Modeling Changes in Molecular Dynamics Time Series as Wasserstein Barycentric Interpolations

Jovan Damjanovic, Yu-Shan Lin, James M. Murphy

Abstract: Molecular dynamics (MD) simulations are a powerful computational tool for elucidation of molecular behavior. These simulations generate an abundance of high-dimensional time series data and parsing these data into a human- interpretable format is nontrivial. Clustering trajectory segments obtained via change point detection has been shown to lower memory complexity and yield improved partitioning resolution of the time series compared to the state of the art. However, accurate change point placement is often inhibited by the presence of gradual changes between long-lived metastable states. The trajectory regions corresponding to these gradual changes are not well-modeled by a single distribution, and therefore are frequently over-segmented. In this work, we model such regions using weighted Wasserstein barycentric interpolations between adjacent metastable states, allowing for gradual changes to be resolved correctly. The improved detection performance of our proposed method is demonstrated on a range of toy and real MD simulation data, showing significant potential for faithfully modeling and compressing complex MD simulations.

Submission Type: Full Paper