Proceedings of the Royal Society B: Biological Sciences
You have accessResearch articles

A phenotypic switch in the dispersal strategy of breast cancer cells selected for metastatic colonization

Published:https://doi.org/10.1098/rspb.2020.2523

Abstract

An important question in cancer evolution concerns which traits make a cell likely to successfully metastasize. Cell motility phenotypes, mediated by cell shape change, are strong candidates. We experimentally evolved breast cancer cells in vitro for metastatic capability, using selective regimes designed to simulate stages of metastasis, then quantified their motility behaviours using computer vision. All evolved lines showed changes to motility phenotypes, and we have identified a previously unknown density-dependent motility phenotype only seen in cells selected for colonization of decellularized lung tissue. These cells increase their rate of morphological change with an increase in migration speed when local cell density is high. However, when the local cell density is low, we find the opposite relationship: the rate of morphological change decreases with an increase in migration speed. Neither the ancestral population, nor cells selected for their ability to escape or invade extracellular matrix-like environments, displays this dynamic behavioural switch. Our results suggest that cells capable of distant-site colonization may be characterized by dynamic morphological phenotypes and the capacity to respond to the local social environment.

1. Introduction

Metastasis is a form of long-range dispersal [1,2], and central to understanding how cancers metastasize is understanding how cells migrate [3,4]. During migration, as cancer cells become more invasive and begin to migrate independently, they adopt an altered morphology, typically taking on elongated shapes characteristic of epithelial–mesenchymal transition [5,6]. This change in cellular morphology is an important marker of migratory state [7,8]. Quantitative measures of cell morphology taken from static images have been shown to effectively differentiate between cancer cell lines with high and low metastatic potential [9,10]. However, there are important aspects of migratory behaviour linked to metastasis that cannot be measured from static images.

Successful metastasis requires a cell to navigate through a series of sequential steps known as the metastatic cascade. The cascade begins with a cell escaping from the primary tumour before migrating through the extracellular matrix (ECM) towards a nearby blood vessel. The cell must then intravasate into the blood before it is carried around the body. After reaching a distant site, the cell then needs to extravasate from the blood and invade the foreign tissue. Finally, the cell must reinitiate aggressive proliferation enabling a secondary tumour to form [11].

In addition to the cellular changes needed for metastatic success, environmental changes are also necessary for a cell to metastasize [12]. This is evident at the onset of cellular dispersal where nearby collagen fibres are straightened perpendicular to the tumour boundary [13]. The straightened fibres then act as a pathway for future migrants in turn improving their migratory success [14]. This dynamic cell–environment interplay continues throughout the metastatic cascade.

To identify the precise changes in cell phenotype that are associated with metastatic success, it is preferable to compare cell lines that differ only in their ability to metastasize. Experimental evolution [15], a powerful approach that has led to major advances in evolutionary biology, is now being applied to cancer evolution and provides the means to generate such cell lines [16,17]. Initially identical populations of cancer cells can be selected in replicate for specific capabilities [18]. We experimentally evolved populations of cancer cells using selective regimes corresponding to three separate stages of metastasis [19,20]: escape from the primary tumour, invasion of foreign tissue and distant-site colonization.

Distant-site colonization, the rate-limiting step of metastasis [21], requires a cell to migrate through the unpredictable microenvironment of the primary tumour [22] and into the novel environment of the distant metastatic site [11]. Success in both stages is achieved, in part, by the cell's capacity to detect and respond to changes in the environment [2326]. Cells selected for distant-site colonization might therefore be expected to be more reactive to environmental changes, and as such display a greater degree of morphological change in response. We would also expect a morphological change to be positively correlated with migration speed in successfully metastasizing cells, because a faster-moving cell will experience a greater degree of environmental variation over a given time period, and therefore change its morphology more rapidly in response.

To test these hypotheses, we have combined an experimental evolution framework with video microscopy and novel statistical analysis that quantifies a morphological change in individual cells over time. This approach has identified unique cell behavioural phenotypes that may be advantageous for successful metastasis.

2. Material and methods

(a) Evolved population summary

We used experimental evolution methods [15] on an initial population of MDA-MB-231 breast cancer cells (figure 1), subjecting them to three separate selective regimes. The experimental selective regimes were designed to be similar to those experienced while traversing the metastatic cascade [11]. We also froze two biological replicate ancestor populations (figure 1) at the start of the experiment to act as a control for comparison with our evolved lines.

Figure 1.

Figure 1. Experimental evolution of cancer cell populations. Ancestor populations were kept frozen throughout. Escape populations were placed in a high-density collagen matrix and surrounded by a low-density outer collagen ring: after 10–14 days cells that had escaped into the outer ring (shown in blue) were released, expanded and reseeded back into a new high-density collagen core; this process was repeated seven times over the course of six months. Invasion populations were seeded around a Matrigel island; after 7 days, cells that had invaded the Matrigel (shown in blue) were released, expanded and reseeded around a new Matrigel island; this was repeated 15 times over the course of six months. Colonization populations were seeded onto a piece of decellularized rat lung, which acted as a novel scaffold for colonization and left to establish for six months. Four replicate lines were maintained for each treatment. (Online version in colour.)

We selected escape populations (figure 1) by tightly packing cells into a high-density core of collagen and then allowing them to escape outwards into a low-density collagen outer ring [27]. After 10–14 days, the cells that had escaped into the outer collagen ring were recovered from the matrix, expanded and then seeded back into a new collagen escape assay, completing one round of selection. In total, seven rounds of selection were applied to each of four biological replicate escape populations. The high-density collagen core and the low-density outer collagen ring were both three-dimensional (3D) culture environments designed to be similar to those experienced during tumour dissemination.

We selected invasion populations (figure 1) following a similar protocol to the escape populations, whereby repeated consecutive rounds of selection were applied. In contrast to the escape assay, however, cells moved from a 2D to 3D environment, similar to the change in the environment experienced during the arrest of a cell at a distant site. The cells were seeded around the outside of a Matrigel island—a synthetic basement membrane matrix widely used in cell culture—and left to invade. After 7 days, the cells were collected from the Matrigel, expanded and seeded around the outside of another Matrigel island. This process was repeated 15 times for each of the four biological replicate populations over the course of the six-month experiment.

We selected colonization populations (figure 1) by culturing cells on a piece of decellularized rat lung, which acted as a scaffold for growth similar to that experienced by cells colonizing a distant site [27]. The protocol involved cells being seeded onto a decellularized scaffold and left to colonize over a six-month period. Decellularized tissue is generated by removing all cells from a piece of tissue such that only the ECM is left. At the end of the experiment, cells were released from the scaffold, ensuring that the population represented cells from within the tissue core as well as the edges. Again, this selection was applied to four biological replicate populations.

Finally, all 12 experimentally evolved cell populations were frozen and then thawed alongside the ancestor populations prior to experimental analysis. This step ensured that any selective pressure from the freezing–thawing process was constant across all treatments and replicate populations.

(b) Experimental assays

(i) Escape assay

Initially, MDA-MB-231 cells (LGC) were encapsulated in a 2 mg ml−1 collagen gel (rat-tail collagen type 1, First Link) and set into a 24-well plate which was used as a mould (750 000 cells per gel, Greiner Bio-One). The collagen gels were compressed for 2 min as described by Keeton et al. [27], then set into a 1 mg ml−1 low-density collagen gel (rat-tail collagen type 1, First Link). Once set, cell culture medium (Dulbeco's Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS), and penicillin 100 µg ml−1, streptomycin 100 U ml−1 (Gibco, Fisher Scientific)) was added over the top. Medium was replaced every 3–4 days. After 10–14 days, the compressed collagen disc was separated from the low-density collagen and collagenase type 1 diluted in phospho-buffered saline solution (Gibco, Fisher Scientific) used to retrieve the cells from the collagen matrix, 200 U ml−1 for compressed collagen and 100 U ml−1 for low-density collagen. Cells in collagenase/phospo-buffered saline (PBS) were incubated at 37°C in a stirred water bath at 45 rpm for 30–60 min, then washed in PBS solution (Gibco Fisher Scientific). Cells extracted from the compressed collagen were placed in liquid nitrogen storage and those collected from the low-density collagen were seeded into 2 mg ml−1 collagen gel with medium over for population expansion. Once expanded, cells were retrieved from collagen using collagenase in PBS then seeded into 2 mg ml−1 collagen for compression or frozen at −80°C and transferred to liquid nitrogen for storage.

(ii) Invasion assay

MDA-MB-231 cells (LGC) were resuspended in PBS, seeded around the outside of a 5 mg ml−1 set Matrigel island in a six-well plate Matrigel (#35623, Corning) and was diluted using DMEM without supplements. Cells were seeded in excess at the island margins, with around 40 000 cells seeded in 200 µl per experiment for the initial set-up. Cells were left to settle and adhere to the 2D surface for 60 min, then cell culture medium added over the top (DMEM supplemented with 10% FBS, and penicillin 100 µg ml−1, streptomycin 100 U ml−1). Medium was changed every 3–4 days, and cells were harvested after 7 days. Cells were retrieved from Matrigel using Cell Recovery solution (#354253, Corning) on ice for 45–60 min, washed with ice-cold PBS and then reseeded into Matrigel at 5 mg ml−1 to expand cell numbers. After 7 days, the cells were released from Matrigel using cell recovery solution as described above (typically 400 000–500 000 per gel), resuspended in PBS and seeded in excess around the outside of a new Matrigel island (5 mg ml−1) for the next round of the 2D/3D invasion assay or cells were frozen at −80°C and transferred to liquid nitrogen for storage.

(iii) Colonization assay

Rat lung was retrieved from nine-week-old Wistar rats (Envigo) and flash frozen. It was then thawed and decellularized using repeated rounds of treatment following an adapted version of the protocol published by Medberry et al. [28]. Briefly, frozen lung was thawed and cut into small pieces of around 100 mg, which were then placed into deionized water (ddH2O) and stirred at 60 rpm for 16 h at 4°C. Lung tissue was treated with 0.02% trypsin/0.05% EDTA for 60 min at 37°C at 60 rpm, 3% Triton-X 100/PBS for 70 min, 1 M sucrose/PBS for 30 min, 4% deoxycholate/ddH2O for 60 min, 0.1% peracetic acid in 4% ethanol for 120 min, PBS for 5 min and finally twice in ddH2O for 15 min. The tissue was washed thoroughly between each treatment with ddH2O. Decellularization was checked between rounds using epifluorescence microscopy and staining with DAPI H1200 Vectashield (Vectorlabs) to identify whether cell nuclei remained within the matrix structure. Decellularized lung tissue was freeze-dried and stored in an airtight container.

Using decellularized lung as a culture matrix, tissue was soaked in 70% ethanol, washed with PBS and then rehydrated in PBS pH 7.2 (Gibco) in a tissue culture incubator for 5 days, then soaked in cell culture medium (DMEM supplemented with 10% FBS and penicillin/streptomycin as described above) for 48 h. Cells grown in 2D tissue culture flasks were trypsinized, resuspended in medium, then 750 000 cells added in low volume of medium (100–150 µl) over the decellularized lung tissue in a six-well plate and left to adhere for 2 h. Medium was then added over the top so that the decellularized lung rafts were floated. Rafts were transferred to new wells when the bottom of the well was confluent with shed and adhered cells. To feed the cells growing in/on the raft, ½ of the medium (2 ml of 4 ml) was aspirated and replaced every 2–4 days. After 140 and 189 days, rafts were retrieved from medium, washed with PBS and cells were harvested by incubating in: collagenase I (170 U ml−1, Gibco 17018-029), collagenase IV (170 U ml−1, Gibco 17104-019) and elastase (0.075 U ml−1, Sigma E7885) (based on the protocol described in [29]) incubated at 37°C 45 rpm in a stirred water bath, then washed twice with PBS before seeding in 2D tissue culture plates for expansion. Expanded cells were then frozen at −80°C and transferred to liquid nitrogen for storage.

(c) Time-lapse microscopy

Cells were retrieved from liquid nitrogen, cultured in 2D tissue culture flasks (25 or 75 cm2 Greiner bio-one), trypsinized and seeded into six-well plates (Greiner bio-one) at 10–15% cell confluence. Time-lapse movies were made for 12 h periods with images taken at 2-min intervals, using a Nikon TiE phase-contrast microscope with an environmental chamber (37°C) and movable platform stage. ×10 Plan Apo DIC L Lens was used in conjunction with an intermediate magnification changer set to ×1.5 to give ×15 magnification. NIS Elements software was used for image capture.

(d) Cell tracking

All cells that were present in each time-lapse video were tracked using the Usiigaci pipeline [30]. The neural network was trained on 300 randomly selected images that were manually annotated using ImageJ [31]. The manually annotated images were than randomly split so that 80% were used for training and a further 20% were used for testing, 240 images in the training set and 60 in the test set. The 240 training images were then further split for training and validation 90:10, so that 216 images were used for training and 24 for validation. We trained three neural networks using the same 240 images; however, different images were used for the training and validation stage each time. All hyperparameter settings were the same as the Usiigaci protocol, except the gradient clip norm was increased to 10. We trained the network on all layers over 300 epochs with the learning rate starting at 0.01 and decreasing by an order of magnitude every 100 epochs.

Once the morphologies had been segmented, we tracked them through time using the inbuilt semi-automated Usiigaci tracker. After tracking, we manually checked the segmented morphologies and corrected any errors. We checked for cases whereby a cell had divided, been misidentified or incorrectly segmented. Finally, we excluded the 30 min prior to and after a cell division to remove the rounded morphologies typical of cell division from our analysis.

(e) Quantifying values

All values were quantified using a custom-built pipeline in Python [32] that can be found on GitHub, https://github.com/george-butler/2d_microscopy; any reference to distance refers to the Euclidean distance. The morphology was quantified using the first 20 Zernike moments. Zernike moments capture the information that is encoded in a shape and translate it into a high dimensional vector, in a similar fashion to spatial location being represented by Cartesian coordinates. When taken to a high enough degree, Zernike moments are capable of representing every shape uniquely and are invariant to rotation, scale and translation [33]. We followed the methods of Alizadeh et al. [10] to pre-process the morphologies and make them invariant to scale and translation. We determined that 20 Zernike moments were adequate to quantify the morphology of each cell by plotting the mean squared error against the number of moments [34] and finding where the gradient approached 0.

(f) Statistical analysis

All statistical analysis was performed in R [35] and figures 35 are made using GGPlot [36]. All code and corresponding data can be found in GitHub, https://github.com/george-butler/2d_microscopy/tree/master/statistical_analysis. A cell needed to appear in at least 30 frames to be included in our analysis and be present for at least 75% of the track. Some cells were not detected in a given frame or had to be removed due to being incorrectly segmented. Throughout our analysis, we used linear mixed models to account for the differences between replicate populations within the four treatments [37]. The mean rate of morphological change and the mean speed of migration were calculated through the use of an intercept-only linear mixed population with independent intercepts for each treatment. The rate of morphological change model is defined below:

rate of morphological change = α+β1(speed of migration)+β2(distance to nearest neighbour)+β3(speed of migration)×(distance to nearest neighbour)+(1|wellid).

The model was selected through forward selection whereby parameters were only included if they were significant at the 5% level. The marginal R2 values were calculated using the method detailed by Nakagawa & Schielzeth [38].

3. Results

(a) Quantifying dispersal in evolved populations

To analyse their dispersal behaviour, cells were placed onto 2D tissue culture plates and their migration was recorded over a 12 h period, with images taken at 2 min intervals. The 2D plastic environment was intentionally chosen as a neutral testing environment and to ensure that the morphology could be clearly seen without the use of fluorescent tags, a factor that might have applied an additional selective pressure [39]. The cells were tracked through the use of a semi-automated pipeline, Usiigaci [30], that combined a convolutional neural network with our own manually annotated images to trace the morphology of each cell at every time point (figure 2a).

Figure 2.

Figure 2. Quantifying dispersal from time-lapse videos. (a) Cells were tracked over a 12 h period with images taken at 2 min intervals using phase-contrast time-lapse microscopy to generate movies from which morphology could be segmented through the use of a convolutional neural network. (b) The rate of morphological change was recorded as the distance between Zernike moments in consecutive frames. (c) The speed of migration is calculated as the distance between the spatial locations of cells in consecutive frames. (d) The distance between neighbouring cells is quantified as the shortest distance between the contour of one cell and the contour of another. The direction of the arrow points from a given cell to the point on the contour of the closest neighbouring cell. (Online version in colour.)

We extracted three quantitative measures per cell per frame of time-lapse video: morphology, speed and the distance to the closest neighbouring cell. Morphology was quantified using Zernike moments. Zernike moments [33] have been used previously to quantify cancer cell morphology in fixed populations [10] and are a method that captures all of the morphological information available rather than needing to make a prior decision about which morphological features might be important (e.g. the length of a cell). The rate of morphological change is then measured as the distance between the vector of moments in frame t and t + 1 relative to the time between frames (figure 2b). Speed of migration was calculated from the change in spatial location between consecutive frames (figure 2c). The distance to the closest neighbour cell was calculated as the shortest distance from the edge of the cell contour to another neighbouring cell contour without crossing the body of the cell (figure 2d). Finally, the average was calculated for each metric over the entire trajectory of the cell, providing a summary of the dispersal phenotype of each cell.

After extracting these three metrics, we sought to evaluate whether the rate of morphological change or the speed of migration was significantly different among the four treatments. We used an analysis of variance (ANOVA) to compare the mean rate of morphological change and the mean speed of migration across all populations; differences in wells were accounted for as a random effect. We found that there was significant variation among populations in their mean rate of morphological change (p = 0.0296, n = 813). We then conducted a post hoc Bonferroni multiple comparison test to identify which populations were different, controlling for any possible between-replicate variation through the use of a random effect. Escape populations had a significantly higher rate of morphological change compared with the invasion populations (p = 0.0152, n = 813; figure 3). There was no significant difference in the mean speed of migration among the four treatments.

Figure 3.

Figure 3. Comparing the mean rate of morphological change among the four treatments. A plot of the natural log-transformed rate of morphological change for each of the four treatments. The centre dot signifies the mean rate of morphological change with errors bars signifying 95% confidence intervals. The escape populations had a significantly faster rate of morphological change compared with the invasion populations (p = 0.0152, n = 813). The mean, standard error and number of observations for each population can be found in electronic supplementary material, table S1. (Online version in colour.)

(b) Speed of migration predicts the rate of cell–morphological change in evolved populations

Next, we investigated how the morphological behaviour of a cell related to its speed and its social environment. We fitted a linear mixed model across our data whereby the rate of morphological change is dependent on the speed of migration, the distance to the nearest-neighbouring cell and the interaction of the two, as detailed in ‘Material and methods’ above. We set the treatment as a fixed effect and allowed intercepts and slopes to vary between treatments. The significant parameters were then used to fit a reduced model to the ancestor, escape and invasion populations (figure 4).

Figure 4.

Figure 4. The rate of morphological change against the speed of migration. The natural log-transformed rate of morphological change plotted against the natural log-transformed speed of migration. The straight lines represent the reduced model for each treatment using only parameters that are significant at the 5% level. The ancestor populations have an intercept-only model fitted (n = 88). The speed of migration is the only significant variable in the escape (n = 230, p = 1.765 × 10−3) and invasion (n = 283, p = 0.018) populations. For both escape and invasion populations, the rate of morphological change is positively correlated with the speed of migration, the faster the speed of migration the higher the rate of morphological change. (Online version in colour.)

In the ancestor populations, neither the speed of migration nor the distance to neighbouring cells significantly affected the rate of morphological change. We proceeded by fitting an intercept-only model to our data (figure 4). However, the intercept model explained only a small proportion of the variance (marginal R2 = 0). This might suggest that the rate of morphological change is highly stochastic, or that it depends on factors not included in our model.

By contrast, in both escape and invasion populations, the rate of morphological change is significantly positively correlated with the speed of migration, (β = 0.680 and 0.319, respectively; figure 4). Furthermore, the escape and invasion models both explain a significant proportion of the variation (marginal R2 = 0.347 and 0.099, respectively). To ensure that our results were not affected by a small cluster of potential outliers, we repeated the same analysis after having removed influential data points, defined by a Cook's distance > (4/n) where n is the sample size [40].

The slope of the relationship is steeper for escape than for invasion populations, suggesting that selection for escape may favour cells that can change their morphology rapidly when migrating at a high speed. This might be a result of the collagen escape assay being a 3D to 3D environment compared with the 2D to 3D environment of the Matrigel invasion assay. However, this also could be due to the different number of rounds of selection between the two assays, or difference in the strength of selection within each.

(c) Spatial density affects morphological dynamics

The colonization populations displayed a complex morphological behaviour dependent on the speed of migration, the distance to the nearest-neighbouring cell and the interaction of the two: as the distance between neighbouring cells increases, the relationship between the rate of morphological change and the speed of migration becomes negative (figure 5a). When close to a neighbouring cell, the rate of morphological change is positively correlated with the speed of migration: a faster speed of migration results in a higher rate of morphological change. However, when the distance between neighbouring cells is large and a cell is isolated, the rate of morphological change is negatively correlated with the speed of migration: a faster speed of migration has a lower rate of morphological change. We repeated the same analysis after the removal of any influential data points and found that the interaction term was still significant in these colonization populations (electronic supplementary material, figure S1). We also found that the colonization model explained a significant proportion of the variation in the rate of morphological change (marginal R2 = 0.236).

Figure 5.

Figure 5. A dynamic switch in the morphological behaviour within cells selected for colonization. Data points have been removed to highlight the behaviour of the model, and the same model with data points can be seen in electronic supplementary material, figure S2. The speed of migration (p = 5.418 × 10−14), the distance to the nearest neighbouring cell (p = 2.207 × 10−10) and the interaction of the two (p = 2.219 × 10−11) were significant in the colonization population (n = 212). (a) The predicted natural log-transformed rate of morphological change against the natural log-transformed speed of migration. The shaded lines indicate the natural log-transformed nearest-neighbour percentile. The lighter the line, the further away from a neighbouring cell with distance values ranging from 2 to 477 µm. (b) The predicted natural log-transformed rate of morphological change against the natural log-transformed nearest-neighbour distance. The shaded lines indicate the speed of migration percentile. The lighter the line the faster the speed of migration. The shaded region indicates the range of distances over which there is no significant relationship in the rate of morphological change and the speed of migration when the data are centred at these distances between 57.9 and 147.2 µm. (Online version in colour.)

Next, we sought to determine whether the switch in morphological behaviour with distance was gradual or sudden. To investigate this hypothesis, we centred the nearest-neighbour data at a distance x and then refitted the same morphological change model. After fitting the model, we evaluated whether the speed of migration was significant in the model. If the speed of migration is not significant, then we know that at a distance x there is not a significant difference in the rate of morphological change for cells migrating at different speeds. We can then repeat the same method for different values of x to find a range of distances over which the speed of migration is not significant. The smaller the range, the more sudden the switch.

We found that for nearest-neighbour distances between 57.9 and 147.2 µm, the speed of migration is not significant in our model, as seen by the shaded region in figure 5b. Therefore, at distances less than 57.9 µm or greater than 147.2 µm, the speed of migration is significantly related to the rate of morphological change. The small range of distance values suggests that the cells have a high degree of sensitivity to the location of neighbouring cells. Interestingly, the range of distance values coincides with values from the literature whereby cells within a tumour core have been seen to display a correlated mode of migration at spatial distances less than 50 µm compared with distances greater than 250 µm [41].

4. Discussion

We have conducted novel phenotypic analysis across experimentally evolved populations of MDA-MB-231 breast cancer cells to investigate their behaviour during dispersal. Combining experimental evolution with computer vision, we have generated a multidimensional dataset that quantifies single-cell dispersal dynamics within each population. In turn, we have built a continuous data-driven morphological model that has uncovered fundamental dispersal behaviour at a cellular level and is capable of distinguishing cells selected for colonization.

The flow of migratory cells through the microenvironment creates a landscape that is heterogeneous both spatially and temporally [42]. This landscape variability might, in turn, explain the correlation between the rate of morphological change and the speed of migration for both the escape and invasion populations (figure 4). The collagen escape and Matrigel invasion assays used to select the escape and invasion populations are porous and complex [43], but yet they are also malleable. The malleability of these two environments means that large structural changes can occur and thus migration routes that were previously accessible may become blocked. Therefore, a cell may need to respond to its environment by changing its morphology to ensure that it can continue to migrate and does not become trapped. Likewise, as the speed of migration increases, an increase in the rate of morphological change might be necessary to ensure that the cells are not temporarily stuck by any potential obstacles. This would also explain why there is no correlation in the ancestor populations where the environment remains constant and there would therefore be no selective advantage to this behaviour.

Distant-site colonization requires a cell to switch from a mode of long-range dispersal and focus on re-initiating aggressive proliferation; the subsequent increase in local cell density may reduce available space and thus intensify competition. A similar selective pressure can be seen in our experimental assays. In contrast to the ancestor, escape and invasion populations, where cells are periodically moved to a new expansive environment, the colonization population remains fixed. As such in addition to the structural changes that occurred in the microenvironment, there was a high density of cells migrating locally so cells themselves could block potential migration routes, and therefore might explain the significance of the neighbour location in our model. This hypothesis would also explain the interaction that is observed between neighbouring cells. If a cell is migrating at a high speed and is close to other neighbouring cells, then changing its morphology rapidly might be necessary to avoid other cells that are changing location dynamically. However, when isolated, the location of neighbouring cells is no longer of concern and thus a reduction in the rate of morphological change might allow a cell to conserve resources.

The significance of the neighbour sensitivity may also suggest that the ability of a cell to sense contact has been re-acquired within the colonization population. A loss of contact inhibition is seen as one of the earliest developments in cancer progression as it allows aggressive proliferation to ensue, which in turn gives rise to the formation of a primary tumour [44]. However, the high degree of neighbour sensitivity seen in figure 5 questions whether contact sensing is, in fact, lost, or instead downregulated earlier in the metastatic cascade. If true, this could suggest that cells selected for distant-site colonization are able to vary their own contact sensing ability dependent on the exogenous environmental stresses they encounter.

In summary, we have shown that evaluating cell morphology as a dynamic process provides a novel insight into the behaviour of breast cancer cells, and furthers our understanding of the phenotypic route to metastasis. A pivotal next step will be evaluating morphological dynamics within a native 3D environment [45] and in the vicinity of stromal cells such as fibroblasts which are known to have a critical role in metastasis [46]. The presence of stromal cells might also change the relationship seen within our escape and invasion populations, as cells would then be able to interact via matrix metalloproteinases. Thus, rather than needing to change their morphology quickly to prevent being trapped, they could exploit the matrix metalloproteinases to cut them free, as seen previously during metastatic dispersal [47]. It would also be of value to subject multiple starting cell lines to a similar selective regime, in case the MDA-MB-231 line used here behaves atypically. However, we believe that this work highlights the power of phenotypic analysis in discovering the complex emergent behaviours that would not have been apparent from genetic data.

Data accessibility

Pre-processed data are available on GitHub at: https://github.com/george-butler/2d_microscopy/tree/master/statistical_analysis.

Authors' contributions

G.B, S.J.K., L.J.J. and P.R.D. conceived and designed the study. S.J.K. evolved the populations and collected the time-lapse data. G.B. developed the methodology and performed the formal analysis. L.J.J. and P.R.D. supervised the work. G.B. wrote the manuscript. All authors gave final approval for publication and agree to be held accountable for the work performed therein.

Competing interests

We declare we have no competing interests.

Acknowledgements

We thank Chris Venditti and Mark Pagel for critical reading of the manuscript.

Footnotes

Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5212562.

Published by the Royal Society. All rights reserved.

References