Spatiotemporal feature learning for actin dynamics.

The social amoeba Dictyostelium discoideum is a standard model system for studying cell motility and formation of biological patterns. D. discoideum cells form protrusions and migrate via cytoskeletal reorganization driven by coordinated waves of actin polymerization and depolymerization. Assembly and disassembly of actin filaments are regulated by a complex network of biochemical reactions, exhibiting sensitivity to external physical cues such as stiffness, composition and surface topography of the extracellular matrix, as well as the presence of external electric fields. In this study, we investigate whether the cellular microenvironment, and in particular the presence of electric fields and the nano-topography type, can be directly inferred from images or videos of actin waves. We employ three machine learning techniques to analyze the resulting videos: dictionary learning, scattering transforms, and optical flow. We predict the type of the extracellular environment by observing actin waves frame-by-frame and identifying key visual features that help classify cell motion by the microenvironment type. Our analysis reveals that the decomposition of static images into an adaptive basis of visual primitives provides a robust approach to classifying cells by the nano-topography type. In contrast, predicting whether cells are moving under the influence of an external electric field requires tracking of stable cellular features such as corners and edges over a period of time. We expect our computational approach to be useful in many settings where non-trivial collective dynamics is observed with the help of fluorescent labeling and video microscopy.