Application of Anomaly Detection to Identify Important Features of Protein Dynamics

Molecular dynamics (MD) simulations are a popular tool for the study of protein dynamics. Recent machine-learning-based structure prediction methods, such as AlphaFold, can provide a broad variety of initial protein structures for MD simulation. Hence, the development of methods to enhance the practicality of MD simulation (such as efficient sampling or detection of collective variables) is increasingly important. Identifying a small number of elements or features that can describe biological phenomena from MD trajectories serves as a basis for these methods. In this study, we applied the anomaly detection method based on sparse structure learning of the element correlation within MD trajectories to identify important features associated with state transitions. This approach was tested on the correlation of residue-residue distances from the open- and closed-state simulations of T4 lysozyme and the holo- and apo-state simulations of the PDZ3 domain. This has clear implications for understanding cooperative motion through its combination with a dimension reduction technique.