DeepCycle reconstructs a cyclic cell cycle trajectory from unsegmented cell images using convolutional neural networks

Abstract The advent of single‐cell methods is paving the way for an in‐depth understanding of the cell cycle with unprecedented detail. Due to its ramifications in nearly all biological processes, the evaluation of cell cycle progression is critical for an exhaustive cellular characterization. Here, we present DeepCycle, a deep learning method for estimating a cell cycle trajectory from unsegmented single‐cell microscopy images, relying exclusively on the brightfield and nuclei‐specific fluorescent signals. DeepCycle was evaluated on 2.6 million single‐cell microscopy images of MDCKII cells with the fluorescent FUCCI2 system. DeepCycle provided a latent representation of cell images revealing a continuous and closed trajectory of the cell cycle. Further, we validated the DeepCycle trajectories by showing its nearly perfect correlation with real time measured from live‐cell imaging of cells undergoing an entire cell cycle. This is the first model able to resolve the closed cell cycle trajectory, including cell division, solely based on unsegmented microscopy data from adherent cell cultures.

Appendix Figure S2: Virtual class classification accuracies and confusion matrices of the DeepCycle network classifier . During the training phase of the DeepCycle network, dual channel cell images are classified into the four virtual classes defined on the FUCCI2 fluorescence intensity plane (mAG-Geminin and mKO2-Cdt1). We report the classification accuracies as well as the confusion matrices for each class with the DeepCycle network trained from cell images with both the brightfield and hoechst channels as input (A), brightfield alone (B) and Hoechst alone (C). The confusion matrices and classification accuracies were calculated on test data. Figure S3 : FUCCI2 fluorescence signal heterogeneity among pre-and postmitotic times . A . Fluorescence intensities of the FUCCI2 mAG-Geminin and mKO2-Cdt1 reporters for 1024 tracks as a function of the time relative to the division event (cytokinesis). Each track has been manually validated and the division frame has been labeled. The X coordinate value is the time of measurement subtracted by the time of division. The fluorescence intensities are the mean pixel values over a square of 13.2x13.2 μm centered around the centroid of the cell nucleus in the GFP or the Cy3 channels for the mAG-Geminin or the mKO2-Cdt1 readout, respectively. Typically, all cells measured before the division event (negative time values) are committed to divide and present the typical FUCCI2 cell cycle trends. On the other hand, not every cell which has already divided (positive time values) will divide again, introducing more heterogeneity in the FUCCI2 fluorescence readout. B. Quantification of the FUCCI2 readout heterogeneity for the pre-and post-mitotic times. The pre-mitotic median standard deviation of the FUCCI2 intensities are lower than the post-mitotic intensities (median standard deviation of pre-and post-mitotic times are 187.9, 340 for mKO2-Cdt1 ***; 216.0, 225.6 for mAG-Geminin N.S.). C. Distribution of mKO2-Cdt1 intensities (red) and mAG-Geminin (green). The cell distributions are unimodal for the cells in their pre-mitotic times. A bimodal trend appears in both the mKO2-Cdt1 and mAG-Geminin intensities in the post-mitotic times further suggesting a separation between the dividing and non-dividing cells. This observation provides additional information on the increasing heterogeneity after cellular division.

Appendix
Appendix Figure S4: Initiation of the DeepCycle trajectory.
To quantitatively characterize the progression of the fluorescence intensities along the UMAP projection, the neighborhood graph based SOM algorithm was employed to define the path which captures the main circular axis of this projection and was called DeepCycle trajectory. A . The DeepCycle trajectory (blue line) is a trajectory derived by the SOM algorithm from the UMAP coordinates. B. Average FUCCI2 intensities (mKO2-Cdt1: red, mAG-Geminin: green) as a progression over the SOM clusters of the training set (average n=325, total n=9753) . The FUCCI trends in function of the DeepCycle pseudotime of all cells is shown in Figure 2B.
Appendix Figure S5 : Results from a DeepCycle model trained on single brighfield channel cell images. A . Representation of the four virtual classes from the FUCCI 2 marker intensities used for the network training. This figure is a replica of Figure 1C, added here for completeness. B . Confusion matrix and virtual class prediction accuracies. The confusion matrix was calculated on test data. This figure is a replica of Appendix Figure S2B, added here for completeness.
C . Low dimensional UMAP projection with the inferred DeepCycle trajectory (SOM, see Methods for more details). The colorbar indicates the SOM clusters index. D . Projection of the FUCCI 2 fluorescence intensities (red: mKO2-Cdt1, green: mAG-Geminin), in function of the DeepCycle pseudotime (SOM cluster index). E . Evaluation of the model performance at estimating the CC time quantified as the correlation between the DeepCycle pseudotime and the CC time (bold central line shows the mean, thin lines show the standard deviation, spearman r=0.94 , two-sided p-value<0.001, ***, n=50 ). The correlation is computed for 50 cells undergoing a full cell cycle, unseen during the network training.
Appendix Figure S6 : Results from a DeepCycle model trained on single Hoechst channel cell images.
A . Representation of the four virtual classes from the FUCCI 2 marker intensities used for the network training.
This figure is a replica of Figure 1C, added here for completeness. B . Confusion matrix and virtual class prediction accuracies. The confusion matrix was calculated on test data. This figure is a replica of Appendix Figure S2C, added here for completeness. C . Low dimensional UMAP projection with the inferred DeepCycle trajectory (SOM, see Methods for more details). The colorbar indicates the SOM clusters index. D . Projection of the FUCCI 2 fluorescence intensities (red: mKO2-Cdt1, green: mAG-Geminin), in function of the DeepCycle pseudotime (SOM cluster index). E . Evaluation of the model performance at estimating the CC time quantified as the correlation between the DeepCycle pseudotime and the CC time ( bold central line shows the mean, thin lines show the standard deviation, spearman r=0.92 , two-sided p-value<0.001, ***, n=50 ). The correlation is computed for 50 cells undergoing a full cell cycle, unseen during the network training. Figure S7 : Results from a DeepCycle model trained on two channel images (Hoechst and brightfield) with FUCCI2 intensities obtained from segmented nuclei. A . Representation of the four virtual classes from the FUCCI 2 marker intensities averaged over the segmented cell nuclei. B . Confusion matrix and virtual class prediction accuracies. The confusion matrix was calculated on test data. C . Low dimensional UMAP projection with the inferred DeepCycle trajectory (SOM, see Methods for details). The jet colormap indicates the SOM cluster indexes. D . Projection of the FUCCI 2 fluorescence intensities (red: mKO2-Cdt1, green: mAG-Geminin), in function of the DeepCycle pseudotime (SOM cluster index). E . Evaluation of the model performance at estimating the CC time quantified as the correlation between the DeepCycle pseudotime and the CC time ( bold central line shows the mean, thin lines show the standard deviation, spearman r=0.96 , two-sided p-value<0.001, ***, n=50 ). The correlation is computed for 50 cells undergoing a full cell cycle, unseen during the network training.