Estimation and Inference of Diffusion Coefficients in Complex Biomolecular Environments.

Abstract

The 1-D diffusion coefficient associated with a charged atom fluctuating in an ion-channel binding pocket is statistically analyzed. More specifically, unconstrained and constrained molecular dynamics simulations of potassium in gramicidin A are studied. Time domain transition density based inference methods are used to fit simple stochastic differential equations and also to carry out frequentist goodness of fit tests. Particular attention is paid to varying the time between adjacent time series observations due to the well-known "non-Markovian noise" that can appear in this system due to inertia and other unresolved coordinates influencing the dynamics. Different types of non-Markovian noise are shown by the goodness of fit tests to be statistically significant on vastly different time scales. On intermediate scales, a Markovian model is not rejected by the tests; models calibrated at these intermediate scales demonstrate a predictive capability for some physical quantities. However, in this intermediate regime, ergodic sampling does not occur over the length of a time series, but a local diffusion coefficient is deemed statistically acceptable for the observed raw data. It is demonstrated that a linear mixed effects model can be used to summarize the variation induced by slow unresolved degrees of freedom acting as a non-Markovian noise source. The utility of quantitative criteria for assessing low-dimensional stochastic models calibrated from time series generated by high-dimensional biomolecular systems is briefly discussed. Less coarse-grained data summaries of this type show promise for better understanding the kinetic signature of unresolved degrees of freedom in time series coming from simulations and single-molecule experiments.