Abstract
Live-cell imaging in microfluidic devices now allows the investigation of cellular heterogeneity within microbial populations. In particular, the mother machine technology developed by Wang et al. has been widely employed to investigate single-cell physiological parameters including gene expression, growth rate, mutagenesis, and response to antibiotics. One of the advantages of the mother machine technology is the ability to generate vast amounts of images; however, the time consuming analysis of these images constitutes a severe bottleneck. Here we overcome this limitation by introducing MMHelper (10.5281/zenodo.3254394), a publicly available custom software implemented in Python which allows the automated analysis of brightfield or phase contrast, and any associated fluorescence, images of bacteria confined in the mother machine. We show that cell data extracted via MMHelper from tens of thousands of individual cells imaged in brightfield are consistent with results obtained via semi-automated image analysis based on ImageJ. Furthermore, we benchmark our software capability in processing phase contrast images from other laboratories against other publicly available software. We demonstrate that MMHelper has over 90% detection efficiency for brightfield and phase contrast images and provides a new open-source platform for the extraction of single-bacterium data, including cell length, area, and fluorescence intensity.
Subject terms: Lab-on-a-chip, Biological fluorescence, Bacterial techniques and applications, Imaging techniques
Introduction
Phenotypic heterogeneity is a common feature within isogenic bacterial populations1–3. Cell-to-cell variations have been observed in bacterial growth rate3, virulence4, and resistance to stress1. As a result, it has been suggested that such heterogeneity may allow some cells to survive within fluctuating environments1,5–8 and hence promote evolutionary adaptation9,10. Traditional microbiological assays are based on ensemble measurements and thus unable to measure cell-to-cell differences within microbial populations. In contrast, microfluidics allows the precise manipulation of fluids at the submillimetre level11 and when used in combination with microscopy can be utilised for biological assays with single-cell resolution12,13. Microfluidics has already been adapted for investigating heterogeneity across multiple domains of life. For instance, Hansen et al. developed a protocol which enables measurement of signalling dynamics in single yeast cells14, Li et al. investigated heterogeneity in the migration ability of a population of lung cancer cells15, Yuan et al. looked at the effects of genome deletions on bacterial growth16, Pagliara et al. showed that embryonic stem cells exhibit auxetic properties17, and Otto et al. measured the mechanical deformability of single cells to identify cell sub-populations in whole blood samples18. There are a multitude of microfluidic designs and devices available for investigating single bacterial cells. One popular example is the mother machine19, which provides an ideal platform for tracking single bacterial cells over time while continuously supplying growth nutrients or compounds to be tested such as antibiotics.
Wang et al. designed the mother machine (MM) to allow the trapping of a single mother cell at the dead-end of each of thousands of microfluidic channels and the tracking of its daughter cells over hundreds of generations19. This tool has since been employed to investigate a variety of research questions with single-cell resolution. Tanouchi et al. and Kaiser et al. used the MM to investigate gene regulation20–22. Robert et al. and Uphoff investigated the emergence of mutations in single cells and the dynamics of mutagenesis23,24. Moolman, et al. utilised it to explore protein stoichiometry and dynamics25 whereas Chait et al. used it to engineer bacterial population behaviour26. Multiple groups have use it to investigate single cell response to antibiotics4,27,28, and Yang et al. studied bacterial adaptation under physical confinement29.
Some research groups have developed software which can be used for the analysis of images of bacteria confined in the mother machine30, although most still use scripts customised around their experimental and imaging set-up23,26,27,31. Initially, Arnoldini, et al. developed mmj, a semi-automatic ImageJ plug-in which facilitates the analysis of mother machine images4. However, it is inefficient to use this semi-automated approach on thousands of images. Sachs et al. developed Molyso an unsupervised software implemented in Python30. Molyso, provides a fast and efficient framework capable of analysing 90 GB of mother machine images in 30 min. Nonetheless, their program has limitations which prevent its use by the wider mother-machine community, including not being suitable to analyse standard brightfield images, and constraints on initial channel orientation. Another ImageJ plug in, MoMA, is also available and the authors claim to achieve unprecedented accuracy in segmenting and tracking bacteria22. However, we were unable to install and run MoMA, on any datasets, within a reasonable (2 hour minimum) time period. Using the suggested installation method we successfully installed MoMA but always encountered a FIJI exception error when trying to run the application due to its dependency on Gurobi, even when running on MoMA’s own image set.
In order to overcome the limitations above, we introduce MMHelper, an analysis framework that, to the best of our knowledge, is the first fully automated program applicable to multiple imaging modalities of the mother machine. MMHelper is implemented as a user-friendly python module which detects bacteria confined within the MM and tracks their progeny and fate through time. These detected bacterial regions can then be used to access information on length and area as well as any accompanying fluorescence intensity data. We demonstrate that by using MMHelper, brightfield imaging can be used for extracting phenotypic information from individual bacteria (e.g. length, width, morphology) in the mother machine as well as phase contrast imaging; with the added value that brightfield imaging does not rely on the use of specialised optical components. Furthermore, we have recently used MMHelper to analyse the response to antibiotics of 11,823 single bacteria thus generating novel insight on the physiology of phenotypic variants28. Therefore, we believe that the efficiency and accuracy of MMHelper will assist the investigation of a variety of biological questions by significantly improving the throughput and reliability of mother machine experiments.
Methods
Our image analysis pipeline can be decomposed into two core stages, detection and tracking, which are followed by the extraction of the temporal changes of single-cell parameters including length, width, area, and fluorescence intensity. After determining the imaging modality (1A), each stage is comprised of channel-centric (Fig. 1B,D) and bacteria-centric (Fig. 1C,E) sub-stages. The detection stages (Figs 1B,C and S1) take place independently of the time-point of the experiment and are shown in more detail in Fig. 2A–D and Fig. 3A–D, respectively. In comparison, the tracking stages (Fig. 1D,E) are performed relative to the previous time point (i.e. the t = 0 h left hand panel images are used as a reference for the tracking on the t = 1 h right hand panel images).
Data organisation and loading
Each image is loaded as a multi-dimensional numpy array using the scikit-image module. For experiments including fluorescence images, these arrays are split such that detection is only performed on the brightfield (or phase contrast) images. MMHelper can be run specifically on single images or on image time-series and it also contains a batch run mode. This mode allows the analysis of a whole folder that contains images from tens of different time points and areas of the MM. In this instance, a naming protocol is used to associate images with areas on the chip. Specifically, a string is used at the start followed by an underscore that identifies which MM area the respective image is from. After this underscore, a time stamp is used in order to sort the images in chronological order (e.g. a suitable filename for an image of area 1 of the MM acquired at 12:33:01 on the 16th October 2017 would be: “Area01_171016_123301.tiff”).
Detection
The first stage of the detection process is to determine whether the image is a phase or brightfield image (Fig. 1A). We noted that the pixel intensity distributions of brightfield and phase-contrast images, obtained with similar N.A. objectives, are significantly different. Therefore, we used the skewness of the pixel intensity distribution to detect the imaging modality. As we have a large sample size in terms of pixel count (for a square image that is 1000 pixels in length: ), we used the uncorrected expression for the skewness G132, with the samples third and second central moments of the pixel data, m3 and m3 respectively.
If this equation returns G1 as a positive value, the image is assumed to be phase contrast, whereas a negative value suggests the input image was acquired in brightfield.
After determining the imaging modality, the input image is filtered (Fig. 2A) using a gradient magnitude Sobel edge for phase contrast images or Frangi ridge filter for brightfield images33.
The edge or ridge filtering accentuates the channel outlines, and is followed by Li’s iterative minimum cross-entropy based automated thresholding34 to binarize the image. This mask image is labelled using connected-component labelling, and the labelled regions are filtered based on area to remove non-channel regions.
The resulting channel-outlines are morphologically dilated to close small gaps in the outline, and the subsequent closed regions are filled using a region-filling algorithm. These inner channel regions are extracted as the difference between the outlines and the filled regions (Fig. 2B). The inner-channel perimeters are converted to pixel locations and, by determining the pixel locations that are farthest apart, channel vectors are generated. These vectors are filtered for length to select only regions in a predetermined range (default: 100–400 pixels) based on the images acquired from our typical experimental set-up, however they can be adjusted using a scale factor (see additional parameters section). The resulting vectors correspond to the long channel edges, therefore the perpendicular distance between them is also filtered to ensure that the selected channels correspond to single channels. The resulting channel regions form the basis for a subsequent interpolation stage (Fig. 2C). First, the aforementioned channel regions are analysed to determine the single channel-to-channel spacing, to allow the identification of undetected channels. Using this spacing, the positions of eventually undetected channels are interpolated from the detected channel positions. The detected average channel shape is stamped into each interpolated position. Using the channel contours, the perimeter of each detected channel can be seen in the final output images (Fig. 2D). Note that at least three channels must be detected in any given image to allow the algorithm to attempt interpolation. If two or less channels are detected the algorithm warns the user that it was unable to accurately detect channels in this image, and the frame is not considered for further detection.
The next sub-stage in detection is to detect bacteria within the channels identified from the process above. In these images, the bacteria initially appear darker than the background (Fig. 2D). Therefore, the images are inverted to allow for the use of standard algorithms to detect bright objects on dark background. To do this, the background intensity for each channel is estimated using a rolling ball filter and subtracted from its respective image35 (Fig. 3A). Furthermore, by subtracting the background intensity, the watershed segmentation can remain the same for bacteria located anywhere along the channel profile (Fig. S2).
These channel images are then processed as follows: first each channel image is scale-space filtered36 using a Laplace of Gaussian convolution at multiple scales, and maximum-projected along the scale axis (Fig. 3B). Using these filtered channel images, a threshold value is determined using Li’s algorithm to avoid over-segmentation of empty channels. Each filtered channel image is then binarized using this threshold value and outlines generated by taking the difference between the dilation (grow) and the erosion (shrink) of the initial binary image. An initial crude region-splitting stage is included as occasionally multiple bacteria are detected as a single region, which reduces the accuracy of the region size filtering step. For this, the algorithm uses the marker-controlled Watershed transform37. Markers are generated from all regions greater than a predefined distance from the mask background, and used to delineate bacteria. These regions are finally filtered for width and size (Fig. 3C). Following the initial bacteria segmentation, a second dedicated bacteria-splitting stage was included to improve the segmentation quality of adjacent bacteria (Fig. 3C). The initially detected bacteria are skeletonised and “splits” identified using a combination of distance transformation and pixel intensity, with the threshold values determined using the median and median absolute deviation of all the initially detected bacteria from the original image.
Tracking
The detected channels and bacteria are tracked in two stages. First global frame shift is determined for whole images using cross-correlation based template matching38. This allows channels from consecutive timepoints to be matched using simple distance-based greedy assignment, which matches each point to its nearest neighbour as long as it is also the nearest neighbour to that point. To do so, channel centroid positions are extracted and channels in consecutive frames are linked if each is the nearest neighbour to the other (Fig. 1D). Once channels have been tracked in adjacent time frames, bacteria can be tracked in each channel. This proceeds according to a simple multiple-hypothesis tracking where probabilities of all possible assignments are calculated. These assignments take into account the centroid position and area of each bacterium, as well as adjustable probabilities that each bacterium remains an individual entity (no-change, Fig. 4A), or fades away from the channel (cell death, Fig. 4B), or gives rise to progeny (cell division, Fig. 4C). These events can occur in a number of different combinations to produce the number of bacteria detected in the current frame relative to the preceding frame (t = 1 h compared to t = 0 in Fig. 4A–C). Therefore, a list of all these possible combinations is generated and for each of these possibilities the total number of bacterial divisions that would be required is determined. A probability based on the change in area between the bacteria and its offspring is determined and normalised by the number of divisions. A second probability based on the change in centroid, is calculated taking into account that for each division the change in centroid location is expected to move by half the length of an average bacterium. Finally, the algorithm calculates the likelihood of a cell dividing, lysing, or remaining a single cell between consecutive time points. All three of these probabilities are then multiplied together to determine the overall likelihood that the given event occurred for an individual bacterium. The determined probability for each bacterium within the channel is multiplied to produce an overall probability for the respective combination of events. The resulting, most probable, combination is then used to correctly relabel each bacterium in each image (e.g. second channel from the left in Fig. 4D), with newly generated bacteria assigned a new unique label (e.g. first and second channels from the left in Fig. 4E).
Extraction of single-cell parameters
Once bacteria detection and tracking has been completed, extraction of all quantities of interest can be achieved through the detected and tracked region-based properties. Each bacterium’s length, width, and area are determined using the various standardised algorithms presented via the regionprops function. The binary masks can then be used to extract the raw fluorescence intensity values from the corresponding fluorescence images reporting for example the activity of transcriptional reporters or the intracellular accumulation of spectrally distinct substrates. The background fluorescence is obtained from the empty areas (parts of the channels not containing detected bacteria) of each channel and subtracted from each respective bacterium’s fluorescence intensity. These quantities are then saved in a csv file. We have recently used MMHelper to measure the temporal changes in promoter activity in 11,823 individual Escherichia coli28. Figure 5A–C report the temporal changes in area, length and GFP fluorescence for three representative bacteria, and their progeny, growing in lysogeny broth. The fluorescence reported in Fig. 5C is the mean pixel intensity and the gradual decline in the fluorescence values reported is not due to photobleaching, but is a genuine proxy for the expression of the multi efflux pump tolC (the promoter upstream of GFP in the plasmid carried by the strain), due to the reduction of cellular stress upon continuous exposure to fresh media, similar to the profile we previously observed28.
Additional parameters and module usage
MMHelper can be used to analyse image time series acquired with different microscopy setups (e.g. different objective magnification and numerical aperture, different cameras) by adjusting a single “Scale factor” parameter. Furthermore, the user can specify how many fluorescence image channels are acquired for each brightfield (or phase contrast) image. More information on parameters and how to adjust them will be available on the repository wiki page (https://github.com/jmetz/mmhelper/wiki).
Due to the 2D nature of MMHelper’s detection, it performs the analysis on any image orientation and there is no need for tilt correction. Furthermore, the modular nature of MMHelper makes it suitable for future adaptation to slightly different experimental set ups such as microchemostat devices39.
Statistical comparison
In order to compare the performances of MMHelper and Molyso, we manually drew ground truth detection masks in the images using the freely available GIMP drawing program and used them to quantify three parameters: the Jaccard index, precision and recall values of the automated detection (Fig. S3). We ran both software programs on our own brighfield images, and three independent sets of phase contrast images from (i) the the Locke’s laboratory40, (ii) the work by Sachs, et al. (Molyso)30 and (iii) the work by Kaiser, et al. (MoMA)22. In order to use Molyso on brightfield images, we inverted these images before analysis since the authors did not develop this software for brightfield imaging. We then directly compared the respective values for each parameter, statistical significance was tested by unpaired t test with Welch’s correction, where p ≤ 0.05 is *p ≤ 0.01 is **p ≤ 0.001 is *** and p ≤ 0.0001 is ****respectively.
Results and Discussion
We developed MMHelper to work on both brightfield and phase contrast images with high detection efficiency and accuracy, this also allowing accurate extraction of data from any associated fluorescence images. In order to quantify the performances of our software, we randomly selected 5 of our brightfield datasets28 and analysed image time-series for 4 consecutive time-points, resulting in the analysis of 14 frames containing between 18 and 120 bacteria each. We characterised the detection efficiency as the percentage of bacteria which were detected and, from a total of 562 bacteria across all of the brightfield images, the efficiency was determined as 98 ± 1%. However, in some cases one bacterium was labelled as multiple bacteria or multiple bacteria detected as an individual bacterium. In these circumstances the detection cannot be said to be accurate, therefore we termed detection accuracy as the percentage of bacteria correctly identified by a single label and calculated it to be 80 ± 3% across the 14 previously mentioned brightfield images. Furthermore, we used MMHelper to analyse an image dataset acquired with a phase contrast microscope in the Locke’s laboratory40, obtaining a bacterial detection efficiency of 95 ± 1% and an accuracy of 65 ± 1%. This demonstrates i) the capability of MMHelper to detect bacteria in mother machine images in both brightfield and phase contrast modalities and ii) the capability to work equally well across independent experimental setups.
For each software and each dataset we then measured three different parameters: detection precision as the overlap area between the detected and ground truth masks divided by the detection mask; detection recall as the overlap area divided by the ground truth mask41,42 (Fig. S3); and finally the Jaccard index, defined as the overlap area divided by the total combined area42. The use of precision and recall allows a comparison of the trade-off between ensuring no areas are missed (recall) and how precise the algorithm is, with the Jaccard index representing a combination of these values42. We compared these parameters for MMHelper and Molyso applied to the detection of 310 channels from our brightfield and Locke’s phase contrast images (Fig. S4 and Table 1). The corresponding Kernel Density Estimation for channel detection precision v recall is reported in Fig. 6A for brightfield and Fig. 6B for phase contrast datasets, respectively. Noteworthy, the multi-modal distribution of density for channel detection in brightfield is probably due to small variations in the quality (e.g. focus) of images acquired, resulting in the precision values varying slightly for individual images. For instance, of the total 14 birghtfield frames, the majority clustered around 0.8, one frame had a precision level of 0.9 and two frames had precision levels close to 1 (Fig. S6).
Table 1.
Pipeline | Bright field | Phase | |||||
---|---|---|---|---|---|---|---|
Precision (%) | Recall (%) | Jaccard index (%) | Precision (%) | Recall (%) | Jaccard index (%) | ||
Channels | MMHelper | 77.8 ± 1.9 | 97.6 ± 1.4 | 77.2 ± 3.1 | 53.8 ± 0.2 | 99.4 ± 0.6 | 53.7 ± 0.3 |
Molyso | 64.3 ± 25.7* | 58.9 ± 22.2* | 42.1 ± 21.7* | 79.9 ± 9.0 | 77.1 ± 9.4 | 64.6 ± 12.8 | |
Bacteria | MMHelper | 78.8 ± 14.6 | 76.3 ± 14.0 | 57.1 ± 14.1 | 47.3 ± 15.0 | 96.5 ± 3.5 | 43.9 ± 14.3 |
Molyso | 43.8 ± 21.7* | 12.7 ± 7.9* | 11.4 ± 6.5* | 39.0 ± 17.1 | 19.5 ± 8.2 | 15.2 ± 7.1 |
*Molyso was not specifically developed for brightfield imaging.
As expected, according to the Jaccard index, MMHelper shows significantly better channel detection than Molyso on the brightfield datasets (p ≤ 0.0001, Fig. S4E), although Molyso performed better on the phase contrast dataset (p ≤ 0.0001, Fig. S4B). MMHelper shows a channel detection recall close to 100% for both phase contrast and brightfield images as a result of the detected channels being slightly larger than the ground truth masks, and was significantly better than Molyso (p ≤ 0.0001, Figs S4C and S4D). The detected channels being larger than the ground truth masks was also reflected in the precision values which were slightly lower, with Molyso being significantly better than MMHelper for phase contrast (p ≤ 0.0001, Fig. S4B), although MMHelper was still significantly better for brightfield (p ≤ 0.0001, Fig. S4A). This over-estimation, however, does not affect the level of accuracy of bacterial detection, see below, which is the ultimate aim of this pipeline. Figure 6C reports the Kernel Density Estimation obtained on the precision and recall values for 434 bacteria from brightfield images whereas Fig. 6D shows the Kernel Density Estimation measured for 494 bacteria from phase contrast images.
Secondly, we compared the Jaccard index of the Molyso and MMHelper performances in detecting channels from phase contrast images from the works by Sachs, et al. (Molyso)30 and by Kaiser, et al. (MoMA)22. Surprisingly, in terms of Jaccard index channel detection MMHelper performed better than Molyso on the Molyso image sets (p < 0.0001, Fig. S5F), whereas Molyso performed slightly better than MMHelper on the MoMA’s dataset (p = 0.0104, Fig. S5E). Similar to the results on our datasets, this appeared to be a result of MMHelper detecting slightly larger channels than the ground truth masks. MMHelper performed better in terms of detection recall for both MoMA (p ≤ 0.0001, Fig. S5C) and Molyso image sets (p ≤ 0.0001, Fig. S5D). Finally, Molyso performed better than MMHelper in terms of detection precision on the Molyso (p ≤ 0.0001, Fig. S5B) and MoMA image sets (p = 0.0073, Fig. S5A).
The next set of comparisons was done in terms of bacterial detection which is the ultimate goal of both Molyso and MMHelper. Therefore, ground truth masks were produced for bacteria allowing for the evaluation of bacterial detection precision, recall and Jaccard index for both Molyso and MMHelper. Bacteria detection is more difficult than channel detection, due to the inherent heterogeneity in bacterial shape and size within a clonal population. As a result, the levels of the three parameters are lower relative to channel detection (Table 1). However, according to the Jaccard index, MMHelper demonstrates superior performances compared to Molyso for both imaging modalities on our brightfield and Locke’s lab phase contrast datasets (p ≤ 0.0001, Figs S7E and S7F). In fact, MMhelper also performed significantly better in terms of recall (p ≤ 0.0001, Fig. S7C and Fig. S7D) and precision (p ≤ 0.0001, Fig. S7A for and p = 0.0044, Fig. S7B) on our brightfield and Locke’s lab phase contrast datasets (Table 1). We then compared the two pipelines in detecting bacteria from the MoMA and Molyso image sets. Interestingly, according to the Jaccard index, MMHelper again performed better than Molyso on the MoMA dataset (p = 0.0041, Fig. S8E) and their own dataset (p ≤ 0.0001, Fig. S8F). All the median values for the three parameters are listed in Table 1.
The fact that MMHelper outperformed Molyso in terms of Jaccard index for bacterial detection for all datasets further emphasises the flexibility of MMHelper for use on different experimental set ups as well as different bacterial species. The superior performances of MMhelper are probably due to the fundamental difference in the approaches to detection: the MMhelper algorithm is applied to the 2D images, whereas Molyso reduces 2D images to 1D by using line profiles and projections for channel and bacteria detection, respectively.
Finally, in order to determine the efficiency of our tracking algorithms we quantified the number of correctly tracked channels or bacteria in consecutive frames. In order to decouple tracking accuracy from detection accuracy, we excluded from the image datasets illustrated above any channels or bacteria that were incorrectly detected. MMHelper returned 100% and 94 ± 2% efficiency in channel and bacteria detection on the brightfield image datasets and 100% and 67 ± 4% efficiency in channel and bacteria detection on the phase contrast image datasets.
An obvious benefit of automated image analysis is the removal of human error. In order to demonstrate the superior performances of MMHelper, we analysed a brightfield image and the corresponding fluorescence image both via MMHelper and via a semi-automated approach based on ImageJ and requiring user input. Briefly, three different users measured each bacterium length from the brightfield image by drawing a straight line through the bacterium and using the corresponding intensity plot to determine where the line crossed the edges of the bacterium thus deducting the bacterial length (Fig. S9A). They then drew a box around each bacterium to measure its area (Fig. S9B). Using this same box, the fluorescence pixel intensity was extracted from the corresponding fluorescence image (Fig. S9C). For each bacterium we calculated the mean and standard deviation of these semi-automated measurements (red shaded areas in Fig. S8) and compared these values to the ones obtained via MMHelper (blue circles in Fig. S9). Whereas MMHelper is able to accurately detect the bacterial contour, the semi-automated approach consistently overestimates the area of individual bacteria and underestimates the GFP fluorescence from single bacteria. Therefore, in order to allow a direct comparison between the values obtained via the two approaches, Fig. S4 reports each single-bacterium value normalised to the corresponding mean of all the single-bacterium values. This allows us to demonstrate that MMHelper robustly and accurately extracts single-cell data with 69% of MMHelper measurements falling within 1 S.D. of the mean, 97% within 2 S.D., and 100% within 3 S.D. of the mean of the values obtained via the semi-automated approach (Fig. S9C).
Input images can vary in quality and magnification and the bacterial geometry can vary depending on species, phase of growth, and due to the phenotypic heterogeneity inherent in clonal populations. In order to account some of these variations, some of the input parameters for MMHelper can be varied accordingly. For example, tuning the scale factor accounts for changes in image magnification. Furthermore, we are also developing a graphic interface for the manual correction of MMHelper output, where needed, which aims to make this process both easier and more efficient.
MMHelper, to the best of our knowledge, is the only automated analysis pipeline that has been designed for the analysis of both brightfield and phase contrast images acquired with the mother machine. Some researchers use fluorescent tags in order to perform their image analysis24,29, but this requires exposure to strong light sources that are known to be extrinsic damage-producing agents43. Conversely, MMHelper allows the extraction of single-bacterium length and area measurements from brightfield or phase contrast images, allowing measurements of single-cell parameters such as growth rate and elongation time that are crucial when investigating phenomena such as ageing19,44,45, bacterial susceptibility28,46 and cell size regulation20.
When needed, fluorescence can be used as a reporter for intracellular pH, gene expression, or substrate accumulation. Therefore, MMHelper, will facilitate the study of mutagenesis, gene regulation, and cellular homeostasis at the single cell level. Furthermore, when current microbiological assays are performed at the population level, viable but non-culturable bacteria are overlooked. VBNC cells are a subpopulation of cells which enter a dormant state allowing them to survive otherwise lethal concentrations of antibiotics but they do not resuscitate immediately upon exposure to fresh media47. As a result, they can be responsible for the recalcitrance of chronic infections and act as a stepping stone in the development of antibiotic resistance47. In contrast, our high-throughput system can be used to ensure that non-growing phenotypes can be detected for example during the testing of new antimicrobials or exposure to stress. In this respect, we have recently used MMHelper to demonstrate that persister and viable but non culturable E. coli cells differentially regulate genes associated with tryptophan metabolism before exposure to ampicillin28 opening new opportunities to map the detailed biochemical makeup of these clonal subpopulations.
Conclusion
MMHelper provides an automated framework for the analysis of any type of microscopy images acquired with the mother machine. This automated approach provides large amounts of data with a high level of accuracy in both a time efficient and reproducible manner. For instance, on average it would take a user approximately an hour to analyse a series of 8 consecutive images using ImageJ, whereas MMHelper can acquire the same information in approximately one minute, requiring only a limited amount of manual editing of the output data thanks to the high level of accuracy provided. After thoroughly testing MMHelper to analyse our own mother machine experiments performed on different experimental set-ups and different bacterial strains we are now making this open-source software available for all the research groups already using the mother machine around the world. Finally, we believe that, thanks to the ease of installation and use, MMHelper will be an incentive for researchers from a variety of scientific backgrounds to employ this powerful technology for investigating biological questions with single cell resolution.
Supplementary information
Acknowledgements
We thank C. Schwall from the Locke’s laboratory at the University of Cambridge for providing us with image datasets acquired on their phase contrast microscope. We thank E. Attrill and O. Goode for manually analysing images as part of the comparison process. This work was supported by a Royal Society Research Grant (RG180007), a Wellcome Trust Strategic Seed Corn Fund (WT097835/Z/11/Z), and a MRC Proximity to Discovery EXCITEME2 grant (MCPC17189) awarded to S.P. A.S. acknowledges support from the BBSRC through a SWBio-DTP studentship (BB/M009122/1). J.M. was generously supported by a Wellcome Trust Institutional Strategic Support Award (WT097835MF).
Author Contributions
S.P. developed the project and designed the research. A.S. and J.M. developed the code and implemented it in Python. A.S., J.M. and S.P. wrote the paper. All authors read and approved the final manuscript.
Competing Interests
The authors declare no competing interests.
Footnotes
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Ashley Smith and Jeremy Metz contributed equally.
Contributor Information
Jeremy Metz, Email: j.metz@exeter.ac.uk.
Stefano Pagliara, Email: s.pagliara@exeter.ac.uk.
Supplementary information
Supplementary information accompanies this paper at 10.1038/s41598-019-46567-0.
References
- 1.Balaban NQ, Merrin J, Chait R, Kowalik L, Leibler S. Bacterial persistence as a phenotypic switch. Science. 2004;305:1622–1625. doi: 10.1126/science.1099390. [DOI] [PubMed] [Google Scholar]
- 2.Nikolic N, Barner T, Ackermann M. Analysis of fluorescent reporters indicates heterogeneity in glucose uptake and utilization in clonal bacterial populations. BMC Microbiol. 2013;13:258. doi: 10.1186/1471-2180-13-258. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Kotte O, Volkmer B, Radzikowski JL, Heinemann M. Phenotypic bistability in Escherichia coli’s central carbon metabolism. Mol. Syst. Biol. 2014;10:736. doi: 10.15252/msb.20135022. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Arnoldini M, et al. Bistable expression of virulence genes in salmonella leads to the formation of an antibiotic-tolerant subpopulation. PLoS Biol. 2014;12:e1001928. doi: 10.1371/journal.pbio.1001928. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Ackermann M. A functional perspective on phenotypic heterogeneity in microorganisms. Nat. Rev. Microbiol. 2015;13:497–508. doi: 10.1038/nrmicro3491. [DOI] [PubMed] [Google Scholar]
- 6.Venturelli OS, Zuleta I, Murray RM, El-Samad H. Population Diversification in a Yeast Metabolic Program Promotes Anticipation of Environmental Shifts. PLoS Biol. 2015;13:1–24. doi: 10.1371/journal.pbio.1002042. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Schreiber F, et al. Phenotypic heterogeneity driven by nutrient limitation promotes growth in fluctuating environments. Nat. Microbiol. 2016;1:1–7. doi: 10.1038/nmicrobiol.2016.55. [DOI] [PubMed] [Google Scholar]
- 8.Bódi Z, et al. Phenotypic heterogeneity promotes adaptive evolution. PLoS Biol. 2017;15:1–26. doi: 10.1371/journal.pbio.2000644. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Beaumont HJE, Gallie J, Kost C, Ferguson GC, Rainey PB. Experimental evolution of bet hedging. Nature. 2009;462:90–93. doi: 10.1038/nature08504. [DOI] [PubMed] [Google Scholar]
- 10.New AM, et al. Different Levels of Catabolite Repression Optimize Growth in Stable and Variable Environments. PLoS Biol. 2014;12:17–20. doi: 10.1371/journal.pbio.1001764. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Sackmann EK, Fulton AL, Beebe DJ. The present and future role of microfluidics in biomedical research. Nature. 2014;507:181–189. doi: 10.1038/nature13118. [DOI] [PubMed] [Google Scholar]
- 12.Hol FJH, Dekker C. Zooming in to see the bigger picture: Microfluidic and nanofabrication tools to study bacteria. Science (80-.). 2014;346:1251821–1251821. doi: 10.1126/science.1251821. [DOI] [PubMed] [Google Scholar]
- 13.Okumus B, Yildiz S, Toprak E. Fluidic and microfluidic tools for quantitative systems biology. Curr. Opin. Biotechnol. 2014;25:30–38. doi: 10.1016/j.copbio.2013.08.016. [DOI] [PubMed] [Google Scholar]
- 14.Hansen AS, Hao N, OShea EK. High-throughput microfluidics to control and measure signaling dynamics in single yeast cells. Nat. Protoc. 2015;10:1181–1197. doi: 10.1038/nprot.2015.079. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Li Y, et al. Cell migration microfluidics for electrotaxis-based heterogeneity study of lung cancer cells. Biosens. Bioelectron. 2017;89:837–845. doi: 10.1016/j.bios.2016.10.002. [DOI] [PubMed] [Google Scholar]
- 16.Yuan X, et al. Single-Cell Microfluidics to Study the Effects of Genome Deletion on Bacterial Growth Behavior. ACS Synth. Biol. 2017;6:2219–2227. doi: 10.1021/acssynbio.7b00177. [DOI] [PubMed] [Google Scholar]
- 17.Pagliara S, et al. Auxetic nuclei in embryonic stem cells exiting pluripotency. Nat Mater. 2014;13:638–644. doi: 10.1038/nmat3943. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Otto O, et al. Real-time deformability cytometry: on-the-fly cell mechanical phenotyping. Nat. Methods. 2015;12:199. doi: 10.1038/nmeth.3281. [DOI] [PubMed] [Google Scholar]
- 19.Wang P, et al. Robust growth of Escherichia coli. Curr. Biol. 2010;20:1099–1103. doi: 10.1016/j.cub.2010.04.045. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Tanouchi Y, et al. A noisy linear map underlies oscillations in cell size and gene expression in bacteria. Nature. 2015;523:357–60. doi: 10.1038/nature14562. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Tanouchi Y, et al. Long-term growth data of Escherichia coli at a single-cell level. Sci. Data. 2017;4:170036. doi: 10.1038/sdata.2017.36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Kaiser M, et al. Monitoring single-cell gene regulation under dynamically controllable conditions with integrated microfluidics and software. Nat. Commun. 2018;9:212. doi: 10.1038/s41467-017-02505-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Robert L, et al. Mutation dynamics and fitness effects followed in single cells. Science (80-.). 2018;359:1283–1286. doi: 10.1126/science.aan0797. [DOI] [PubMed] [Google Scholar]
- 24.Uphoff S. Real-time dynamics of mutagenesis reveal the chronology of DNA repair and damage tolerance responses in single cells. Proc. Natl. Acad. Sci. 2018;115:E6516–E6525. doi: 10.1073/pnas.1801101115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Moolman MC, et al. Slow unloading leads to DNA-bound beta2-sliding clamp accumulation in live Escherichia coli cells. Nat Commun. 2014;5:5820. doi: 10.1038/ncomms6820. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Chait, R., Ruess, J., Bergmiller, T., Tkačik, G. & Guet, C. C. Shaping bacterial population behavior through computer-interfaced control of individual cells. Nat. Commun. 8 (2017). [DOI] [PMC free article] [PubMed]
- 27.Bergmiller T, et al. Biased partitioning of the multidrug efflux pump AcrAB-TolC underlies long-lived phenotypic heterogeneity. Science (80-.). 2017;356:311–315. doi: 10.1126/science.aaf4762. [DOI] [PubMed] [Google Scholar]
- 28.Bamford RA, et al. Investigating the physiology of viable but non-culturable bacteria by microfluidics and time-lapse microscopy. BMC Biol. 2017;15:1–12. doi: 10.1186/s12915-017-0465-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Yang D, Jennings AD, Borrego E, Retterer ST, Männik J. Analysis of factors limiting bacterial growth in PDMS mother machine devices. Front. Microbiol. 2018;9:1–12. doi: 10.3389/fmicb.2018.00001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Sachs CC, et al. Image-based single cell profiling: High-throughput processing of mother machine experiments. PLoS One. 2016;11:1–15. doi: 10.1371/journal.pone.0163453. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Norman TM, Lord ND, Paulsson J, Losick R. Memory and modularity in cell-fate decision making. Nature. 2013;503:481–486. doi: 10.1038/nature12804. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Zwillinger, D. & Kokoska, S. CRC Standard Probability and Statistics Tables and Formulae. (Chapman & Hall: New York, 2000).
- 33.Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel enhancement filtering. Med. Image Comput. Comput. Interv. 1998;1496:130–137. [Google Scholar]
- 34.Li CH, Lee CK. Minimum cross entropy thresholding. Pattern Recognit. 1993;26:617–625. doi: 10.1016/0031-3203(93)90115-D. [DOI] [Google Scholar]
- 35.Sternberg S. Biomedical Image Processing. Computer (Long. Beach. Calif). 1983;16:22–34. [Google Scholar]
- 36.Lindeberg T. Feature Detection with Automatic Scale Selection. Int. J. Comput. Vis. 1998;30:79–116. doi: 10.1023/A:1008045108935. [DOI] [Google Scholar]
- 37.Vincent L, Soille P. Watersheds in digital spaces: an efficient algorithm based on\nimmersion simulations. IEEE Trans. Pattern Anal. Mach. Intell. 1991;13:583–598. doi: 10.1109/34.87344. [DOI] [Google Scholar]
- 38.Lewis JP. Fast Template Matching. Vis. interface. 1995;10:120–123. [Google Scholar]
- 39.Nobs JB, Maerkl SJ. Long-term single cell analysis of S. pombe on a microfluidic microchemostat array. PLoS One. 2014;9:e93466. doi: 10.1371/journal.pone.0093466. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Patange, O. et al. Escherichia coli can survive stress by noisy growth modulation. Nat. Commun. 9 (2018). [DOI] [PMC free article] [PubMed]
- 41.Tung F, Wong A, Clausi DA. Enabling scalable spectral clustering for image segmentation. Pattern Recognit. 2010;43:4069–4076. doi: 10.1016/j.patcog.2010.06.015. [DOI] [Google Scholar]
- 42.Taha, A. A. & Hanbury, A. Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool. BMC Med. Imaging15 (2015). [DOI] [PMC free article] [PubMed]
- 43.Rang CU, Peng AY, Poon AF, Chao L. Ageing in Escherichia coli requires damage by an extrinsic agent. Microbiol. (United Kingdom) 2012;158:1553–1559. doi: 10.1099/mic.0.057240-0. [DOI] [PubMed] [Google Scholar]
- 44.Rang CU, Peng AY, Chao L. Temporal dynamics of bacterial aging and rejuvenation. Curr. Biol. 2011;21:1813–1816. doi: 10.1016/j.cub.2011.09.018. [DOI] [PubMed] [Google Scholar]
- 45.Stewart EJ, Madden R, Paul G, Taddei F. Aging and death in an organism that reproduces by morphologically symmetric division. PLoS Biol. 2005;3:0295–0300. doi: 10.1371/journal.pbio.0030045. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Greulich P, Scott M, Evans MR, Allen RJ. Growth-dependent bacterial susceptibility to ribosome-targeting antibiotics. Mol. Syst. Biol. 2015;11:796. doi: 10.15252/msb.20145949. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Ayrapetyan, M., Williams, T. C., Baxter, R. & Oliver, J. D. Viable but Nonculturable and Persister Cells Coexist Stochastically and Are Induced by Human Serum. 83, 4194–4203 (2015). [DOI] [PMC free article] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.