Classification of first embryonic division stages of multiple Caenorhabditis species by deep learning.

Abstract

The first embryonic division of Caenorhabditis elegans is a model for asymmetric cell division, and identifying the stages of cell division across related species could improve our understanding of the divergence of cellular events and mechanisms. Comparative microscopy of evolutionarily divergent species continues to rely on label-free differential interference contrast (DIC) microscopy due to technical challenges in molecular tagging, with the identification of cell division stages still relying on label-free microscopy. Here, we compare multiple deep convolutional neural networks (CNNs) trained to automate cell stage classification in DIC microscopy movies and interpret the results, with code and classification weights released as OpenSource. The networks are trained to identify if a single frame of a time-series belongs to one of the four morphologically distinct stages: (i) pro-nuclear migration, (ii) centration and rotation, (iii) spindle displacement and (iv) cytokinesis, that had been manually labeled. Three previously described networks, ResNet, VggNet, and EfficientNet, and a customized shallow network, which we refer to as EvoCellNet, achieved 91% or greater accuracy in test data from 23 different nematode species. We find activation vectors of the CNNs of the sparse EvoCellNet correlate with spatial localization of intracellular features of multiple species, such as pro-nuclei, spindle, and spindle-poles. While the pipeline is robust when applied to comparable DIC time-series of C. elegans and C. briggsae embryos, distinct from those on which it was trained and tested, successful classification is limited to images with conserved morphological features. Thus, deep learning networks can be used to generalize the morphological changes across species of nematode embryos, capturing chronology based on low-level intracellular features with biological relevance.