Single cell polarity in liquid phase facilitates tumour metastasis

Dynamic polarisation of tumour cells is essential for metastasis. While the role of polarisation during dedifferentiation and migration is well established, polarisation of metastasising tumour cells during phases of detachment has not been investigated. Here we identify and characterise a type of polarisation maintained by single cells in liquid phase termed single-cell (sc) polarity and investigate its role during metastasis. We demonstrate that sc polarity is an inherent feature of cells from different tumour entities that is observed in circulating tumour cells in patients. Functionally, we propose that the sc pole is directly involved in early attachment, thereby affecting adhesion, transmigration and metastasis. In vivo, the metastatic capacity of cell lines correlates with the extent of sc polarisation. By manipulating sc polarity regulators and by generic depolarisation, we show that sc polarity prior to migration affects transmigration and metastasis in vitro and in vivo.

M etastases are the major cause of cancer-related deaths 1,2 . Despite novel promising targeted cancer therapies, patients diagnosed with systemic metastatic disease are no longer eligible for curative treatment options in many cancer subtypes [3][4][5] necessitating research on additional, broadly applicable strategies for metastasis intervention. Metastasis is a multistep process comprising dedifferentiation, dissociation and local invasion of primary tumour cells, intravasation into blood or lymph vessels, survival and transport in circulation, arrest in microvessels of distant organs and extravasation and metastatic outgrowth 6 .
Throughout the metastatic process, solid tumour cells establish distinct types of polarity, such as apical-basal polarity in the tissue context of established primary or metastatic tumours or front-back polarity during migratory phases 7,8 . The metastatic cascade thus involves dynamic depolarisation and repolarisation of metastasising cells, reflecting their high plasticity. However, the polarisation of cells during liquid or detached phases and the relevance of such polarisation for metastasis have remained unclear.
Here we identify a distinct type of polarity termed single-cell (sc) polarity that tumour cells maintain in liquid phase. Sc polarity is defined by the intrinsic presence of an ezrin-and actinrich pole in absence of an extracellular stimulus in non-adhering, non-migrating cells. We characterise sc polarity in tumour cell lines and human tumour specimens from biopsies collected in liquid phase and investigate the role of sc polarity in human tumour cells, mouse models of metastasis and ex vivo. We find that sc polarity affects attachment, adhesion, transmigration and metastasis.

Results
Tumour cells maintain their polarity in liquid phase. To investigate sc polarity in tumour cells in liquid phase, polarity markers of different polar structures of single cells [9][10][11][12][13] were imaged in human SkMel2 melanoma cells in suspension (Fig. 1a). Ezrin-green fluorescent protein (GFP) as well as endogenous ezrin, moesin, Radixin-GFP and phosphorylated ezrin/radixin/ moesin proteins accumulated at one pole of single cells in suspension ( Fig. 1a and Supplementary Fig. 1a). Additionally, polar accumulation of F-actin and the plasma membrane (PM) receptors CD44, β1-Integrin, melanoma cell adhesion molecule (MCAM) and intercellular adhesion molecule-1 (ICAM-1) was observed (Fig. 1a). The PM itself was accumulated at the pole and enriched with phosphatidylinositol 4,5-bisphosphate (PIP 2 , Fig. 1a and Supplementary Fig. 1a) while the polarity regulator Protein Kinase C ζ did not co-localize with the ezrin pole (Fig. 1a). Interestingly, the apical marker podocalyxin was polarised in detached cells, however, independently of the ezrin pole, localising to a PM area located distal to the nucleus (Fig. 1a), demonstrating that sc polarity is distinct from apical-basal polarity. Furthermore, in suspension the ezrin pole was not aligned with the nuclear-centrosomal axis, distinguishing it from uropod-like structures of amoeboid migrating cells 12,13 (Supplementary Fig. 1b).
The PM is convoluted at the site of ezrin accumulation as demonstrated by correlative light and electron microscopy (CLEM) of SkMel2 cells expressing ezrin-GFP in suspension (Fig. 1b). The PM folds were larger than microvilli or shedded microvesicles and connected to the cell body, as confirmed by serial tomography (Fig. 1c, Supplementary Movie 1). Ribosomes, endoplasmic reticulum (ER) and mitochondria were identified and visualised inside the PM folds (Fig. 1c, d). Furthermore, an actin-rim was observed at the base of the pole structure (Fig. 1e), corroborating the phalloidin staining (Fig. 1a). Analysis of pole morphologies revealed two subpopulations of polarised cells, with cap-like or spot-like, sometimes protruding poles ( Supplementary  Fig. 1c, Supplementary Movies 2 and 3).
To assess the dynamics of sc polarisation, we performed polarity assays with detached SkMel2 cells maintained in suspension (Fig. 2a). The fraction of polarised cells, which is defined as the percentage of cells forming an ezrin pole, decreased with a T 1/2 of 1 h. Even after 6 h of depolarisation, 39% of cells remained polarised (Fig. 2a). When maintained in suspension on a non-adhesive (poly-HEMA) substrate, most cells maintained their poles over hours, while only few cells displayed dynamic poles (Supplementary Movie 4). Polarised cells were detected throughout a period of 10 weeks (Fig. 2b). After 4 h in suspension, cells started to aggregate ( Supplementary Fig. 1d). The distribution of cell-cycle stages as measured by ki67positivity 14,15 remained equal for polarised and unpolarised cells in suspension (Fig. 2c). Polarisation was observed among ki67 high , ki67 low and ki67 neg cells (Fig. 2c, d), indicating that sc polarity is not specifically associated with a certain cell-cycle stage. Apoptotic cells always displayed cytoplasmic ezrin distribution (Fig. 2e). Imaging of detaching SkMel2 cells showed that the pole was frequently generated from a retracting protrusion where ezrin accumulated and remained as a stable spot upon cell rounding (Fig. 2f, Supplementary Movie 5), similar to ezrin accumulation in physiological processes such as morphological switch to amoeboid shape (Supplementary Movie 6) or tail retraction 11 . In cells with more than one retracting protrusion, multiple ezrin spots fused into one pole (Supplementary Movie 7). A minor population of cells developed the pole from a preformed ezrin-rich site at the flank of the cell (Supplementary Movie 8). In dividing cells, ezrin accumulated at the cleavage furrow 10 (Fig. 2d, g) and remained polarised in daughter cells after division ( Fig. 2g and Supplementary Movie 9). These data indicate that single tumour cells in liquid phase maintain a polarisation of ezrin, actin and the PM that constitutes a 'memory' of a previous polarised state, irrespective of the type of initial polarisation event.
Sc polarity is observed in tumour cells from patients. Polarisation of ezrin-GFP and F-actin was also observed in detached cells from different tumour entities, including melanoma, colorectal adenocarcinoma, cervical adenocarcinoma, hepatocellular carcinoma and breast invasive ductal carcinoma ( Fig. 3a and Supplementary Table 1), demonstrating that sc polarity constitutes a general mechanism by which detached tumour cells of different entities maintain their polarity.
To ascertain whether our observations in cell lines also hold true for human specimens, we investigated tumour cells from liquid-phase biopsies of cancer patients for the presence of ezrin caps or spots (Supplementary Table 2). We observed polarisation of ezrin and CD44 in cells from pleural effusion (PE) or ascitic effusion (AE) of pancreatic and breast cancer patients (Fig. 3b, Supplementary Fig. 2 and Supplementary Movie 10) showing that sc polarity occurs in human patients. Additionally, we detected sc polarity in single circulating tumour cells (CTCs) isolated from peripheral blood of metastatic breast cancer patients (Fig. 3c, Supplementary Fig. 2 and Supplementary Movie 11) and in a cluster of CTCs from a small cell lung cancer (SCLC) patient ( Supplementary Fig. 2c). Polarisation of ezrin, phalloidin, β1-Integrin and MCAM was observed in CTCs ( Supplementary  Fig. 2d-f). Malignant origin of CTCs was confirmed by single-cell array comparative hybridization 16 (Supplementary Figs. 4 and 5). Ezrin polarisation was also confirmed in single, disseminated melanoma or non-SCLC cells in lymph node specimens of patients (Fig. 3d) and even observed in single, round melanoma cells within the tumour mass of established primary and metastatic melanomas (Fig. 3e), however, at a very low frequency (<0.1%). The fraction of polarised cells was increased (~1-20%) in cell lines derived from these tumours ( Supplementary Fig. 3d) when cells were singularised in suspension, indicating that detachment favours sc polarity. Altogether, our data show that sc polarity occurs in different tumour entities in the liquid phase and in single cells at different steps of the metastatic cascade.
Sc polarity is correlated with tumour metastasis. In order to investigate the role of sc polarity in metastasis in vivo, an MDA-MB-231 xenograft mouse model 17 was used to isolate GFPlabelled CTCs, assess their polarisation state and measure the metastatic burden in the lung (Fig. 4a) (Fig. 4a), indicating that increased sc polarity is associated with a higher metastatic activity of CTCs. Furthermore, sc polarity was significantly lower in parental A375 human melanoma (Fig. 4b), TD-2 mouse pancreatic ductal adenocarcinoma (Fig. 4c) (Fig. 4e). Moreover, melanoma cell lines derived from metastases also showed increased sc polarity as compared to cell lines derived from primary melanoma 21 (Fig. 4f). Altogether, these data show a correlation between sc polarity and the cell-intrinsic metastatic potential of tumour cells in mouse models and human tumour cells.
The sc pole is involved in early cell attachment. To measure the orientation of the pole during attachment, SkMel2 cells expressing ezrin-GFP were seeded onto plastic or a layer of human umbilical vein endothelial cells (HUVEC). Cells were fixed after 5 or 30 min (Fig. 5a) and non-attached cells were removed by gentle washing. The ezrin pole was predominantly oriented towards the substrate (plastic or HUVEC) at 5 min suggesting an active role of the pole in attachment. At 30 min after seeding, the pole shifted towards the distal side (Fig. 5a, b). This change in orientation was also confirmed in SkMel28 and A375 melanoma cells (Fig. 5c). Interference reflection microscopy (IRM) demonstrated close contact of the pole to the substrate at 5 min (Fig. 5d). Attachment with the pole and reorientation was also observed by confocal live-cell imaging (Fig. 5e (6 h) in suspension (n = 6, mean ± SD). The line shows an exponential decay fit. b Fluorescence images of SkMel2 cells expressing ezrin-GFP maintained in suspension on poly-HEMA-coated plates for 10 weeks. Scale bars: 50 µm. c SkMel2 cells expressing ezrin-GFP maintained in suspension for 15 min, 3 h or 6 days were stained against ki67. Polarised and unpolarised cells within each population were graded by intensities into high, low and negative ki67expression. In all, 17-90 single cells were analysed per measurement (n = 3, mean ± SD, unpaired t-tests ), similar to a previously described moesin cap 9 . The finding that the early orientation towards the substrate was also observed on plastic or glass surfaces suggests that these initial interactions between sc pole and substrate are unspecific and not receptor-mediated.
We next investigated the polarisation of human melanoma cells during seeding in the human liver, one of the target sites for melanoma metastasis 22 , using an ex vivo liver perfusion system ( Fig. 5f and Supplementary Fig. 7 perfusion. After injection of tumour cells, the liver was perfused for 5 min in order to remove non-attached cells before fixation. Distribution, attachment and orientation of retained melanoma cells were assessed by immunofluorescence and immunohistochemistry ( Fig. 5f and Supplementary Fig. 7). Most melanoma cells were mechanically arrested in liver sinusoids and deformed ( Supplementary Fig. 7b, c). In addition, single, round cells predominantly oriented with the pole towards the attachment site were identified in larger vessels ( Fig. 5f and Supplementary  Fig. 7d). Altogether, these experiments show that melanoma cells are initially oriented with the pole towards their attachment site in vitro as well as in the human liver ex vivo and that the pole is in close contact with the substrate, indicating that the pole is directly involved in early attachment.
MCAM, Merlin and MLC regulate sc polarity and transmigration. To investigate the regulation of sc polarity, we first analysed the effects of ezrin overexpression, knockdown (kd) or mutant ezrin on sc polarity. Kd or overexpression of ezrin showed no relevant effect on the fraction of polarised SkMel2 cells ( Supplementary Fig. 8a, b). The phospho-mimetic ezrin-TD and the non-phosphorylable ezrin-TA mutants both displayed decreased fractions of polarised cells (Fig. 6a), indicating that ezrin phosphorylation turnover but not ezrin expression levels play a role in the regulation of the polar localisation of ezrin. However, secondary effects due to overexpression of ezrin or mutants cannot be excluded 23 . We therefore identified other cellular factors that affect sc polarity and investigated their effect on transmigration and experimental metastasis.
One of the proteins that co-localised with the ezrin pole was MCAM (Figs. 1a and 6b), which is a melanoma progression marker 24 involved in adhesion, intracellular signaling and regulation of the rear of migrating cells [25][26][27][28][29][30] . In SkMel2 cells, overexpression of MCAM increased the fraction of polarised cells (Fig. 6c). This effect is not limited to MCAM, as overexpression of the cell adhesion molecules (CAMs) L1-CAM or ICAM also increased the fraction of polarised cells (Fig. 6d), although to a lesser extent. Consistently, kd of MCAM using small interfering RNAs (siRNAs) reduced the fraction of polarised cells (Fig. 6e). Ezrin was evenly localised and not decreased at the PM of MCAM kd cells ( Supplementary Fig. 8e). MCAM kd also reduced adhesion of cells under flow conditions (Fig. 6f). In transwell assays, overexpression of MCAM enhanced transmigration through a layer of HUVEC (Fig. 6g), whereas MCAM kd reduced transmigration (Fig. 6h). Interestingly, these effects were also observed without HUVEC, indicating a cell-autonomous mechanism (Fig. 6g, h). A positive feedback mechanism between MCAM and ezrin has been proposed involving direct MCAM-ezrin interaction 26 . However, deletion of the ezrininteraction motif KKGK in MCAM (MCAMΔKKGK) 26 did not alter the subcellular distribution of MCAM (Fig. 6b) or revert the effect of MCAM overexpression on ezrin polarisation (Fig. 6c) and transmigration (Fig. 6g). Moreover, ezrin overexpression did not decisively increase the fraction of cells displaying polarised MCAM (Fig. 6i), indicating that the regulation of sc polarity by MCAM is independent of positive feedback or direct interactions between MCAM and ezrin.
In vivo, overexpression of wild-type (wt) Mcam as well as McamΔKKGK increased the number of experimental metastases (Fig. 7a) as well as early metastatic seeding (Fig. 7b) of B16-F0 mouse melanoma cells to mouse lungs. While overexpression of wt Mcam had no effect on tumour size, overexpression of McamΔKKGK reduced the size of experimental metastases ( Supplementary Fig. 9c), indicating that the interaction between Mcam and ezrin plays no role in tumour seeding but in tumour growth. Consistently, kd of Mcam decreased experimental metastasis (Fig. 7c) as well as early metastatic seeding (Fig. 7d) of B16-F1 mouse melanoma cells, corroborating our in vitro observations. The effect of Mcam kd on long-term experimental metastasis (Fig. 7c) was more pronounced than the effect on metastatic seeding (Fig. 7d), suggesting that MCAM also plays a role during later steps of metastasis following metastatic seeding. Altogether, these results demonstrate that MCAM positively regulates sc polarity, transmigration and metastasis and reveal an additional, cell-autonomous mechanism by which overexpression of CAMs can favour metastasis.
Merlin, the product of the tumour--suppressor gene neurofibromatosis type II (NF2), is a regulator of sc polarity during cell division 10 . In detached SkMel2 cells, endogenous merlin was mostly cytoplasmic, but weak staining at the PM and pole was also observed (Fig. 8a). In polarity assays, kd of NF2 decreased the fraction of polarised cells (Fig. 8b). Ezrin was evenly distributed and not decreased at the PM in NF2 kd cells ( Supplementary  Fig. 11b). Consistently, NF2 kd also reduced adhesion under flow (Fig. 8c) and transmigration of SkMel2 cells through HUVEC or membrane (Fig. 8d). This result contrasts previous findings in cell lines, where kd of NF2 resulted in increased transmigration 32 . We therefore performed additional polarity and transmigration assays in three different melanoma cell lines A375, SkMel28 and WM1366 with NF2 kd ( Supplementary Fig. 11c), demonstrating that the effect of loss of merlin is cell type-dependent. Importantly, the effect on transmigration in each of the cell lines was correlated with the effect on sc polarity ( Supplementary  Fig. 11c). B16-F1 mouse melanoma cells with kd of Nf2 displayed significantly decreased experimental metastasis in vivo (Fig. 8e), correlating with reduced sc polarity and transmigration (Supplementary Fig. 11d, e). Nf2 kd had no effect on tumour size ( Supplementary Fig. 11f). Our data indicate that merlin, while being a negative regulator of tumourigenesis 33 , can be a positive regulator of transmigration and metastasis of tumour cells. Myosin light chain (MLC) was previously shown to regulate polarisation of amoeboid migrating cells 34,35 . In SkMel2 cells, phosphorylated MLC (P-MLC) strongly accumulated at the pole (Fig. 8f), while GFP-MLC was distributed cytoplasmic, demonstrating local activation of MLC at the pole. Overexpression of wt MLC (Fig. 8g) or the increase of phosphorylated MLC by 50 nM calyculin A ( Fig. 8h and Supplementary Fig. 13a) enhanced sc polarity, whereas overexpression of a dominant-negative (dn) MLC mutant (dnMLC, Fig. 8i) or the decrease of MLC activity by blebbistatin ( Fig. 8j and Supplementary Fig. 13b) reduced sc polarity in SkMel2 cells.
We next investigated the connections between merlin, MCAM and MLC (Fig. 8k-m). Reduction of merlin levels in SkMel2 cells overexpressing MCAM reverted the fraction of polarised cells to control levels (Fig. 8k), indicating that merlin is required for the polarity-inducing effect of MCAM. Co-expression of wt MLC together with MCAM did not further increase sc polarity as compared to MCAM expression alone, but co-expression of dnMLC decreased the fraction of polarised cells to control levels ( Fig. 8l), showing that MLC activity is required for the polarityinducing effect of MCAM. Furthermore, activation of MLC by 10 nM calyculin A had no relevant effect on control cells but reverted sc polarity in NF2 and MCAM kd cells to control levels ( Fig. 8m and Supplementary Fig. 13c). We therefore conclude that phosphorylation of MLC is a major regulatory mechanism for sc polarity in liquid phase.
Generic depolarisation decreases transmigration and seeding in vivo. As additional effects of MCAM or Merlin levels on transmigration and metastasis cannot be excluded 27,32 , we investigated the effect of generic depolarisation (without genetic or chemical manipulation) on transmigration and in vivo seeding of melanoma cells into the lung. SkMel2 cells were depolarised by incubation in suspension for 3 h, which reduced the fraction of polarised cells by 30% compared to 30 min in suspension (Fig. 9a). Depolarisation had no decisive effect on cell viability, apoptosis or growth ( Supplementary Fig. 15a-c) or on the expression of PM receptors at the cell surface ( Supplementary  Fig. 15d). Using polarised and depolarised cells from the same preparation, transmigration assays were performed either through HUVEC (Fig. 9b) or through membranes (Supplementary Fig. 15f, g). Cells depolarised prior to seeding showed strongly reduced transmigration after 4 h and slightly reduced     Fig. 9b and Supplementary Fig. 15f), demonstrating that the state of polarisation prior to seeding affects transmigration. The experiment was also extended to 24 h of transmigration time to allow seeded cells to (re-) polarise upon attachment to the substrate 9 (Supplementary Movie 15). After 24 h, no clear difference between polarised and depolarised cells was observed ( Fig. 9c and Supplementary Fig. 15g), showing that generic depolarisation does not relevantly affect the migratory capacity of cells. We therefore investigated whether sc polarity affects adhesion of cells  (Fig. 9c, d). Adhesion of depolarised cells was delayed as compared to polarised cells up to 60 min after seeding (Fig. 9c). Furthermore, adhesion of single, unpolarised cells was delayed over the same time course compared to single, polarised cells within the same, untreated population of cells ( Fig. 9d and Supplementary Movie 16). However, after 90 min, polarised and depolarised cells had adhered to the same extent. Under flow conditions, the slower adhesion kinetics of depolarised cells led to a reduced number of attached/adhered cells compared to polarised cells (Fig. 9e). The transient delay in adhesion over time (Fig. 9c, d) was consistent with the effect of generic depolarisation on transmigration (Fig. 9b). The transient, transmigrationreducing effect of generic depolarisation was also observed in U87 glioblastoma 36,37 (Fig. 9f, g) and semi-adherent SNU-1 stomach carcinoma cells 38 (Fig. 9h, i), which can both grow in adherent as well as in suspension phase. Depolarisation had no relevant effect on cell viability, apoptosis, growth or expression of PM receptors at the cell surface of U87 or SNU-1 cells (Supplementary Fig. 17).
To investigate the effect of generic depolarisation on seeding of cells into the mouse lung in vivo, polarised (30 min in suspension) and depolarised (3 h in suspension) SkMel2 cells from the same preparation expressing ezrin-GFP were injected into the tail vein of mice and the number of GFP + cells was quantified in the lung after 30 min (Fig. 9j, k). Histological images and a three-dimensional (3D) reconstruction using step-cuts from lung tissue (Supplementary Movies 17 and 18) demonstrated that the seeding of cells was higher in the polarised fraction than in the depolarised fraction (Fig. 9j). The amount of GFP + cells was quantified by three different methods ( Fig. 9k and Supplementary Fig. 18a) to correct for effects of differences in tissue density. All three quantifications revealed a significantly decreased number of seeded cells from the depolarised fraction as compared to the polarised fraction ( Fig. 9k and Supplementary  Fig. 18a). The same effect of generic depolarisation on seeding into the lung was observed using U87 or SNU-1 cells (Fig. 9l, m and Supplementary Fig. 18b, c). Altogether, these experiments show that sc polarisation prior to attachment directly advances adhesion, transmigration and in vivo seeding of melanoma, glioblastoma and gastric cancer cells.
Sc polarity affects attachment, adhesion and metastatic seeding. In order to investigate which steps of early metastatic seeding are affected by sc polarity, a basic compartmental model was generated (Fig. 10a and Supplementary Fig. 19) comprising circulation, attachment, adherence and tissue residence of unmixed populations of 100% polarised or 100% depolarised cells using parameters for transfer rates between compartments measured for mixed populations of polarised and depolarised cells in Fig. 9d, e, k and Supplementary Fig. 18a.
This model was used to simulate the individual steps of early metastatic seeding and assess the differences in attachment, adhesion and tissue residence between polarised and depolarised cells (Fig. 10c-e and Supplementary Fig. 19). The model predicts an approximately 1.3-fold higher adherence rate for polarised cells (3.95%/s) as compared to depolarised cells (3.02%/s), which is in agreement with our experimental data on adhesion kinetics (Fig. 9d). Furthermore, the model predicts a three-fold attachment rate for polarised cells (10.74%/s) compared to depolarised cells (3.81%/s) as suggested by our data showing attachment of cells via the sc pole (Fig. 5). The predicted tissue residence rate of polarised cells (0.06%/s) is also slightly enhanced as compared to depolarised cells (0.05%/s) but considerably lower than attachment and adhesion rates and therefore contributing to the difference in tissue residence only to a minor extent. Simulation of the full chain of events taking all compartments into account demonstrates that a slight delay in attachment (Fig. 10c) and adhesion (Fig. 10d) of depolarised cells results in considerably decreased tissue residence (Fig. 10e) thus predicting a strong effect of sc polarity on metastatic seeding as observed in our in vivo experiments with mixed populations (Fig. 9k-l).
In summary, the model allows us to confirm our hypothesis that sc polarity increases attachment and adhesion rates (and to a minor extent tissue residence rate), leading to enhanced tissue residence of polarised cells.

Discussion
We have performed a comprehensive investigation of sc polarity in detached tumour cells and its role in metastasis. We define sc polarity as a type of polarity that cells adopt in liquid phase without directional stimulus in vitro and in vivo. Sc polarity is characterized by a round morphology of cells and formation of a pole containing ezrin, F-actin, P-MLC, PIP 2 , CAMs, integrins and PM folding. Despite distinct functions, morphologies and origins, the molecular composition of the sc pole overlaps with that of uropod-like structures 12,13 , cell capping during cell division 10 , adhesion 9 or leucocyte capping 11 , invadosomes 39 , the immunological synapse (IS) 40 or the distal pole of IS-forming cells 41 , which might favour the transition between the sc pole and such structures. While cells forming such structures are fully polarised by cell-cell attachment, cell-matrix attachment or migration and centrosome positioning, cells displaying sc polarity in liquid phase show no alignment of the pole with the nuclear-centrosomal axis. We propose that the sc pole constitutes not only in cancer patients in the liquid phase (CTCs, PE and AE cells) but also at other steps of the metastatic process and might favour metastasis at different steps during tumour progression by maintaining a generic type of polarity (Fig. 10f). Our in vitro, in vivo and in silico data suggest that sc polarity affects transmigration and metastasis by two distinct mechanisms, favouring early attachment by unspecific interaction with substrate and advancing adhesion by 'pre-polarisation'. First, local accumulation of PM rich in highly glycosylated proteins may lead to a high tendency for unspecific interaction of the sc pole with any substrate and thereby enhance early attachment as suggested by the preferential orientation of the pole towards the substrate (Fig. 5). Second, pre-polarisation of the PM and actin cytoskeleton might advance the formation of a distal cap 9 and thereby favour adhesion. During the metastatic process in vivo, these two effects of sc polarity could lead to increased vascular retention and accelerated adhesion of CTCs.
The extent of sc polarity correlated with metastatic properties of human tumour cell lines as well as CTCs in a mouse model. Our findings agree with previous concepts 43, 44 suggesting that not only quantity but rather quality of CTCs and disseminated tumour cells influence the chance of successful metastatic colonisation and that sc polarity may constitute such a metastatic quality of CTCs. The therapeutic potential of inhibition of sc polarity to target cancer metastasis will need to be evaluated in clinical studies. Research into molecular determinants involved in regulation of sc polarity will provide further insight into the underlying cellular mechanisms.

Methods
Definition of sc polarisation. Throughout this study, a sc pole was defined as a single, continuous PM area encompassing less than half of the circumference of the cell image enriched in ezrin (or phalloidin) fluorescence by two-fold or higher as compared to the remaining area of the PM.  45 by Dr. S. Stangl (TU Munich) and SNU-1 by Professor M. Gerhard (TU Munich). Cell lines were not specifically authenticated. U87 and TD-2 cells were maintained in RPMI (Gibco); all other tumour cell lines in Dulbecco's modified Eagle's medium (DMEM; Gibco) with Glutamax and 10% foetal calf serum (FCS, Gibco), 100 IU/ml penicillin and 100 µg/ml streptomycin (Gibco). HUVEC were maintained in medium 199 (Sigma), with 20% FCS, 2 mM ultraglutamine (Lonza), 25 IU/ml penicillin and 25 µg/ml streptomycin (Gibco), 62.5 µg/ml Heparin (Sigma) and 30 µg/ml endothelial growth factor supplement (Sigma). SkMel2 cells stably expressing ezrin-GFP were fluorescence-activated cell sorted (FACS) for high green fluorescence and maintained in selection medium with 1 mg/ml G418/geneticin (Sigma). B16-F0 cells stably transfected with pEpi-GFP, pEpi-MCAM or pEpiMCAM-ΔKKGK were maintained in selection medium containing 5 µg/ml blasticidin (Applichem). B16-F1 cells stably transduced with short hairpin RNAs (shRNAs) were maintained in selection medium containing 625 ng/ml puromycin (Calbiochem). To image amoeboid morphology, cells were grown on top of a thick layer of 1.7 mg/ml bovine collagen I (Gibco A10644-01). Cell lines were routinely tested for mycoplasma infection using the PlasmoTest Kit (InvivoGen). All cells were split 1 day before use. Transfections with plasmids and/or siRNAs were performed using Lipofectamine 2000 or 3000 (Thermo Fisher Sci.); transductions with viral particles were performed according to the manufacturer's protocol. Growth and viability of polarised and depolarised fractions of SkMel2, U87 and SNU-1 cells was assessed by (i) trypan blue staining and counting (ViCell XR, Beckman Coulter), (ii) plating of cells for 24 h, fixation in 4% formaldehyde, staining with 4,6-diamidino-2-phenylindole (DAPI), imaging of nuclei and counting using ImageJ 46 , (iii) AnnexinV staining (BD), (iv) live-dead staining using EMA (Life Technologies) and (v) LDH assay (Cytotoxicity Detection Kit, LDH, Roche) and (vi) soft agar colonyformation assay (CytoSelect 96-Well Cell Transformation Assay, Cell Biolabs, Inc.). For aggregation assays, cells were stained with 1 µM Cell Tracker Green CMFDA (Life Technologies) for 30 min, recovered in DMEM with 10% FCS for 24 h, detached with Versene (EDTA, Lonza) and fixed in suspension after the indicated time. Aggregation (direct contact of 2, 3 or >3 cells) was quantified from 20× widefield fluorescence images of fixed suspension cells. Growth of B16 cells was assessed using the Cell Proliferation Kit II (XTT, Roche). All n values represent biological replicates.
Establishment and detachment of ccRCC cell lines. Primary clear cell renal cancer-derived cell lines were established from tumour resection specimens of ccRCC patients. Primary xenografts were established by implanting tumour pieces of 1-2 mm 3 into the renal capsule of NOD.Cg-Prkdcscid Il2rgtm1Wjl (NSG) mice bred in the animal facility of the German Cancer Research Center. Primary xenografts were resected at a volume of approximately 1 cm 3 , tumour pieces were minced, dissociated by incubation with 1 μg/ml collagenase IV and DNase for 1 h at 37°C and filtered through a 100-and 70-µm mesh. In all, 5 × 10 6 cells were seeded in T75 flasks in serum-free medium. Suspension and adherent cultures were maintained at 37°C and 5% CO 2 . Established KIKA lines 47 were authenticated (Multiplexion) and tested for mycoplasma contaminations. After separation with StemPro Accutase (Gibco), cells were resuspended in CO 2 -independent medium (Gibco) supplemented with 1% bovine serum albumin (BSA), incubated 1 h at 37°C in a tube rotator and fixed in 4% paraformaldehyde (PFA).
Adhesion assays. SkMel2 cells expressing ezrin-GFP were detached for 10 min using Versene/EDTA, (Lonza) and incubated in DMEM for 30 min. Cells were seeded onto 24-well cell culture dishes, fixed with 4% formaldehyde after 10, 30, 60 or 90 min and imaged using a 20× objective on an Olympus CKX41 with an Olympus XM10 camera using the CellSens Standard software (Olympus). The area per cell was measured using ImageJ 46 Supplementary Fig. 9a, 9e and 11d were fixed in suspension, mounted on a glass slide and imaged on a Leica SP5 confocal microscope using a 20× objective. The fraction of cells with a polarised cap or spot per total round cells was counted manually, blinded. For each experiment, images of at least five random areas were acquired and at least 100 cells were analysed. All n values represent biological replicates.
Transmigration assays. Transmigration was measured in transwell dishes (Corning-Costar, Lowell MA, 8 µm pore size, 6.5 mm diameter). Inserts were either used uncoated or coated with 0.1% gelatine (Sigma) for 30 min and 3.5 × 10 4 HUVEC (Cellworks, ZHC2301) for 3 days. For transmigration through HUVEC, cells were stained with 1 µM (SkMel2) or 10 µM (B16) Cell Tracker Green CMFDA (Life Technologies) for 30 min and recovered in DMEM with 10% FCS for 3 h (SkMel2) or 12 h (B16). Cells were detached using Versene and incubated in DMEM or HUVEC medium for 30 min. In all, 5 × 10 4 SkMel2 or B16 or 2 × 10 5 U87 cells were seeded on top of HUVEC, 10 6 (SkMel) or 5 × 10 5 (SNU-1) on the transwell insert, incubated as indicated and fixed with 4% formaldehyde. Cells on the top of the membrane were removed, transmigrated cells were stained with DAPI and membranes were mounted on a glass slide using Mowiol (Merck). Seven random images were obtained of DAPI (for transmigration through membrane only) or Cell Tracker Green (for transmigration through HUVEC) using a 10× objective on an Olympus CKX41 with an Olympus XM10 camera using the CellSens Standard software (Olympus). Cell numbers were quantified using Ima-geJ 46 . All n values represent biological replicates performed as technical duplicates and averaged; all experiments were normalised to an untransfected control.
Histology and immunohistochemistry. Paraffin sections were stained manually with haematoxylin and eosin or on a Bond MAX immunohistochemistry robot (Leica) with antibodies at concentrations listed in Supplementary Table 4 Table 4 for 12-18 h at 4°C, washed twice with PBS, incubated with secondary antibody and phalloidin as listed under 'Antibodies and reagents' for 1 h, washed with PBS, incubated with DAPI or ToPro3 for 5-15 min and washed 4 times with PBS. Cells were mounted on glass slides with Vectashield (Vector Laboratories) and imaged on a Leica SP5 confocal microscope with a 63× oil objective using the LAS software (Leica Microsystems, Wetzlar, Germany).
Pole orientation in vitro. SkMel2, SkMel28 or A375 cells expressing ezrin-GFP were seeded on chamber slides (µ-Slide 8-well, ibiTreat, Ibidi) uncoated or coated with 0.1% gelatine and HUVEC grown to confluence for 3 days. Cells were fixed with 4% formaldehyde 5, 30 or 45 min after seeding, washed with PBS and imaged. Stacks of 25 or 50 images of random round cells with a pole were acquired on a Leica SP5 confocal microscope with a 63× oil objective using the LAS software (Leica Microsystems, Wetzlar, Germany). Orientation of cells was assessed by two independent researchers not involved in the project. All n values represent biological replicates.
Widefield live cell imaging. SkMel2 cells stably expressing ezrin-GFP were grown on chamber slides (µ-Slide 8-well, ibiTreat, Ibidi) to 70-80% confluence. Versene/ EDTA was added directly at the microscope and acquisition started immediately. Fluorescence images were recorded using an Axiovert 200M and a Plan-neofluar 20×/0.5 NA objective equipped with a motorised heating stage, incubator (Zeiss Microimaging) and an AxioCam MRm camera. Using the AxioVision Software (Zeiss Microimaging), images were acquired every 30 s. Movies were exported and processed using imageJ 46 .
Confocal live cell imaging. SkMel2 cells stably expressing ezrin-GFP were detached using Versene/EDTA, resuspended and maintained in medium for 1 h, seeded directly onto imaging slides (µ-Slide VI 0.4 ibiTreat, Ibidi) or imaging slides coated with 0.1% gelatin and HUVEC grown to confluence and imaged directly after seeding. HUVEC were stimulated with 2 ng/ml tumour necrosis factor-α (Invitrogen) for 5 h before seeding of SkMel2. Stacks of 50 confocal images over 25 µm were recorded every minute using a 63× objective (HC PL APO CS2/1.40 oil) on a Leica TCS SP8 confocal microscope equipped with 405, 488, 552 and 638 nm diode lasers and a heating chamber. Top and side views of the stacks were reconstructed using the Leica LAS software and further processed using ImageJ 46 . For temporal colour coding, the Time Series Color Coder by Kota Miura (CMCI, EMBL Heidelberg) was used.
Interference reflection microscopy. SkMel2 cells stably expressing ezrin-GFP were detached using Versene/EDTA, resuspended and maintained in DMEM for 1 h, seeded onto imaging slides (µ-Slide 8-well, glass bottom, Ibidi) coated with 2% BSA, fixed with 4% formaldehyde and carefully washed twice with PBS. Interference reflection imaging 48 was performed on a Leica TCS SP8 confocal microscope. Reflection and fluorescence images were obtained using 488 nm excitation and 480-497 nm or 503-559 nm emission wavelengths, respectively.
To produce an IRM image, the focus was adjusted near the glass surface to obtain a zero-order interference image and the detector was adjusted for best contrast. For the reflection image, the pinhole was opened to 4 airy units (AU). In addition to the reflection image, a stack of 25 confocal fluorescence images was recorded in the GFP channel with the pinhole set to 1 AU. Top and side views of the stacks were reconstructed using the Leica LAS software and further processed using ImageJ 46 .
Fluorescence-activated cell sorting. Cells were detached with Versene/EDTA, incubated for 30 min and fixed in 4% formaldehyde immediately or 2.5 h later. Cells were washed three times with FACS buffer (PBS containing 2% FCS and 20 mM EDTA), stained with the respective antibody as listed in Supplementary  Table 4  Electron microscopy. SkMel2 cells were detached using Versene/EDTA, incubated in DMEM for 30 min and fixed in 2.5% glutaraldehyde (GA, Electron Microscopy Sciences) in 0.1 M PHEM buffer. Cells were pelleted and the subsequent EM processing steps (OSO 4 , uranyl acetate (UA), dehydration, Epon embedding) were performed on the entire pellet using a PELCO Biowave Pro microwave processor containing ColdSpot (Ted Pella, Inc.). This instrument uses microwaves under vacuum condition and controlled temperature. This has proven beneficial for EM sample preparation, as it improves diffusion of the fixatives as well as the resin during infiltration, giving better morphological results in a faster way 49 . In order to achieve this, 2 min on-and-off cycles of 100 W were applied under vacuum during glutaraldehyde and osmium steps and 1 min on-and-off cycles at 150 W for UA. For each dehydration step, samples were subjected to 40 s incubation at 250 W. Finally, infiltration was obtained with 3 min at 250 W for each step. For the entire processing, the temperature was controlled and set to 21°C. After polymerisation of the resin at 60°C for 48 h, blocks were sectioned with a Leica Ultracut S microtome. Seventy or 300-nm-thick sections were collected on formvar-coated slot grids and imaged with a Philips Biotwin CM120 (thin sections) or a FEI Tecnai F30 (tomography) electron microscope. For EM tomography experiments, tomogram reconstruction, segmentation and 3D rendering were carried out with the IMOD software package 50 .
Correlative light and EM. SkMel2 cells expressing ezrin-GFP were detached using Versene/EDTA, incubated in DMEM for 30 min, pelleted, resuspended in 20% BSA (Sigma 20 µl/10 cm dish) and cryo-immobilised using a high-pressure freezing machine (HPM 010, BAL-TEC). Freeze-substitution with 0.1% UA in acetone and embedding in Lowicryl resin 51 was performed in an AFS2 machine (Leica). The samples were incubated with 0.1% UA in acetone at −90°C for 6 h. The temperature was then raised to −45°C (5 degrees/h) and the samples incubated for 5 more hours. The freeze-substitution solution was then washed out with dry acetone, before infiltration in Lowicryl HM20 resin (steps: 10%-25%-50%-75%, 2 h each, followed by 3× rinses in 100% resin for 10 h each, during which the temperature was slowly raised to −25°C). Polymerization of the resin was carried out with ultraviolet light at −25°C. Sections (300 nm) were cut from the polymerised resin block and picked up on carbon-coated mesh grids. For fluorescence imaging, the grids were sandwiched with water between two coverlips 51 and imaged using an Olympus ScanR microscope and a 60× oil objective. Grids were then stained with UA and Reynolds lead citrate and tilt series was acquired with a FEI Tecnai F30 electron microscope. Tomograms were reconstructed using the IMOD software package 50 . Correlation between fluorescence microscopic and EM images (single tomographic slide) was performed by alignment of the grid bars and cell and nuclear outlines visible in both imaging modalities.
Antibodies and reagents. Primary antibodies, suppliers, catalogue numbers and concentrations used for immunofluorescence, FACS or western blotting are shown in Supplementary Table 4 Thirty minutes after injection, mice were sacrificed, the lungs perfused, fixed in 4% formaldehyde for 3-5 days and embedded in paraffin in 2-4 pieces. Two-µm tissue sections were stained with haematoxylin/eosin and with GFP, ki67 or gp100 antibody. All n values represent biological replicates.
Quantification of metastatic seeding to the mouse lung. Slides were scanned using a Leica SCN400 slide scanner at 20× magnification. Analysis was performed using Tissue IA (Slidepath, Leica). All slides within an experiment were analysed with the same algorithm and settings. For each slide, 2-3 annotations were made to cover most of the tissue avoiding staining artefacts and merged in the results. For generic depolarisation experiments, 2 step cuts of 100 µm, and for MCAM overexpression and kd experiments, 2-3 lung sections of each lung were analysed using an algorithm for stained area with optimised colour definition files with deconvoluted intensity threshold of 161 (generic depolarisation) or 165 (MCAM manipulation). Tissue intensity threshold was set to 211 or 215, respectively. For generic depolarisation experiments, the haematoxylin-positive area was measured using a predefined deconvolution-haematoxylin colour definition file with a deconvoluted intensity threshold of 230. All slides were analysed as GFP+ area per total tissue area, GFP+ area per haematoxylin+ area and GFP+ cells per total amount of cells. SkMel2 were in addition analysed by manual counting of GFP+ cells in 10 randomly chosen 40× view fields per slide. For MCAM overexpression or kd experiments, all slides were analysed as melanoma gp100+ area per total tissue area.
3D reconstruction of GFP + cells seeding in the lung. 3D reconstruction of an anatomical structure from 30 serial sections was generated using the Voloom software (microDimensions GmbH, München, Germany). The virtual slide data was loaded into Voloom where individual sections were automatically detected and separated. For 3D reconstruction, consecutive sections were registered to each other avoiding error propagation due to micro-artefacts like folding or tears. An initial reconstruction result on low-resolution images was propagated from low-to high-magnification levels inherently using the pyramidal structure of virtual slides. The anatomical structure was then segmented from the volumetric histology data using global thresholding and rendered with a green pseudo colour. The rendering process is shown in Supplementary Movies 17 and 18.
Generation and analysis of CTCs from MDA-MB-231 xenografts. Six-week-old NSG immunocompromised mice (Jackson Laboratories) underwent surgery under isofluorane anestesia. A cell suspension containing 100,000 GFP-positive MDA-MB-231 cells in 50 µl PBS was mixed with one volume of growth factor-reduced Matrigel (Corning) and injected into the fourth right mammary fat pad. Mice were housed under specific pathogen-free conditions in individually ventilated cages and maintained according to the German federal law, authorization number G-240/11. Mice were monitored weekly for tumour growth by palpation, and after 5 weeks, tumours were surgically resected from mice upon anesthesia by intraperitoneal injection of 10 ml/g of a solution of 4.5 mg/kg xylazinhydrochloride and 90 mg/kg ketamine. Surgically resected mice were closely monitored until clear signs of pain appeared; hence, they were then euthanized by anesthetic overdose according to local regulations. The peripheral blood was collected by heart puncture in EDTA tubes and mononuclear cells were separated by Ficoll-Paque gradient. The lungs were washed with PBS and digested enzymatically, using type IV collagenase and dispase, and mechanically using a GentleMACS device (Miltenyi). Circulating or disseminated cells were FACS-purified from these preparations as live intact GFPpositive cells that were negative for a mouse lineage antibody cocktail (including antibodies specific for mCD45, mCD31, mCD11b, Ly-6G and TER-119) using a FACSAria Fusion (BD) cytometer. Cells were immediately fixed in a 1.6% paraformaldehyde solution upon sorting. Ex vivo human liver perfusion. Fresh surgical human liver biopsy (26 g) was perfused via portal vein branches through two G20 cannules in closed circuit at 37°C/5% CO 2 with perfusion medium (50 ml Leibovitz's L15 medium without phenol red (Life Technologies) supplemented with 5% FCS, 1 µg/ml insulin (Serva), 23 mM HEPES pH 7.4, 5 mM L-glutamine, 50 IU/ml penicillin and 50 µg/ml streptomycin (Gibco), 100 µg/ml Gentamycin and 4.4 µg/ml Hydrocortisol at a flow rate of 1 ml/min/g. In all, 10 6 SkMel2 cells stably expressing ezrin-GFP were injected via the perfusion cannules at 1 ml/min/g and the liver perfused with medium for 5 further minutes. The liver was then fixed in 4% formaldehyde for 3 days and embedded in paraffin. Liver integrity was confirmed by haematoxylin/eosin staining, and sections were stained for GFP and vimentin and imaged on a Leica SP5 confocal microscope with a 63× oil objective using the LAS software (Leica Microsystems, Wetzlar, Germany). To quantify the orientation of cell polarisation, random round cells in vessels were imaged, where not more than half of the circumference of the cell was attached to vimentin-positive, GFP-negative cells. The orientation of cells was assessed by two independent researchers not involved in this project. Only cells with consistent orientation in both assessments were included in the quantification (40 out of the 58).
Human CTCs. CTCs of breast cancer patients were enriched and detected using the CellSearch® Epithelial Cell Test (Janssen Diagnostics, LLC). In brief, 7.5 ml of blood was collected in CellSave tubes and CTCs captured by anti-epithelial cell adhesion molecule (anti-EpCAM) antibody-bearing ferrofluid. CTCs were defined by positivity for cytokeratin, negativity for the leukocyte common antigen CD45 and DAPI positivity to ensure integrity of the nucleus. After quantification, stained cells were flushed from the cartridge and one aliquot was used to isolate single CTCs and white blood cells by DEPArray TM (Menarini Silicon Biosystems) for subsequent whole-genome amplification by Ampli1 TM according to the manufacturer's protocol (Menarini Silicon Biosystems) 54 . The remaining cells were stored in glycerol 55 , sedimented on adhesion slides (Thermo Fisher Scientific), stained for ezrin/Alexa 488 as described above and imaged on a Leica TCS SP8 confocal microscope.
Single-cell array comparative hybridisation. Ampli1TM whole-genome amplification libraries were checked for genome integrity index and then reamplified 54 . Labelling of CTC DNA and corresponding reference DNA from WGA libraries of single leukocytes of a healthy female donor and hybridisation on SurePrint G3 Human CGH 4×180K microarray slides (Agilent Technologies, design code 022060) were performed according to the manufacturer's protocol (Agilent Oligonucleotide Array-Based CGH for Genomic DNA Analysis, version 7.1, December 2011) 56 . Slides were scanned using an Agilent Microarray Scanner Type C, and images were processed with Agilent Genomic Feature Extraction Software (version 10.7) and imported and analysed with the Agilent Genomic Workbench Software (version 6.5 lite). For defining aberrant regions, we used the ADM-2 algorithm with threshold set to 7.0 and a centralisation of 6.0. To avoid false positive calls, the minimal number of probes in an aberrant interval was set to 50 probes and minimum log2 ratio to 0.25.
Tumour cells from PEs and AEs. Fresh pleural or peritoneal punctates were processed immediately. The whole cellular fraction was frozen and stored or immediately depleted from most CD45+ cells using magnetic human CD45 microbeads (Miltenyi Biotec 130-045-801). Cells were stained with EpCAM-FITC (Miltenyi Biotec) and CD45-APC (Miltenyi Biotec) in suspension or further purified by sorting of singlet-gated EpCAM+CD45−DAPI− cells on a FACSAria TM Fusion device (BD Biosciences) and fixed in 1.6% PFA. Cells were stained as indicated, attached to microscopy slides and imaged on a Leica TCS SP8 confocal microscope.
Human disseminated cancer cells. Adhesion slides with preparations of human lymph nodes from five melanoma patients and four non-SCLC patients were prepared 57 . Lymph nodes were mechanically disintegrated using a Medimachine (Beckham Coulter) and cells were subsequently enriched using percoll 60% gradient centrifugation. A total of 2 × 10 6 cells (at a density of 1 × 10 6 cells/ ml PBS) were transferred onto adhesion slides (Thermo Fisher Scientific). After 1 h sedimentation, PBS was removed. Slides were air-dried overnight at room temperature and fixed for 10 min with 4% formaldehyde prior to permeabilisation and staining as described above. Samples were stained for gp100 or EpCAM, ezrin and DAPI and imaged on a Leica TCS SP8 confocal microscope. In two out of the three melanoma samples and in two out of the two lung cancer samples that showed sufficiently high ezrin expression, polarisation of ezrin was observed. Control (non-cancer) nodes were obtained from patients with chronic venous insufficiency from whom a lymph node was removed during crossectomy. For each experiment, three negative controls were stained and analysed.
Human melanoma specimens. TMAs 21 comprised 70 primary melanomas, 216 melanoma metastases and melanoma patient's cell lines consisting of 41 cell lines derived from primary and metastatic melanomas. Primary tumours and metastases were mostly not related. TMA specification and construction were previously performed 21 . In all, 2 µm sections from these TMAs were immunohistochemically stained for melan-A and MCAM or fluorescently for melan-A and ezrin. Ezrin polarisation was imaged on a Leica SP5 confocal microscope with a 63× oil objective.
Evaluation of MCAM expression. To determine the expression frequencies of MCAM, a semi-quantitative scoring system was applied following the German immunohistochemical scoring (GIS) system in which the final immuno-reactive score equaled the product of the percentage of positive cells times the average staining intensity. Percentage of positive cells was graded as follows: 0 = negative, 1 = up to 10% positive cells, 2 = 11-50%, 3 = 51-90%, 4 = > 90%. Staining intensity of 0 = negative, 1 = weakly positive and 2 = moderately or strongly positive 58 . GIS score ≥4 was defined as high and <4 as low. All stainings were evaluated by an experienced pathologist (D. M.-P.). MCAM expression was compared between different patient groups using Pearson's Chisquared Test. p-values <0.05 were considered statistically significant.
Statistical analyses. Descriptive statistics of quantitative data are mean ± standard deviation or median and other quantiles, depending on the distribution of data. Likewise, qualitative data are presented by absolute and relative frequencies. Hypothesis testing was performed in an exploratory manner on two-sided 5% significance levels. Group differences of normally distributed data were assessed by t-tests. One-sample t-tests and paired-samples t-tests were used for dependent data. Relations of the latter were assessed by the Pearson correlation coefficient. Mann-Whitney U-tests were used to infer on group differences of non-normally distributed data. Differences in the distribution of qualitative data were investigated by Pearson's Chi-squared Test. Error bars are mean ± standard deviation. SPSS 22 software (SPSS Inc.) and GraphPad Prism were used for statistical analysis.
Analysis of pole morphology. Sixteen-bit images of SkMel2 cells expressing ezrin-GFP were converted to 32-bit without interpolation and the background was subtracted. Yen thresholding was used to segment cellular regions of low intensity and ezrin clusters of high intensity. Object-based regions were defined and utilised to measure the area and intensity of the identified cells and ezrin clusters.
'Aggregation score' for ezrin was defined as: Aggregation¼ where I is the intensity and A is the area of large ezrin clusters per cell. This agregation score is maximised when ezrin is strongly confined to a single cluster in a cell (spot-like) and minimised when it is dispersed over the cell (cap-like). A histogram of the aggregation scores was generated and fitted to a double modified-Gaussian peak model (Supplementary Table 5) to disentangle the two populations of cells. Image processing and calculation of aggregation score were performed using a single automated Python script provided as 'EzrinClassifier.py'. Dependencies: tifffile, NumPy, SciPy, pandas, Pillow, and scikits-image.
In silico modelling. Model set-up: An ODE-based compartmental model was constructed using the Simbiology toolbox in MATLAB R2016a. Three independent compartments corresponding to blood circulation, attachment/adherence substrate for cells and tissue parenchyma were defined in the model (Fig. 10a). Each compartment contains populations of polarised and depolarised cells. Transfer of cells between compartments was defined as occurring according to first-order reaction kinetics as follows: The first-order reaction rates for each transfer were determined by linked-fitting the appropriate steps in the model to three independent sets of experiments, i. in vitro experiments of attached polarised or depolarised cells to measure the adherence rate (Fig. 9d), fitted to Eqs. 3 and 4 ii. in vitro experiments of mixed populations of polarised and depolarised cells in suspension to measure their combined 'attachment and adherence' rates ( Fig. 9e) and fitted to Eqs. 1-4 iii. in vivo experiments of a mixed population of polarised and depolarised cells injected into blood circulation to measure their combined 'attachment, adherence and tissue residence' (Fig. 9k and Supplementary Fig. 18a), fitted to Eqs. 1-5 These experiments allowed us to fit all the required rate parameters and obtain predictions for the attachment and tissue residence rates of polarised and depolarised cells individually.
For each compartment, the polarised and depolarised cells are represented as distinct species. The only exception is for the tissue compartment since the experimental data do not provide information regarding the previous polarisation status of cells that are already tissue resident.
Fitting procedures: The data from all three independent experiments described above was pooled into a single matrix that contained the observed circulating, adhered, attached and tissue-resident cell populations for polarised and depolarised cells. The model fitting procedure did not contain any information regarding the biological/technical nature of the experiment, effectively making the result a linked fit of all three experiments. Linked fitting constrains the parameter space to only those ranges which fit data from all three experiments simultaneously. Fitting was performed using the 'constrained mixed effects non-linear least squares optimization' (nlmm-lsqnonlin) algorithm 59 , which is particularly suited for data from mixed populations, coming from disparate experiments. Termination tolerances were set as ResidualSumofSquares = 1.0E-8 or MaxIterations = 400. The linked-fit-derived rates were then tested against the three experiments individually to see if they predicted the observations to within 95% confidence bounds.
Simulations: The experimental data providing the fitted first-order rate parameters typically consist of mixed but biased populations of polarised and depolarised cells. Once obtained, the rate parameters can be used to simulate the behaviour of 'pure' populations of polarised and depolarised cells in terms of tissue residence. Simulations were then performed using the derived rate parameters for an interval of 15,000 s (250 min), setting the populations of polarised or depolarised cells to 100% and then computing the tissue residence.
The model is provided as an SBML file, datasets are provided as annotated CSV files and the results of simulations as an OPJ file (the viewer can be obtained under //www.originlab.com/viewer/). All data are also provided as one single SBPROJ file that is readable in MATLAB with Simbiology toolbox.
Results and predictions: The compartmental model fitted the data to within 95% confidence limits for all parameters (Supplementary Fig. 19a). The obtained transfer rates are provided in Supplementary Figure 19b. In agreement with experimental indications (Fig. 5), the model predicts a three-fold higher attachment rate of polarised cells compared to depolarised cells. Similarly, in agreement with experiments (Fig. 9d), the adherence rate of polarised cells is approximately 1.3fold the adherence rate of depolarised cells. Both cell populations are predicted to have much slower tissue invasion rates. The tissue invasion rate of polarised cells is only slightly enhanced compared to depolarised cells. These rates were used to simulate the entire chain of events from circulation to tissue residence for pure unmixed populations of polarised and depolarised cells (Fig. 10b-e). Such unmixed populations cannot be obtained experimentally, justifying the development of the aforementioned compartmental model. The changes in the circulating, attached, adhered and tissue-resident populations are shown in Supplementary Fig. 19c. Direct comparisons for all rates between polarised and depolarised cells are shown in Fig. 10b-e. Code availability. Custom code used for analysis and simulations are available as Supplementary Information. Data availability. The authors declare that the data supporting the findings of this study are available within the paper and its Supplementary Information files or available from the authors upon request.