Testing the differential adhesion hypothesis across the epithelial−mesenchymal transition

Cadherin densities were measured with molecular resolution using stimulated emission depletion (STED) microscopy (see methods section 4.2). An example illustrating the increased image resolution of STED is shown in figure 1. The determined surface densities of the three individual cadherin types presented in figure 2(a) are well within the range of densities found for other cells with methods like quantitative flow cytometry or AFM surface scanning [25, 47]. Our data show a qualitative difference in cell line specific contribution of E-, N- and P-cadherin, mirroring the typical characteristics of an EMT. For the MCF-10A cell line, the main contribution to the overall cadherin density is the high amount of E-cadherins, indicating an epithelial phenotype, while for the MDA-MB-231 and MDA-MB-436 cell lines N- and P-cadherins are more dominant, indicating a more mesenchymal state.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Moreover, a switch of cadherin types is often assumed to be characteristic when normal cells become cancerous [9, 11–14]. We observe a loss in E-cadherins and a modest increase in N-cadherins between the non-malignant MCF-10A cells and the breast cancer cell lines MDA-MB-231 and MDA-MB-436, while P-cadherin levels remain quite similar between all three cell lines. Although all of them are sometimes classified epithelial, the two cancer cell lines are obtained from pleural effusions, thus stemming from cells that have undergone EMT. Moreover, the cancer cell lines behave highly invasive in Boyden chamber essays [44].

Unfortunately, it is not clear a priori how each cadherin type contributes to the mechanical force required to separate cells. One approach would be to study only E-cadherins which are often assumed to dominate the interaction. However, recent results contradict the often voiced assumption that E-cadherin mediated contacts are particularly strong [48]. Previously reported cross-type cadherin binding [25, 49, 50] with binding strengths close to self-type binding [4, 51] motivates the assumption that all three types may contribute equally to the total energy of binding. The summation of all E-, N- and P-cadherins per surface area for each cell type is shown in figure 2(b). MCF-10A has the highest density with (23.5 ± 1.5) μm⁻², followed by MDA-MB-436 with (19.8 ± 1.1) μm⁻² and MDA-MB-231 with (17.4 ± 1.2) μm⁻². If adhesiveness is, in fact, proportional to the total number of cadherins at the interface, our measurements predicts that MCF-10A should have the highest adhesiveness, while MDA-MB-231 should have the lowest. If TST is also proportional to the total density of cadherins, these data suggest $\sigma (10{\rm{A}})\gt \sigma (436)\gt \sigma (231)$.

Next, we used the AFM-based CellHesion technique to measure the maximum effective adhesion between cell pairs for the three cell lines, i.e. the forces required for cell separation (see figure 3(a) and methods section 4.3). Since the contact area between the cells cannot be precisely determined, it is not possible to normalize the measured unbinding forces (see supplementary section S.1). Nevertheless, simple evaluations are sufficient to estimate when cellular size or stiffness effects dominate over changes in adhesive strength. The calculation of an adhesive surface energy is also problematic since cell detachment is more like a highly nonlinear rupture event. Figure 3(b) summarizes the measured maximum cell adhesion forces. Unlike the conventional dogma, it is not the normal epithelial MCF-10A cell line that adheres the strongest. With (976 ± 290) pN, MDA-MB-231 cells show the highest adhesion force, approximately twice as high as that of MCF-10A cells with (508 ± 62) pN. MDA-MB-436 cells have the lowest adhesiveness, only a quarter of that of MDA-MB-231 cells, with (257 ± 109) pN (mean ± STD). The cell–cell contacts strengthen with contact time, which makes the differences in maximum cell adhesion force even more pronounced (see figure 4). This ripening phase seems to level off after about 90 s.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Since desmosomes form very slowly [52–54], the investigated cadherin types should be mostly responsible for the cell–cell contact on the time scale of the AFM experiments. In that light, the AFM results are quite surprising, as they demonstrate that the force required to separate cells is not proportional to the density of cadherins at interfaces. Even assuming that one of the cadherin types might be the main contributor to cell–cell adhesion, neither the distribution of E-, N- or P-cadherins fits our AFM adhesion data. For instance, E-E-cadherin bonds are reported to be stronger as the N-N-cadherin bonds [49, 55], but this cannot resolve the discrepancy.

One explanation for the discrepancy is that mechanical properties of cells other than adhesion or cell size are affecting the AFM measurement. For example, softer cells permit deeper indentation and correspondingly larger surface area in contact, resulting in a larger measured adhesiveness. In supplementary sections S.1 and S.2, we discuss several measurements of cell stiffness and cell contact area, which were used to estimate the adhesive force per unit contact area. However, we find that these effects are not sufficient to change the ordering of cell–cell adhesiveness among the three cell lines studied.

A second explanation is that different cadherin molecules regulate different spatiotemporal aspects of adhesion. For example, it has recently been shown that P-cadherin densities predict the overall magnitudes of intercellular forces while E-cadherin densities affect the rates at which these forces build up [15]. In this case, we might expect the relationship between adhesion and cadherin density to depend sensitively on the ratios between cadherin molecule densities and the time scale of the adhesion measurement.

A third possible explanation for the discrepancy is a difference in the spatial organization or regulation of cadherin molecules. Cadherins are able to nonlinearly alter adhesion by clustering [56], independent of contributions either from cytoskeletal or cytoplasmic interactions [57]. Unfortunately, it is not known how adhesion forces scale with cluster size, so we are unable to investigate this hypothesis quantitatively. Therefore, this represents a plausible but unverified distinction between different tissue types.

A related explanation is that cadherins primarily act as signalling molecules that down-regulate cortical tension over long timescales [30, 32, 33]. If this explanation is correct, we would expect cell shapes and TST to correlate strongly with cadherin expression but not with cell adhesion measurements. In contrast, if short time cell–cell adhesion does correlate with TST (as suggested by the original version of the DAH), our data would indicate $\sigma (231)\gt \sigma (10{\rm{A}})\gt \sigma (436)$.

We first imaged liquid-overlay cultures of homogeneous cell suspensions for each of the three cell types. As shown in figure 5, all three cell lines do form aggregates, which is a bit unanticipated given that MDA-MB-231 and MDA-MB-436 cells exhibit mesenchymal features. However, the aggregates differ qualitatively. While the MCF-10A cells form compact aggregates with a smooth, rounded, fluid-like surface, the MDA-MB-436 and MDA-MB-231 aggregates are increasingly less compact and display rough surfaces, where individual cells sit on the surface like sand grains. MDA-MB-231 aggregates are even less compact than MDA-MB-436, with branches, protrusions and even holes. It is important to note that any fluid-like droplet must have smooth, continuously rounded surfaces; therefore these observed macroscopic shapes already demonstrate that these aggregates are not acting as Newtonian fluids. Instead, the surface is more reminiscent of dense colloidal clusters or complex yield-stress fluids. This strongly suggests that cell activity and kinetics (as opposed to passive mechanical properties) are playing an important role.

Zoom In Zoom Out Reset image size

Nevertheless, we can study the mechanical forces acting on each cell and compare them to predictions of the DSCH and the regular as well as the extended DAH. The DSCH suggests that TST is governed by cortical tension. For cells described as an active fluid, a stiffer cell translates into a higher cortical tension and consequentially in a high TST, while softer cells have a lower cortical tension and lower TST. As discussed in supplementary section S.2, we used an optical cell stretcher to quantify cells' compliance. We find that MCF-10A cells are slightly stiffer than MDA-MB-436 cells, and both are significantly stiffer than MDA-MB-231 cells (see supplementary figure S.5). Because this stiffness correlates with cell cortical tensions, the DSCH would predict $\sigma (10{\rm{A}})\geqslant \sigma (436)\gt \sigma (231)$.

As discussed in [32, 33], the extended DAH suggests that cell shapes at the boundary of an aggregate are determined by a balance between the effective tension on contacting and non-contacting interfaces, and should correctly account for downregulation of cortical tension due to adhesion molecule signalling at interfaces. Qualitatively, this means that for cell types with large numbers of adhesion molecules, cells at the surface share significant contact interfaces with neighbours and a nearly flat interface with the cell culture medium, as we have found for MCF-10A. In contrast, surface cells on MDA-MB-436 and MDA-MB-231 aggregates are much more rounded and share very little contact with other surface cells. MDA-MB-231 cells are clearly more loosely bound to the surface than the other two cell lines, if contact area is a measure how strongly the cells stick to the surface.

To quantitatively test the extended DAH, we analyzed the contact angle θ between the curved cell-medium interface for each cell and a vector tangent to the macroscopic curvature of the aggregate, see figure 6(a). The contact angles were determined manually from the fluorescence pictures of the cell aggregates. The results are highlighted in figure 6(b): the angles are (41 ± 36)° for the MDA-MB-231 (N = 68), (22 ± 12)° for the MCF-10A (N = 36) and (32 ± 11)° for the MDA-MB-436 (N = 40). This suggests that the mean values of the contact angles are mostly insensitive to the clearly visible surface roughening. The loss of the smooth surface for the mesenchymal cells becomes obvious when looking at the measured values for the variance which considerably increases for the MDA-MB-231 cells.

Zoom In Zoom Out Reset image size

For cell–cell contacts near the surface, the cell shape and simple force balance can be used to calculate the local tension as an estimate for the TST. As discussed in supplementary section S.3, the contact angles together with the data for the cortical tension can be used to predict $\sigma (10{\rm{A}})\gt \sigma (436)\gt \sigma (231),$ which correlates perfectly with overall cadherin levels, but not with AFM adhesion data.

We performed cell sorting experiments with all three possible combinations of two cell lines. (The method is illustrated in figure 7 and described in detail in section 4.5.) Our first observation is that 50:50 mixtures of cells (with about 2000 cells total) do, in fact, segregate. In all cases, an inner 'core' was formed by one cell type, surrounded by an outer 'shell' of the other cell type. A clear hierarchy was reliably reproduced in all experiments with the three cell lines: MDA-MB-231 envelops MCF-10A, which in turn envelops MDA-MB-436. Figure 8 shows the segregation processes after approximately 20 h, which is still an intermediate time point. As seen in figure 8(b), there are occasionally smaller satellite clusters that have not yet fused to the main cluster. For longer observation times, these clusters will eventually fuse, forming a single cell type phase embedded in the other cell type. This robust segregation behaviour was not necessarily expected because all cell lines originate from the same developmental compartment (mammary gland). This suggests that segregation is a universal process, and importantly, that even cells that underwent EMT, such as MDA-MB-231 and MDA-MB-436, can spontaneously segregate.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Not surprisingly, if one cell type makes up less than 50% of the mixture then there is not a high enough density to generate a single main cluster [58]. Additional experiments with mixing ratios of 70:30 and 60:40 generated final states where small clusters of the minority are embedded in a spheroid of the majority type, and so we focus on 50:50 mixtures for the remainder of this manuscript.

According to both the DSCH and the extended DAH, the tissue with the highest surface energy should always sort to the interior, while the tissue with the lowest energy should sort to the exterior. If this is true, the cell sorting order would indicate $\sigma (436)\gt \sigma (10{\rm{A}})\gt \sigma (231).$ However, this is not consistent with predictions based on either the cadherin density ($\sigma (10{\rm{A}})\gt \sigma (436)\gt \sigma (231)),$ cortical tension ($\sigma (10{\rm{A}})\gt \sigma (436)\gt \sigma (231)),$ or surface cell shapes ($\sigma (10{\rm{A}})\gt \sigma (436)\gt \sigma (231))$ or with predictions based on cell–cell detachment force ($\sigma (231)\gt \sigma (10{\rm{A}})\gt \sigma (436)$).

Initially, this observation was very surprising, as previous experiments confirm that embryonic tissues sort according to these mechanical properties associated with surface energy [21, 36]. However, embryonic tissues behave like Newtonian fluids, while our observations demonstrate that our three cell lines undergoing EMT are not behaving as simple liquids. Because DSCH, regular DAH and extended DAH are all based on the assumption that tissues are liquids on long time scales (and therefore kinetic traps and cell activity can be neglected), it is not surprising that they fail.

In order to test whether cell activity and dynamics play an important role in cell sorting, we analyze and quantify the temporal evolution of coarsening and segregation in our sorting experiments. Figures 9–11 show consecutive still images illustrating the typical domain coarsening we observe during cell sorting. In the beginning, cells sediment to the bottom of the moulded spherical cavity and then start to form connected regions. The clusters of the same cell type grow in size and fuse with one another. Gradually, they form one large cluster of an inner core and an outer shell. Time-lapse videos of the segregations can be found in the supplementary material. Using these videos, we quantify the length scale L_m associated with these domains as it evolves in time (see methods section 4.6). The temporal evolution of L_m for all three cell line combinations is plotted in figure 12. A theory for Newtonian fluids developed by Siggia [59] and extended to tissues by Beysens et al [60] predicts that the rate at which domains coalesce is linear in time; we fit the linear regime in each plot in figure 12 to estimate the domain coalescence rate m. We find that ${m}_{10{\rm{A}}/231}\gt {m}_{10{\rm{A}}/436}\gt {m}_{\mathrm{231/436}}$.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

According to Newtonian fluid theory, cell mixtures with greater surface tension differences should have domains that coalesce faster (see supplementary section S.4). In combination with TST values concluded from the final sorted states ($\sigma (436)\gt \sigma (10{\rm{A}})\gt \sigma (231),$ see figure 8), this generates a testable prediction for the domain coalescent rates:

$$\begin{eqnarray}&&{m}_{10{\rm{A}}/436}~\displaystyle \frac{|{\sigma }_{436}-{\sigma }_{10{\rm{A}}}|}{{\eta }_{436}}\lt \displaystyle \frac{|{\sigma }_{436}-{\sigma }_{231}|}{{\eta }_{436}}~{m}_{\mathrm{231/436}}\end{eqnarray}$$

Obviously, this prediction does not fit to the coalescence rates m determined from the pseudo period data as the two mixtures containing MCF-10A cells have higher slopes m than the mixture containing no MCF-10A cells. This discrepancy again indicates that our epithelial cell lines are not behaving like a mixture of two fluids.

We can also use the TST predictions from our other measurements to make predictions for domain coalescence rates, as shown in the last column of table 1.

Table 1.  Summary of the predictions for the tissue surface tension σ and coalescence rate m resulting from our different measurements. The order of m can be concluded from the order of σ among the three cell types according to Newtonian fluid theory that predicts that cell mixtures with greater difference of σ should coalesce faster.

	Predicted tissue surface tension σ:			Predicted coalescence rate m:
Criteria	low		high	low		high
Cadherin density	231	436	10A	10A/436		10A/231
(STED)
Cell–cell adhesion forces	436	10A	231	10A/231		231/436
(AFM)
Cellular stiffness	231	436	10A	10A/436		10A/231
(optical stretcher)
Cell shapes and cell–medium	231	436	10A	10A/436		10A/231
interfacial tension
Sorting hierarchy	231	10A	436	10A/436		231/436
(segregation experiments)	(outside)		(inside)
Pseudo periods				231/436	10A/436	10A/231
(segregation dynamics)

Breast carcinoma is a key example where normal epithelial cells transform to malignant mesenchymal cells. For our studies, we chose three breast cell lines: MCF-10A (Cat.No. CRL-10317, ATCC) as an example of normal breast epithelial cells as well as MDA-MB-231 (Cat.No. HTB-26, ATCC) and MDA-MB-436 (Cat.No. HTB-130, ATCC) as particularly motile and invasive cancer cell lines [44]. Thus, we consider an enhanced motility as the main feature of the EMT and do not focus on a particularly strong change in cadherin expression. Another reason for choosing these three cell lines was their similar elastic properties and cell size (see supplementary section S.1), yielding a scenario in which the segregation process should not be affected by differences in these properties.

MCF-10A cells were cultured in DMEM/Ham's F12 medium containing l-glutamine (Cat.No. E15-813, PAA Laboratories GmbH, Austria) supplemented with 5% horse serum (Cat.No. A15-151, PAA), 20 ng ml⁻¹ human epidermal growth factor (Cat.No. E9644, Sigma-Aldrich), 10 μg ml⁻¹ insulin (Cat. No.I9278, Sigma-Aldrich), 100 ng ml⁻¹ cholera toxin (Cat.No. C8052, Sigma-Aldrich), 500 ng ml⁻¹ hydrocortisone (Cat.No. H0888, Sigma-Aldrich) and 100 U ml⁻¹ penicillin/streptomycin (Cat.No. P11-010, PAA).

MDA-MB-231 and MDA-MB-436 cells were cultured in DMEM containing 4.5 g l⁻¹ glucose, l-glutamine, but without sodium pyruvate (Cat.No. E15-810, PAA) supplemented with 10% foetal bovine serum (Cat.No. A15-151, PAA) and 100 U ml⁻¹ penicillin/streptomycin.

All cell lines were incubated at 37 °C in a 95% air/5% CO₂ atmosphere. The culture medium was changed every 2–3 days and cells were passaged every 4–5 days. To detach the cells, a PBS solution containing 0.025%(w/v) trypsin and 0.011%(w/v) EDTA (Cat.No. L11-004, PAA) was applied for several minutes.

STED microscopy introduced in [64, 65] was used to visualize E-, N- and P-cadherins on the cell surface, which were immunofluorescently labelled with cadherin type specific primary antibodies and a STED-compatible secondary antibody.

First, cells were seeded in glass bottomed tissue culture treated microwell plates (μ-Plate 96 well ibiTreat, Cat.No. 89626, ibidi GmbH, Germany) suitable for high resolution immersion microscopy. The adherent cells were fixed by applying a PBS solution containing 4%(w/v) paraformaldehyde and 1%(w/v) sucrose for 10 min at room temperature. Afterwards, cells were washed three-times with PBS, followed by incubation with blocking buffer composed of 0.1% Triton X-100 (Cat.No. X100, Sigma-Aldrich) and 10% goat serum (Cat.No. G9023, Sigma-Aldrich) in PBS for 15 min at room temperature. Then, cells were incubated with 5 μg ml⁻¹ mouse anti-E-cadherin (Cat.No. 610181, BD Biosciences), 5 μg ml⁻¹ mouse anti-N-cadherin (Cat.No. 610920, BD Biosciences), or 5 μg ml⁻¹ mouse anti-P-cadherin (Cat.No. 610227, BD Biosciences) in blocking buffer for 1 h at room temperature. Afterwards, cells were again washed three-times with PBS, followed by incubation with blocking buffer for 15 min at room temperature. Then, cells were incubated with 0.2 μg ml⁻¹ ATTO 647N conjugated goat anti-mouse IgG (Cat.No. 610-156-121S, Rockland Immunochemicals, USA) in blocking buffer for 1 h at room temperature. After that, cells were washed four-times with PBS and rinsed once in purified water (Millipore Milli-Q). Finally, cells were mounted in ProLong Gold Antifade (Cat.No. P36930, Invitrogen/Molecular Probes) to reduce photo-bleaching. Curing was allowed at least 48 h at room temperature before measurement.

Cadherins in the interface between the cell surface and the glass substrate were imaged with a confocal laser scanning microscope (Leica TCS SP5 STED, Leica Microsystems GmbH, Germany) in STED mode using a 635 nm laser for excitation. The conventional confocal laser scanning mode with a 633 nm laser was just used for control. The STED resolution in the xy-plane was measured to be 70 nm for the wavelength of the cadherin bound ATTO 647N fluorophore. The fluorescence signal was faint; thus, 32 to 64 line averages were done to increase the image contrast. Images were taken with at least half the maximal resolution of the STED system, in order not to lose any information.

For further image analysis, only the super-resolution STED images were used. A customized peak finding algorithm written in MATLAB applies a local maximum based filter to detect cadherins. If connected regions are larger than the maximum STED resolution, they are regarded to be clusters of multiple cadherins. Based on the cluster size and the reported size of single cadherin complexes [47], the number of cadherins in the cluster was approximated. For each of the different combinations of cell type and cadherin type, at least three cells were imaged and analyzed.

Note that the cadherins of individual adhered cells have been imaged and quantified at the cell–substrate interface, while during the segregation process, the situation of cadherins in the cell–cell interface might be different. We preferred this approach to quantify cadherins over real time quantitative PCR methods because they do not even quantify actually present receptors, but only the 'underlying' gene expression.

An atomic force microscope (AFM) (JPK CellHesion 200, JPK Instruments AG, Germany) in combination with an inverted phase contrast microscope (Leica DM IRB, Leica Microsystems GmbH, Germany) was used in order to probe the adhesion forces between pairs of cells, as introduced in [66]. To allow cell attachment, tipless cantilevers (Arrow TL2, NanoWorld AG, Switzerland) with spring constants around 30 mN m⁻¹ were functionalized by coating them with 20 μg ml⁻¹ fibronectin (Cat.No. F1141, Sigma-Aldrich) in PBS overnight at 4 °C. For the measurement, suspended cells were freshly flushed into a surface-treated plastic Petri dish (Cat.No. 93040, TPP Techno Plastic Products AG, Switzerland). While the cells settled down and started to adhere to the substrate, one cell had to be attached to the cantilever. To 'capture' a cell, the coated cantilever was pushed with a force of 500 pN for 5–10 s onto a cell still loosely laying on the substrate and subsequently, it was retracted again with the cell slightly sticking to the cantilever. This was followed by a rest time of 20–30 min to let the one cell firmly adhere to the surface of the cantilever and the other cells to the substrate.

To measure cell–cell adhesion forces, the cell on the cantilever was brought into contact with a cell on the substrate with an approach speed of 5 μm s⁻¹, pushed against it with a constant force of 500 pN for 5 s (constant force mode) and finally separated again with a pulling speed of 5 μm s⁻¹. For the whole cycle, a force–distance curve was recorded. In this manner, each cell on the substrate was probed five times with a rest time of 10 s between cycles. A self-made Petri dish heater was used to keep the cells at approximately 37 °C during the whole AFM measurement. We used the same media which were used during cell culture (see section 4.1).

Our analysis concentrated on the maximum adhesion force which was extracted as the minimum of the retract segment of a force–distance curve, see figure 3(a). We ignored single (molecular) unbinding events which are sometimes visible as steps in the retract segment. As they only appear rarely, it is doubtful if they are representative.

Note that the AFM measurements have to be done on a different, much shorter, time scale in comparison to the segregation experiments which run over several hours to more than a day. On the one hand, this is necessary to gain acceptable statistics with the AFM by probing multiple cells. On the other hand, there are technical limitations impairing long-term AFM measurements such as drifts. Due to the limited pulling range of the AFM setup (in our case 100 μm), at one point, it becomes practically impossible to completely separate the cells after the contact (at least in a well-defined and measurable manner). Admittedly, short-term AFM measurements cannot cover the full dynamic range of cell–cell adhesion with respect to the segregation results. We tried to address this issue with a second set of AFM measurements in which a smaller amount of cells was systematically probed with varying contact times ranging from 2.5 to 90 s. The adhesion forces clearly reach a saturation level. Nevertheless, the formation of desmosomes at a later stage may further increase cell–cell adhesion in the normal epithelial cells.

Moreover, there is no generally accepted reliable method for normalizing the measured adhesion forces by means of the contact area between the two cells. Although our setup (at least to some extent) permits simultaneous usage of transmitted light microscopy techniques during a measurement, it is not able to precisely visualize the contact area. We tried several approaches to indirectly gain estimates for the contact area. But none of them turned out to be fully convincing (see supplementary section S.1). Still, our three cell lines do not drastically differ in cell size or elasticity (see supplementary section S.2), both of which would generally affect the contact area.

In order to create homogeneous cellular aggregates, a standard liquid overlay technique was used [67]. In brief, cells were passaged into plastic Petri dishes (Cat.No. 93060, TPP Techno Plastic Products AG, Switzerland) whose bottom surfaces had been coated with a 2%(w/v) agar gel (Cat.No. A6686, Sigma-Aldrich), preventing cell attachment. Thus, cells remained in suspension, only interacting with each other. After one day of culture, cell aggregates had formed. Then, aggregates were stained with 1x working solution of CellMask Deep Red plasma membrane stain (Cat.No. C10046, Molecular Probes/Life Technologies) for 10 min at 37 °C. A confocal laser scanning microscope (Leica TCS SP2 and Leica DM IR2, Leica Microsystems GmbH, Germany) with standard fluorescence excitation (633 nm HeNe laser) was used to image the aggregates.

For segregation experiments with heterogeneous mixtures of two cell populations, cells were first fluorescently labelled with unspecific cytoplasmic dyes, CellTracker Green CMFDA (Cat.No. C7025, Invitrogen/Molecular Probes) and CellTracker Red CMTPX (Cat.No. C34552) according to manufacturer's protocol. In brief, cell cultures were incubated with the staining solution (2 μM dye in serum-free medium) for 45 min at 37 °C and subsequently washed two-times with PBS, before detaching the cells by trypsinization (see section 4.1). Cells were then counted and mixed in the ratio of 50:50. As medium, we now used a 50:50 mixture of the two culture media of the individual cell lines (see section 4.1).

For initial experiments, the traditional method of hanging droplet cultures [68] was used, see figure 7(a). Droplets of 15 μl cell culture medium containing approximately 20 000 cells were deposited to the hydrophobic inner surface of the lid of a small plastic Petri dish (Cat.No. 93040, TPP Techno Plastic Products AG). The dish was sealed with Parafilm to prevent too fast drying. However, this method has several disadvantages: The amount of nutrients in such a small droplet of medium is very limited and turned out to be insufficient to observe the final stage of segregation for one of the cell population pairs. It seems that if all nutrients are depleted, the segregation process arrests prematurely. There is also the danger of cells altering their normal properties and behaviour under nutrient shortage. Furthermore, the setup is not suitable for long-term observations while keeping acceptable ambient conditions, in particular temperature and pH of the medium. Additionally, it is limited to upright microscopes. Since most culture media are carbonate buffered, CO₂ supply is required to keep the physiological pH of 7.4 in the medium.

To solve these issues, a novel technique for studying multicellular spheroid formation and cell segregation was developed, see figure 7(b).Using home-made stamps, hemispherical cavities with diameters of either 1 or 2 mm were moulded into a 2%(w/v) agar gel (Cat.No. A6686, Sigma-Aldrich) resembling the essential geometry of a hanging droplet. Due to inverting the setup, now plenty of medium, a heating cup and CO₂ supply can be used, allowing long-term observation of the segregation dynamics. As an alternative to the agar gel, a PDMS gel (Sylgard 184 Silicone Elastomer Kit, Dow Corning, USA) with a 0.1 mg ml⁻¹ PLL-PEG (Cat.No. PLL(20)-g[3.5]-PEG(2), SuSoS AG, Switzerland) coating can be used instead, increasing the storage life. Cell cultures in the moulded cavities were observed with the same confocal laser scanning microscope (Leica TCS SP5 STED) used for imaging the surface cadherins, but without STED mode. Multiphoton excitation with a 780 nm laser was used to reduce photo bleaching and phototoxic effects compared to standard fluorescence excitation (with 488 nm and 561 nm lasers). Every 35 min, image stacks containing sections of the culture were recorded, thus providing 3D image information. An automated stage allowed parallel imaging of multiple cavities with cells, increasing the throughput of experiments considerably.

In order to quantitatively characterize the dynamics of segregation process, the pseudo period L_m was determined for each time step from the binarized confocal image stacks with a 3D autocorrelation based algorithm using customized MATLAB scripts (The MathWorks, USA). The pseudo period L_m is defined as the mean spacing between domains of fluid of the same type. Because we study nearly 50:50 mixtures of two cell populations, the average cluster radius of one phase is the mean spacing between clusters of the other phase. Here, we use the cell population that will eventually form the inner cluster for the evaluation of L_m for each time step. To measure the average cluster radii in the confocal image stacks, we borrowed a technique from the analysis of porous media [69, 70] and calculated a 3D two-point autocorrelation function of the binarized images via ${\hat{S}}_{2}({r}_{1},{r}_{2})=\langle f(x+{r}_{1})f(x+{r}_{2})\rangle ,$ where f is a characteristic function: 1 for pixels corresponding to the cell type forming the inner cluster and 0 for the rest. Brackets denote an average over the spatial coordinate x and radially outwards from the central point. Then, the specific surface area is ${S}_{1}/{V}_{{\rm{tot}}}=-4{{S}_{2}}^{\prime }(0)$ and the volume fraction is ${V}_{1}/{V}_{{\rm{tot}}}={S}_{2}(0)$ [69]. ${S}_{1}/{V}_{1}=\frac{1}{3}\langle {r}^{3}\rangle /\langle {r}^{2}\rangle ,$ and if there is little variation in droplet sizes ${L}_{m}~{r}_{{\rm{avg}}}~-\frac{3}{4}{S}_{2}(0)/{{S}_{2}}^{\prime }(0).$ We carefully checked the formalism for several test stacks and found good agreement to the test data, even for larger variations in droplet sizes.