Novel metastatic models of esophageal adenocarcinoma derived from FLO-1 cells highlight the importance of E-cadherin in cancer metastasis

There is currently a paucity of preclinical models available to study the metastatic process in esophageal cancer. Here we report FLO-1, and its isogenic derivative FLO-1LM, as two spontaneously metastatic cell line models of human esophageal adenocarcinoma. We show that FLO-1 has undergone epithelial-mesenchymal transition and metastasizes following subcutaneous injection in mice. FLO-1LM, derived from a FLO-1 liver metastasis, has markedly enhanced proliferative, clonogenic, anti-apoptotic, invasive, immune-tolerant and metastatic potential. Genome-wide RNAseq profiling revealed a significant enrichment of metastasis-related pathways in FLO-1LM cells. Moreover, CDH1, which encodes the adhesion molecule E-cadherin, was the most significantly downregulated gene in FLO-1LM compared to FLO-1. Consistent with this, repression of E-cadherin expression in FLO-1 cells resulted in increased metastatic activity. Importantly, reduced E-cadherin expression is commonly reported in esophageal adenocarcinoma and independently predicts poor patient survival. Collectively, these findings highlight the biological importance of E-cadherin activity in the pathogenesis of metastatic esophageal adenocarcinoma and validate the utility of FLO-1 parental and FLO-1LM cells as preclinical models of metastasis in this disease.


INTRODUCTION
Metastatic esophageal cancer has an exceedingly poor prognosis with a median survival of under 12 months [1]. Treatment options for these patients are limited and often ineffective [2]. The fact that over 70% of patients present with de novo metastatic disease or develop metastases after their diagnosis [3], highlights the need to improve our understanding of this pathological process in order to deliver better patient care.
Epithelial to mesenchymal transition (EMT) has been shown to play an important role in promoting metastasis in epithelium-derived carcinomas [4]. EMT involves changes at the genomic, epigenomic,

Research Paper
transcriptomic and proteomic levels both intrinsic and extrinsic to the cancer cell [5]. These alterations affect signaling pathways that ultimately enable cancer cells to invade locally, traverse the systemic circulation and colonize distant sites [4]. In esophageal cancer, how these molecular events interact to promote metastasis remains poorly understood.
Metastatic models of esophageal cancer are scarce and difficult to establish. As a result, most investigators typically use in vitro assays only [6,7]. Of those that are conducted in animals, intravenous or intracardiac injections are often used to seed cancer cells into distant organs [8,9]. These methods however, fail to mimic the full metastatic process which occurs in patients and thus risk obscuring translatable insights into the biology of metastasis. Therefore, spontaneously metastatic models of human esophageal cancer would be extremely valuable for understanding the metastatic process.
To date, a limited number of spontaneously metastatic animal models of esophageal cancer have been reported [10][11][12][13]. These models however, pose several key challenges. Firstly, they involve surgery to the esophagus which may result in heavy bleeding, organ perforation, anastomotic leakage and death. Indeed, the reported postoperative mortality for Levrat's rodent surgical reflux model is at least 30% [13]. Secondly, the metastatic phenotype is not robust or reproducible, with the rate of metastasis varying between 0-78% across different studies [11,[13][14][15][16]. Thirdly, the duration from surgery or cancer cell inoculation to micro-metastasis is over 40 weeks in some models [13,15]. These limitations therefore significantly hinder the use of these models for scientific discovery.
Models that develop timely and robust spontaneous metastasis without the need for invasive surgery would have significant preclinical utility. In this study, we show that FLO-1, a human esophageal adenocarcinoma (EAC) cell line, develops spontaneous metastasis following subcutaneous inoculation in mice. From this, we derived a highly metastatic and aggressive subline which, in combination with parental FLO-1, provides important insights into potential mechanisms underlying metastasis in esophageal cancer.

FLO-1 spontaneously metastasizes in NOD-SCID IL-2Rγ KO (NSG) mice
Spontaneously metastatic models of human esophageal cancer are lacking. To address this area of need, we subcutaneously injected 8 human esophageal cancer cell lines into mice with different levels of immunocompetency to determine whether they are tumorigenic and spontaneously metastatic (Table 1). A cell line was deemed non-tumorigenic if the injection site remained tumor-free 6 months post injection. Once subcutaneous tumors reached endpoint volume, necropsy was performed on all animals to search for evidence of macro-metastasis. We found that all 8 cell lines were tumorigenic in NSG mice. However, depending on the cell line, tumorigenicity decreased with increasing host immunocompetency (Table 1). Notably, macrometastases were only evident in NSG mice injected with the EAC cell line, FLO-1 ( Figure 1A). The location of these metastases mirrored those seen in EAC patients, with tumors predominately present in the lung, liver, peritoneum and mediastinal lymph nodes ( Figure 1A). Interestingly, we observed that the mammary artery ipsilateral to the subcutaneous tumor was consistently wider (Supplementary Figure S1A-S1B) and had more distributaries (Supplementary Figure S1C) than its contralateral counterpart. Furthermore, we also noted that metastases to the axillary lymph node, whilst relatively uncommon, always occurred ipsilateral to the subcutaneous tumor. These findings suggest that FLO-1 cells are able to metastasize via both lymphatic and haematological routes. To verify that these macrometastases were indeed derived from FLO-1 cells, we demonstrated that tumors in the liver and lung stained positively for human mitochondria and pan-cytokeratin ( Figure 1B). As NSG mice are at risk of developing de novo lymphomas [17], we also performed CD45 immunohistochemistry to exclude the possibility that these metastatic deposits were murine lymphoma in origin ( Figure 1B). To enhance the metastatic phenotype of FLO-1, we subcutaneously passaged liver metastases over 5 consecutive generations in NSG mice ( Figure 1C). We observed that with each successive generation the metastatic frequency increased, reaching 100% by the third passage ( Figure 1D). Furthermore, the distribution of metastases became more homogenous with increasing passages ( Figure 1E), and consequently, animals succumbed much earlier to their disease ( Figure 1F).

FLO-1 cells exhibit a mesenchymal phenotype
To understand why FLO-1 is more metastatic than the other esophageal cancer cell lines, we examined H&E stained sections of all subcutaneous xenografts established in NSG mice (Supplementary Figure S2A). We noted that FLO-1 xenografts were poorly differentiated ( Figure 2A) and had obvious evidence of lymphovascular invasion ( Figure 2B-2C, Supplementary Figure S2B). Importantly, given that EMT may result in tumor de-differentiation, invasion and metastasis [4,5], we next compared the mRNA and protein levels of well validated EMT genes across our cell line panel. We found that FLO-1 expressed significantly higher mRNA levels of the mesenchymal markers SNAI1, ZEB1 and ZEB2 compared to the other cancer cell lines ( Figure 2D). Furthermore, the level of the mesenchymal proteins N-cadherin, Vimentin, SNAI1 and ZEB1 were elevated in FLO-1 cells ( Figure 2E). In contrast, the expression of the key epithelial marker, E-cadherin, was lowest in FLO-1 cells and xenografts ( Figure 2E-2F). Taken together, these findings suggest that FLO-1 cells have undergone EMT, and thus posses an inherent predisposition for metastasis.

FLO-1 LM : a highly metastatic and aggressive derivative of parental FLO-1
To generate a robust and highly metastatic model of EAC, we derived a cell line designated FLO-1 LM from a FLO-1 liver macro-metastasis following repeated passages through NSG mice ( Figure 1C). Using short tandem repeat (STR) analysis, we firstly confirmed that FLO-1 LM is indeed derived from FLO-1 parental cells (Supplementary  Table S1). In comparison to parental cells, we found that FLO-1 LM were smaller in size, and exhibited greater migratory, invasive, proliferative and clonogenic capacity in vitro ( Figure 3A-3E). Additionally, we noted that FLO-1 LM were more resistant to apoptosis, both under basal growth conditions and following cisplatin treatment, than parental cells ( Figure 3F).
with a luciferase-eGFP reporter and compared their metastatic behavior in vivo in a limiting dilution assay using bioluminescence imaging ( Figure 4A). We found that regardless of cell numbers injected into NSG mice, FLO-1 LM tumors grew faster ( Figures 4A-4B, Supplementary Figure S3A-S3C), metastasized earlier ( Figure 4C) and more widely ( Figure 4D-4E, Supplementary Figure S3D) than FLO-1 parental tumors. Interestingly, we found that FLO-1 LM cells were also tumorigenic and metastatic in nude mice ( Figure 4F-4H), whereas FLO-1 parental cells completely failed to grow in these mice with relatively higher immunocompetency.

FLO-1 LM
We next performed genome-wide RNA-sequencing (RNAseq) comparing FLO-1 LM and FLO-1 parental cells to gain molecular insights into the phenotypic differences between these two cell lines. Using an adjusted p-value < 0.05 as a cut-off, a total of 375 (307 downregulated, 68 upregulated) genes were differentially expressed in FLO-1 LM compared with parental cells (Supplementary Tables S2-S3). We found that the gene ontology (GO) terms most strongly associated with differentially expressed genes in FLO-1 LM were processes linked to cancer metastasis ( Figure 5A, Supplementary  Table S4). These included regulation of cell adhesion, migration, differentiation, cytoskeletal components, cellular proliferation, apoptosis, immune system processes and angiogenesis. To extend these findings, we performed qRT-PCR and western blot analysis and identified increased expression of multiple anti-apoptotic and EMT genes in FLO-1 LM compared with parental cells ( Figure 5B-5C, Supplementary Figure S4A-S4D). Strikingly, on review of our RNAseq data, we found that CDH1, which encodes for E-cadherin, was the most significantly downregulated gene in FLO-1 LM ( Figure 5D). This was subsequently validated using qRT-PCR and western blotting ( Figure 5C and 5E). To determine whether FLO-1 LM is a clinically relevant model, we next compared GO processes enriched in patient samples with those found in FLO-1 LM . The GSE19417 Gene Expression Omnibus (GEO) microarray dataset was chosen for comparison, as it is one of the largest publicly available treatment-naive EAC cohorts with clinical annotation in humans [19]. We stratified this cohort based on CDH1 expression, with low CDH1 defined a priori as the bottom 25% (and high CDH1 the top 75%). Significantly differentially expressed genes (Benjamini-Hochberg adjusted p-value < 0.05) between CDH1 high and low tumors were identified (Supplementary Table S5) and the derived list subjected to GO enrichment analysis. Top enriched GO processes found in FLO-1 LM (vs. FLO-1 parental) and CDH1 low tumors (vs. CDH1 high tumours) were subsequently compared. Of 234 GO processes (top 20%) enriched in FLO-1 LM , 132 (56.4%) overlapped with those found in the patient derived samples of the GEO dataset ( Figure 5F, Supplementary Table S6), and many of these have immediate biological relevance to EMT, cancer invasion and metastasis ( Figure 5G). Taken together, these results demonstrate that we have derived a clinically relevant isogenic subline of FLO-1 with heightened mesenchymal characteristics that is more aggressive, less immunogenic, and highly metastatic in nature.

Low E-cadherin expression is associated with increased FLO-1 metastasis and reduced patient survival
Given that E-cadherin is a key inhibitor of EMT in other tumor types [4], we next examined the role of E-cadherin in metastasis using our FLO-1 model. Based on our finding that E-cadherin was significantly reduced in FLO-1 LM relative to parental cells ( Figures 5E  and 6A), we hypothesized that FLO-1 cells with low E-cadherin expression are intrinsically more metastatic than E-cadherin high expressers. To address this, we isolated subpopulations of FLO-1 parental cells with high and low E-cadherin levels ( Figure 6A) and injected them separately into NSG mice. Using a TaqMan qPCR assay which quantifies the proportion of human versus mouse vimentin DNA in an organ, we analyzed the metastatic burden in the liver and lungs from these animals once their subcutaneous primary tumors reached the same endpoint weight ( Figure 6B). Strikingly, we found that metastatic burden from cells with low E-cadherin expression was on average 160,000-fold greater in the liver and 12-fold greater in the lungs than high E-cadherin expressing cells ( Figure 6C). Extending these findings, we found that E-cadherin knockdown in FLO1-parental cells ( Figure 6D) resulted in smaller cells, enhanced migration, and increased the expression of SNAI2, ZEB2 and TWIST1 in vitro ( Figure 6E-6G, Supplementary Figure S5A), effectively phenocopying FLO-1 LM cells.
To investigate the clinical importance of CDH1/Ecadherin dysregulation in esophageal cancer, we firstly analyzed TCGA datasets (publicly available at cBioportal for Cancer Genomics: www.cbioportal.org) and reviewed the literature to determine the extent of CDH1/E-cadherin aberration in esophageal EAC. Across multiple studies, we found that E-cadherin levels were commonly lost, reduced or heterogeneously expressed in EAC ( Figure 6H) [20][21][22][23][24][25][26][27][28][29][30][31]. This is in part explained by the frequent hyper-methylation of the CDH1 promoter resulting in impaired gene transcription ( Figure 6I) [32][33][34][35][36]. Additionally, genomic abnormalities such as mutations and deletions, albeit uncommon, have also been reported ( Figure 6I) [28,37]. Interestingly, these findings in EAC were also evident in  . Finally, we correlated CDH1 expression with patient survival using the GSE19417 dataset whilst simultaneously controlling for clinicopathological variables (Supplementary Table S7). We found that low CDH1 expression was independently associated with significantly worse overall survival ( Figure 6J). In keeping with this, multiple studies have reported that low E-cadherin expression in esophageal cancer is significantly associated with increased lymph node and distant organ metastasis ( Figure 6H, Supplementary Figure S5B). Collectively, these findings illustrate the biological importance of CDH1 expression in esophageal cancer metastasis, and further validate the utility of FLO-1 parental and FLO-1 LM cells as preclinical models of metastasis.

DISCUSSION
There is a profound need for preclinical metastatic models of esophageal cancer. Here, we report the identification of FLO-1, and derivation of FLO-1 LM , as two complementary spontaneously metastatic cell line models of human EAC in NSG mice. Using a combination of genomic, transcriptomic, proteomic and histopathological approaches, we have characterized these two cell lines and demonstrated that they recapitulate key aspects of metastatic disease in patients. These include the presence of lymphovascular invasion and high histological grade in primary subcutaneous xenografts. Additionally, the distribution of metastases seen in vivo, with a predilection for the liver, lungs and mediastinal lymph nodes, are in line with what is commonly observed in patients [63]. In support of this, metastasis-related pathways active in FLO-1 LM were also enriched in EAC samples. Furthermore, since parental FLO-1 cells were originally established from a treatment-naive patient (Table 1), these pathway alterations are likely to reflect innate tumor biology absent of therapy-related selection pressures. Collectively, these findings strongly support the utility of FLO-1 and FLO-1 LM as representative preclinical models of metastasis.
From an experimental standpoint, FLO-1 and FLO-1 LM overcome several key challenges inherent in existing metastatic models of esophageal cancer that have significantly limited their use. Firstly, FLO-1 and FLO-1 LM cells metastasize in a timely manner, approximately 12 and 6 weeks after subcutaneous injection of tumor cells respectively. In comparison, metastases only occur after 40 weeks post surgery in Levrat's surgical reflux model [13] or with EC9706 cells [15]. Secondly, the metastatic frequency of both FLO-1 parental (92%) and FLO-1 LM (100%) lines are much higher than those reported for PT1590 cells (78%), OE19 cells (70%), EC9706 cells (50%) and Levrat's reflux model (0-40%) [11,[13][14][15][16]. Consistent with this, in our hands, OE19 cells failed to metastasize in all animals despite robust primary tumor growth. Thirdly, spontaneous metastasis from a subcutaneous site eliminates the need for invasive surgery to generate such models [10,13,64]. This would minimize procedural related morbidity and mortality [13] and increase experimental reproducibility of the model.
Interestingly, parental FLO-1 cells were metastatic in NSG mice but not in NOD-SCID mice. The main difference between these two transgenic strains is the knockout of IL-2 receptor gamma-chain in NSG mice [65], which results in loss of cytokine signalling and impaired T, B, and natural killer cell function [66]. This likely facilitates metastasis of FLO-1 cells in NSG mice. Consistently, other studies have also reported that engraftment rates of humanderived tissues and cells are higher in NSG mice than in any other immuno-deficient mouse models [67,68].
A key strength of our study is the establishment of FLO-1 LM , which in contrast to parental FLO-1, is significantly more invasive in vitro and more metastatic in vivo. Therefore, these two models enable further understanding and investigation of mechanisms involved in cancer invasion and metastasis. To harness this aspect, we conducted a genome-wide comparison of both cell lines using RNAseq and gene ontology enrichment analysis. Consistently, the most differentially altered pathways between FLO-1 LM and parental FLO-1 were processes involved in regulating cell adhesion, migration, differentiation, proliferation, cytoskeletal organization, angiogenesis and apoptosis. Importantly, CDH1 was identified as the most significantly downregulated gene in FLO-1 LM .
The tumor suppressor gene CDH1 encodes the cell adhesion molecule E-cadherin [69]. E-cadherin is crucial for the establishment and maintenance of epithelial cell polarity, mediating intercellular adhesion, regulating cytoskeletal organization and preventing EMT [69]. Therefore, E-cadherin expression is a hallmark of epithelial differentiation, and its repression is associated with tumorigenesis, invasion and metastasis in many tumor types [69]. Whilst numerous retrospective clinical studies of EAC have correlated low CDH1 or E-cadherin expression with metastatic disease and poor survival [70], our study is the first to provide functional in vivo evidence of this process.
Here, we showed that FLO-1 parental cells with low levels of E-cadherin expression have increased metastatic capacity compared with E-cadherin high expressers. Additionally, E-cadherin knockdown in FLO-1 parental cells induced a mesenchymal-like cell morphology, increased cell migration and upregulated EMT genes.
Reduced E-cadherin expression is common in EAC and typically occurs early during carcinogenesis [71]. The cause of reduced E-cadherin expression is unclear. It is unlikely however, that genomic aberrations alone will explain this phenomenon, as the incidence of CDH1 mutation is relatively low in EAC. Whilst frequent promoter hypermethylation have been reported [35,36], these results are not consistent across all studies [32,33]. Therefore, it is likely that multiple factors are at play to suppress CDH1 expression. Indeed, SNAI1, ZEB1 and ZEB2, well established transcriptional repressors of CDH1 [4] are overexpressed in parental FLO-1. Their expression is further elevated in FLO-1 LM , which concordantly has lower CDH1 levels. Furthermore, as reported in other tumors types, upregulation of DAB2 [72], RND3 [73,74], TMEM97 [75], CSE1L [76], MYO1C [77], and PTPRK [78] have been shown to suppress or functionally redistribute E-cadherin expression in cancer cells, resulting in EMT, increased invasion, advanced tumor stage and poor patient survival. Consistently, our RNAseq analysis demonstrates that these genes are all significantly upregulated in FLO-1 LM . Thus, these findings support our hypothesis that multiple factors conspire to repress E-cadherin expression in EAC.
Although E-cadherin low expressing FLO-1 cells were significantly more metastatic than E-cadherin high expressers, we observed that FLO-1 cells with high E-cadherin levels were still able to metastasize. It is likely that other factors may also promote metastasis. Consistent with this, our RNAseq analysis have also identified significant upregulation of TRIP6, RAP2A and CCL2 in FLO-1 LM . Elevated expression of these genes has been functionally shown to promote key hallmarks of metastasis in different cancers independent of E-cadherin [79][80][81][82][83][84][85][86][87][88][89][90][91]. Uniquely, our isogenic FLO-1 and FLO-1 LM cell lines thus enable identification and functional interrogation of these genes and other pathways underlying esophageal cancer metastasis.
Beyond highlighting mechanisms of metastasis, our study of FLO-1 and FLO-1 LM also suggest that the metastatic process may endow or select for cells with additional attributes. These include increased proliferative, clonogenic, anti-apoptotic and immune-tolerant potential. These findings provide further insights into why metastatic esophageal cancer is often refractory to conventional treatment. Therapeutic strategies designed to inhibit metastatic processes may thus be of important clinical value [92]. In this way, FLO-1 and FLO-1 LM are ideal platforms for testing novel anti-metastatic agents.
In summary, we have identified and characterized two spontaneously metastatic cell line models of human esophageal cancer. These models suggest that E-cadherin expression is a key inhibitor of metastasis in this disease, and also implicates a complex interplay of many other factors that may modulate this process. We anticipate that these two cell lines will shed new light into the pathogenesis and treatment of metastatic disease in EAC.

Cell lines and culture
OE19, OE33 and HEK-293T cells were purchased from the American Type Culture Collection. FLO-1, Eso26, TE7 and OE21 cells were gifts from Rebecca Fitzgerald (University of Cambridge, UK). JH-EsoAd1 cells were provided by James Eshleman (John Hopkins University, MD). OANC1 cells were established in our laboratory [93], and immortalized human esophageal epithelial cells (NES) were a gift from Rhonda Souza (University of Texas Southwestern Medical Centre, TX). All cells were maintained at 37°C with 5% CO 2 . Unless otherwise stated, all culture media contained 50 U/ml penicillin, 50 mg/ml streptomycin (Life Technologies) and 10% fetal bovine serum (FBS). HEK-293T, FLO-1 and OANC1 cells were grown in Dulbecco's Modified Eagle Medium (DMEM) containing 2.5 mmol/L L-glutamine and 4.5 g/L D-glucose (Life Technologies). Eso26, OE33, JH-EsoAd1, OE19, TE7 and OE21 cells were cultured in Roswell Park Memorial Institute (RPMI) 1640 medium containing 2.5 mmol/L L-glutamine (Life Technologies). NES cells were maintained in modified MCDB-153 medium as previously reported [94]. All cell lines were authenticated by STR analysis using the PowerPlex ® 16 genotyping system (Promega) and confirmed mycoplasma free by PCR (Cerberus Sciences, Australia).

Mice
All animal experiments were performed in accordance with the National Health and Medical Research Council Australian Code of Practice for the Care and Use of Animals for Scientific Purposes and approved by the Peter MacCallum Cancer Centre (PMCC) Animal Experimentation Ethics Committee. Female BALB/c nu/nu (Nude), SCID and NOD-SCID mice were obtained from the Animal Resource Centre (Western Australia). NOD-SCID IL-2Rγ KO (NSG) mice were bred in-house.

Tumor xenografts
To assess tumorigenicity and metastatic potential of esophageal cancer cell lines, 5 million cells suspended in 100 µl of 1:1 phosphate buffered saline (PBS) and matrigel (BD Bioscience) were subcutaneously injected into the flank of female Nude, SCID, NOD-SCID or NSG mice. FLO-1 metastatic deposits from the liver, approximately 2 mm in size, were implanted subcutaneously into NSG mice. Subcutaneous tumor volume was determined weekly with calipers and calculated using the formula (length × width 2 )/2. All mice were euthanised when subcutaneous tumors reached ≥ 1500 mm 3 or at first signs of ill health (labored breathing, bloated abdomen or excessive weight loss: > 10% of baseline body weight).

Cancer cell line isolation and establishment
Macro-metastatic deposits in the liver were dissected away from the surrounding parenchyma and washed in PBS containing 12 µg/ml penicillin and 50 mg/ml streptomycin. Tumor pieces were finely chopped and incubated in Hank's Balance Salt Solution (Thermo Fisher Scientific) containing 6 mg/ml dispase II (Roche) and 3 mg/ml collagenase A (Roche) at 37°C on an orbital shaker for 2 hr. The resultant cell suspension was pelleted by centrifugation (1400 rpm, 4 min) and cultured in DMEM. Cells were passaged using 0.25% trypsin and verified using STR analysis. Formal experimentation with these cells began after 10 passages in culture.

RNA-sequencing (RNAseq)
Cells were harvested at 80% confluency, 24 hr after a media change. Total RNA was extracted using the Qiagen RNeasy Kit. RNA quality was confirmed using a Fragment Analyzer TM (Advanced Analytical Technologies). cDNA library preparation was performed using the NEBNext Ultra directional RNA library Prep Kit (Illumina). DNA sequencing was conducted on the Illumina NextSeq 500 system. Image analysis, base calling and quality checks were undertaken with the Illumina data analysis pipeline RTA v2.4.11 and Bcl2fastq v2.17. Read data was mapped to the reference sequence Homo_sapiens.GRCh37.75 using a short read aligner after trimming for adapter sequences with Trimmomatic v0.30. A default mismatch rate of 2% was used. Read counts per gene were imported into R Statistical Package v3.2.3 (www.r-project.org), analyzed using LIMMA package v3.26.9 and voom transformed. Genes differentially expressed between cell lines were identified using linear regression models. p-values were adjusted using the Benjamini and Hochberg's method. GO process enrichment was computed using MetaCore TM v6.27 for all differentially expressing genes with a FDR p-value < 0.05.

Gene expression with quantitative real-time PCR (qRT-PCR)
Total cell RNA was isolated using the Qiagen RNeasy Kit and reverse transcribed using the Transcriptor First Strand cDNA Synthesis Kit (Roche). SYBR green qRT-PCR was performed on a Lightcycler ® 480 (Roche). Gene expression was normalized against GAPDH and analyzed using the ∆∆C t method. The ∆C t method was used for plotting heatmaps. PCR primers are detailed in Supplementary Table S8.

Western blot analysis
Cells were lyzed and processed as described previously [94]. Protein densitometric analysis was conducted using Image J software (http://imagej.nih.gov/ ij/). The antibodies used are detailed in Supplementary  Table S9.

Cell morphology assessment
Cell morphology was assessed using the AMG EVOS FL (Advanced Microscopy Group) phase-contrast microscope.
Proliferation assay 5000 cells/well were seeded into 96 well plates and allowed to adhere overnight. Cellular proliferation was assayed with AlamarBlue (Life Technologies) reagent and measured using a FLUOstar OPTIMA microplate reader (BMG Labtech) every 24 hr over a 96 hr period as previously reported [94].

Clonogenic survival assay
1000 cells/well were seeded into 6 well plates and cultured for 7 days. Cell colonies were fixed and processed as previously described [94]. Discrete colonies www.impactjournals.com/oncotarget (> 50 cells/colony) were counted using MetaMorph software (Molecular Devices). 25,000 cells/well were seeded into 96 well plates and allowed to adhere overnight to achieve a confluent monolayer. Following wounding with a robot-assisted (Sciclone ALH3000 Workstation, Caliper Life Science) 1.67 mm diameter pin tool (FP3-WP, V&P Scientific), cells were cultured in DMEM containing 1% FBS for the remainder of the assay. At 0 and 40 hr after wounding, cells were fixed with 4% paraformaldehyde (Santa Cruz Biotechnologies), stained with DAPI (Thermo Fisher Scientific), and imaged using a Cellomics ArrayScan VTI HCS reader (ThermoFisher Scientific). The extent of wound closure was quantified using MetaMorph software.

Invasion assay
Following 24 hr of serum starvation, 400,000 cells suspended in 200 µL of 3:2 serum free DMEM and Matrigel™ were placed into the upper compartment of a Transwell (Corning) chamber. 800 µL of DMEM containing 10% FBS was added to the lower compartment. Cells were allowed to invade across Matrigel™ and an 8 µm pore membrane for 24 hr at 37°C and 5% CO 2 . Cells remaining on the membrane's upper surface were removed with a cotton swab, whilst invaded cells on the membrane's under surface were fixed in 4% paraformaldehyde and stained with DAPI. Membranes were mounted and visualized using a BX51 microscope. Five representative fields per membrane were imaged and invaded cells counted using Image J software. 25,000 cells/well were seeded into 24 well plates and cultured for 48 hr. Cells were stained with annexin V-APC antibody (BD Pharmigen) and propidium iodide (Molecular Probes) as previously described [94]. The extent of apoptosis was measured by flow cytometry (BD FACSCanto™ II, BD Bioscience) and analyzed using Flowlogic software (Inivai Technologies).

Luciferase-eGFP ectopic expression
Luciferase cDNA was stably expressed in FLO-1 parental and FLO-1 LM cells using the pFUGW lentiviral vector kindly provided by Dr. Mark Shackleton (PMCC, Australia). In this system, luciferase cDNA has been cloned downstream of the ubiquitin-C promoter and contains eGFP as a reporter gene [95]. Lentiviral particles were produced from HEK-293T cells using Lenti-X (Clontech) packaging mix according to manufacturer's instructions. FLO-1 cells were transduced with pFUGW and FACS sorted (BD FACSAria™ Fusion, BD Bioscience) for eGFP + cells.

Bioluminescence imaging
Luciferase-eGFP + FLO-1 cells suspended in PBS and matrigel were subcutaneously injected into NSG (cohorts of 5 million, 1 million, 100 thousand and 10 thousand cells per mouse) and nude (5 million cells per mouse) mice. Animals were imaged fortnightly on a Xenogen IVIS 100 Imaging System (Caliper Life Science) to detect metastatic onset. 100 µl of 20 mg/ml luciferin (Promega) in PBS was subcutaneously injected into each mouse 5 min before imaging. Imaging was performed under general anesthesia. At experimental endpoint (either tumour volume >1500 mm 3 or signs of ill health), the whole mouse and its organs were imaged to determine the extent and distribution of metastases. The imaging exposure times were 60 s for whole animal and 5 min for organs. The bioluminescence signal was quantified using the 'region of interest' function in Living Image software.

E-cadherin cell sorting
FLO-1 parental cells were lifted using 0.5 mM EDTA/PBS mixture (30 min), filtered through a 40 µm nylon mesh, pelleted by centrifugation (1400 rpm, 4 min) and resuspended in blocking buffer (2% BSA and 2% FBS in PBS) for 1 hr at 4°C. Cells were then incubated in labeling buffer (2% BSA in PBS) with either anti-IgG (Dako) or anti-E-cadherin (EP700Y, Abcam) antibodies for 1 hr at 4°C. After two washes, cells were stained with Alexa Fluor 700 conjugated goat anti-rabbit IgG (Invitrogen) for 1 hr at 4°C in the dark. Cells were then washed thrice before the addition of 2% FluoroGold™ (Sigma-Aldrich). Viable FLO-1 cells were FACS sorted for E-cadherin high and low expressing populations, and separately injected into NSG mice (500,000 cells/mice). Animals were sacrificed once their subcutaneous tumours reached 1500 mm 3 and lungs and livers were harvested for quantification of metastatic burden by PCR.

Assessment of metastatic burden with qPCR
Genomic DNA from whole lungs and livers were extracted using the Qiagen DNA Blood and Tissue Kit. A multiplex TaqMan (Applied Biosystems) qPCR assay was performed on a StepOnePlus RT-PCR system (Applied Biosystems). Metastatic burden was quantified by normalizing the amount of human vimentin against mouse vimentin DNA per organ. PCR primers and probe sequences are detailed in Supplementary Table S10 www.impactjournals.com/oncotarget Microarray dataset analysis CDH1 expression from 64 treatment naive EAC cases were extracted from an Agilent 44K 60-mer oligomicroarray dataset (GSE19417) deposited at the NCBI Gene Expression Omnibus [19]. This cohort consisted of patients recruited from the Bristol Royal Infirmary, Bristol, UK between 1992 and 2000. Clinicopathological and outcome data were correlated with CDH1 expression. Samples were stratified by CDH1 high and low expression, and analyzed using the GEO2R platform from NCBI to identify significantly differentially expressed genes (FDR p-value < 0.05) for GO process enrichment analysis using MetaCore TM .

Statistical analysis
Data were analyzed with Student's t-test or Mann-Whitney U test for parametric and non-parametric variables respectively. ANOVA with Dunnett's posttest was performed for multiple comparisons. Categorical variables were analyzed using the Chi-square test. Survival differences were compared using Kaplan-Meier log-rank analysis. A multivariate Cox regression analysis adjusting for clinicopathological variables (GSE19417: gender, nodal status, tumor location and grade) was also performed to interrogate differences in survival outcomes. Statistical analyses were conducted using R Statistical Packages and Prism 6 (GraphPad) with p < 0.05 considered statistically significant.

Supplementary data
Additional supplementary information are available online including 5 figures and 10 tables. RNAseq data for FLO-1 parental and FLO-1 LM cell lines are available through the NCBI Gene Expression Omnibus public database (GSE88802).

ACKNOWLEDGMENTS
We thank Elizabeth Christie and the Peter MacCallum Cancer Centre FACS, microscopy, functional genomics and animal core facilities for their technical assistance.

CONFLICTS OF INTEREST
No conflicts of interest to disclose FUNDING This work was supported by a National Health and Medical Research Council (NHMRC) of Australia