Motion sensing superpixels (MOSES) is a systematic computational framework to quantify and discover cellular motion phenotypes

To develop MOSES, we chose to investigate in vitro the boundary formation dynamics between squamous and columnar epithelia at the esophageal squamous-columnar junction (SCJ) (Figure 1A). To recapitulate features of the in vivo boundary formation, we used three epithelial cell lines in pairwise combinations and an experimental model system with similar characteristics to wound-healing and migration assays but with additional complexity. Together the resulting videos pose a number of analytical challenges that require the development of a more advanced method beyond the current capabilities of PIV and CIV.

Figure 1 with 5 supplements see all

Download asset Open asset

To model the relevant esophageal interfaces, we used three epithelial cell lines: EPC2, an immortalised squamous epithelial cell line from the normal esophagus (Harada et al., 2003); CP-A, an immortalised BE cell line with gastric columnar epithelial properties (Merlo et al., 2011); and OE33, derived from EAC (Boonstra et al., 2010). We co-cultured these lines in the following combinations: 1) EPC2:EPC2 (squamous:squamous, as a normal control); 2) EPC2:CP-A (squamous:columnar, as in Barrett’s esophagus); and 3) EPC2:OE33 (squamous:cancer, as in EAC) (Figure 1B). Two epithelial cell populations, each labelled with a different lipophilic membrane dye (Progatzky et al., 2013), were co-cultured in the same well of a 24-well plate, separated by a divider with width 500 µm. The divider was removed after 12 h and cells were allowed to migrate towards each other (Figure 1C).

We first compared the effects of the two dyes on proliferation and migration using monolayer combinations of the three cell types (Video 1). Proliferation was assessed by cell density, automatically counted from DAPI staining using convolutional neural networks (CNN; see Figure 1—figure supplement 1, Materials and methods). Migration (diffusive behaviour) was assessed by mean squared displacement (Park et al., 2015). Green and red fluorescent-labelled EPC2, CP-A, and OE33 cells proliferated at the same rate (Figure 1D, for EPC2 cells: slope = 0.976, Pearson correlation coefficient 0.978, Figure 1—figure supplement 1E), and migrated in a similar way (Figure 1—figure supplement 2). Compared to non-dyed cells, the mobility of CP-A cells were unaffected however both dyes equally reduced the mobility of EPC2 and OE33 cells. The diffusion modes of all cells were unaffected (Figure 1—figure supplement 2) and the boundary formation behaviour between dyed and non-dyed cells was identical.

Video 1

Download asset

We next analysed boundary formation in the three different combinations of co-cultured epithelial cell lines (EPC2:EPC2, EPC2:CP-A, and EPC2:OE33), initially in serum-containing media (5% fetal bovine serum, FBS). In all three combinations, as expected, both populations moved as epithelial sheets (Video 2). Firstly, in the squamous combination using the same EPC2 cell line labelled with two different colours, the green- or red-labelled EPC2:EPC2 cells met and coalesced into a monolayer as expected. Secondly, in the squamous-columnar EPC2:CP-A combination, we observed boundary formation between the two populations after 72 h, following a short period of CP-A ‘pushing’ EPC2. Thirdly, in the squamous-cancer EPC2:OE33 combination, the cancer cell line OE33 expanded continuously, resulting in the disappearance of EPC2 from the field of view (Video 2, Figure 1E) as assessed by the motion field and confocal images (Figure 1—figure supplement 3). The forces that govern the behaviour of the two cell lines on contact are unknown and traction force microscopy is required to investigate the ‘retracting’ or ‘pushing’ behaviour of EPC2 or OE33 cells respectively in future studies. Nonetheless, the observed boundary formation in the EPC2:CP-A combination and disordered interactions between EPC2:OE33 cells suggest that the three chosen epithelial cell lines in the tested combinations may model certain motion dynamics of the in vivo cell-cell interactions. Interestingly, cell counting suggests the observed phenotypes were not due to differential proliferation rate between the individual cell lines in combination (Figure 1—figure supplement 4C,D).

Video 2

Download asset

Evidence from model systems, including Drosophila melanogaster embryonic parasegment (Monier et al., 2010) and anteroposterior and dorsoventral wing disc boundaries (Major and Irvine, 2005; Major and Irvine, 2006; Landsberg et al., 2009), suggests the importance of cell-cell interactions and collective migration for stable boundary formation between epithelial populations in vivo. Thus to create a dataset in which we predict that cell motion would be dramatically disrupted we repeated the assays in serum-free conditions. We used convolutional neural networks (Materials and methods) to determine the mean cell density and mean change in cell density from video frames over the first 48 h of cells grown in serum (5% FBS) and no serum (0% FBS) conditions. We observed that culturing of the same cell combinations in serum-free versus 5% serum medium had undetectable impact on cell density at the very confluent cell conditions investigated (Figure 1—figure supplement 4). However, as expected, serum-free condition induced large global changes in cell migration (Figure 1F, Figure 1—figure supplement 5A) with observed loss of cell contacts, collective sheet migration (Figure 1—figure supplement 5B,C) and the absence of boundary formation. We also observed reduced overall boundary displacement (Figure 1F,G) and all cell line combinations exhibited similar motion dynamics (Figure 1F,G, Video 2). These results illustrate that serum-free medium has a profound impact on cell motion dynamics and the generated video datasets in serum-free medium are ideal as an experimental condition for testing the ability of MOSES to detect motion dynamics under different experimental conditions. It is important to note that the serum-free condition is not used as a biological negative control of boundary formation but only as a computational negative control for the development of our method.

We next developed a computational workflow, MOtion SEnsing Superpixels (MOSES), to quantify cellular motion from video datasets. MOSES was formulated modularly with three components (Figure 2): 1) motion extraction; 2) construction of long-time motion tracks; and 3) capture of local dynamic context.

Figure 2 with 2 supplements see all

Download asset Open asset

Component 2: Construct long-time motion tracks

To capture all spatial motion dynamics in the video, we construct long-time motion tracks that continuously track the cellular motion for any amount of time using superpixels. Local neighbouring image pixels are grouped into a pre-specified (see below) number of image regions or ‘superpixels’ to enable spatial averaging of the pixel-based motion fields (Figure 2A, Step 1). At each time point, the (x,y) position of a given superpixel is updated by averaging the optical flow (Farnebäck, 2003) over the image area covered by the superpixel to obtain the displaced direction and distance (Figure 2A, Steps 2-4). Subsequently, long-time superpixel motion tracks are generated by concatenating the positions of each superpixel at all time-points (Figure 2A, Step 5). All long-time superpixel motion tracks together encode the entire spatio-temporal motion history within the video. By default, the number of superpixels used are user-specified and fixed to sufficiently subsample only the field-of-view of the initial video frame. Where a video exhibits highly dynamic movement and coverage of the full spatio-temporal motion, including the movement of new cells not initially present but which later move into the field of view, is necessary for analysis, additional superpixels can be dynamically added during tracking. Tracks produced by the latter approach are known as dense trajectories (Wang et al., 2011). In our experiments, we used a fixed number of 1000 superpixels to cover an image size of 1344 x 1024 (with this setting an average superpixel is 37 pixels x 37 pixels covering $\approx$ 14,000 μm² (2x lens) or $\approx$ 3600 μm² (4x lens)) to monitor epithelial sheet dynamics. For multichannel images, each channel representing different coloured cell populations (i.e. red and green in our assay) was independently tracked in this manner (Figure 2B).

Depending on the specific application, background artefacts such as stage drift or floating debris, can contribute tracks that are biologically irrelevant and need to be filtered in post-processing (Figure 2B, middle panel). In this filtering step, superpixels are assigned to cover only the dynamic motion of all ‘objects’ of interest. The ‘object’ for single-cell tracking is each individual cell, and for epithelial sheets is the entire sheet within the video frame. To assign superpixels to each sheet, we use motion information (Figure 2—figure supplement 1, Materials and methods). This approach avoids relying on image intensity features whose variation commonly leads to segmentation errors. We validated our motion extraction and long-time track construction using published single-cell tracking datasets from the cell tracking challenge (Maška et al., 2014) (Video 3), and also experimentally, by spiking one of the dyed cell populations in the in vitro setup described above with a third, sparse population of blue-dyed cells. Our method produced very similar tracks (similarity score of 0.8) to an open software tool for single-particle tracking, TrackMate (Tinevez et al., 2017) (Figure 2—figure supplement 2, Video 4).

Video 3

Download asset

Video 4

Download asset

To verify the enhanced ability of MOSES over standard PIV analysis to quantitatively assess relevant biological features of interest in tissue, we applied it to analyse the formation of boundary dynamics between squamous and columnar cells. In total, 125 videos (48 with 0% serum and 77 with 5% serum) were collected from four independent experiments and jointly analysed. These videos are highly heterogeneous, creating a challenging dataset for analysis (Supplementary file 1, Figure 3—figure supplement 1).

Standard velocity kymographs were able to firstly confirm our observations in Figure 1E–G of the differences in cell motion between 0% and 5% serum conditions. As shown in Figure 3A, all cell combinations grown in 0% serum exhibited a similar global motion pattern, minimal sheet-like motion and interface movement following gap closure. By contrast, in 5% serum the same three cell line combinations exhibited very different dynamics, amongst which we observed the formation of a stable equilibrium following contact only between EPC2 and CP-A cell sheets. However, subsequent computation of the mean sheet speed, as performed during routine PIV analysis, highlighted the analytical need for a computational framework to extract more descriptive motion measures. As shown in Figure 3B, aside from an indication of a global speed increase in 5% serum, speed was an insufficient metric to characterise the observed phenotypes, exhibiting high variance between replicates and failure to highlight the uniqueness of the EPC2:CP-A interaction. We thus used MOSES to construct more discriminative phenotypic measures to characterise the boundary formation of EPC2:CP-A.

Figure 3 with 10 supplements see all

Download asset Open asset

Using MOSES, we constructed a motion saliency map for visualising global motion patterns (Figure 3—figure supplement 2) and designed four different additional measures as summarised in Table 1 to quantitatively distinguish between different combinations of the three cell lines (EPC2, CP-A and OE33) used in our assay: i) boundary formation index, based on the degree that motion concentrates into a local spatial region (Figure 3—figure supplements 2 and 3); ii) mesh stability index, which measures the degree of movement between neighbouring cell groups locally within the sheet at the endpoint (Figure 3—figure supplement 4); iii) mesh order to measure the degree of collective motion of the whole sheet (Figure 3—figure supplement 5, Video 5) and iv) the maximum velocity cross-correlation, which measures the average correlation in motion between the two epithelial sheets from their movement history (Figure 3—figure supplement 6). The first three are based on the mesh; the fourth uses the individual superpixel tracks. These measures represent examples of statistics derivable using MOSES and are not limited to wound-healing type assays. For technical details, see Materials and methods. Cells grown in 0% serum were used as a computational negative control to set appropriate cut-offs for detecting boundary formation in our computational analysis (see Materials and methods for details).

Video 5

Download asset

Computing the proposed indices for all videos, the boundary formation index was ranked on a continuous scale for all 5% serum videos (Figure 3—figure supplement 7) by MOSES. The boundary formation index was highest (median 0.74) for EPC2:CP-A grown in 5% serum (n = 30/77) (Figure 3C, Figure 3—figure supplement 8B), whilst EPC2:EPC2 and EPC2:OE33 in 5% serum were below the boundary formation cut-off. In 5% serum, EPC2:CP-A had the highest mesh stability index (median 0.90), compared to EPC2:EPC2 (0.76) and EPC2:OE33 (0.25) (Figure 3D, Figure 3—figure supplement 8C). Together these results illustrate that only the squamous-columnar combination (EPC2:CP-A) formed a final stable boundary.

We next measured sheet-sheet interactions using the maximum velocity cross-correlation (VCC) before and after gap closure for the three cell line combinations. Gap closure was automatically determined (Figure 3—figure supplement 9 , Materials and methods). For two initially diametrically opposed migrating sheets, a significant increase in VCC after gap closure compared to before gap closure is suggestive of increased sheet interaction. In 0% serum, no difference in velocity cross-correlation across all combinations was found before and after gap closure (Figure 3E ) - the two sheets do not move cohesively as a unit, as expected with minimal cell-cell contact. In serum, the difference for EPC2:CP-A (0.03 before and 0.20 after gap closure) was $~$3-6 times larger than for EPC2:EPC2 (0.01 to 0.08) and EPC2:OE33 (0.02 to 0.05) (c.f. left and right violin plots in Figure 3E ), suggesting potential physical interaction. Interestingly, CP-A:OE33 (Barrett’s:cancer, n = 6) also exhibited a substantial increase in VCC following gap closure (0.03 to 0.17) (Figure 3—figure supplement 8D,E) but no substantial increase was observed for CP-A:CP-A (0.01 to 0.06) (Figure 3—figure supplement 8D,E), confirming this observation was not simply CP-A cell line-specific. Both the velocity order parameter and our mesh order parameter (Figure 3F,G) indicate clear overall reduction in collective motion in cells grown in 0% serum compared to 5% serum.

Finally, we used MOSES with image segmentation and convolutional neural network (CNN) cell counting to assess morphological variations of the interface among all cell line combinations to check for ‘invasive fingers’ or boundary breakdowns that often exist in normal and tumour cell-formed boundaries (Figure 3—figure supplement 10A,B). Interestingly, we found non-invading boundaries in all cell line combinations except CP-A:OE33, where we observed clear infiltration with ‘finger-like’ protrusions of CP-A into the OE33 cell sheet grown in 5% serum (Figure 3—figure supplement 10B). No such ‘finger-like’ protrusions or significant intermixing of cells (Figure 3—figure supplement 10D,F–H) were detected in EPC2:CP-A under the same conditions. Altogether, our characterisation suggests that, among all tested cell-type combinations, we detect an ‘interacting’ final stable boundary uniquely in the squamous-columnar EPC2:CP-A combination grown in 5% serum (Video 6).

Video 6

Download asset

The ability to assess subtle phenotype changes with a minimal number of replicates is critical for high-content screening applications. We thus tested the ability of MOSES to detect changes in EPC2:CP-A boundary formation using an external perturbation. The epidermal growth factor receptor (EGFR) is frequently mutated in EAC, sometimes overexpressed in BE and is activated by bile acid reflux, the main observed cause of BE clinically (Dixon et al., 2001; Souza, 2010). In our experiments we thus used epidermal growth factor (EGF), the ligand of EGFR, as a biologically relevant external perturbation. Increasing amounts of EGF (0 ng/ml to 20 ng/ml) were added to the culture medium to assess incremental effects on cellular motion and boundary formation in the EPC2:CP-A combination. A total of 40 videos (each 144 h, one frame per h) were collected from three independent experiments, in a 24-well plate medium-throughput screen (Supplementary file 2). With increasing EGF, the boundary position was displaced a distance farther from the initial point of contact between the two cell populations, with slightly enhanced cell speed and decreased boundary coherence (Figure 4A–D). This is quantitatively reflected in the shape of the mean normalised strain curve (Figure 4E) which measures the average distance between neighbouring cell groups (Materials and methods): at 0 ng/ml EGF, the curve linearly increases before plateauing around 72 h; as EGF concentration increases, the curve becomes more linear and the plateau is lost above 5 ng/ml.

Figure 4 with 2 supplements see all

Download asset Open asset

The boundary formation index decreased with increasing EGF (0.74 at 0 ng/ml to 0.46 at 20 ng/ml, comparable to EPC2:OE33 without EGF (0.46)), indicating loss of the boundary (i.e. index below the 0.69 cut-off) (Figure 4F). The mesh stability index decreased from 0.94 (stable, 0 ng/ml EGF) to 0.72 (unstable, 20 ng/ml) (Figure 4G), suggesting increased movement between neighbouring cells and the loss of interaction between the two cell populations since the maximum VCC difference before and after gap closure decreased from 0.16 (0 ng/ml EGF) to 0.04 (20 ng/ml EGF) (Figure 4H). The maximum VCC after gap closure was similar to that for EPC2:OE33 (0.02) (Figure 4H), but the mesh stability index remained higher (Figure 4G). Together these measures show that above 5 ng/ml EGF, the phenotype of EPC2:CP-A becomes similar to that between EPC2 and the EAC cell line, OE33 (Video 7). Cell counting and quantification of fluorescence decay suggest cell division is not the primary factor influencing the boundary at the cell density used in our experiments in 5% serum (Figure 4—figure supplement 1). Of note, titrating EGF did not rescue the effect of serum absence, with non-significant changes in the boundary formation index (0 ng/ml: 0.60 ± 0.07, 20 ng/ml: 0.65 ± 0.03) and maximum velocity cross-correlation, although the mesh stability index decreased, likely due to increased cell movement (Figure 4—figure supplement 2 (n = 25)).

Video 7

Download asset

Using both the velocity order parameter and mesh order index to characterise collective motion, we found little change in collective motion upon EGF addition in 0% serum (Figure 4—figure supplement 2J,K), which might explain the lack of boundary formation. However, in 5% serum plus EGF, the two measures exhibited opposite results: raising EGF concentration increased velocity order but decreased mesh order, Figure 4I,J. We note however that the mesh order, by explicitly accounting for the motion of neighbouring cells, better reflects human observation of motion in videos (Videos 5–7). This highlights the pitfalls of only quantifying the individual alignment of velocity vectors computed from one video frame. Cell counting for 0% serum again suggested minimal influence of cell proliferation (Figure 4—figure supplement 2L,M). In summary, this example with EGF demonstrates that MOSES enables robust continuous-scale quantification of motion phenotype following systematic perturbation.

High-content imaging screens are often explorative, with the aim of screening for unknown differences in complex cellular motions from a large number of videos in an unbiased manner (Zaritsky et al., 2017). In general it is therefore not known a priori the behaviour of the imaged cells. MOSES addresses this need for an unbiased phenotyping approach by enabling the systematic generation of unique ‘motion signatures’ for individual videos in a manner similar to the relatively automatic generation of geometric features for cell shape quantification in high-content image screens (Boutros et al., 2015; Bray et al., 2016). Below we demonstrate that unsupervised machine learning techniques requiring no manual user annotation can be applied to MOSES generated signatures to visualise all videos onto a 2D motion phenotype map. This advanced feature of MOSES enables easy visual assessment of motion phenotype and the generation of hypotheses without the need to individually interrogate each video.

The general process for the motion map generation is illustrated in Figure 5A. To position each video on a 2D map, we applied principal component analysis (PCA) to the normalised mesh strain curves of the 77 videos of all cell line combinations cultured in 5% serum conditions to learn the principal component vectors that define the x-y axis of the 2D map. The normalised mesh strain curve for each video was used here as an example 1D motion signature to summarise the entire video motion (see Materials and methods for constructing more descriptive signatures). The constructed PCA map of the 77 videos from cells cultured in 5% serum (Figure 5B) shows that this unbiased approach automatically clusters all videos corresponding to each cell line combination. Furthermore, the videos were ordered in a continuous manner, as shown by the increasingly linear shape of the mean normalised strain curve when looking left to right across the plot in Figure 5B from CP-A:CP-A to EPC2:OE33. This clustering was not achieved with root mean squared displacement (RMSD), the non-mesh version of the MOSES-normalised mesh strain curve (Figure 5—figure supplement 1). Moreover, it appears independent of the particular dimensionality reduction technique used (Figure 5—figure supplement 2), indicating that the signatures constructed using MOSES are intrinsically informative. Finally, the 1D MOSES-based motion signatures trained a machine learning classifier with no further processing to predict cell combination identity better than RMSD (Figure 5—figure supplement 2).

Figure 5 with 2 supplements see all

Download asset Open asset

To demonstrate how such 2D motion phenotype maps can be used to compare videos, we next mapped the 48 videos from 0% serum cultures on the same axes as the videos from 5% serum cultures (Figure 5C,D). The videos from 0% serum mapped to a different area of the 2D plot, whilst preserving the continuous ordering of the previous videos. Therefore, without having watched the videos, it is easy to predict that the cells have markedly different motion dynamics in 0% serum compared to 5% serum. Furthermore, since the points for the 5% serum videos cover a larger area of the 2D plot than the 0% serum videos, one can predict more diversity of motion in 5% serum (Figure 5D).

The motion map can also capture subtle changes in dynamic behaviour. This is demonstrated by mapping the mean video motion for each concentration of EGF from 0 to 20 ng/ml (represented by the respective mean normalised strain curves for each concentration (one per concentration from a total n = 40 videos, Figure 4E) onto the same axis as the 5% serum videos in the absence of EGF (square points in Figure 5E). With increasing EGF, the squamous-columnar EPC2:CP-A motion dynamics become increasingly similar to squamous-cancer EPC2:OE33 above 5 ng/ml, as evidenced by the square points moving from the area of blue circular EPC2:CP-A points into the area of orange circular EPC2:OE33 points. Thus our motion map is consistent with the result using the specific derived measures (above). These results illustrate the ability to detect biological and technical variability across independent experiments and that MOSES possesses the required features for an algorithm to be used in an unbiased manner in high-content screens with minimal prior knowledge.

Finally, to further illustrate the full potential of MOSES and to demonstrate its application, we compared MOSES with the widely used PIV method on two published timelapse microscopy datasets of epithelial monolayers. In the original publication, Malinverno et al. (2017) used PIV to describe the induction of large-scale coordinated motility in MCF-10A RAB5A expressing cells compared to MCF-10A control cells. In the publication associated with the second dataset, Rodríguez-Franco et al. (2017) used PIV to show the detection of deformation waves that propagate away from the cell boundary between two epithelial monolayers. The MDCK cell monolayers expressed EphB2 and its ligand ephrinB1, respectively.

Reanalysing the datasets with MOSES, we found that motion fields inferred from optical flow by MOSES were similar to those from PIV, yielding both similar speed curves and velocity kymographs. However, MOSES exhibited greater sensitivity to salient motion events (indicated in Figure 6A). Compared to PIV velocity vectors, MOSES superpixel tracks are a more data-efficient (see Discussion) encoding of the spatio-temporal velocity distribution that naturally enhances and preserves the salient motion. Reconstructed velocity kymographs from the MOSES motion trajectories capture not only the pattern of the full velocity kymograph but further selectively enhanced the detection of the deformation wave signature formed at the interface between EphB2/ephrinB1 epithelial monolayers (as indicated in Figure 6A, right panel). Thus, all velocity-based statistics that can be derived from PIV, such as the velocity order parameter, are fully preserved in MOSES. Yet, MOSES offers additional advanced possibilities. Firstly, instantaneous velocity-based measures from single videos commonly derived from PIV are noisy. For example, the velocity order parameter variation for the slower MCF-10A control cells is non-smooth and highly variable between consecutive time points (Figure 6B). This leads to the misinterpretation that at certain time points, MCF-10A control cell motion is more collective than MCF-10A RAB5A expressing cells following induction. In contrast, the MOSES mesh order exploits long-time continuity and neighbourhood relations to robustly capture collective motion in a manner consistent with human observation (Video 5). Secondly, long-time MOSES tracks and superpixel mesh strain curves can unbiasedly cluster the global motion pattern into small spatio-temporal groups (Figure 6C, Video 8). This provides a systematic approach to the interrogation of motion sources (Figure 6D, Video 9). The computation of motion saliency maps from forward and backward tracked frames effectively highlight the spatial concentration of motion, and the boundary formation index attests to the efficacy of the MOSES-enabled measures across independent datasets. Finally, and uniquely, MOSES’ meshes and tracks present a systematic framework for users to define extensive sets of custom measures to comprehensively characterise complex motion phenotypes, such as the investigated boundary formation behaviour between epithelial cells. By exploiting this, we successfully constructed feature descriptors to generate interpretable motion maps for our large video dataset. Figure 6E summarises the key points of comparison between MOSES and PIV.

Video 8

Download asset

Video 9

Download asset

Figure 6

Download asset Open asset

EPC2 (from Prof. Hiroshi Nakagawa, University of Pennsylvania, Perelman School of Medicine, Department of Gastroenterology, USA) and CP-A (ATCC) cells were grown in full KSFM (Thermo Fisher), AGS (ATCC) and OE33 (ECACC) in full RPMI with 10% FBS. Both were supplemented with glutamine and Penicillin streptomycin at 37°C and 5% CO₂ until 80% confluent. To passage EPC2 and CP-A, cells were resuspended after trypsinization for 5 min with PBS supplemented with soybean trypsin inhibitor (0.25 g/L, Sigma) to prevent cell death prior to resuspension in KSFM. To store, cells were resuspended at a concentration of 10⁶ cells/ml with 90% FBS +10% DMSO freezing media following centrifugation and stored at −80°C before passing to liquid nitrogen storage. All cell lines were tested monthly for mycoplasma infection using MycoAlertTM PLUS Mycoplasma Detection Kit (Catalog #: LT07-705 from Lonza) at the Ludwig Cancer Institute, Oxford, UK and have not shown evidence of Mycoplasma. Cell lines have been authenticated by Eurofins.

Cells were labelled using Celltracker Green CMFDA and Celltracker Orange CMRA dyes (Life Technologies) according to protocol. Two different concentrations 2.5 µM and 10 µM were used. The lower concentration still permits tracking but has fewer toxicity concerns.

Samples were washed twice with PBS and fixed with 4% PFA for 15 min. The samples were then washed twice and permeabilised with 0.1% Triton-X for 15 min. Samples were blocked with image-iT Fx signal enhancer for 1 h before incubation with E-cadherin (610181 Becton Dickinson U.K. Limited) overnight at a 1:400 dilution in Agilent Dako Antibody Diluent with Background Reducing Components (S3022, Agilent Dako) at 4°C. Following 3 washes of 10 min in PBS, samples were incubated for 2 h in the dark with Alexa 488 goat anti-mouse (A11001, Thermo Fisher) secondary antibody and Alexa Fluor 647 Phalloidin (A22287, ThermoFisher), both at a 1:400 dilution and DAPI (1 mg/mL) at 1:1000 dilution (62248, Thermo Fisher) in Agilent Dako Antibody dilutent (S0809, Agilent Dako). Finally, samples were washed three times with PBS and mounted using Fluoromount-G (0100–01, SouthernBiotech).

In 70 µL of culture media, 70,000 labelled cells were seeded into each side of a cell culture insert (Ibidi) in one well of a 24-well plate. After 12 h, inserts were removed and the well washed three times with PBS to remove non-attached cells before adding the desired media (KSFM (0% serum in the text) or 1:1 mixture of KSFM:RPMI + 5% FBS (5% serum in the text)) for filming. For the perturbations, the effector, for example EGF, was also added to the media at the stated concentrations.

A third population of cells dyed with 2.5 µM of CellTracker DeepRed (Life Technologies) was diluted 1:200 into one of the other two populations dyed with either Celltracker Green CMFDA (Life Technologies) or Celltracker Orange CMRA (Life Technologies) of the same cell line. The mix was then added to one side of the insert (Ibidi) as with the two cell population experiments described in the main text.

The different conditions were filmed on a Nikon microscope for 96 or 144 h at a frequency of 1 image per hour. 2x and 4x objectives were used. The microscope filter wavelengths used to visualise the red and green dyes were 546 nm and 488 nm, respectively.

The convolutional neural network (CNN) density counting approach of Xie et al., 2016 was used to automatically count cells in confocal DAPI-stained nuclei images or fluorescent and phase contrast video frames. Once trained, given an input image the CNN model outputs a dot-like image with the property that the sum of all pixel intensities in the output dot-like image equals the number of cells within the image (Figure 1—figure supplement 1A). We describe the training method for DAPI images only. Other image modalities were trained in a similar manner details of which are given below under ‘Migration independent cell counting in videos’. To generate the training data for DAPI cell counting, 200 image patches (size 256 pixels x 256 pixels) were first randomly extracted from the acquired DAPI images of resolution 4096 pixels x 4096 pixels (Figure 1—figure supplement 1A). For each extracted 256 × 256 image patch, the stained nuclei centroids were manually marked using a ‘dot’. Then for each 256 × 256 image patch, 50 randomly sampled 64 pixel x 64 pixel patches were extracted to yield a total training set of 10,000 image patches (Figure 1—figure supplement 1A). The CNN training settings used for a 70:30 train-test split were 200 epochs, batch size 100, RMSprop (lr = 0.001, rho = 0.9, epsilon = 1e-08, decay = 0.0) with a mean absolute error (MAE) loss. The final test accuracy was reported as the mean absolute deviation (MAD) between manually counted and predicted cell counts on the result of applying the learnt CNN to the larger 256 × 256 manually labelled image patches (Figure 1—figure supplement 1B,C). To count specific cell types stained with different coloured dyes from confocal images, epithelial sheets were segmented from their respective channels by applying k-means clustering on the RGB image pixel intensity values with k = 3 (retaining the two clusters of highest intensity) after downsampling the full-size images (4096 × 4096 pixels) by a factor of 4 (to 1024 × 1024 pixels). Small objects (<200 pixels) were removed, holes were filled and the largest connected component kept before upsampling the binary mask to its original resolution (4096 × 4096). The respective final binary mask was used to mask and count cells specific to each colour channel (Figure 1—figure supplement 1C). Counting of specific cell types in timelapse video frames, which are more variable in quality, is similar but uses the more optimized image segmentation protocol detailed below to segment individual epithelial sheets.

To count the migrating cells in the videos, two different approaches were used. The first and default approach used in this paper is the automatic CNN counting described above trained on manual annotations of the fluorescent video frames (Figure 4—figure supplement 1A,B). To bypass the issue of moving areas, we first produced a binary mask based on image segmentation as described below to identify the respective ‘red’ and ‘green’ cell sheet areas to quantify. Using the mask, equal sized image patches of 64 × 64 were randomly sampled. These were then fed into the trained CNN to produce cell count estimates for each image patch. The average cell density over 100 random such image patches were taken as an estimate of cell density for the entire frame. This operation was repeated for each channel separately and for every time point to yield a temporal cell count profile. The proliferation rate was subsequently estimated as the average absolute cell density change in successive frames normalised by the mean cell density (computed over the desired time frame) (Figure 4—figure supplement 1C). To check results were unaffected by the imaging modality, CNN based counting was also applied to the corresponding phase contrast images where the fluorescence channels were used to identify the individual cell types (Figure 4—figure supplement 1D). The second approach exploits the change in dye fluorescence as cells proliferate (the cell dye intensity decreases as they divide) (Figure 4—figure supplement 1E–H). For our videos, the image intensity exhibits too discrete a transition between time points (Figure 4—figure supplement 1F) for accurate extraction of the fluorescence decay. Instead, given that all videos of EGF addition to EPC2:CP-A were of the same temporal length (144 h), we found that the modal image intensity and modal normalised frequency of the intensity histogram peak at every time point served as an alternative quantification of fluorescence decay that yields the same conclusions (data not shown) but could be robustly fitted with a linear best-fit line (Figure 4—figure supplement 1I). The faster the fluorescence decay, the higher was the cell proliferation and the higher the absolute value of the fitted linear gradient that we used as a proxy proliferation coefficient (Figure 4—figure supplement 1J,K).

Red and green channel images were anisotropically filtered (Perona et al., 1994) to enhance image edges and suppress stochastic image noise. The entropy image was then computed to enhance the epithelial sheet. Otsu thresholding was subsequently applied to obtain red and green binary masks of the sheet. Finally ‘holes’ in the resultant masks were filled using a line rastering approach.

A custom image edge finding script using Sobel filters to detect edges on downsampled fluorescent images was used to find the leading sheet edge (in terms of an average x-coordinate) of both cell populations for each time frame (the averaged y-coordinate of the boundary did not change significantly over the filmed duration). The gap closure point was determined by the intersection of the two x-coordinate curves. The displaced distance of the boundary following gap closure was computed as the absolute distance between the final frame averaged x-position and the gap-closure frame averaged x-position.

MOSES was developed using the Python Anaconda 2.7 distribution, in particular it uses Numpy, the Scipy-stack and OpenCV libraries. It comprises separately a cell tracking and data analysis component.

Regular superpixels were generated by applying the SLIC (Achanta et al., 2012) algorithm in scikit-image on a blank image, the same size as the video frame. 1000 superpixels were used throughout in this paper. Motion fields for updating the superpixel centroid position over time were computed with OpenCV Farnebäck optical flow (Farnebäck, 2003) using default parameters. For ease of implementation, displacement vectors were rounded to the nearest integer. Superpixels passing out of the frame progressively lose pixels and retain their last motion position for the tracking duration. For simplicity, lost area is not recovered. PIV was computed using the Python openpiv package, using the closest equivalent parameters to MOSES; window size of 16 pixels, overlap of 8 pixels and search area of 32 pixels.

Step 1: Given the tracks of one colour for example red, identify all superpixels that initially move by thresholding on the cumulatively moved distance within the first few frames, (here we used 2 frames, equivalent to 2 h). Step 2: Form the superpixel neighbourhood graph by connecting together all identified moving superpixels from step 1 to any other identified moving superpixel within a specified radial distance cutoff (1.2 x average superpixel width here) using their initial (x,y) centroid positions. The largest connected graph component is then found to approximate the covered area of the epithelial sheet at frame 0. Step 3: In some videos, image artefacts such as autofluorescence or the presence of isolated cells contributes superpixel tracks that are biologically irrelevant and affects quantification of the migrating sheet dynamics. Tracks associated with these noise sources must be removed. The need for such removal is automatically evaluated through a user-set cutoff based on prior knowledge of the expected maximum fraction of the field-of-view covered by any one of the red or green populations at time $t$ = 0 (e.g. for a 50:50 plating of red and green cells, a conservative cut-off fraction of 0.70 is used here). Assuming that no red and green superpixel can jointly occupy the same (x,y) position in frame 0, joint filtering based on the degree of movement is applied to clean segmentation errors from previous steps. Step 4 (not required if running MOSES using dense tracking): The kept superpixels after steps 1-3, are then iteratively dynamically propagated to identify ‘activated’ superpixels (those that lie in the joint area occupied by the kept superpixels) frame-by-frame. Step 5 (potentially optional): In case the dynamic superpixel propagation identifies superpixels that do not move much over the entire video and to ensure the temporal continuity of superpixel positions, steps 1-2 are repeated. Step 6: To ensure the same number of superpixel tracks across all videos for statistical comparison, finally we assign constant tracks for all unused or inactivated superpixel tracks where for all frames their (x,y) positions are fixed to their initial centroid position. The procedure described is illustrated more concisely in Figure 2—figure supplement 1A.

To generate meshes, each superpixel is connected to its nearest ‘neighbours’. The notion of ‘neighbour’ is mathematically defined by the user (see below). For the MOSES mesh used here for visualisation and stability analysis, we defined ‘neighbours’ using a pre-set distance cut-off based on the distance between individual superpixel centroids at the start of tracking. The mesh strain curve thus measures the relative distortion between connected superpixels with respect to their initial mesh geometry over time. For computation of the boundary formation index, a different mesh was used where neighbours were independently determined frame-by-frame by a pre-set distance threshold. The specified threshold in both meshes are given in Euclidean distance as a multiplicative factor of the average superpixel width used. A factor of 1.2 for the mesh strain and 5.0 was used for the boundary formation index throughout.

Numerous ways exist to connect a collection of (x,y) points to form a graph or mesh. Depending on the mesh formed, different aspects of the movement can be enhanced and probed in interesting ways. In this paper, the presence of collective motion biologically motivates the mesh concept. More generally, meshes are ‘abstract’ constructs to assess relationships. In terms of motion analysis, in this paper, we recognise that different spatially located points may be correlated particularly if they are spatially close. In view of this, we first describe and explain the rationale of the two meshes constructed in the main text for visualising and quantifying epithelial sheet dynamics before describing possible extensions and how these may be more useful for specific experiments. The first and the primary mesh used throughout is what we term the MOSES mesh. It is constructed by joining each superpixel track with all superpixel tracks whose initial (x,y) centroids are within a user-specified constant Euclidean distance cut-off. Implicitly this assumes that under collective sheet migration, initially spatially close superpixels continue to remain spatially close. Violation of this assumption leads to large stretching or compression of the MOSES mesh, which can be used to derive a measure of motion collectiveness (see mesh order below). The second mesh is used to generate the motion saliency maps for localising boundary formation, a dynamic state that varies frame to frame. Each superpixel track is again connected to all superpixel tracks whose (x,y) centroids are within a user-specified Euclidean distance cut-off. However, contrary to the first mesh, which uses only the centroid positions in frame 0, here the distances are determined using the current position of all superpixels in the current frame. Thus the ‘neighbours’ continuously change frame-to-frame. We refer to this mesh as the radius neighbour graph and the motion saliency map is the result of counting the number of neighbours for each superpixel (i.e. the node degree). The number of surrounding neighbours increases in spatial areas where motion concentrates e.g. there exists a local chemoattractant or a physical impedance to cell movement such as a boundary. Thus boundaries are natural attraction motion centres; leading superpixels at the boundary cannot advance whilst those behind continue to move towards the boundary. This leads to an overall accumulation of superpixels at the boundary (c.f. motion saliency maps in Figure 3—figure supplement 2). In both meshes, tuning the distance cut-off threshold tunes the length-scale of the spatial interaction one wishes to analyse. For boundary formation, a phenomenon that spans the entire height of the image, one should choose a relatively large distance cut-off such as 5x the average superpixel width compared to, for example, identifying the localisation of a macrophage to a cell apoptosis site. In the latter, the attraction site is more point-like and a radius cut-off 1x the average superpixel width may be more accurate. Finally, whilst both of the presented meshes in our paper only utilise physical distance to define neighbours this is by no means the only possibility. In some applications such as flocking, the topological distance (i.e. the number of points away) not the physical distance may be more relevant (Ballerini et al., 2008). In this situation, it is more common to construct the local kNN (k-nearest neighbour) graph, designating the closest k superpixels as neighbours. The kNN is also frequently used when the magnitude of the ‘interaction’ between points is not known a priori as a way to propagate local information and mine patterns in data c.f. t-SNE for dimensionality reduction, spectral clustering and similarity network fusion (Wang et al., 2014) for combining multimodal datasets. Finally, superpixels may be joined not only spatially but also temporally to enforce consistent temporal neighbour relations (Chang et al., 2011) as well as according to more ‘semantic’ notions of similarity for example similar image appearance, similar instantaneous velocities (Chang et al., 2011). In short, by formalising motion analysis under the framework of dynamical meshes that connects together ‘neighbouring’ superpixels, we can analyse complex motion not just in terms of instantaneous speed and orientation but can additionally leverage powerful established tools developed in the fields of computational graph theory, network theory, topology etc. to effectively quantify and mine increasingly complex motion phenotypes in high-content screens.

As a measure of cellular motions, MSD was computed as a function of time interval, $\mathrm{\Delta }t$ as in (Park et al., 2015).

where ${\displaystyle {\mathbf{r}}_i\left(t\right)}$ is the position of the superpixel $i$ at time ${\displaystyle t}$ and $⟨\cdot ⟩$ is the average over all time $t$ and all superpixels. For small $\mathrm{\Delta }t$, the MSD increases as a power law $\mathrm{\Delta }{t}^{\alpha }$, where the exponent ${\displaystyle \alpha }$ is determined empirically by fitting. For unity exponent ($\alpha =1$), the movement is uncorrelated random Brownian motion and cellular motion is diffusive. When ${\displaystyle \alpha >1}$, cellular motions are super-diffusive, and when $\alpha =2$, motions are ‘ballistic’.

As a summary of the whole video motion and a measure of movement, the root mean squared displacement was computed as a function of time relative to the initial time ${\displaystyle t_0}$, ${\displaystyle \text{RMSD}(t)=\sqrt{⟨{\left|{\mathbf{r}}_i(t)-{\mathbf{r}}_i(t_0)\right|}^2⟩}}$ where ${\displaystyle {\mathbf{r}}_i(t)}$ is the position of the superpixel ${\displaystyle i}$ at time ${\displaystyle t}$ and ${\displaystyle ⟨\cdot ⟩}$ is the average over all time ${\displaystyle t}$ and all superpixels. For multi-channel videos, the average RMSD was used to describe video motion. Unless otherwise stated in the text, the normalised RMSD (here division by maximum value within the common time window of comparison) was plotted to permit comparison across different conditions. To compare across videos of different duration, in Figure 1—figure supplement 4B we instead compute the RMSD divided by its value at 96 h, the maximum timepoint shared by both our 96 h and 144 h videos.

For each superpixel $i$ we define its neighbourhood, $N_i$ as all superpixels $j$ which lie within a specified distance, $r$. Given the time-dependent velocity function $V\equiv V\left(t\right)=\mathbf{r}\left(t+1\right)-\mathbf{r}\left(t\right)$ where $\mathbf{r}\left(t\right)$ is the track (all (x,y) positions) up to time $t$, the normalised spatial correlation of a video with a total of $n$ superpixel tracks is defined as

where $\mathbb{E}\left[\cdot \right]$ is the mean function averaging over the superpixel neighbourhood, $cov(\cdot )$ is the covariance function and ${\sigma }_{V_i}$ is the standard deviation of $V_i\left(t\right)$. Computing spatial correlation as a function of $r$ for our videos yields an exponential decay which can be fitted to an equation of the form $y=a{e}^{-x/b}$ from which the initial correlation ${\displaystyle a}$ and characteristic correlation distance ${\displaystyle b}$ can be determined for plotting, (Figure 1—figure supplement 5). In our plots, the distance $r$ is in terms of normalised units (i.e. the number of average superpixel widths away).

MOSES does not explicitly handle individual cells leaving and entering the field of view or cell divisions during filming. For fair comparison of motion capture ability with manually annotated tracks, the tracks were only compared for cells present in the initial frame as depicted with coloured masks (Figure 2B). Single-cell tracks were generated from MOSES (1000 superpixels) by identifying the superpixel track that has moved the greatest distance over the video duration amongst all superpixel tracks whose initial (x,y) position lies within the area of the individual cell of interest as marked out by manual annotation at t = 0 (Video 3). 

The Fiji TrackMate plugin (Tinevez et al., 2017) was run on the third blue image channel containing only the sparse population of DeepRed dyed cells with the following settings: estimated blob diameter, 10 pixel (default); threshold, 2.5; linking max distance, 50; gap-closing max distance, 50; and gap-closing max frame gap, 100. All other parameters were left at their default values.

As with the cell tracking challenge dataset for fair comparison of motion capture ability with single cell trackers like TrackMate, tracks were only compared for cells present in the initial frame as detected by TrackMate. To generate single-cell tracks using MOSES (with 10,000 superpixels) for the sparse DeepRed dyed cells, the nearest four superpixel tracks to each cell centroid were averaged to produce a single track. Similarly, to generate single cell tracks using MOSES (with 1000 superpixels) from the Green CMFDA dyed or Orange CMRA dyed sheet, for each cell, the nearest four green/red superpixel tracks were found to compute a mean track. Track similarity was computed by evaluating the normalised velocity cross-correlation (value between 0 and 1 as defined below) between each MOSES track and its corresponding TrackMate track with the average normalised velocity cross-correlation over all tracks reported for each video (Figure 2—figure supplement 2B, denoted M. for matched tracks). To assess the statistical significance of the resultant value, the track similarity from random pairing of the tracks were computed and the average of 10 permutations were reported (labelled P. for permuted in Figure 2—figure supplement 2B). All three combinations of cell types (EPC2:EPC2, EPC2:CP-A and EPC2:OE33) and all red/green dye combinations were tested, a total of 23 videos (each 144 h acquired with one image per h). The frame size of each acquired video was 512 × 672 pixels. As such, the mean spiked-in cell diameter was 5 pixels, the average superpixel width was 6 pixels (10,000 superpixels) and 19 pixels (1000 superpixels). From Figure 2—figure supplement 2B and Video 4, MOSES achieves near perfect similarity compared to TrackMate. In some cases, the produced MOSES tracks are more desirable, guaranteeing a continuous track whereas TrackMate requires explicit linking of cell detections across frames and thus often tends to produce many ‘broken’ tracks when the same cell is unable to be detected across all time points.

The motion saliency map illustrates in a heat map format spatial areas of motion sources and sinks, and was constructed using the MOSES dynamic meshes and superpixel tracks. It is inspired by Lagrangian fluid mechanics (Shadden et al., 2005; Ali and Shah, 2007). To compute this map for each frame, the radius neighbour graph was constructed (see above paragraph) using the spatial positions of superpixels in that frame and a blank image was populated at the (x,y) centroid position of each superpixel with the count of the number of surrounding neighbours according to the radius neighbour graph. This yielded an ‘image’ of size n_frames x n_rows x n_cols, where n_rows, n_cols are the video frame dimensions. We now have a multidimensional spatial heat map for each frame that captures the spatial-temporal motion saliency. To reveal long-time temporally persistent behaviour, we averaged the heatmap both spatially and temporally using the superpixel partition of the initial video frame as illustrated in Figure 3—figure supplement 2. By construction, this spatial map is general for studying any phenomenon where spatial localisation plays a role.

We assume normally/t- distributed statistics for all measures. Boundary formation cut-off was set one standard deviation above the pooled mean of 0% serum samples. Mesh stability index was set one standard deviation below the pooled mean of 0% serum EPC2:EPC2, EPC2:CP-A samples.

The boundary formation index (Figure 3—figure supplement 2) quantifies the extent to which motion concentrates into localised spatial regions in the motion saliency map as a signal-to-noise ratio with value from 0 to 1 suitable for global comparison across video datasets. For example a boundary concentrates motion along a ‘line’ whilst cell death may generate multiple spot-like concentrations (Figure 3—figure supplement 3). The higher the index, the more that motion is concentrated into a single spatial region. To compute the boundary formation index from the visual motion saliency image, the motion saliency image was segmented into ‘high’ and ‘low’ intensity using Otsu thresholding and the normalised signal-to-noise ratio was computed, defined by ${\displaystyle \frac{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left(\text{high}\right)-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left(\text{low}\right)}{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left(\text{high}\right)}}$

(Figure 3—figure supplement 2). The mean was used as the motion saliency map was computed from a sparse set of points given by the number of superpixels. Individual pixel statistics such as the maximum intensity are therefore noisy and not robust measures. The denominator was set to be the mean of the ‘high’ intensity region in order to give a numerical value bounded between 0–1 for standardised comparison. As this measure captures the ‘peakiness’ of the spatial distribution, it can also be used to quantify other localised spatial processes with adaptation for example cell death.

For a superpixel, $i$ at time ${\displaystyle t}$ we define the mesh strain, ${\epsilon }_i\left(t\right)$ of the local neighbourhood, $N_i$ with ${\displaystyle n}$ neighbours as the mean of the absolute difference in the distance between superpixel $i$ and superpixel $j$ in its neighbourhood at time $t$, $r_{ij}\left(t\right)$ and at the start at $t=0$, $r_{ij}\left(0\right)$ so that ${\displaystyle {\epsilon }_i\left(t\right)=\frac{1}{n}{\mathrm{\Sigma }}_{j\in N_i}\left|r_{ij}\left(t\right)-r_{ij}\left(0\right)\right|}$ where $|\cdot |$ is the absolute value or L1-norm. The time-dependent mesh strain for one mesh is the mean local neighbourhood mesh strain over all superpixels for each time frame. The result is a vector the same length as the number of frames in the video. For multi-channels, the average vector is used to describe video motion. The absolute value of the strain curve is susceptible to the image acquisition conditions and geometry whilst the motion information is primarily encoded by the shape of the resulting curve. To permit comparison across different conditions, the normalised strain curve (here division by maximum strain within the common time window of comparison, 0-96 h here) up to 96 h (the maximum timepoint shared between 96 h and 144 h videos in this paper) was used as a simple signature to describe the global video motion pattern when computing motion maps.

Here we provide more details as to why we chose the L1-norm for computing the mesh strain. When computing the ‘distance’ between two vectors, ${\mathit{x}}_1,{\mathit{x}}_2$ both of length $n$ one can define the notion of distance in different ways. A popular family of distances or norms to use is the $L^p$-norm denoted ${\displaystyle {\left|\left|{\mathit{x}}_1-{\mathit{x}}_2\right|\right|}_p}$ defined mathematically as follows:

${\displaystyle |\cdot |}$ denotes the absolute difference. Within this family the most popular is the L1 ${\displaystyle \left(p=1\right)}$ and L2 ${\displaystyle \left(p=2\right)}$ norms. The L1-norm is also called the mean absolute deviation and the L2-norm the Euclidean or mean squared distance. The quadratic function of L2 amplifies large differences (> 1) and reduces the effect of smaller differences close to zero thus where differentiability is not a concern L1 is the more robust choice. In our case where most of the distances are larger than 1 (as all superpixels are initially seeded at a distance > 1), we chose to use L1 which is more resistant to the effect of extremal values as a result of errors in motion extraction.

The mesh stability index attempts to quantify the global stationarity (Figure 3—figure supplement 4) across a moving collective such as an epithelial sheet by measuring the change in the average distance between each superpixel and its neighbours. Specifically, it measures the flatness of the normalised mesh strain curve as defined above. Note whilst a curve is flat only if its gradient is flat that is 0, it is more visually appealing to report increasing stability with increasing values therefore we define the mesh stability index as 1 minus the end gradient. The end gradient is most stable when it has a value of 0 therefore this index is upper bounded by 1. To standardise the value of the gradient for comparison, we normalise also across time such that for any length video, time is from 0 to 1. To compute the end gradient stably without curve fitting procedures, we assume the end point is locally linear with respect to time and average the first-order differences over the last few frames. For 96 h videos, the period of stability given by the curve plateau is shorter, therefore the last 10 frames (10 h) were used for computing the gradient. For 144 h videos, the last 24 frames (24 h) were used.

Inspired by the definition of the velocity order parameter, the mesh order is identically defined but uses the local resultant mesh strain vector instead of instantaneous velocity vectors for computation (Figure 3—figure supplement 5). In this paper, we used the MOSES mesh but any similar mesh construction is equally valid. Given a mesh, the mesh strain vector for a superpixel is the sum of the displacement vector of the superpixel relative to each neighbouring superpixel (Figure 3—figure supplement 5). The mesh order is computed for each frame accordingly for a video. The mean value over all frames was reported for statistical comparison.

Cross-correlation measures the similarity between two signals taking into account time delay, (Figure 3—figure supplement 6). As ‘signals’, we use the time-dependent velocity ${\displaystyle V\left(t\right)}$ computed from the spatial displacement between the location $r\left(t+1\right)$ at time $t+1$ and location $r\left(t\right)$ at time $t$, $V\left(t\right)=r\left(t+1\right)-r\left(t\right)$ which is spatially location-independent. Velocity is a vector quantity with both $x$ and $y$ components. Letting $\cdot$ denote the scalar product, the normalised velocity cross-correlation (VCC) of the tracks of red superpixel $i$ and green superpixel $j$ at time $t$ and time lag $m$ such that the extremas are bound by [-1,1] is

where ${\displaystyle \widehat{V_i}=\frac{V_i-\stackrel{-}{V_i}}{{\sigma }_i}}$ , ${\displaystyle \stackrel{-}{V_i}}$ and ${\sigma }_i$ is the mean and standard deviation of $V_i$ respectively. $T$ is the maximum time lag and is set to be the length of $V_i$. VCC can be either positive or negative. We report the maximum absolute value for all red-green pairings and the average over all pairings as evidence of interaction between two epithelial sheets. In the main text, this is computed with tracks before (up to - 5 frames) and after (from + 5 frames) the gap closure point. The offset of 5 frames either side is based on the accuracy to which we could determine the gap closure point automatically (see below).

The frame of gap closure when the two epithelial sheets contact was found by finding the video frame in which the average distance between the migrating fronts of the two epithelial layers was minimised. To compute this distance as a function of time (frame number), first each epithelial sheet was independently segmented based on their respective colour channel pixel intensity (Figure 3—figure supplement 9A). For each sheet, images were preprocessed using a median filter (square kernel, the same size as the average superpixel width) and segmented using two class k-means clustering (for 0% serum, three class k-means was used to include the weaker pixel intensity of leading cells). The resulting segmentation was post-processed by binary morphological operations (binary closing with disk kernel of 5 pixels, removal of small objects (< 5% total image size) and binary filling). Sheet boundary points were efficiently identified using a sweepline algorithm. The image was evenly divided into 100 horizontal strips or sweeps in the y-direction. For each sweep, one boundary point was identified by selecting either the right-most point (largest x-coordinate) if the sheet was moving right, or left-most point (smallest x-coordinate) if the sheet was moving left (Figure 3—figure supplement 9B). Doing this for both sheets, each of the boundary points was paired in the red/green sheet to the closest in the opposing colour by physical distance. The average distance between the migrating fronts of the two epithelial layers for a particular frame was then the average euclidean distance of all red-green boundary point pairs in the frame. Computing the average distance between the two sheets frame-by-frame, there was a change in the rate of decrease as the gap closed (Figure 3—figure supplement 9C). To estimate the frame at which this transition occurs, asymmetric least means squares (Eilers and Boelens, 2005) was used to first fit the baseline (representing the contribution to the distance measurement due to image segmentation errors) and second a linear spline (smoothing factor 0.1*variance(curve)) was used to approximate the temporal average distance curve. The gap closure frame was then found as the first frame for which the fitted spline value falls below the fitted baseline + 2*std(fitted baseline), where std is the standard deviation operation. We validated the method using a total of n = 246 videos of different cell combinations in different media by comparing the inferred frame to the consensus (average frame) of two manual annotators. A strong Pearson correlation coefficient of r = 0.902 (Figure 3—figure supplement 9D) and an accuracy of 94% within ± 5 frames (Figure 3—figure supplement 9E) compared to an accuracy of 97% within ± 5 frames between two human annotators (Figure 3—figure supplement 9F) was found.

Boundaries were detected either i) by image segmentation (early timepoints) or ii) from the MOSES tracks (late time points). For image segmentation, red and green epithelial sheets were segmented as described above for timelapse video frames. The boundary binary mask was then found as the mathematical set intersection of red and green binary masks. To derive a boundary line, the non-zero image coordinates of the binary boundary mask were forced to form a line given by unique (x,y) coordinates by returning the average x-coordinate (along image horizontal) for each unique y-coordinate (along image vertical). A piece-wise cubic spline was subsequently fitted to enable interpolation of the boundary line. To derive the boundary line from the MOSES superpixel tracks, first all non-moving tracks that is all tracks that do not move a total distance greater than a predefined threshold was removed. Then at each frame, $t$ we considered all superpixels that have moved since the previous frame $t-1$. We then attempted to match the red superpixels to green superpixels where we defined a match when the distance between two points is smaller than a predetermined cut-off. All red and green superpixel points that have at least one successful match were kept. Together these points form the boundary candidate points. As discussed, a boundary is a motion attractor thus we can robustly find the boundary points from the candidate points using an asymmetric least squares filter (Eilers and Boelens, 2005). We modified the original formulation of Eilers and Boelens, 2005 to enforce density based filtering by using a cut-off based on the number of neighbours within a predefined distance. Finally, given the boundary points the boundary line was found as in the case of image segmentation.

For a single video frame, the infiltration of the first colour cell type into the second colour cell type is defined as the fraction of first colour cells that lie on the side of the boundary line of the second coloured sheet. The boundary line was located from image segmentation as described above.

The boundary shape is defined as the total length of the boundary line ($L$) divided by the equivalent straight line length ($L_0$) illustrated as solid white and yellow lines respectively in Figure 3—figure supplement 10C.

We computed two different measures to reflect the ‘intermixing’ behaviour of cells at the boundary, which we denoted either ‘(Image)’ or ‘(MOSES)’ in Figure 3—figure supplement 10. The image intermixing coefficient reflects the ‘spread’ of the wound after we perform image segmentation on the video frame. A heavily infiltrated and wavy boundary yields a larger binary mask and occupies a larger fraction of the image than a sharp, linear boundary. We capture this idea by defining the intermixing coefficient (image) as the area of the boundary mask relative to the total image area. This is however a static measure of intermixing. It does not account for the fact that the boundary is not static and can be highly dynamic such as in EPC2:EPC2 or how EPC2 cells appear to ‘flee’ OE33 cells with little coordination. We thus propose a second dynamic measure of intermixing coefficient based on measuring the spread in the motion behaviour. This intermixing coefficient (MOSES) is the area of the binary mask after thresholding the motion saliency map relative to the total image area. Of note, the two intermixing coefficients are identical if the final boundary is stable and the motion at the boundary during boundary formation is coordinated such as in EPC2:CP-A (Figure 3—figure supplement 10E).

Boundaries were detected using the MOSES tracks as described. The kymograph of one time slice shows the median values of the projected x-direction velocities of all pixels as a function of the x-position. For superpixel tracks, the x-axis was binned with two times the number of unique x-coordinates based on the initial superpixel centroids (in line with Nyquist sampling theorem). This procedure was repeated for each timepoint to build the full kymograph over time.

A feature descriptor for an image/video is a 1D numerical vector designed to summarise the image/video content in a compact code or signature. No universally optimal method exists to generate such a descriptor for all applications. In the main text, for simplicity to avoid the introduction of new concepts we used the normalised strain curve used to compute the mesh stability index as an example video feature vector for PCA. As we showed, this was sufficient to distinguish the different migration behaviour of the investigated cell combinations. However, the normalised strain curve is a coarse 1D approximation of the local MOSES mesh stretching and only one characteristic of the complex mesh dynamics. More generally, one can exploit other mesh constructions as discussed above to derive a plethora of graph theoretic measures such as the algebraic connectivity, the Laplacian spectrum and centrality, or supplement mesh-based statistics with trajectory-based measures such as turning angle and speed for a more comprehensive unbiased description of the spatio-temporal motion. In circumstances where cellular appearance exhibits large temporal changes such as in the case of migrating cells with lamellipodium, motion features alone may not provide a sufficiently descriptive signal for quantifying phenotypic differences. Here one could additionally supplement motion signatures with appearance-based features such as image texture descriptors (e.g. LBP, Haralick, HoG (histogram of oriented gradients) and SIFT (and its variants)) and construct the mesh semantically with the augmented motion-appearance descriptor.

We used the Python scikit-learn implementation of PCA (principal components analysis), MDS (multi-dimensional scaling) and TSNE (t-distributed stochastic neighbour embedding) and applied them to the ‘raw’ normalised mesh strain curves (no pre-processing). PCA was applied with n_components = 2 without input whitening. MDS used default scikit-learn parameters with n_components = 2, random_state = 0. TSNE used n_components = 2, learning_rate = 100, random_state = 0, init = ‘random’. A 97-20-2 fully connected neural network autoencoder was implemented and trained with Keras (Tensorflow backend) with mean squared error loss using the Adam optimizer (lr = 0.001, beta_1 = 0.5, beta_2 = 0.999, epsilon = 1e-08, decay = 0.0). 48/77 of the 5% serum videos were used for training and the remaining 29/77 for validation to check for model overfitting. Tanh activations were used throughout. To maximise gradient propagation in the linear range of the ‘tanh’ activation function, we subtract 0.5 from the input normalised strain curves (values between 0 and 1) as a pre-processing step for the neural network.

To automatically cluster superpixel tracks, we first computed for each superpixel its local mesh strain curve with respect to its neighbours. This yields a matrix N rows by T columns for N superpixel tracks and a total of T timepoints. Gaussian mixture model (GMM) was then used to generate clusters using BIC (Bayesian Information Criterion) to select the optimal number of clusters (Fraley, 1998).

The MOSES code is available open-source under a Ludwig non-commercial and academic license at GitHub, https://github.com/fyz11/MOSES.git (Zhou, 2019; copy archived at https://github.com/elifesciences-publications/MOSES) where it is maintained and updated. Example data for testing can be downloaded from Google Drive, https://drive.google.com/open?id=0BwFVL6r9ww5BaTh6NExLR1JMUXM. The full video dataset can be found under DOIs, https://dx.doi.org/10.17632/j8yrmntc7x.1, https://dx.doi.org/10.17632/vrhtsdhprr.1 for videos with and without EGF addition, respectively.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank Prof. Hiroshi Nakagawa for generous donation of EPC2 cells and Profs. Roberto Cerbino and Xavier Trepat for making available the raw videos in their publications. We thank Dr. Mary Muers, Dr. Françoise Howe, Profs. Sebastian Nijman, Francis Szele, Colin Goding and Shankar Srinivas for critical reading of the manuscript; Mark Shipman for technical assistance with timelapse microscopy. This work is mainly funded by the Ludwig Institute for Cancer Research (LICR) with additional support from a CRUK grant to XL (C9720/A18513). FYZ is mainly funded through the EPSRC Life Sciences Interface Doctoral Training Centre EP/F500394/1, CRP and XL are funded by LICR, RPO is supported by LICR, the Oxford Health Services Research Committee and Oxford University Clinical Academic Graduate School supported by the National Institute for Health Research (NIHR) Biomedical Research Centre based at the Oxford University Hospitals Trust, Oxford, MJW is supported by CRUK (C5255/A19498, through an Oxford Cancer Research Centre Clinical Research Training Fellowship), and JR is funded by LICR and the EPSRC SeeBiByte Programme Grant (EP/M013774/1). The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health.

© 2019, Zhou et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.