On the Use of Local Diffusion Models for Path Ensemble Averaging in Potential of Mean Force Computations

TL;DR

This paper introduces a method to create surrogate stochastic models from steered molecular dynamics data, enabling efficient estimation of potential of mean force with fewer simulation paths by bootstrapping local diffusion models.

Contribution

The paper presents a novel approach to construct global surrogate diffusion models from local models fitted to SMD data, facilitating efficient PMF computations with limited samples.

Findings

Global models resemble a single family of diffusion processes.

Surrogate models enable bootstrapping to generate many work paths.

Method is applicable to expensive simulations and optical tweezers experiments.

Abstract

We use a constant velocity steered molecular dynamics (SMD) simulation of the stretching of deca-alanine in vacuum to demonstrate a technique that can be used to create surrogate stochastic processes using the time series that come out of SMD simulations. The surrogate processes are constructed by first estimating a sequence of local parametric models along a SMD trajectory and then a single global model is constructed by piecing the local models together through smoothing splines (estimation is made computationally feasible by likelihood function approximations). The calibrated surrogate models are then "bootstrapped" in order to simulate the large number of work paths typically needed to construct a potential of mean force (PMF) by appealing to Jarzynski's work theorem. When this procedure is repeated for a small number of SMD paths, it is shown that the global models appear to come from a single family of closely related diffusion processes. Possible techniques for exploiting this observation are also briefly discussed. The findings of this paper have potential relevance to computationally expensive computer simulations and experimental works involving optical tweezers where it difficult to collect a large number of samples, but possible to sample accurately and frequently in time.