Mother machine workflow, schematic, and applications.

(1.1) Mother machine schematic. Growth channels flank a central flow cell that supplies fresh media and whisks away daughter cells. In a typical experiment, numerous fields of view (FOVs) are imaged for several hours. (1.2) Fluorescence images of E. coli strains expressing cytoplasmic YFP [1] (left) and markers for the replisome protein DnaN and division protein FtsZ (right) [10]. (1.3) The mother machine setup allows long-term monitoring of the old-pole mother cell lineage [1] and has other versatile applications, including (1.4) the study of the mechanical properties of bacterial cells by applying controlled Stokes forces [9].

MM3 workflow and example images.

(2.1) The MM3 image analysis pipeline takes raw mother machine images and produces cell objects. Processes (rounded rectangles) are modular; multiple methods are provided for each. (2.2) Example images from the processing of one growth channel in a single FOV. The growth channel is first identified, cropped, and compiled in time. All cells are segmented (colored regions). Lineages are tracked by linking segments in time to determine growth and division (solid and dashed lines, respectively), creating cell objects.

napari-MM3 interface.

The napari viewer enables interactive analysis of mother machine data with real-time feedback and fast debugging. Raw data shown is from MG1655 background E coli expressing the fluorescence protein YPet fused to the replisome protein DnaN [10].

Performance metrics for napari-MM3.

Processing times were measured on an iMac with a 3.6 GHz 10-Core Intel Core i9 processor with 64 GB of RAM and an AMD Radeon Pro 5500 XT 8 GB GPU. Tensorflow was configured to use the AMD GPU according to [40]. The GPU was used in U-Net training and segmentation steps. The dataset analyzed is from [10] and consists of 26 GB of raw image data (12 hours, 262 time frames, 2 imaging planes, 34 FOVs, and ∼35 growth channels per FOV). Note that while the Otsu segmentation method is slightly faster than the U-Net, it also requires a background subtraction step, such that the total runtimes of the two methods are comparable.

Testing napari-MM3 on external datasets.

Quality of segmentation masks produced by running napari-MM3 on a subset of published datasets from other groups [13,15,22,41]. As exact boundaries are difficult to determine by eye, we considered a cell to be correctly segmented if the Intersection over Union of the predicted mask and ground truth mask was greater than 0.6 (Methods). To evaluate the quality of the segmentation, we report the Jaccard index [42,43].

Overview of mother-machine image analysis tools.

A comparison of several published imaging methods. ‘2D’ or ‘1D’ segmentation indicates whether the cells are labeled in an image and analyzed in two dimensions, or projected onto a vertical axis and analyzed in one dimension. Several tools support the use of deep learning (in place of or in addition to classical computer vision techniques).

Comparison of various image analysis approaches.

(4.1) A time series of a typical cell growing in a nutrient-rich medium. The birth size, division size, and added size are indicated. (4.2) The adder principle ensures cell size homeostasis via passive convergence of cell size to the population mean. (4.3) We analyzed multiple datasets from our lab using MM3, DeLTA, and BACMMAN, and obtained robust correlations between birth length, doubling time, elongation rate, and added length. Representative results from one dataset [10] for MG1655 background E. coli grown in MOPS glycerol + 11 amino acids are shown, with 9,000 – 13,000 cells analyzed depending on the method. (4.4) Distributions of key physiological parameters are independent of the analysis methods. The data and code used to generate this figure are available at [46]

Effect of systematic deviation in segmentation output from different methods.

5.1. Otsu / random walker and U-Net segmentation masks. The classical method systematically yields masks that are 5%-10% larger than the other methods. 5.2. We confirmed that this discrepancy occurs consistently across the cell cycle. 5.3 We trained the Omnipose model on masks generated by either napari-MM3-Otsu or napari-MM3-U-Net separately. 5.4. The systematic discrepancy in the training data masks propagated to the output of the trained models.

MM3 parameter values for processed external datasets.

Parameters which were changed from the default values are shaded in yellow. Ollion et al., Jug et al. and Sachs et al. datasets were segmented with the non-learning method, while the Lugagne et al. dataset was segmented using the U-Net method.

Inexpensive fabrication of cell loader with 3D printing.

Evaluating segmentation output of napari-MM3 Otsu and U-Net methods

Old-pole aging phenotype is strain specific.

Cells imaged with fluorescence often show signs of aging in the old-pole “mother” cell. For instance, in the dataset analyzed in Figure 4 (E. coli MG1655 with the fluorescent protein YPet fused to DnaN), we observed systematic differences in cell elongation rate and size between the old-pole cell at the end of the growth channel and its sisters, which inherit the new pole (top center). However, this asymmetry is not universal. Using napari-MM3’s Otsu segmentation method, we re-analyzed previously published data obtained without fluorescence illumination [32], and found that the old-pole and new-pole cell elongation rates varied only on the order of 1% (lower center), while in the dataset obtained under fluorescence imaging, the old-pole mother cells grow 7-10% slower than the new pole cells. Cells born third or fourth from the closed end of the channel also grow slower than the old-pole mother (right). The asymmetry in growth rate between old-pole and new pole cells persists across time (right, inset). These results are consistent with a previous survey [63], which found that most evidence for aging in E. coli comes from studies utilizing fluorescent proteins for visualization.