Continued hardware and software improvements have established molecular dynamics (MD) simulations as a vital tool for molecular behavior prediction. One of the greatest challenges brought on by these advances is the abundance of data generated by these simulations. These data may be difficult to parse in a timely manner, and in a way that reveals all of the important information about the system at hand. Cluster analysis is a tool that's very commonly used to partition the output of MD simulations into structural states. Currently available clustering algorithms, however, tend not to take advantage of the time-evolution of the MD trajectory, instead treating the trajectory as a collection of independent data points. Moreover, popular algorithms that rely on stored pairwise distances of data points run into memory complexity issues that require coarse-graining to resolve, leading to loss of resolution. We present a novel clustering method that builds upon existing density-peak based clustering to resolve both of these issues. Using change detection to demarcate trajectory segments, we allow ourselves to take into account the temporal dynamics of the system, as well reduce memory complexity by orders of magnitude by storing only inter-segment distances. Benchmarking this method on a variety of synthetic and biomolecular datasets, we find it to be highly performant, allowing for overlap of assigned states, and improving resolution to the point of being able to pick up on dimensions omitted from the dataset. We also show how this method may be used to distinguish metastable states from transition regions, as well as be applied to systems without a priori determined intrinsic coordinates.

Change-detection-aided processing of molecular dynamics trajectories

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.