Transferable Generative Models Bridge Femtosecond to Nanosecond Time-Step Molecular Dynamics

TL;DR

The paper tackles the sampling bottleneck in atomistic molecular dynamics by introducing Transferable Implicit Transfer Operators (TITO), a deep generative framework that learns long-lag transition densities across diverse chemistries. By using a continuous normalizing flow with equivariant flow matching, TITO can generate statistically faithful trajectories at lag times Δt much larger than femtosecond steps, achieving speedups up to ~10^4 while preserving Boltzmann equilibrium and relaxation kinetics. It demonstrates transferability across small molecules and peptides, reproducing thermodynamics and kinetics, uncovering metastable states, and extrapolating to larger peptides with corrective priors. The work establishes a new paradigm for multi-timescale molecular dynamics, enabling accelerated, thermodynamically consistent exploration of conformational landscapes and kinetic processes with potential impact in chemistry and biophysics.

Abstract

Understanding molecular structure, dynamics, and reactivity requires bridging processes that occur across widely separated time scales. Conventional molecular dynamics simulations provide atomistic resolution, but their femtosecond time steps limit access to the slow conformational changes and relaxation processes that govern chemical function. Here, we introduce a deep generative modeling framework that accelerates sampling of molecular dynamics by four orders of magnitude while retaining physical realism. Applied to small organic molecules and peptides, the approach enables quantitative characterization of equilibrium ensembles and dynamical relaxation processes that were previously only accessible by costly brute-force simulation. Importantly, the method generalizes across chemical composition and system size, extrapolating to peptides larger than those used for training, and captures chemically meaningful transitions on extended time scales. By expanding the accessible range of molecular motions without sacrificing atomistic detail, this approach opens new opportunities for probing conformational landscapes, thermodynamics, and kinetics in systems central to chemistry and biophysics.

Figures (17)

• Figure 1: Transferable Implicit Transfer Operators (TITO): A multi–time-scale surrogate model for molecular dynamics that is transferable across systems. Starting from an initial condition (black cross), TITO generates molecular dynamics ensembles for diverse molecules at arbitrary lag times.
• Figure 2: TITO accurately predicts both thermodynamics and kinetics. Small molecules (top) and tetra-peptides (bottom). Top row: Projection onto the first two TICA components and comparison of VAMP timescales between MD and TITO-generated samples for a representative molecule. Bottom row: Aggregated evaluation across systems: Jensen–Shannon divergence of TICA projections (left), VAMP-2 gap (center), and top-10 relative error (right). Black arrows denote the position of the example molecule within each histogram.
• Figure 3: TITO accurately samples two orders of magnitude slower dynamics than the training data. (A) Coverage and (B) precision, see \ref{['sec:metrics']} section, w.r.t to RE of projections onto the first two TICs for training set-like simulations (MD, 36.5 $\mathrm{ns}$), TITO (32 $\mu \mathrm{s}$) and an ensemble of ultra-short MD initialized with TITO samples (10 $\mathrm{ps}$ per sample). TITO achieves better coverage of the conformational space than MD. (C) Example propiolamide test-set molecule: MD fails to sample the most meta-stable basin, while TITO (160 $\mu\mathrm{s}$, initialized from MD) successfully transitions into the dominant basin, recovers the Boltzmann distribution and estimates the right order of magnitude for the corresponding VAMP implied time-scale, estimated from ultra-long MD (16 $\mathrm{ms}$). Structural insights of this transition and alternative TICA projections are provided in Suppl. Figs. \ref{['fig:si-mol21-torsions']} and \ref{['fig:si-long-tica']}, respectively.
• Figure 4: TITO recapitulates transients, fast vibrations and potential energies. (A) Free energy of the transition probability estimated with TITO for AYTG (test set) at increasing lag times (top to bottom). The first column shows conditional free energies projected onto the first two TICA components (TICs). Contours represent TITO samples, while 2D histograms correspond to Markov state model estimates from MD data. The red cross marks the simulation’s initial state. The second and third columns display marginal distributions along each TIC. Nested samples are generated with 5 steps. (B) Free energy profiles of bond distances (top) and angles (bottom). (C) Probability density of the non-harmonic potential energy.
• Figure 5: Extrapolation to larger systems. Free energy landscape and VAMP time-scales of a TITO model trained on tetra-peptides and performing 5 $\mathrm{ns}$ single-step sampling for penta, hexa and hepta and octa- peptides.