Tight contacts with neighbours to form a cohesive sheet of cells is a fundamental property of multicellular organisms and underpins organ development and function. Conversely, signalling pathways necessary to maintain junctions are often targeted by pathogens and underlie key mechanisms of diseases of the vasculature, heart and different epithelial organs.

Attachment to neighbouring cells has a distinct configuration in different cell types and is dynamically remodelled in homeostasis and diseases. In epithelial tissues, the characteristic cell–cell adhesion site appears as a straight and tense or stiff junction, represented by an apparent contiguous, adjoining staining of E-cadherin receptors (the epithelia-specific cadherin protein). Following stimuli such as growth factor treatment or oncogene expression in epithelial cells, the dynamic nature of cell–cell contacts is manifested in a variety of ways: disturbances in the configuration of the contacting interface between cells, fragmentation of cadherin staining or thinning out of the distribution of receptors from contacting cell borders (Braga et al., 2000; Erasmus et al., 2016; Frasa et al., 2010; Lozano et al., 2008; Nola et al., 2011).

However, a linear junction appearance does not apply to other cell types that also require strong cell–cell attachment. Intercalated discs, specialised junctions in cardiomyocytes, sustain considerable mechanical stress with heart beating (Ehler, 2016; Vermij et al., 2017). Extensive remodelling of the intercalated discs composition (Estigoy et al., 2009) and architecture is observed in cardiac `aging (Sessions and Engler, 2016; Tribulova et al., 2015), diabetes-induced cardiopathies (Adeghate and Singh, 2014), arrythmogenic cardiomyopathies (Calore et al., 2015) and cardiac hypertrophy and failure (Lyon et al., 2015). During vascular homeostasis, endothelial cell–cell contacts may have a similar appearance as epithelial junctions (Dejana and Orsenigo, 2013; Malinova and Huveneers, 2018). Upon stimulation with inflammatory agonists (Radeva and Waschke, 2018), endothelial contacts undergo adjustments to increase permeability, changing from a linear to a zig-zag configuration and the appearance of gaps between cells (Malinova and Huveneers, 2018). Collectively, the above data demonstrate that distinct patterns of junctions are stimulus-dependent and reflect the specific destabilization (or strengthening) of cadherin receptors at contact sites in various cell types.

Despite the extensive scientific progress in our understanding of how cell–cell contacts are modulated, how these distinct phenotypes of junction modulation are fully attained is still unclear. A major road block to furthering mechanistic studies on junction regulation is the restricted capability and efficiency of the quantitative image analysis currently available. Existing imaging platforms (i.e. Cell Profiler) (Carpenter et al., 2006; McQuin et al., 2018) are fantastic resources for cell biologists. For Drosophila or nematode epithelia, several software are available for quantification of morphogenesis that robustly detect cell boundaries during morphogenesis (i.e. during dorsal closure or germ band extension) (Curran et al., 2017; Sumi et al., 2018). In contrast, for mammalian epithelial cells, available systems are not always suitable for the precise delineation of cell borders and the output of current pipelines is mostly morphometric parameters (i.e. cell size, shape, number or texture) (Campbell et al., 2017; McQuin et al., 2018; Yin et al., 2013). Cell–cell borders of mammalian cells are particularly difficult to detect because of cytoplasmic noise, irregular shapes (Held et al., 2011) and variable junction phenotypes, particularly when junctions are severely disrupted. This hitherto prevents an objective approach to analyse regulation of cell–cell contacts of mammalian cells and tissues.

Nevertheless, previous computer vision studies of mammalian junctions report the automated quantification of a single heterotypic junction (e.g. tumour-immune cell contact or host-pathogen contact) (Graus et al., 2014; Merouane et al., 2015), morphometry of mammary gland spheroids (Härmä et al., 2014) or dynamics of VE-cadherin contacts during cell rearrangements in angiogenesis (Bentley et al., 2014). Disruption of cell–cell contacts has been assessed in high-throughput manner by coupling junction segmentation with cell tracker and endothelia stimulation (Seebach et al., 2015), cadherin intensity at junctions (Erasmus et al., 2016) or indirectly, by increased inter-nuclear distance as cells scatter (Loerke et al., 2012).

Notwithstanding these successful studies, available methodology does not enable quantification of distinct patterns of organization of receptors or junction morphometry that are readily identified by the human eye. In addition, manual methods available for junction quantification rely on intensity levels and thresholding, which are not appropriate to detect junction attributes such as alterations in shape, length, fragmentation or continuity of cell–cell contacts. Non-intensity-based attributes of junctions are usually defined visually and/or painstakingly analysed via laborious user-dependent quantification of individual junctions. For example, the switch between a straight to undulated cell–cell contact occurs without apparent changes in receptors levels at contact sites (Otani et al., 2006). In this case, rather than alterations of receptor levels at junctions, it involves impaired signalling of the small GTPase Cdc42 to modulate the amount of contraction, making cell–cell contacts less stiff and tense (Oda et al., 2014; Otani et al., 2006).

To address the above issues, we developed a semi-automated pipeline, Junction Mapper, which can fully capture the distinct patterns of junction perturbation by diverse stimuli in various cell types. It can efficiently (i) identify cell boundaries and cell–cell corners, (ii) describe phenotypes of junction architecture and (iii) quantify parameters that reflect the distribution and organization of junctional markers along the contacting interface. To broaden the software suitability to different models, we validate the robustness of the Junction Mapper software in endothelial cells and cardiomyocytes that show distinct receptor organization and architecture when compared to epithelial cells. The repertoire of parameters distinguishes subtle differences of junction disruption and provides a fingerprint for each stimulus, with insights into modes of action and how efficient and functional junctions are.

We envisage that the analytical capabilities of Junction Mapper will be invaluable for the scientific community to perform quantitative image analysis in mechanistic and translational studies of cell–cell contacts. Most importantly, the generation of tools to facilitate unbiased phenotype identification will be a major step forward to understand how junction dynamics are modulated in homeostasis and pathologies of different tissues.

In normal epithelia, a junction between neighbouring cells usually appears as a straight, taut line, with E-cadherin receptors uniformly distributed along the contacting interface (Figure 1A). Junctions are delimited by corners between three or more cells, where a specialised type of contacts are formed (tricellular junctions) (Figure 1A). Distinct stimuli disrupt the above junction architecture in different ways, from minor reduction in levels to complete removal of adhesion receptors from contacting cells (Figure 1B). Concomitant with changes in levels, junction configuration and architecture are also compromised, which are not always captured by intensity measurements.

Figure 1

Download asset Open asset

We designed and validated a semi-automated system (Figure 2), Junction Mapper, that builds from our quantitative analysis of images from RNAi screens (Erasmus et al., 2016). Our previous software defines an E-cadherin mask to calculate the intensity specifically around junctions as a percentage of thresholded area of the whole original image. The E-cadherin mask is also used to subtract an ROI from an image of a distinct marker (i.e. F-actin), so that mostly the signal localized at contacts is considered. Junction Mapper implements novel quantification tools (corners, length and area) and a variety of novel primary and secondary parameters expressed per individual junction.

Figure 2 with 2 supplements see all

Download asset Open asset

The main advantage of Junction Mapper is to efficiently segment junctions in a variety of cell types and of different junction patterns, from linear to fragmented or disrupted contacts. The software detects the contacting interface between two cells and obtains a skeletonized edge map (summarised stepwise in Figure 2—figure supplement 1; see Materials and methods). Although semi-automated, users can adjust the outline manually by removing incorrect or adding missing lines, with subsequent refinement of the line location and geometry by the software algorithm (detailed in Appendix 1). A dilation step is applied (user controlled) to define the area to be quantified around the skeletonized map (Figure 2—figure supplement 1). To identify individual junctions for further quantification, the software then automatically identifies each cell–cell corner (i.e. point of contact between three or more cells, see its mathematical definition in Appendix 1). The number and location of corners in each cell can also be manually adjusted by the user (Appendix 1).

Images obtained at different resolutions can be used for analyses in Junction Mapper (Supplementary file 1). However, the highest quality images possible should be used, as resolution may impact the ability to detect individual clusters of the junction marker. For defining the automated skeletonized edge of each image, a suitable signal-to-noise ratio is necessary to ensure enough contrast to differentiate staining at junctions from the cytoplasm (Figure 2—figure supplement 2). The higher the Peak Signal to Noise ratio, the easier it is for the program to automatically delineate the cell outline skeleton. For efficient skeleton definition the ratio should be above 22 dB (Figure 2—figure supplement 2). Using the automated skeleton as a start point, the user can refine the cell outline map by manually drawing or removing lines to close any gaps or correct deviations, particularly in cells with reduced staining at cell–cell contacts (Figure 2—figure supplement 1 to 2). The higher the fragmentation of junctions in the image, the more user input is necessary to define the final skeleton outline for quantification.

User-controlled threshold selection is done by inspection of the image with a slider component on the software interface (Appendix 1). Thresholding aims to reduce background without removing pixels from contact areas. The skeleton obtained is then projected onto the original thresholded image to segment the area of interest and proceed with quantification of each junction for up to two different junctional markers per image (see below). The measurement of the different parameters is performed simultaneously and automatically, in an unbiased manner (Figure 2). The output per junction is produced as an Excel file that contains the selected image, skeleton and the quantification of pre-designed measurements and parameters (see below).

Primary parameters and validation

We envisage that different phenotypes of junction perturbation may require a distinct set of parameters to appropriately capture the disruption features (Figure 1). Towards this goal, we propose a variety of parameters to be used as bona fide readouts and measure information that is based on intensity, area and length of the interface and cell–cell contact (Figure 3A; Appendix 2). The following concepts are defined and used herein. Interface is the contacting membrane between two neighbouring cells and delimited by cell–cell corners. A cell–cell contact or junction is the region of the interface covered by adhesion receptors (e.g. cadherin staining). Junctions may or may not extend to the whole of the interface (corner-to-corner), and can appear fragmented or dotted (Figure 1, Figure 3A). Area is the dilated region around the skeleton and is set to encompass the width of a junctional marker staining (which can vary in thickness). Contour is the length measurement of the outline of the skeleton between defined points (interface or junction). Finally, the straight-line length is defined as the shortest distance (Euclidian distance) between corners that form the boundaries of one junction or interface.

Figure 3 with 2 supplements see all

Download asset Open asset

Primary parameters obtain the basic metrics (area, contour and straight-line length) of each interface and junction selected in the monolayer (Figure 3A, Appendix 2). A key innovation of the Junction Mapper is its ability to quantify junction marker staining that is not contiguous and that does not extend to both corners. New parameters assess area or contour occupied by fragments of the junctional staining (Fragmented Junction Contour and Junction Area). To validate the length-based measurements, individual contacts disrupted by H-Ras expression were quantified and, as predicted, decreasing values for the contour of the interface, junction or fragmented junction were revealed for each assessed junction (Figure 3B–C). Although measurements of contacts of control cells also decline between these three parameters, they did not decline to the same extent as the ones of junctions from expressing cells (Figure 3C). The area-based measurements Interface Area and Junction Area of the same junctions followed similar pattern (Figure 3D). We envisage that the primary parameters are useful to show extension or retraction of the contacting interface/junctions, increased fragmentation of the staining of junction marker or fluctuations of global staining intensity at cell–cell contacts.

We next assessed the impact of user-defined settings (dilation and threshold) on the primary parameters. The variable dilation settings define areas for quantification and permits the user to account for different thickness of the staining at cell–cell contacts, wavy or undulated junctions. Increasing the dilated area did not affect the length of contacting interface or fragmented junction (Figure 3—figure supplement 1A–C), but positively correlated with E-cadherin intensity and area (Figure 3—figure supplement 1D). Similarly, larger dilation values better captured the amount of VE-cadherin present in endothelial junctions as they acquire a zig-zag conformation after thrombin stimulation (Figure 3—figure supplement 1E–F,H). A trade-off is necessary between the amount of dilation and thresholding, so that the contribution of cytoplasmic staining is minimized with larger dilation values.

The effect of thresholding was tested using the same images (Figure 3—figure supplement 2A–B,E–F). Of note is that the precise outline of VE-cadherin zig-zag staining was not recognized by the edge map produced (red line, Figure 3—figure supplement 2E–F). Increasing the threshold applied to the original images decreased the measured cadherin area and intensity (Figure 3—figure supplement 2D, H) and the length of junction fragments was severely reduced (Figure 3—figure supplement 2C,G; see definition in Appendix 2). The interface contour values were not affected by thresholding, as the interface is delimited by corners and independent of intensity values (Figure 3—figure supplement 2C,G). Taken together, we concluded that the primary parameters measure the length and area of various features as predicted and are selectively modulated by user-controlled settings in accordance to their definition (Appendix 1).

Secondary parameters

The availability of a repertoire of parameters is a unique advantage of the software for mapping distinct phenotypes. Secondary parameters (Scheme 1, Appendix 2) are derived from the primary metrics above and aim to normalise the measurements to the size of each junction or contacting interface (length or area). The Linearity Index measures how straight an interface between two cells is, as proposed by Takeichi and colleagues (Otani et al., 2006). Coverage Index is a length measurement of the percentage of the interface length that is covered by the junction marker staining and has been previously used manually in the lab (Lozano et al., 2008). Three different parameters quantify the distribution of a junctional marker along the interface. First, Interface Occupancy, measures the area occupied by the marker within the Interface Area of each junction. Second, Junction Protein Intensity per Interface area calculates the fluorescence intensity within the Interface Area. Finally, Cluster Density is the junction protein intensity level within the area delimited by its staining (i.e. Fragmented Junction Area, which considers any fragmentation of the staining; additional parameters are described in Appendix 2). We predict that the secondary parameters are more useful to compare the accumulation or removal of specific markers, their density and relative changes when comparing across different samples.

Scheme 1

Download asset Open asset

The Coverage Index parameter calculated by Junction Mapper was validated by manual quantification using FIJI on cells expressing active Rac1 (Supplementary file 2A-B). The manual quantification used straight-line measurements (Lozano et al., 2008), rather than the more precise contour measurement by Junction Mapper (Scheme 1). Using either quantification, manual or Junction Mapper, a significant statistical difference was observed between values of control and Rac1-expressing cells (Supplementary file 2C). Lower values were obtained with the manual quantification method as predicted when using straight-lines for measurement. However, when controls values were compared between the two methods, no statistical difference was obtained (Supplementary file 2C).

Thus, the parameters give results as expected from previous manual methods. In addition, the reproducibility of secondary parameters was assessed in independent biological replicates by the same user. Across biological replicates of cells overexpressing active Rac1, the absolute values obtained for each sample were slightly different (Scheme 1B-D). However, the overall result is the same between replicates: a reduction of Coverage Index and Interface Occupancy of E-cadherin at junctions from Rac1 expressing cells (Scheme 1B-D). Comparison of control values between biological replicates (or between Rac1 expressing cells) was not statistically different.

Distinct oncogenes trigger different patterns of junction disruption in epithelia

We addressed whether Junction Mapper parameters could identify distinct features and patterns of junction disruption by different stimuli. We tested images of epithelial cells transfected with oncogenic Rac (H-Ras^G12V; Figure 4A), constitutively activated Rac1 (Rac^Q61L) or activated Src (Src^Y527F; Figure 4—figure supplement 1). Previous visual analysis indicated that perturbation of cell–cell contacts occurs more efficiently between two Rac^Q61L-expressing neighbouring keratinocytes (Braga et al., 1997; Lozano et al., 2008). To validate this quantitatively, junctions were classified into two groups (i) between two expressing cells (ee) and (ii) between one expressing and one non-expressing cell (en). Control junctions from the same image were quantified from cells without oncogenic H-Ras expression. Measurements of the primary parameters (Figure 3) confirmed that junction contour and area were more severely disrupted when two neighbouring cells contained exogenous H-Ras^G12V (Figure 4D–E). The interface of mosaic contacts shared by expressing and non-expressing cells (en) were not significantly different from controls (Figure 4B,C,F).

Figure 4 with 1 supplement see all

Download asset Open asset

The interface between cells containing oncogenic H-Ras (ee) was significantly longer and larger than control cells (Interface Contour and Interface Area, Figure 4B,C). Irrespective of the longer contacting interface, the contour and the area of E-cadherin stained fragments were considerably altered upon expression of H-Ras^G12V (Fragmented Junction Contour and E-cadherin Area, Figure 4D,E). Finally, the length between cell–cell corners was significantly longer between controls and two adjacent transfected cells (Straight-line Interface Length, Figure 4F). Based on the quantification of the primary parameters, oncogenic H-Ras expression induces a longer contacting interface between cells and a progressive fragmentation of E-cadherin staining.

The primary parameter measurements following expression of activated Rac1^Q61L or Src^Y527F (Figure 4—figure supplement 1) showed distinct alteration profiles when compared between each other and to activated H-Ras (Figure 4). For example, active Src expression did not promote elongation of the contacting interface or an increase in interface area (Figure 4—figure supplement 1H–I), while active Rac1 did not induce fragmentation of cadherin staining (Figure 4—figure supplement 1D–E). The above data may indicate that distinct subsets of parameters can differentiate alterations by different stimuli, thereby providing a unique disruption profile (Figure 5). While data are from one technical replicate only, the high number of junctions quantified per sample is still sufficient to indicate significant differences among groups (Supplementary file 1). However, further experimentation is required to confirm the patterns observed.

Figure 5

Download asset Open asset

The secondary parameters (Scheme 1) were designed so that the distribution of a junctional marker is normalised to the area or length of the interface or junction between neighbouring cells. Although Intensity is a primary parameter, it was also included here for comparison with other published studies. The patterns of the secondary parameters Interface Occupancy and Cadherin Intensity at Interface Area essentially followed the corresponding primary parameters E-cadherin Area and E-Cadherin Intensity (Figures 4 and 5). Yet, the differences between groups are more apparent in the secondary parameters: data are less scattered with fewer outliers when compared to primary parameters (Figure 5).

We decided to focus on junctions that were shared by two transfected cells (ee), where phenotypes are clearer (Frasa et al., 2010; Lozano et al., 2008). Upon transfection of activated forms of H-Ras, Src or Rac1, a progressive disappearance of E-cadherin from the interface between cells was observed in different patterns (Figure 5A,C,E). When compared to controls, the undulation of the interface was increased among cells expressing H-Ras^G12V or Src^Y527F (higher Interface Linearity Index, Figure 5B,D), but remained unchanged for Rac1^Q61L-perturbed junctions (ee, Figure 5F). These data indicate that, as E-cadherin is removed from junctions, the interface between H-Ras^G12V and Src^Y527F expressing cells becomes less tensile. The percentage of the interface area occupied by E-cadherin receptors was decreased in all samples (Interface Occupancy, Figure 5B,F), but did not reach significance in Src-expressing cells (Figure 5D).

The intensity levels of cadherin at junctions was significantly reduced following transfection of activated H-Ras or Src when measured as raw values (E-cadherin Intensity) or corrected per contacting area between two cells (E-cadherin Intensity per Interface Area) (Figure 5B,D). In contrast, after active Rac1 expression, neither parameter was significantly changed. Consistent with the distinct phenotype of junction perturbation seen in Rac1^Q61L-transfected cells, the density of cadherin clusters was decreased in H-Ras^G12V and Src^Y527F, but slightly augmented in Rac^Q61L.

These data are summarized as a diagram in Scheme 2. We concluded that Junction Mapper quantification can capture the various phenotypes of junction perturbation and collectively define a specific profile for each oncogene. Our preliminary observations suggest that activated Rac1 does not significantly reduce the overall levels of E-cadherin at junctions, since receptor intensity at the interface area is not reduced. Rather, Rac1 activation progressively redistributes receptors at the junction (reduced Interface Occupancy and higher Cluster Density), maintaining straight, linear junctions. In contrast, activated H-Ras or Src disrupt junctions via fragmentation and removal of E-cadherin consistently throughout the contacting interface (decreased in Interface Occupancy and Cluster Density), with concomitant undulation of cell–cell contacts (i.e. reduced tension or stiffness at junctions).

Scheme 2

Download asset Open asset

Dynamic range of the measurements of cell–cell contact phenotypes

Disassembly of junctions by oncogenes is a severe phenotype, often leading to complete dissolution of contacts. However, other stimuli (i.e. differentiation, protein depletion, growth factor or drug treatment) may induce a milder phenotype that is not easily quantified. We asked whether the Junction Mapper could efficiently detect small changes in E-cadherin levels or distribution under different conditions (Figure 6). Datasets were obtained where the role of actin-regulatory proteins in junction formation was investigated in normal keratinocytes (Erasmus et al., 2016). Depletion of CIP4 (a regulator of cadherin trafficking) (Leibfried et al., 2008; Rolland et al., 2014), VAV2 (an exchange factor for Rho, Rac2 and Cdc42) (Vigil et al., 2010) or EEF1A (an actin bundling protein) (Mateyak and Kinzy, 2010) results in modest, but significant fluctuations in E-cadherin at junctions (20%–30%) using a thresholding method (Erasmus et al., 2016).

Figure 6

Download asset Open asset

When analysed with Junction Mapper, CIP4, VAV2 or EEF1A siRNA did not interfere with the linearity of the contacting interface (Figure 6B,D,F), in line with the appearance of normal, linear junctions. Consistent with our previous findings (Erasmus et al., 2016), a small decrease in E-cadherin intensity was observed in both CIP4- and VAV2-depleted cells (Figure 6B,F), while EEF1A siRNA promoted the unusual phenotype of higher levels of cadherin receptors (Figure 6D). These distinct patterns were also seen when E-cadherin intensity and area were normalized to the interface area (Intensity per Interface Area and Interface Occupancy, respectively). Strikingly, despite the similar reduction in E-cadherin intensity levels following VAV2 and CIP4 depletion, receptors were removed in different ways from junctions. The clusters of cadherin receptors were less dense with lower levels of CIP4, while upon VAV2 siRNA, the density of the clusters slightly increased (Cluster density, Figure 6B,F).

Thus, the discrete changes in junctions result from reduced cadherin levels throughout the contacting interface by lower density of receptor clusters (CIP4 RNAi) or localised E-cadherin removal and redistribution into denser clusters (VAV2 RNAi; Figure 6B,F). In contrast, EEF1A depletion augmented E-cadherin Intensity and Cluster Density along the interface. Taken together, these data highlight the ability of the Junction Mapper to detect phenotypes reproducibly and with good dynamic range. Scheme 3 summarizes the changes and distinct profiles detected by Junction Mapper in our validation experiments.

Scheme 3

Download asset Open asset

Endothelial junctions and cardiomyocyte intercalated discs

We next asked whether the software would be applicable to endothelial cells and cardiomyocytes that can have junctions differently shaped when compared with epithelial contacts. Both types of junctions are considerably more fragmented than epithelial contacts, and thus it was not clear whether the parameters quantified by the Junction Mapper would be suitable.

Thrombin is a serine protease that stimulates protease-activated receptor (PAR) in endothelial cells to increase vascular permeability in inflammation and injury (Figure 7A) (Kumar et al., 2009). Visually, the junctions of endothelial cells treated with thrombin are quite distinct (Malinova and Huveneers, 2018), but so far it has not been possible to evaluate differences quantitatively. There were no significant changes in the Interface Linearity Index (Figure 7A,B), consistent with the limitation of Junction Mapper to skeletonize zig-zagged junctions and thus measure their length (Figure 3—figure supplement 2E–F).

Figure 7 with 1 supplement see all

Download asset Open asset

However, with appropriate dilation values (Supplementary file 1, Figure 3—figure supplement 1), alterations were detected for area-based measurements. In thrombin-treated cells, VE-cadherin distribution along the interface area (Interface Occupancy), intensity of VE-cadherin staining and normalised intensity per contacting interface were substantially higher (Figure 7C–E). The density of VE-cadherin clusters was also enhanced upon treatment with thrombin (Figure 7F), implying that a higher number of receptors is recruited per contact area. Of note is that the analysis of stimulated endothelial cells, with their typical junction morphology and gaps, was strongly influenced by the user-controlled settings (Figure 7—figure supplement 1A–C). Yet, although raw values of each junction varied with different user settings, the overall trend and conclusion remained the same (Figure 7—figure supplement 1D–F), consistent with our previous comparative analyses (Supplementary files 2–3).

Rat neonatal cardiomyocytes were treated with phenylephrine (PE) as a model to induce hypertrophy (Miragoli et al., 2011; Simpson, 1985). Cells were co-stained with β-catenin as a junctional marker and connexin 43 (Figure 8A), a protein found in gap junctions, a structure necessary for synchronization of cardiomyocyte beating. The appearance of both markers at junctions was considerably fragmented (dotted appearance; Figure 8A), suggesting that parameters that consider the staining area would be the most appropriate.

Figure 8

Download asset Open asset

The software used the skeleton and dilated area of β-catenin as a mask as described by Erasmus et al. (2016) to segment the area where connexin 43 was localised. The Interface Linearity Index of β-catenin (Figure 8B) or connexin 43 (Figure 8C) was not significantly altered upon induction of hypertrophy in cellulo. Instead, hypertrophy stimulation promoted higher Interface Occupancy and higher levels of β-catenin and connexin 43 at junctions (raw intensity or corrected by the interface area; Figure 8B,C). However, the density of β-catenin clusters was significantly reduced, while connexin 43 density was augmented in PE-treated cells, in a small but significant manner. Thus, the two markers are modulated differently by hypertrophy stimulation in cellulo. Increased levels of β-catenin are achieved by less dense clusters spread along the interface, while connexin 43 molecules are clustered more densely, suggesting localised stimulation of gap junction formation. These data demonstrate the use of Junction Mapper for multiple cell types and its power to correlate the distribution of different proteins at cell–cell junctions. The data from cardiomyocytes and endothelial cells are summarized in Scheme 4. A heuristic approach on how to define the user-dependent settings and the parameters to use for these cell types are described in Supplementary file 4.

Scheme 4

Download asset Open asset

Junction configuration and adhesion receptor organization at contact sites are maintained via a complex interplay of distinct cytoskeletal filaments and associated regulatory proteins, with an exquisite regulation by diverse pathways and cellular contractility. At cell–cell contact sites, the challenge remains to translate the regulation of junctional components into functional and appropriately shaped cell–cell contacts. Junction Mapper facilitates the profiling of cell–cell contact behaviour with a variety of novel parameters in a fast, robust and reliable manner. Collectively, the parameter repertoire indicates how effective cell–cell adhesion is, identifies altered patterns of receptor distribution and guides experimental design to unravel the underlying molecular mechanisms.

Junction Mapper provides a semi-automated computer vision solution with broad applicability. The most innovative aspects of Junction Mapper are, first, the measurement of receptor density and occupancy, via the normalization of the junction marker intensity, length and area to the available contacting interface and cell–cell contact area. Second, the automatic quantification of fragmented staining at a junction, which has not been feasible previously and, third, the correlation of the parameters from two junctional markers along the same junction. Underpinning the above aspects is the ability of Junction Mapper to detect cell borders and corners when considerable disruption is observed. Machine learning algorithms certainly have the potential to improve and automate segmentation success, especially of highly fragmented contacts. However, the availability of training image datasets with boundaries manually-generated is still a bottleneck (Häring et al., 2018). One possible use of the software is therefore the generation of precise cell outline skeletons that can be used to train machine-learning algorithms in the future.

While semi-automation has been implemented during the image processing by Junction Mapper, user contribution is necessary and essential to perform the quality control of skeleton outline, corner positioning and setting up the dilation and threshold levels in a given dataset. User bias is particularly relevant in images with very disrupted and irregular junctions, and absolute values are not suitable for comparison across experiments in some cases. However, user bias can be minimized. Analysis of biological replicates or independent analysis by different users show that (i) the secondary parameters are more robust against user bias and (ii) obtained results are similar across replicates when comparing control and treated samples.

The software can detect both major and minor changes at junctions in different experiments. Unexpected and distinct profiles emerge from junction disruption by oncogenes that potently remove E-cadherin from contacts. The junctional defects caused by H-Ras^G12V or Src^Y527F illustrate the novel parameters that measure junction fragmentation and the specific density of E-cadherin in remaining fragments. In contrast, the perturbation profile induced by expression of Rac1^Q61L is not appropriately quantified by intensity alone and may be better assessed by ‘Interface Occupancy’, ‘Coverage Index’ and ‘Cluster Density’. Among the repertoire of pre-defined parameters, we find that a subset is highly likely to assess the features of a particular junction phenotype. Phenotypes not yet analysed could present additional possibilities to improve the Junction Mapper repertoire in future studies.

Junction Mapper analysis confirms mild phenotypes previously observed with manual, threshold-based quantifications of whole images (Erasmus et al., 2016), that indicate either an increase (EEF1A siRNA) or a decrease (CIP4 or VAV2 siRNA) in the intensity levels of E-cadherin. In addition, new Junction Mapper parameters uncover distinct features, that is that EEF1A depleted cells have higher E-cadherin occupancy and augmented cluster density when compared to controls. The profiling with new parameters underscores the potential of Junction Mapper to differentiate among distinct, subtle modes of junction perturbation. Yet, the conceptual significance of such alterations remains to be consolidated in future experiments and with additional biological replicates.

The plasticity of endothelial junctions is well established (Radeva and Waschke, 2018), but their unique responses to different stimuli have been challenging to quantify (Häring et al., 2018). The remarkable zig-zag pattern of thrombin-stimulated endothelial junctions correlates with increased vascular permeability (Malinova and Huveneers, 2018) but it is not recognized by the automated skeleton definition. Using area-based parameters, we find that the contacting interface is occupied more efficiently, with higher density of VE-cadherin receptors at endothelial junctions upon thrombin treatment. Thus, it seems that the increased levels of VE-cadherin at junctions may promote stronger endothelial adhesion, which is relevant to sustain elevated intracellular tension and contractility induced by thrombin stimulation. These results are consistent with the role of mechanical tension in receptor modulation and integrity of multicellular tissues (Liu et al., 2010), and merit further experimentation.

Because of the fragmented and undulated nature of cardiomyocyte contacts (Ehler, 2016; Vermij et al., 2017), quantitative imaging tools specifically designed for intercalated discs have not been available or systematically used. At steady state, our data show that, in control cardiomyocytes, intercalated discs have clusters of cadherin receptors that are far apart. Hypertrophic stimulus in cellulo (neonatal rat cardiomyocytes) potently increases the levels of β-catenin at contacting interfaces, consistent with what was reported in hamster and human hypertrophic hearts (Masuelli et al., 2003). Connexin 43 is a major connexin isoform found in cardiomyocytes and its total protein and mRNA levels are augmented by hypertrophic signals in cellulo (i.e. phenylephrine) (Salameh et al., 2008; Stanbouly et al., 2008) or in human hearts with compensated left-ventricular hypertrophy by pressure-overloading (Kostin et al., 2004). The initial profiling analysis with Junction Mapper suggest that the formation of gap junctions is enhanced after hypertrophy stimulation in cellulo, confirming the broadening of intercalated disc area and higher number of GAP junctions in compensated hypertrophic hearts of human patients (Kostin et al., 2004). Although hypertrophic stimulus increases both β-catenin and connexin 43 levels at intercalated discs, these markers are regulated in distinct ways. The β-catenin cluster density is decreased leading to a more contiguous distribution, while connexin 43 is localised in clusters of higher density.

Clearly, further investigation is necessary to ascertain the profiling and biological significance of the phenotypes observed in different models. The quantification of very fragmented and zig-zagged junctions such as those in endothelia and cardiomyocytes is a challenge that Junction Mapper has begun to address, but improvements in future computation studies are welcomed. The complexity of these junctions also requires more user input to quality control the definition of the skeleton and corners. In addition, the ability of Junction Mapper to quantify junctions in a stratified epithelium, where the added complexity of multiple epithelial cell layers provides an additional challenge, remains to be tested.

We foresee the potential of Junction Mapper in distinct research areas, due to its innovative and in-depth approach to quantify cell–cell contacts. The detailed fine mapping of junction properties forms a basis to distinguish between disassembly mechanisms and infer cellular processes such as intracellular trafficking, receptor clustering or modulation of contraction at junctions. As multiple cellular processes contribute to junction stability, the fingerprinting of junction phenotypes after different stimuli is a powerful tool in pathway inference and guides rescue and translational experiments. Despite its limitations, Junction Mapper’s broad dynamic range, repertoire of novel parameters and applicability to quantify junctions in various cell types will have a significant impact on studies in numerous model systems.

1. 1
   Structural changes in the myocardium during diabetes-induced cardiomyopathy

   1. E Adeghate
   2. J Singh

   (2014)

   Heart Failure Reviews 19:15–23.

   https://doi.org/10.1007/s10741-013-9388-5

   • PubMed
   • Google Scholar
2. 2
   Parallel thinning of binary pictures

   1. C Arcelli
   2. L Cordella
   3. S Levialdi

   (1975)

   Electronics Letters 11:148.

   https://doi.org/10.1049/el:19750113

   • Google Scholar
3. 3
   The role of differential VE-cadherin dynamics in cell rearrangement during angiogenesis

   1. K Bentley
   2. CA Franco
   3. A Philippides
   4. R Blanco
   5. M Dierkes
   6. V Gebala
   7. F Stanchi
   8. M Jones
   9. IM Aspalter
   10. G Cagna
   11. S Weström
   12. L Claesson-Welsh
   13. D Vestweber
   14. H Gerhardt

   (2014)

   Nature Cell Biology 16:309–321.

   https://doi.org/10.1038/ncb2926

   • PubMed
   • Google Scholar
4. 4
   The small GTPases rho and rac are required for the establishment of cadherin-dependent cell-cell contacts

   1. VM Braga
   2. LM Machesky
   3. A Hall
   4. NA Hotchin

   (1997)

   The Journal of Cell Biology 137:1421–1431.

   https://doi.org/10.1083/jcb.137.6.1421

   • PubMed
   • Google Scholar
5. 5
   Activation of the small GTPase rac is sufficient to disrupt cadherin-dependent cell-cell adhesion in normal human keratinocytes

   1. VM Braga
   2. M Betson
   3. X Li
   4. N Lamarche-Vane

   (2000)

   Molecular Biology of the Cell 11:3703–3721.

   https://doi.org/10.1091/mbc.11.11.3703

   • PubMed
   • Google Scholar
6. 6
   Arrhythmogenic cardiomyopathy: a disease of intercalated discs

   1. M Calore
   2. A Lorenzon
   3. M De Bortoli
   4. G Poloni
   5. A Rampazzo

   (2015)

   Cell and Tissue Research 360:491–500.

   https://doi.org/10.1007/s00441-014-2015-5

   • PubMed
   • Google Scholar
7. 7
   Using CellProfiler to analyze and quantify vascular morphology

   1. L Campbell
   2. M Kumar
   3. S Turner

   (2017)

   Methods in Molecular Biology 1544:179–189.

   https://doi.org/10.1007/978-1-4939-6722-3_13

   • PubMed
   • Google Scholar
8. 8
   CellProfiler: image analysis software for identifying and quantifying cell phenotypes

   1. AE Carpenter
   2. TR Jones
   3. MR Lamprecht
   4. C Clarke
   5. IH Kang
   6. O Friman
   7. DA Guertin
   8. JH Chang
   9. RA Lindquist
   10. J Moffat
   11. P Golland
   12. DM Sabatini

   (2006)

   Genome Biology 7:R100.

   https://doi.org/10.1186/gb-2006-7-10-r100

   • PubMed
   • Google Scholar
9. 9
   Myosin II controls junction fluctuations to guide epithelial tissue ordering

   1. S Curran
   2. C Strandkvist
   3. J Bathmann
   4. M de Gennes
   5. A Kabla
   6. G Salbreux
   7. B Baum

   (2017)

   Developmental Cell 43:480–492.

   https://doi.org/10.1016/j.devcel.2017.09.018

   • PubMed
   • Google Scholar
10. 10
    Endothelial adherens junctions at a glance

    1. E Dejana
    2. F Orsenigo

    (2013)

    Journal of Cell Science 126:2545–2549.

    https://doi.org/10.1242/jcs.124529

    • PubMed
    • Google Scholar
11. 11
    Cardiac cytoarchitecture — why the “hardware” is important for heart function!

    1. E Ehler

    (2016)

    Biochimica Et Biophysica Acta (BBA) - Molecular Cell Research 1863:1857–1863.

    https://doi.org/10.1016/j.bbamcr.2015.11.006

    • Google Scholar
12. 12
    Defining functional interactions during biogenesis of epithelial junctions

    1. JC Erasmus
    2. S Bruche
    3. L Pizarro
    4. N Maimari
    5. T Pogglioli
    6. C Tomlinson
    7. J Lees
    8. I Zalivina
    9. A Wheeler
    10. A Alberts
    11. A Russo
    12. VM Braga

    (2016)

    Nature Communications 7:13542.

    https://doi.org/10.1038/ncomms13542

    • PubMed
    • Google Scholar
13. 13
    Intercalated discs: multiple proteins perform multiple functions in non-failing and failing human hearts

    1. CB Estigoy
    2. F Pontén
    3. J Odeberg
    4. B Herbert
    5. M Guilhaus
    6. M Charleston
    7. JWK Ho
    8. D Cameron
    9. CG Dos Remedios

    (2009)

    Biophysical Reviews 1:43–49.

    https://doi.org/10.1007/s12551-008-0007-y

    • PubMed
    • Google Scholar
14. 14
    ImageJ User Guide - IJ 1

    1. T Ferreira
    2. WS Rasband

    (2012)

    ImageJ User Guide - IJ 1, 46, https://imagej.net/docs/guide/user-guide.pdf.

    • Google Scholar
15. 15
    Armus is a Rac1 effector that inactivates Rab7 and regulates E-cadherin degradation

    1. MA Frasa
    2. FC Maximiano
    3. K Smolarczyk
    4. RE Francis
    5. ME Betson
    6. E Lozano
    7. J Goldenring
    8. MC Seabra
    9. A Rak
    10. MR Ahmadian
    11. VM Braga

    (2010)

    Current Biology 20:198–208.

    https://doi.org/10.1016/j.cub.2009.12.053

    • PubMed
    • Google Scholar
16. 16
    A new tool to quantify receptor recruitment to cell contact sites during host-pathogen interaction

    1. MS Graus
    2. C Pehlke
    3. MJ Wester
    4. LB Davidson
    5. SL Steinberg
    6. AK Neumann

    (2014)

    PLOS Computational Biology 10:e1003639.

    https://doi.org/10.1371/journal.pcbi.1003639

    • PubMed
    • Google Scholar
17. 17
    Automated segmentation of epithelial tissue using Cycle-Consistent generative adversarial networks

    1. M Häring
    2. J Großhans
    3. F Wolf
    4. S Eule

    (2018)

    bioRxiv.

    https://doi.org/10.1101/311373

    • Google Scholar
18. 18
    Quantification of dynamic morphological drug responses in 3D organotypic cell cultures by automated image analysis

    1. V Härmä
    2. HP Schukov
    3. A Happonen
    4. I Ahonen
    5. J Virtanen
    6. H Siitari
    7. M Åkerfelt
    8. J Lötjönen
    9. M Nees

    (2014)

    PLOS ONE 9:e96426.

    https://doi.org/10.1371/journal.pone.0096426

    • PubMed
    • Google Scholar
19. 19
    Comparison of parameter-adapted segmentation methods for fluorescence micrographs

    1. C Held
    2. R Palmisano
    3. L Häberle
    4. M Hensel
    5. T Wittenberg

    (2011)

    Cytometry Part A 79A:933–945.

    https://doi.org/10.1002/cyto.a.21122

    • Google Scholar
20. 20
    Connexin 43 expression and distribution in compensated and decompensated cardiac hypertrophy in patients with aortic Stenosis

    1. S Kostin
    2. S Dammer
    3. S Hein
    4. WP Klovekorn
    5. EP Bauer
    6. J Schaper

    (2004)

    Cardiovascular Research 62:426–436.

    https://doi.org/10.1016/j.cardiores.2003.12.010

    • PubMed
    • Google Scholar
21. 21
    Molecular mechanisms of endothelial hyperpermeability: implications in inflammation

    1. P Kumar
    2. Q Shen
    3. CD Pivetti
    4. ES Lee
    5. MH Wu
    6. SY Yuan

    (2009)

    Expert Reviews in Molecular Medicine 11:e19.

    https://doi.org/10.1017/S1462399409001112

    • PubMed
    • Google Scholar
22. 22
    Rac and Cdc42 induce actin polymerization and G1 cell cycle progression independently of p65PAK and the JNK/SAPK MAP kinase cascade

    1. N Lamarche
    2. N Tapon
    3. L Stowers
    4. PD Burbelo
    5. P Aspenström
    6. T Bridges
    7. J Chant
    8. A Hall

    (1996)

    Cell 87:519–529.

    https://doi.org/10.1016/S0092-8674(00)81371-9

    • PubMed
    • Google Scholar
23. 23
    Drosophila Cip4 and WASp define a branch of the Cdc42-Par6-aPKC pathway regulating E-cadherin endocytosis

    1. A Leibfried
    2. R Fricke
    3. MJ Morgan
    4. S Bogdan
    5. Y Bellaiche

    (2008)

    Current Biology 18:1639–1648.

    https://doi.org/10.1016/j.cub.2008.09.063

    • PubMed
    • Google Scholar
24. 24
    Mechanical tugging force regulates the size of cell-cell junctions

    1. Z Liu
    2. JL Tan
    3. DM Cohen
    4. MT Yang
    5. NJ Sniadecki
    6. SA Ruiz
    7. CM Nelson
    8. CS Chen

    (2010)

    PNAS 107:9944–9949.

    https://doi.org/10.1073/pnas.0914547107

    • PubMed
    • Google Scholar
25. 25
    Quantitative imaging of epithelial cell scattering identifies specific inhibitors of cell motility and cell-cell dissociation

    1. D Loerke
    2. Q le Duc
    3. I Blonk
    4. A Kerstens
    5. E Spanjaard
    6. M Machacek
    7. G Danuser
    8. J de Rooij

    (2012)

    Science Signaling 5:rs5.

    https://doi.org/10.1126/scisignal.2002677

    • PubMed
    • Google Scholar
26. 26
    PAK is required for the disruption of E-cadherin adhesion by the small GTPase rac

    1. E Lozano
    2. MA Frasa
    3. K Smolarczyk
    4. UG Knaus
    5. VM Braga

    (2008)

    Journal of Cell Science 121:933–938.

    https://doi.org/10.1242/jcs.016121

    • PubMed
    • Google Scholar
27. 27
    Mechanotransduction in cardiac hypertrophy and failure

    1. RC Lyon
    2. F Zanella
    3. JH Omens
    4. F Sheikh

    (2015)

    Circulation Research 116:1462–1476.

    https://doi.org/10.1161/CIRCRESAHA.116.304937

    • PubMed
    • Google Scholar
28. 28
    Sensing of cytoskeletal forces by asymmetric adherens junctions

    1. TS Malinova
    2. S Huveneers

    (2018)

    Trends in Cell Biology 28:328–341.

    https://doi.org/10.1016/j.tcb.2017.11.002

    • PubMed
    • Google Scholar
29. 29
    Beta-catenin accumulates in intercalated disks of hypertrophic cardiomyopathic hearts

    1. L Masuelli
    2. R Bei
    3. P Sacchetti
    4. I Scappaticci
    5. P Francalanci
    6. L Albonici
    7. A Coletti
    8. C Palumbo
    9. M Minieri
    10. R Fiaccavento
    11. F Carotenuto
    12. C Fantini
    13. L Carosella
    14. A Modesti
    15. P Di Nardo

    (2003)

    Cardiovascular Research 60:376–387.

    https://doi.org/10.1016/j.cardiores.2003.08.005

    • PubMed
    • Google Scholar
30. 30
    eEF1A: thinking outside the ribosome

    1. MK Mateyak
    2. TG Kinzy

    (2010)

    Journal of Biological Chemistry 285:21209–21213.

    https://doi.org/10.1074/jbc.R110.113795

    • PubMed
    • Google Scholar
31. 31
    CellProfiler 3.0: next-generation image processing for biology

    1. C McQuin
    2. A Goodman
    3. V Chernyshev
    4. L Kamentsky
    5. BA Cimini
    6. KW Karhohs
    7. M Doan
    8. L Ding
    9. SM Rafelski
    10. D Thirstrup
    11. W Wiegraebe
    12. S Singh
    13. T Becker
    14. JC Caicedo
    15. AE Carpenter

    (2018)

    PLOS Biology 16:e2005970.

    https://doi.org/10.1371/journal.pbio.2005970

    • PubMed
    • Google Scholar
32. 32
    Automated profiling of individual cell-cell interactions from high-throughput time-lapse imaging microscopy in nanowell grids (TIMING)

    1. A Merouane
    2. N Rey-Villamizar
    3. Y Lu
    4. I Liadi
    5. G Romain
    6. J Lu
    7. H Singh
    8. LJ Cooper
    9. N Varadarajan
    10. B Roysam

    (2015)

    Bioinformatics 31:3189–3197.

    https://doi.org/10.1093/bioinformatics/btv355

    • PubMed
    • Google Scholar
33. 33
    Scanning ion conductance microscopy: a convergent high-resolution technology for multi-parametric analysis of living cardiovascular cells

    1. M Miragoli
    2. A Moshkov
    3. P Novak
    4. A Shevchuk
    5. VO Nikolaev
    6. I El-Hamamsy
    7. CMF Potter
    8. P Wright
    9. SHSA Kadir
    10. AR Lyon
    11. JA Mitchell
    12. AH Chester
    13. D Klenerman
    14. MJ Lab
    15. YE Korchev
    16. SE Harding
    17. J Gorelik

    (2011)

    Journal of the Royal Society Interface 8:913–925.

    https://doi.org/10.1098/rsif.2010.0597

    • Google Scholar
34. 34
    Ajuba is required for rac activation and maintenance of E-cadherin adhesion

    1. S Nola
    2. R Daigaku
    3. K Smolarczyk
    4. M Carstens
    5. B Martin-Martin
    6. G Longmore
    7. M Bailly
    8. VM Braga

    (2011)

    The Journal of Cell Biology 195:855–871.

    https://doi.org/10.1083/jcb.201107162

    • PubMed
    • Google Scholar
35. 35
    Tricellulin regulates junctional tension of epithelial cells at Tricellular contacts through Cdc42

    1. Y Oda
    2. T Otani
    3. J Ikenouchi
    4. M Furuse

    (2014)

    Journal of Cell Science 127:4201–4212.

    https://doi.org/10.1242/jcs.150607

    • PubMed
    • Google Scholar
36. 36
    Cdc42 GEF tuba regulates the junctional configuration of simple epithelial cells

    1. T Otani
    2. T Ichii
    3. S Aono
    4. M Takeichi

    (2006)

    The Journal of Cell Biology 175:135–146.

    https://doi.org/10.1083/jcb.200605012

    • PubMed
    • Google Scholar
37. 37
    Reflection essay: algorithms for the reduction of the number of points required to represent a digitized line or its caricature

    1. T Poiker
    2. DH Douglas

    (2011)

    Classics in Cartography pp. 29–36.

    https://doi.org/10.1002/9780470669488.ch3

    • Google Scholar
38. 38
    Mind the gap: mechanisms regulating the endothelial barrier

    1. MY Radeva
    2. J Waschke

    (2018)

    Acta Physiologica 222:e12860.

    https://doi.org/10.1111/apha.12860

    • Google Scholar
39. 39
    The CDC42-interacting protein 4 controls epithelial cell cohesion and tumor dissemination

    1. Y Rolland
    2. P Marighetti
    3. C Malinverno
    4. S Confalonieri
    5. C Luise
    6. N Ducano
    7. A Palamidessi
    8. S Bisi
    9. H Kajiho
    10. F Troglio
    11. OG Shcherbakova
    12. AR Dunn
    13. A Oldani
    14. L Lanzetti
    15. PP Di Fiore
    16. A Disanza
    17. G Scita

    (2014)

    Developmental Cell 30:553–568.

    https://doi.org/10.1016/j.devcel.2014.08.006

    • PubMed
    • Google Scholar
40. 40
    Teaching image-processing programming in Java

    1. D Sage
    2. M Unser

    (2003)

    IEEE Signal Processing Magazine 20:43–52.

    https://doi.org/10.1109/MSP.2003.1253553

    • Google Scholar
41. 41
    Signal transduction and transcriptional control of cardiac connexin43 up-regulation after alpha ₁-adrenoceptor stimulation

    1. A Salameh
    2. S Krautblatter
    3. S Baessler
    4. S Karl
    5. D Rojas Gomez
    6. S Dhein
    7. D Pfeiffer

    (2008)

    The Journal of Pharmacology and Experimental Therapeutics 326:315–322.

    https://doi.org/10.1124/jpet.108.136663

    • PubMed
    • Google Scholar
42. 42
    Fiji: an open-source platform for biological-image analysis

    1. J Schindelin
    2. I Arganda-Carreras
    3. E Frise
    4. V Kaynig
    5. M Longair
    6. T Pietzsch
    7. S Preibisch
    8. C Rueden
    9. S Saalfeld
    10. B Schmid
    11. JY Tinevez
    12. DJ White
    13. V Hartenstein
    14. K Eliceiri
    15. P Tomancak
    16. A Cardona

    (2012)

    Nature Methods 9:676–682.

    https://doi.org/10.1038/nmeth.2019

    • PubMed
    • Google Scholar
43. 43
    The CellBorderTracker, a novel tool to quantitatively analyze spatiotemporal endothelial junction dynamics at the subcellular level

    1. J Seebach
    2. AA Taha
    3. J Lenk
    4. N Lindemann
    5. X Jiang
    6. K Brinkmann
    7. S Bogdan
    8. HJ Schnittler

    (2015)

    Histochemistry and Cell Biology 144:517–532.

    https://doi.org/10.1007/s00418-015-1357-8

    • PubMed
    • Google Scholar
44. 44
    Mechanical regulation of cardiac aging in model systems

    1. AO Sessions
    2. AJ Engler

    (2016)

    Circulation Research 118:1553–1562.

    https://doi.org/10.1161/CIRCRESAHA.116.307472

    • PubMed
    • Google Scholar
45. 45
    Stimulation of hypertrophy of cultured neonatal rat heart cells through an alpha 1-adrenergic receptor and induction of beating through an alpha 1- and beta 1-adrenergic receptor interaction. Evidence for independent regulation of growth and beating

    1. P Simpson

    (1985)

    Circulation Research 56:884–894.

    https://doi.org/10.1161/01.RES.56.6.884

    • PubMed
    • Google Scholar
46. 46
    Sodium hydrogen exchange 1 (NHE-1) regulates connexin 43 expression in cardiomyocytes via reverse mode sodium calcium exchange and c-Jun NH2-terminal kinase-dependent pathways

    1. S Stanbouly
    2. LA Kirshenbaum
    3. DL Jones
    4. M Karmazyn

    (2008)

    Journal of Pharmacology and Experimental Therapeutics 327:105–113.

    https://doi.org/10.1124/jpet.108.140228

    • PubMed
    • Google Scholar
47. 47
    Adherens junction length during tissue contraction is controlled by the mechanosensitive activity of actomyosin and junctional recycling

    1. A Sumi
    2. P Hayes
    3. A D'Angelo
    4. J Colombelli
    5. G Salbreux
    6. K Dierkes
    7. J Solon

    (2018)

    Developmental Cell 47:453–463.

    https://doi.org/10.1016/j.devcel.2018.10.025

    • PubMed
    • Google Scholar
48. 48
    New aspects of pathogenesis of atrial fibrillation: remodelling of intercalated discs

    1. N Tribulova
    2. T Egan Benova
    3. B Szeiffova Bacova
    4. C Viczenczova
    5. M Barancik

    (2015)

    Journal of Physiology and Pharmacology 66:625–634.

    • PubMed
    • Google Scholar
49. 49
    Refining the molecular organization of the cardiac intercalated disc

    1. SH Vermij
    2. H Abriel
    3. TAB van Veen

    (2017)

    Cardiovascular Research 113:259–275.

    https://doi.org/10.1093/cvr/cvw259

    • Google Scholar
50. 50
    Ras superfamily GEFs and GAPs: validated and tractable targets for Cancer therapy?

    1. D Vigil
    2. J Cherfils
    3. KL Rossman
    4. CJ Der

    (2010)

    Nature Reviews Cancer 10:842–857.

    https://doi.org/10.1038/nrc2960

    • PubMed
    • Google Scholar
51. 51
    A screen for morphological complexity identifies regulators of switch-like transitions between discrete cell shapes

    1. Z Yin
    2. A Sadok
    3. H Sailem
    4. A McCarthy
    5. X Xia
    6. F Li
    7. MA Garcia
    8. L Evans
    9. AR Barr
    10. N Perrimon
    11. CJ Marshall
    12. ST Wong
    13. C Bakal

    (2013)

    Nature Cell Biology 15:860–871.

    https://doi.org/10.1038/ncb2764

    • PubMed
    • Google Scholar

Acceptance summary:

This manuscript provides a novel tool to quantify different parameters of mammalian intercellular junctions, and as such, is aptly called junction mapper. As pointed out by the authors in the Introduction of the manuscript quantification of different mammalian junctional parameters is often rather time consuming as for many programs finding borders and, often, correcting borders, can until now mostly be done manually. This tool will be useful to researchers in many fields for quantifying cell junctions.

Decision letter after peer review:

[Editors’ note: the authors were asked to provide a plan for revisions before the editors issued a final decision, this was subsequently approved by the editors. What follows is the editors’ letter requesting such plan.]

Thank you for sending your article entitled "Junction Mapper: quantitative analysis to decipher cell-cell contact phenotypes" for peer review at eLife. Your article is being evaluated by three peer reviewers, and the evaluation is being overseen by a Reviewing Editor and Anna Akhmanova as the Senior Editor.

Given the list of essential revisions, including new experiments, the editors and reviewers invite you to respond within the next two weeks with an action plan and timetable for the completion of the additional work. We plan to share your responses with the reviewers and then issue a binding recommendation.

While the reviewers recognized the benefits of the tool to analyze junctions, the reviewers felt that the manuscript lacked details regarding the algorithm testing and methods. Furthermore, the reviewers also would like to see whether this tool can be used on mammalian junctions to rigorously test whether this will be broadly used by the field of epithelial junction biologists.

Reviewer #1:

This very nice manuscript provides a novel tool to quantify different parameters of mammalian intercellular junctions, and as such, is aptly called junction mapper. As pointed out by the authors in the Introduction of the manuscript quantification of different mammalian junctional parameters is often rather time consuming as for many programs finding borders and, often, correcting borders, can until now mostly be done manually. Automated analysis is not trivial and in general requires a strong background in image bioinformatics/analysis or access to a core facility that provides such analysis. The exciting point of the work presented here is that this program provides a general and very accessible platform to quantify junctions in a semi-automatic manner to obtain multiple parameters. Moreover, proper detection of fragmented junctions is an even more complex problem which the authors have apparently solved. The program provides reasonable output parameters and the detection quality and sophistication of junctional staining appears to be outstanding judged by the data presented on both epithelial junctions as well as the already much harder endothelial junctions. Finally, the experiments that were performed convincingly demonstrate the applicability of the program and its sensitivity of detecting minor and specific alterations. Taken together, this manuscript provides convincing and exciting evidence that junction mapper is a great open software to quantitatively analyze junctions, and that allows laboratories not specialized in junctional imaging to also more precisely analyze and assess junctional phenotypes.

I have two comments that need to be addressed:

1) The last author is a well-known specialist on keratinocytes, which in vivo and in culture form multi-layered epithelia upon prolonged induction of differentiation. Although the authors have used keratinocytes, it seems they only used early time points after differentiation when these cells did not form multilayers yet. The multi-layering provides an extra complexity to the proper quantification of junctions. Can the program deal with this complexity? If so, would be very nice to include an example. If not, please also more stringently then discuss the limitations of the program.

2) One important point does need to be clarified. It is not exactly clear from the figures, legends or Materials and methods how the statistics were done. The authors name all the appropriate tests but it is not clear e.g. whether means of technical replicates or the collective number of junctions were used for statistical test. Clearly, the high number of junctions that can be analyzed here is an advantage. However, the variance between experiments is not shown and it is not clear whether the statistics take that into account.

Reviewer #2:

1) This manuscript describes a novel semiautomated software, 'Junction Mapper,' for analyzing cell junctions on parameters such as defining the junctional interphase, intensity, and length. Currently, analyzing cell junctions is often done manually which can be time consuming and laborious, even when using automated software, it is usually tailored to a specific cell type or unable to automatically distinguish nuances of perturbed or fragmented junctions. The authors evaluate their software's performance in several cell types, and in multiple conditions to evaluate its ability to detect a range of phenotypes (from severe to mild). This manuscript presents a new tool that address these significant problems and has the potential to be widely applied. However there are numerous flaws and oversights in this presentation that significantly dampen my enthusiasm. There is also a lack of rigor in evaluating the algorithms as presented which makes it difficult to assess the utility of the software thereby raising a major concern.

2) First, the availability of the software and test data sets is not clear. Nor is the computing language or platform that the software is written in. This is a major oversight. Second, the microscopy needs to be fully described. There are not scale bars on images. Are these max intensity projections, single confocal slices, widefield images, deconvovled? What is the resolution? The quality of the input will impact the ability of the algorithms to function and this needs to be defined.

3) How do different user selected parameters impact the output from the program? Since the skeleton can be manually adjusted, does the specific skeleton have an impact on the results? These factors need to be considered and quantified for their impact on the final measurements.

4) Input parameters need to be discussed – dilation and threshold level are important parameters to disclose for each analyzed data set. How do these impact final measurements?

5) The process of interactive editing, blurring, and sharpening is not shown nor well described. Is this process done manually on every image? Can it be automated for a data set? Seeing these steps, including when manual adjustment is required would be helpful. The reader has no feel for how often user intervention is required (every cell, every image, every border?).

Reviewer #3:

The manuscript describes the development of software for quantifying the phenotype of cell-cell contacts from high resolution images. The authors have been able to produce an impressive data set with the tool. However, the new tool is not validated sufficiently, and the methods are not adequately described. This is particularly important, as by the studying design, the biological ramifications of the analysis were not probed.

1) In the Introduction (third paragraph), the authors make the claim that the software developed for analysis of Drosophilia or nematode epithelia are not applicable to mammalian cells. Image analysis algorithm performance are typically based of the nature of the structures to be analyzed and the quality of data. There is no reason for these parameters to be fundamentally different in all mammalian cells. If the authors, are trying to say that tools for analysis of epithelial cells may not be directly applicable to other cells types or in all possible treatments (points made later), they should be more explicit here. Also, the (Held et al., 2011) does not describe the analysis of cell-cell contacts and does not seem to support the authors point.

2) The development of a new algorithm for image analysis is typically associated with several key endeavors. First is validation. Validation must be performed on data where the ground truth is known. For a "first-of-kind" algorithm this is often associated done with simulated data or data that was analyzed by hand. By this standard definition, the Junction Mapper is not validated. This must be corrected for the manuscript to be considered for publication. It seems that the data shown in Figure 6 may be appropriate for this task, as the data was previously analyzed by hand? This reviewer notes that citations within the manuscript are validated in this manner (i.e. Held et al., 2011). Also, the section in the Materials and methods described "Software validation" actually discusses exclusion criteria and naming procedures.

3) Additionally, there is typically an analysis of when the algorithm begins to fail. This is often done in terms of data with varying signal to noise. The limitations of this of the Junction Mapper are not tested. While it would be preferred that a formal analysis of the performance of the algorithm be performed, at a minimum image quality (which is often reported as signal to noise ratios) requirements (or at least typical image characteristics) should be provided so that users know the criteria for the suitability of the algorithm are met being attempting to use it.

4) The Materials and methods also lack sufficient detail. According to the manuscript, the Junction Mapper is a novel stand alone software. This means specific algorithms for background estimation, skeletonization, segmentation, and parameter calculation were used. Precise definitions (including formulae) or citations to the utilized algorithms need to be provided. Also, formulae describing how the parameters were calculated need to be added. Also, how corners were identified needs to be explained in greater detail. If corners are manually detected, then it is improper to state the Junction Mapper detected them. Additionally, are all junctions analyzed at once or individually? If different segmentation parameters are used on different junctions, are the results actually directly comparable? This reviewer also notes that the required details are not available in the often-cited previous work of Erasmus et al., 2016. This is understandable as that was a highly biological study, but that reference should not be used to described methodological details.

5) In several places the authors make claims that junctions that are straighter are more tensed. This is not true, as the junctions could also be stiffer, and therefore deform less. As no measurements of junctional tension were made, these statements should be removed.

6) The algorithm seems to assume constant width for the junctions. This will likely lead to substantial errors in area estimations. Have the authors given thought to this issue? Are the errors inconsequential? Can the algorithm be readily updated to avoid this artefact?

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Junction Mapper is a novel computer vision tool to decipher cell-cell contact phenotypes" for further consideration at eLife. Your revised article has been favorably evaluated by Anna Akhmanova as the Senior Editor, a Reviewing Editor, and three reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

Reviewer #1:

The authors took a lot of effort to address the points that were raised by the reviewers. I do not have the proper expertise to judge whether the authors have sufficiently addressed the points of reviewers 2 and 3.

1) In response to point 2, the authors now describe which statistical tests are used and clarify for which figure panels they used biological or technical replicates. The authors now state that the statistics and the tests used are in Supplementary file 5. Unfortunately, I was unable to find Supplementary file 5 in the present submission, making it impossible whether the statistics were done properly. Importantly, the authors still do not clearly state in the legends and/or Materials and methods the number of replicates on which the data shown are based on. Throughout the manuscript, each measured junction is called a sample and the number of analyzed junctions are now provided in the figures. However, for the statistics the authors should provide information on the number of replicates, meaning number of independent experiments or biological samples, in the figure or legend to clearly show that differences and significance levels are due to the high number of junctions but not because of experimental variation.

2) For siRNA experiments the section on statistical analysis says that only one control and knockdown group was analyzed. If this is not a misunderstanding, then even high number of analyzed junctions from one experiment do not allow strong conclusions as experimental variation upon knockdown of CIP4, EEFA1A or VAV2 has not been assessed. For physiological relevant conclusions one would expect a set of at least 3 independent experiments. I do understand that the authors want to test their very interesting tool for which they provide many different experimental tests. However, the authors should then acknowledge and tune down mechanistic conclusions. For experiments concerning HUVEC cells, there is no information on the number of biological replicates on which the data are based.

3) Thus, at present the way the data are presented now, provides nice evidence that the program allows for quantification of high number of junctions to potentially detect small differences. If the authors want to use the experiments shown in the manuscript to validate the usefulness as well as limits of the software, this approach with limited number of repetitions seems valid but then this has to be made more clearly throughout the manuscript. Not all of the data sets are sufficiently comprehensive to univocally demonstrate a biological relevant junctional outcome upon treatment or knockdown and this should be acknowledged. Having said this, I feel if the authors were able to address the concerns of reviewer 2 and 3, they should be provided the opportunity to properly address and include in the manuscript the comments above.

Reviewer #2:

Overall this tool has the potential to be useful for researchers, and the authors have now included many more rigorous definition of the method and treatments of the parameters both in definition and in comparison of data processed in different ways. This strengthens the manuscript. However, I am still not sure this is sufficient to fully describe the bounds of the junction mapper approach.

1) Most scientific conclusions as presented come from comparisons of values in control/disrupted. When looking at effects of user inputs the graphs are shown for each treatments individually but what are the changes with different thresholds/dilations between treatments (i.e. Figure 3—figure supplements 2 and 3)? Would different biological conclusions be drawn if you push the dilation or threshold "too far"?

2) Authors point out a central issue and why this analysis is challenging – If there is no staining at a cell corner then where the program or user places the junction is arbitrary. Without a second marker of plasma membrane rather than junction this will always be an issue these measurements regardless of what automated algorithm is deployed or with user definition.

3) The authors show secondary characteristics are more robust to error in corner selection and that individual users get same direction of changes but not same absolute values, which does show the value added here. This should be emphasized in the Discussion.

4) For the noise quantification, there should be a metric for measurement accuracy, or deviation from the ground truth. Guidance for user on minimum S/N needed for robust quantification. There is also no information on type of noise added (shot noise, dependent on pixel intensity is dominant noise in most microscopy images and is recommended here). The goal would be for a user to know if a junction has sufficient S/N to be analyzed.

5) This leads me to the quantification of the type "ee" boarders in the H-Ras^G12V overexpressing cells. There appears to be greatly reduced e-cad staining at the boarders (it is difficult to see by eye any in the examples shown). In this case, how are the interfaces defined? Does this lower signal to noise negatively impact the program? How confident should we feel about comparing images with different S/N levels?

Reviewer #3:

The manuscript is greatly improved and nearly ready for publication.

[Editors’ notes: the authors’ response after being formally invited to submit a revised submission follows.]

Reviewer #1:

[…] I have two comments that need to be addressed:

1) The last author is a well-known specialist on keratinocytes, which in vivo and in culture form multi-layered epithelia upon prolonged induction of differentiation. Although the authors have used keratinocytes, it seems they only used early time points after differentiation when these cells did not form multilayers yet. The multi-layering provides an extra complexity to the proper quantification of junctions. Can the program deal with this complexity? If so, would be very nice to include an example. If not, please also more stringently then discuss the limitations of the program.

There could be adaptations to quantify en face confocal Z-sections of a stratified epithelia. However, the complexity of stratified epithelium images (en face or transversal sections (see Figure 3A from Sevilla et al., 2007) means that a potential analysis using Junction Mapper will not be meaningful. For example, it will be very much dependent on the angle of the section, variations in cell sizes and the amount of incomplete outlines of individual cells, which makes comparison across samples difficult. It also remains to be tested whether Junction Mapper can quantify junctions in pseudo-stratified epithelia. This information and limitations were added to the Discussion.

2) One important point does need to be clarified. It is not exactly clear from the figures, legends or Materials and methods how the statistics were done. The authors name all the appropriate tests but it is not clear e.g. whether means of technical replicates or the collective number of junctions were used for statistical test. Clearly, the high number of junctions that can be analyzed here is an advantage. However, the variance between experiments is not shown and it is not clear whether the statistics take that into account.

Apologies for the oversight and unclear text. The results are mostly from technical replicates: 5-10 images were analysed per sample and all junctions from cells in each image were quantified (except those excluded by the quality control criteria). Each point on graphs represents a measurement from one junction. The numbers ('n =') cited include all junctions analyzed (pooled from all available images of a given sample). For some experiments (Src expression and endothelial cells), biological replicates were analysed separately. If the pattern of the parameters were similar across biological replicates, the data was pooled all together and shown in the graph. A new supplementary table contains the statistics performed and where appropriate, the variance (Supplementary file 5). Clarifications on which data are from a technical or biological replicate were added in Supplementary file 1 and in the figure legends.

Reviewer #2:

[…]

1) This manuscript presents a new tool that address these significant problems and has the potential to be widely applied. However there are numerous flaws and oversights in this presentation that significantly dampen my enthusiasm. There is also a lack of rigor in evaluating the algorithms as presented which makes it difficult to assess the utility of the software thereby raising a major concern.

We thank the reviewer for the valid points made. The parameter explanations and methodology could have been made clearer and further validation presented. We have now provided extensive validation of the program in several supplementary figures, main figures and tables (please see those outlined below).

2) First, the availability of the software and test data sets is not clear. Nor is the computing language or platform that the software is written in. This is a major oversight. Second, the microscopy needs to be fully described. There are not scale bars on images. Are these max intensity projections, single confocal slices, widefield images, deconvovled? What is the resolution? The quality of the input will impact the ability of the algorithms to function and this needs to be defined.

The reviewer has a point and we are sorry for the oversight. The test data sets are from our own labs, unpublished data and from published papers; i.e. from Erasmus et al., 2016 – these are referred to in the text. The software was developed using Java and is available as an open access standalone, downloadable format (information added in the subsection “Software development”). A new supplementary table containing detailed information of the acquisition of images for each dataset was added (Supplementary file 1). Explanations of what critically affects the analyses and the limitations of the software were written in the Materials and methods and Results(please see also point 3, 4 below and reviewer 3 points 2, 3, 4 and 6).

3) How do different user selected parameters impact the output from the program? Since the skeleton can be manually adjusted, does the specific skeleton have an impact on the results? These factors need to be considered and quantified for their impact on the final measurements.

The user can control four distinct steps. The first two (skeleton and corner identification) are done automatically but can be revised by the user. While it is important that the skeleton overlays the junctional marker staining, there is some tolerance on the precision of skeleton outline, because there is a dilation step to define the ROI.

The other two user-controlled settings are to choose the dilation value (to encompass the whole width of the staining along the skeleton) and the thresholding levels (to keep noise in the cytoplasm at a minimum, without information loss at junctions). These settings are kept constant for all images of a given experiment. Explanations are added to the text to clarify the relevance of the user-defined settings (Results and in “Figure 2—figure supplement 2”).

4) Input parameters need to be discussed – dilation and threshold level are important parameters to disclose for each analyzed data set. How do these impact final measurements?

Thanks for the suggestion. In the revised manuscript, two new supplementary figures (Figure 3—figure supplement 1 to 2) demonstrate the effects of varying dilation and thresholding on the primary parameters (length and intensity). A supplementary table was added with the information of how each sample was analysed (dilation and thresholding settings; Supplementary file 1).

5) The process of interactive editing, blurring, and sharpening is not shown nor well described. Is this process done manually on every image? Can it be automated for a data set? Seeing these steps, including when manual adjustment is required would be helpful. The reader has no feel for how often user intervention is required (every cell, every image, every border?).

We agree with the reviewer. The process is done in each image because of image-to-image variability and the extent of junction disruption when stimulated. We addressed the concerns in two ways: (i) a new supplementary figure shows the precise steps to obtain the skeleton and corners (Figure 2—figure supplement 2) and (ii) instructions on how the software operates are included in a custom-made video downloadable package of Junction Mapper from a dedicated website.

Reviewer #3:

[…]

1) In the Introduction (third paragraph), the authors make the claim that the software developed for analysis of Drosophilia or nematode epithelia are not applicable to mammalian cells. Image analysis algorithm performance are typically based of the nature of the structures to be analyzed and the quality of data. There is no reason for these parameters to be fundamentally different in all mammalian cells. If the authors, are trying to say that tools for analysis of epithelial cells may not be directly applicable to other cells types or in all possible treatments (points made later), they should be more explicit here.

This is a misunderstanding. What the referred paragraph is meant to say is that algorithms that work fine to identify cell borders (i.e. skeleton) in Drosophila or C. elegans epithelia struggle to perform well in mammalian epithelia. This is because mammalian junctions are thinner, more irregular and have higher signal to noise ratio than those of in vivo images of invertebrate epithelia. The Junction Mapper parameters (novel and established) are universal and are applicable to epithelia from different organisms, cell types and following different treatments.

Also, the (Held, et al., 2011) does not describe the analysis of cell-cell contacts and does not seem to support the authors point.

Apologies, this reference is incorrect as it refers to different ways to identify cells in a monolayer.

2) The development of a new algorithm for image analysis is typically associated with several key endeavors. First is validation. Validation must be performed on data where the ground truth is known. For a "first-of-kind" algorithm this is often associated done with simulated data or data that was analyzed by hand. By this standard definition, the Junction Mapper is not validated. This must be corrected for the manuscript to be considered for publication. It seems that the data shown in Figure 6 may be appropriate for this task, as the data was previously analyzed by hand? This reviewer notes that citations within the manuscript are validated in this manner (i.e. Held et al., 2011). Also, the section in the Materials and methods described "Software validation" actually discusses exclusion criteria and naming procedures.

The reviewer is correct. Many of the parameters are novel and have not been used previously. Unfortunately, data shown in Figure 6 has been quantified by another thresholding method, using the intensity of the whole image as output (i.e. all cells in the image).^[2] The raw data obtained is thus not comparable, as the results are from the intensity of all junctions in the whole image corrected for the total image area.^[2] The raw data from Junction Mapper output measure values per junction corrected for its respective interface area. Nevertheless, in an experiment, when the percentage of rescue is computed with either our previous and current method, similar results are obtained: a mild reduction of E-cadherin intensity at junctions compared to controls after depletion of TRIP10 or VAV2 (20-30%) while EEF1A shows a 50% increase in E-cadherin intensity at junctions (Figure 6).^[2] We provided validation in several ways and revised the text accordingly:

- Length-based and area-based parameters are measured as predicted, i.e. with increasing shorter values for the interface, junctions and fragmented junction staining (new Figure 3B).

- Dilation manipulation interferes with area-based measurements, but not length-based parameters Figure 3—figure supplement 2).

- Increasing levels of thresholding interferes with intensity and pixel area as well as the length of fragments of junction markers (Figure 3—figure supplement 3).

- The validation of the secondary parameter Coverage Index by manual quantification as described in our previous paper (i.e. Supplementary file 2).^[3]

- Comparison of quantification by an independent user (new Supplementary file 3 and Figure 7—figure supplement 1; please see points 3, 4 and 6 below).

3) Additionally, there is typically an analysis of when the algorithm begins to fail. This is often done in terms of data with varying signal to noise. The limitations of this of the Junction Mapper are not tested. While it would be preferred that a formal analysis of the performance of the algorithm be performed, at a minimum image quality (which is often reported as signal to noise ratios) requirements (or at least typical image characteristics) should be provided so that users know the criteria for the suitability of the algorithm are met being attempting to use it.

We thank the reviewer for the suggestion. Additional explanations were added in the Results and a new supplementary figure provide validation of signal-to-noise ratios (new Figure 2—figure supplement 3).

4) The Materials and methods also lack sufficient detail. According to the manuscript, the Junction Mapper is a novel stand alone software. This means specific algorithms for background estimation, skeletonization, segmentation, and parameter calculation were used. Precise definitions (including formulae) or citations to the utilized algorithms need to be provided. Also, formulae describing how the parameters were calculated need to be added. Also, how corners were identified needs to be explained in greater detail. If corners are manually detected, then it is improper to state the Junction Mapper detected them.

The software uses algorithms that are mostly available open access, with some added innovation. The novelty of the Junction Mapper is in the integration of distinct measurements, calculations of new parameters and consolidation of all parameters in a single system. The mathematical definitions of the parameters, the innovation of the automatic corner setting, its correction by users and the calculation of fragmented regions of contacts are now described in a more comprehensive explanation of the software (new Figure 2—figure supplement 2). This information was added to the Materials and methods section.

Additionally, are all junctions analyzed at once or individually? If different segmentation parameters are used on different junctions, are the results actually directly comparable?

Once the skeleton, corners and dilation are set, all junctions in the ROI are quantified simultaneously at a click, and all measurements and parameters exported to an excel file automatically. For results to be comparable, images from an experiment with various treatments would need to be analysed with the same conditions (dilation, thresholding, etc). This information was added to the revised text (Results) and in new Figure 2—figure supplement 2.

This reviewer also notes that the required details are not available in the often-cited previous work of Erasmus et al., 2016. This is understandable as that was a highly biological study, but that reference should not be used to described methodological details.

The reviewer has a point and it needs to be better clarified in the text. The work on Erasmus et al. defines the obtention of the E-cadherin mask as a tool to calculate the intensity specifically around junctions as% thresholded area of the whole image. It also defines the Ecadherin mask to subtract an ROI on an image of a distinct marker (i.e. F-actin), so that only the signal localized at contacts is considered. Junction Mapper builds from this work and improves with novel quantification tools (corners, length and area), calculation of fragments at junctions and a variety of novel primary and secondary parameters. These points were clarified in the text (Results and Materials and methods).

5) In several places the authors make claims that junctions that are straighter are more tensed. This is not true, as the junctions could also be stiffer, and therefore deform less. As no measurements of junctional tension were made, these statements should be removed.

Thanks for point out this over-interpretation – we have indeed shown no data to support either case (tension or stiffness). The reviewer is correct that junctions could also be stiffer. However, the statement is based on numerous papers that: (i) using laser ablation show a striking recoil of junctions, which is abolished by different treatments to inhibit Rho pathway ^[4], (ii) show weaker, undulated contacts upon enhancing signalling from Cdc42 GTPase ^[1], (iii) measurement of tension at junctions with biosensors^[5] and (iv) the wiggly morphology of junctions observed in tumour cells versus in vivo in epithelial tissues. The text was revised throughout as requested.

6) The algorithm seems to assume constant width for the junctions. This will likely lead to substantial errors in area estimations. Have the authors given thought to this issue? Are the errors inconsequential? Can the algorithm be readily updated to avoid this artefact?

This is a misunderstanding. The variable width of staining at junctions can be fully captured by the dilation step (user defined) and will depend on the amplification of the image. The dilation steps increase the numbers of pixels at each side of the skeleton line (range 0 to 9 pixels) and determines automatically the area in which quantification will be done (ROI). The influence of varying the dilation for different parameters is now shown in new Figure 3—figure supplement 2. The dilation value is manually set up and kept constant for each dataset (i.e. H-Ras, Rac1, etc). We provided the dilation values in new Supplementary file 1 along with the details for final thresholding and image dataset characteristics).

References:

1) Otani, T., et al., Cdc42 GEF Tuba regulates the junctional configuration of simple epithelial cells. J. Cell Biol., 2006. 175(1): p. 135-146.

2) Erasmus, J.C., et al., Defining functional interactions during biogenesis of epithelial junctions. Nature Commun., 2016. 7: p. 13542.

3) Lozano, E., et al., PAK is required for the disruption of E-cadherin adhesion by the small GTPase Rac. J. Cell Sci., 2008. 121(Pt 7): p. 933-938.

4) Priya, R., et al., Feedback regulation through myosin II confers robustness on RhoA signalling at E-cadherin junctions. Nature Cell Biol., 2015. 17: p. 1282-1293.

5) Acharya, B.R., et al., Mammalian Diaphanous 1 Mediates a Pathway for E-cadherin to Stabilize Epithelial Barriers through Junctional Contractility. Cell Rep., 2017. 18(12): p. 2854-2867.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Reviewer #1:

[…]

1) In response to point 2, the authors now describe which statistical tests are used and clarify for which figure panels they used biological or technical replicates. The authors now state that the statistics and the tests used are in Supplementary file 5. Unfortunately, I was unable to find Supplementary file 5 in the present submission, making it impossible whether the statistics were done properly. Importantly, the authors still do not clearly state in the legends and/or Materials and methods the number of replicates on which the data shown are based on. Throughout the manuscript, each measured junction is called a sample and the number of analyzed junctions are now provided in the figures. However, for the statistics the authors should provide information on the number of replicates, meaning number of independent experiments or biological samples, in the figure or legend to clearly show that differences and significance levels are due to the high number of junctions but not because of experimental variation.

The information was present in the legends but we acknowledge that it was not very clear. We now revised each relevant figure legend and state whether the data are technical or biological replicates. We changed the legends to state “number of junctions” instead of “number of samples”. The text was also modified to clarify the significance due to the high number of junctions (subsection “Distinct oncogenes trigger different patterns of junction disruption in epithelia”) and revised the correct citations to supplementary materials to be fully compliant with eLife requirements.

2) For siRNA experiments the section on statistical analysis says that only one control and knockdown group was analyzed. If this is not a misunderstanding, then even high number of analyzed junctions from one experiment do not allow strong conclusions as experimental variation upon knockdown of CIP4, EEFA1A or VAV2 has not been assessed. For physiological relevant conclusions one would expect a set of at least 3 independent experiments. I do understand that the authors want to test their very interesting tool for which they provide many different experimental tests. However, the authors should then acknowledge and tune down mechanistic conclusions. For experiments concerning HUVEC cells, there is no information on the number of biological replicates on which the data are based.

We thank the reviewer for the suggestion and we endeavour to eliminate any over-interpretation. We are acutely aware that the above distinction should be made very clear to the readers, and are happy to clarify further. While there were caution notes in the text, we have revised the text to explain this point better, tone down the conclusions and extensively altered the Discussion section to focus more on Junction Mapper properties, advantages and caveats.

3) Thus, at present the way the data are presented now, provides nice evidence that the program allows for quantification of high number of junctions to potentially detect small differences. If the authors want to use the experiments shown in the manuscript to validate the usefulness as well as limits of the software, this approach with limited number of repetitions seems valid but then this has to be made more clearly throughout the manuscript. Not all of the data sets are sufficiently comprehensive to univocally demonstrate a biological relevant junctional outcome upon treatment or knockdown and this should be acknowledged. Having said this, I feel if the authors were able to address the concerns of reviewer 2 and 3, they should be provided the opportunity to properly address and include in the manuscript the comments above.

Please see our reply to point 2 above and to the other reviewers below.

Reviewer #2:

[…]

1) Most scientific conclusions as presented come from comparisons of values in control/disrupted. When looking at effects of user inputs the graphs are shown for each treatments individually but what are the changes with different thresholds/dilations between treatments (i.e. Figure 3—figure supplements 2 and 3)? Would different biological conclusions be drawn if you push the dilation or threshold "too far"?

We understand the reviewer point. In the analysis, we keep the threshold at a minimum to avoid loss of signal at junctions (subsection “Software development”, fourth paragraph). Very high thresholding levels are clearly not appropriate for the analyses, as too much signal is lost (i.e. threshold 100-150, Figure 3—figure supplement 3). Similarly, very high dilation settings may not detect any further increase in measurement values in stimulated samples, but not in controls, in some (Figure 3—figure supplement 2A-D), but not all experiments (Figure 3—figure supplement 2E-H).

As shown in our validation experiments, different users, using distinct dilation and thresholding settings obtain different absolute values for the parameters using Junction Mapper (Figure 7—figure supplement 1; Supplementary file 3). However, the conclusions are statistically similar when compared with controls quantified under the same conditions. As the settings are applied to both control and treated images, we expect that the results are not changed overall. Higher threshold levels will remove intensity at junctions and cytoplasm in both control and treated groups. Conversely, we expect that by increasing dilation much wider than the junctional staining, the background or noise will be computed as junctional signals for all samples (control and treated).

2) Authors point out a central issue and why this analysis is challenging – If there is no staining at a cell corner then where the program or user places the junction is arbitrary. Without a second marker of plasma membrane rather than junction this will always be an issue these measurements regardless of what automated algorithm is deployed or with user definition.

The reviewer is correct: a membrane or cell marker is very helpful to guide corner positioning. This can be done using cytoskeletal staining (i.e. F-actin) or a marker that labels the whole cell; i.e. Cell Mask. The above point is stressed in the third paragraph of the subsection “Software development”.

3) The authors show secondary characteristics are more robust to error in corner selection and that individual users get same direction of changes but not same absolute values, which does show the value added here. This should be emphasized in the Discussion.

Thanks for the suggestion – this information is now discussed in the subsection “User-bias validation” and in the first paragraph of the Discussion.

4) For the noise quantification, there should be a metric for measurement accuracy, or deviation from the ground truth. Guidance for user on minimum S/N needed for robust quantification. There is also no information on type of noise added (shot noise, dependent on pixel intensity is dominant noise in most microscopy images and is recommended here). The goal would be for a user to know if a junction has sufficient S/N to be analyzed.

We thank the reviewer for the recommendation. We performed the determination of peak signal to noise ratio (PSNR) in the original images compared to those where noise was artificially added using FIJI. Our quantification showed that a PSNR above 21.72dB +/- 0.67 is optimal and generally measured in the original images (cardiomyocytes and keratinocyte samples). Reducing the PSNR values to below 16.00dB (epithelia) or 18.00dB (cardiomyocytes) severely compromises the performance of the software. This information is now added to the subsection “Software validation” and in the Results. The PSNR values were added to each modified image (original and modified, Figure 2—figure supplement 3).

5) This leads me to the quantification of the type "ee" boarders in the H-Ras^G12V overexpressing cells. There appears to be greatly reduced e-cad staining at the boarders (it is difficult to see by eye any in the examples shown). In this case, how are the interfaces defined? Does this lower signal to noise negatively impact the program? How confident should we feel about comparing images with different S/N levels?

As stated in point 2 above, for severely perturbed junctions, the junctional remains, staining of neighbouring cells and with a second marker are helpful to delineate the borders. The visualization is also much better in the computer screen. Of note is that we find that samples that are severely disrupted are not easily rescued, as the perturbation goes too far. The experimental design is best when the expression/stimulus is carefully optimized to significantly disrupt in a quantifiable manner, but not completely abolish the junctional staining. The potential of the Junctional Mapper software is to be able to analyse high numbers of junctions and then obtain the sample variability in a statistically significant manner.

Author details

1. Helena Brezovjakova

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Contribution

   Formal analysis, Validation, Investigation, Visualization, Writing - original draft, Performed the formal analyses, software validation and visualization of the data

   Contributed equally with

   Chris Tomlinson

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-3554-6084

2. Chris Tomlinson

   Bioinformatics Data Science Group, Faculty of Medicine, Imperial College London, London, United Kingdom

   Contribution

   Software, Supervision, Validation, Methodology, Writing - original draft, Designed the software and computational solutions, Produced the instructions video and website

   Contributed equally with

   Helena Brezovjakova

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-8634-6111

3. Noor Mohd Naim

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Present address

   PAPRSB Institute of Health Sciences, Universiti Brunei, Darussalam, Brunei

   Contribution

   Data curation, Investigation, Methodology, Provided resources (image datasets)

   Competing interests

   No competing interests declared

4. Pamela Swiatlowska

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Contribution

   Formal analysis, Validation, Visualization, Methodology, Writing - original draft, provided resources (image datasets)

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-3705-7495

5. Jennifer C Erasmus

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Contribution

   Data curation, Supervision, Investigation, Visualization, Methodology, Writing - original draft, provided resources (image datasets) and software validation

   Competing interests

   No competing interests declared

6. Stephan Huveneers

   Department Medical Biochemistry, Amsterdam Cardiovascular Sciences, Amsterdam UMC, University of Amsterdam, Amsterdam, Netherlands

   Contribution

   Resources, Data curation, Validation, Writing - original draft, Writing - review and editing, Provided datasets

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-1091-475X

7. Julia Gorelik

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Contribution

   Conceptualization, Supervision, Funding acquisition, Methodology, Project administration, Writing - review and editing, Provided datasets

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1148-9158

8. Susann Bruche

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Present address

   Department of Anatomy, Physiology and Genetics, Oxford University, Oxford, United Kingdom

   Contribution

   Conceptualization, Software, Supervision, Investigation, Visualization, Methodology, Writing - review and editing, provided resources (image datasets) and software validation

   For correspondence

   susann.bruche@dpag.ox.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-5814-7166

9. Vania MM Braga

   National Heart and Lung Institute, National Institutes of Health, London, United Kingdom

   Contribution

   Conceptualization, Software, Funding acquisition, Writing - original draft, Project administration, Writing - review and editing

   For correspondence

   v.braga@imperial.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0546-7163

• 1,854

  Page views

• 333

  Downloads

• 1

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.