SAPPHIRE-Based Clustering

Molecular dynamics simulations are a popular means to study biomolecules, but it is often difficult to gain insights from the trajectories due to their large size, in both time and number of features. The SAPPHIRE (States And Pathways Projected with HIgh REsolution) plot allows a direct visual inference of the dominant states visited by high-dimensional systems and how they are interconnected in time. Here, we extend this visual inference into a clustering algorithm. Specifically, the automatic procedure derives from the SAPPHIRE plot states that are kinetically homogeneous, structurally annotated, and of tunable granularity. We provide a relative assessment of the kinetic fidelity of the SAPPHIRE-based partitioning in comparison to popular clustering methods. This assessment is carried out on trajectories of n-butane, a beta-sheet peptide, and the small protein BPTI. We conclude with an application of our approach to a recent 100 mu s trajectory of the main protease of SARS-CoV-2.