However, dynamical fidelity rests partly on descriptive diagnostics. The authors explicitly caution that autocorrelations are 'interpreted as a descriptive comparison of temporal persistence rather than as an equilibrium correlation function or an independent kinetic-rate estimate,' which weakens the case that the model accurately resolves kinetic pathways rather than temporally correlated structures. Related work is positioned fairly against single-molecule GenMD methods, though the prior TITO framework already established transferability, making SupraTITO’s incremental contribution more architectural (periodic boundaries) than conceptual.

Experimental detail is unusually thorough. Architecture dimensions, exact distance bin boundaries (Equation 29), optimizer settings ($10^{-4}$ learning rate, zero weight decay, gradient clipping at 0.1), batch size, training duration, and inference protocol are all reported, together with computational cost estimates. The main blockers are that code and data are not yet available—the authors state they 'will be made publicly available upon acceptance'—and exact GROMACS topology and run input files are not included, though the Martini 2.1 protocol is cited. Reproducing the 400-trajectory benchmark from scratch would be expensive without released artifacts.