Size-and-Shape Space Gaussian Mixture Models for Structural Clustering of Molecular Dynamics Trajectories

TL;DR

This paper introduces shape-GMM, a novel clustering method for molecular dynamics trajectories that uses a weighted alignment and Gaussian mixture models to identify structural states with high accuracy, even in complex cases.

Contribution

The paper presents shape-GMM, a new approach combining weighted maximum likelihood alignment with GMMs to improve structural clustering of molecular configurations.

Findings

Successfully distinguishes structures indistinguishable by RMSD.

Supports a 4-state folding/unfolding mechanism in protein simulations.

Achieves kinetic detail comparable to state-of-the-art methods.

Abstract

Determining the optimal number and identity of structural clusters from an ensemble of molecular configurations continues to be a challenge. Recent structural clustering methods have focused on the use of internal coordinates due to the innate rotational and translational invariance of these features. The vast number of possible internal coordinates necessitates a feature space supervision step to make clustering tractable, but yields a protocol that can be system type specific. Particle positions offer an appealing alternative to internal coordinates, but suffer from a lack of rotational and translational invariance, as well as a perceived insensitivity to regions of structural dissimilarity. Here, we present a method, denoted shape-GMM, that overcomes the shortcomings of particle positions using a weighted maximum likelihood (ML) alignment procedure. This alignment strategy is then built into an expectation maximization Gaussian mixture model (GMM) procedure to capture metastable states in the free energy landscape. The resulting algorithm distinguishes between a variety of different structures, including those indistinguishable by RMSD and pair-wise distances, as demonstrated on several model systems. Shape-GMM results on an extensive simulation of the the fast-folding HP35 Nle/Nle mutant protein support a 4-state folding/unfolding mechanism which is consistent with previous experimental results and provides kinetic detail comparable to previous state of the art clustering approaches, as measured by the VAMP-2 score. Currently, training of shape-GMMs is recommended for systems (or subsystems) that can be represented by $\lesssim$ 200 particles and $\lesssim$ 100K configurations to estimate high-dimensional covariance matrices and balance computational expense. Once a shape-GMM is trained, it can be used to predict the cluster identities of millions of configurations.