Developing epithelial tissues as active materials: Tracking cell dynamics in dorsal closure using machine learning

Dorsal closure in Drosophila melanogaster embryos is a key model system for cell sheet morphogenesis and wound healing. Understanding system dynamics, regulation and causal relations requires a quantitative understanding of the mesoscopic mechanical and dynamic properties of this “active soft material”. We utilized deep learning to automatically and robustly detect and temporally track various features, even in noisy microscopy movies: individual cell shapes in two distinct tissue types, cell junctions and edge lengths, and tissue topology. Past studies of epithelial dynamics were restricted to semi-manual segmentation of cell shapes and thus suffered from relatively low statistics. Our automatized algorithm reduces processing time for 1000 frames from weeks to 20 min and allows us to harvest high-quality temporal data from ~1000 cells per embryo. Epithelial cells in dorsal closure exhibit oscillations and contribute to progressive cell sheet movements, while showing a large variability in individual shapes and dynamics. Using spatial and temporal correlations and unsupervised machine learning techniques, we detect subtle behavioral phenotypes and emerging dynamical pattern on basis of the cell’s high-dimensional features.