Accelerating Molecular Dynamics Simulations with Foundation Neural Network Models using Multiple Time-Step and Distillation

TL;DR

A distilled multi-time-step (DMTS) strategy to accelerate molecular dynamics simulations using foundation neural network models that conserves accuracy, preserving both static and dynamical properties, while enabling to evaluate the costly model only every 3 to 6 fs depending on the system.

Abstract

We present a distilled multi-time-step (DMTS) strategy to accelerate molecular dynamics simulations using foundation neural network models. DMTS uses a dual-level neural network where the target accurate potential is coupled to a simpler but faster model obtained via a distillation process. The 3.5 Å-cutoff distilled model is sufficient to capture the fast-varying forces, i.e., mainly bonded interactions, from the accurate potential allowing its use in a reversible reference system propagator algorithms (RESPA)-like formalism. The approach conserves accuracy, preserving both static and dynamical properties, while enabling to evaluate the costly model only every 3 to 6 fs depending on the system. Consequently, large simulation speedups over standard 1 fs integration are observed: nearly 4-fold in homogeneous systems and 3-fold in large solvated proteins through leveraging active learning for enhanced stability. Such a strategy is applicable to any neural network potential and reduces their performance gap with classical force fields.

Figures (5)

• Figure 1: Average velocity autocorrelation spectra of hydrogen atoms for bulk water. The reference single time step (STS) simulation are shown in black. Distilled Multi time step (DMTS) simulations are shown for different integration time steps (2 fs, 3 fs, 4 fs, 5 fs, and 6 fs), with corresponding heavy-mass repartitioning (HMR) variants represented by dashed lines of the same color. Inset zooms in high frequency regions containing MTS integration artifacts.
• Figure 2: Radial distribution function, $g(r)$, as a function of the distance ($r$) in Å. The solid blue curve corresponds to the DMTS simulation with the on-the-fly system-specific model and the orange one to DMTS with the small generic model. Both DMTS simulations used an inner time step of 1fs and an outer time step of 5fs. The dotted green curve corresponds to the reference STS simulation.
• Figure 3: Hydration free energy of small molecules (water, ethanol, benzene, trimethylamine, diethylsulfide and acetic acid) using DMTS with system-specific model (blue points) and with small generic model (orange points) compared to the STS result. MAE is 0.091kcal/mol and 0.103kcal/mol, RMSE is 0.124kcal/mol and 0.138kcal/mol, R$^2$ is 0.996 and 0.995 for respectively the system-specific model and the generic model.
• Figure 4: a) Average velocity autocorrelation spectra of hydrogen atoms for the phenol-lysozyme complex with HMR, comparing STS with MTS 3fs and 4fs. Insets zoom in high frequency regions containing MTS integration artifacts. b) Distribution (in log-scale) of the norm of the force differences for the generic model with respect to the FENNIX-Bio1(M) reference model. The distributions are estimated over the first 500 frames of a phenol–lysozyme in water simulation.
• Figure 5: Time evolution of the protein backbone RMSD and the ligand’s Distance to Bound Configuration (DBC)salari2018streamlined during a 20ns simulation of the lysozyme–phenol complex in water. Results obtained with the implemented MTS integrator (solid lines) are compared to those from a reference STS simulation (dotted lines). DMTS simulations use either the generic or active learned models combined with the FeNNix-Bio1(M) potential with an internal time step of 1.75fs and an external time step of 3.5fs with HMR for generic model and an internal-external time step of 2fs-4fs with HMR for the active learned model.