Deep learning based reconstruction of embryonic cell-division cycle from label-free microscopy time-series of evolutionarily diverse nematodes

Microscopy of cellular dynamics during embryogenesis of non-model organisms can be tech- nically challenging due to limitations of molecular labelling methods. Label-free differential interference contrast (DIC) microscopy of the first embryonic cell division of nematodes related toCaenorhabditis eleganshas been successfully employed to examine the constraints and divergence of intra-cellular mechanisms during this asymmetric cell division. However, identifying stages of the cell division cycle were performed interactively, pointing to a need to automate of cell stage identification from DIC microscopy. To this end, we have trained deep convolutional neural networks (CNNs), both pre-existing such as ResNet, VGGNet and EfficientNet, and a customized shallow network, EvoCellNet, to automatically classify first-embryonic division into the stages: (i) pro-nuclear migration and (ii) centration and rotation, (iii) spindle elongation and (iv) cytokinesis, with all networks performing with 91% or greater accuracy. The activations of the networks superimposed on the images result in segmentation-free detection of intracellular features such as pro-nuclei, spindle and spindle- poles in case of the shallow EvoCellNet, while ResNet, VGGNet and and EfficientNet detect large-scale, features that are less biologically meaningful. The UMAP space representation combined with support vector machines (SVM) allows for stage boundary identification and recovers a cyclical map connecting the states (i) to (iv) of the division. This approach could be used to automate quantification of cell division stages and sub-cellular dynamics without explicit labelling in label-free microscopy.