Clustering algorithms to analyze molecular dynamics simulation trajectories for complex chemical and biological systems

Molecular dynamics (MD) simulation has become a powerful tool to investigate the structure-function relationship of proteins and other biological macromolecules at atomic resolution and biologically relevant timescales. MD simulations often produce massive datasets containing millions of snapshots describing proteins in motion. Therefore, clustering algorithms have been in high demand to be developed and applied to classify these MD snapshots and gain biological insights. There mainly exist two categories of clustering algorithms that aim to group protein conformations into clusters based on the similarity of their shape (geometric clustering) and kinetics (kinetic clustering). In this paper, we review a series of frequently used clustering algorithms applied in MD simulations, including divisive algorithms, agglomerative algorithms (single-linkage, complete-linkage, average-linkage, centroid-linkage and ward-linkage), center-based algorithms (K-Means, K-Medoids, K-Centers, and APM), density-based algorithms (neighbor-based, DBSCAN, density-peaks, and Robust-DB), and spectral-based algorithms (PCCA and PCCA+). In particular, differences between geometric and kinetic clustering metrics will be discussed along with the performances of different clustering algorithms. We note that there does not exist a one-size-fits-all algorithm in the classification of MD datasets. For a specific application, the right choice of clustering algorithm should be based on the purpose of clustering, and the intrinsic properties of the MD conformational ensembles. Therefore, a main focus of our review is to describe the merits and limitations of each clustering algorithm. We expect that this review would be helpful to guide researchers to choose appropriate clustering algorithms for their own MD datasets.