Cancer-associated fibroblasts predict poor outcome and promote periostin-dependent invasion in oesophageal adenocarcinoma

Interactions between cancer cells and cancer-associated fibroblasts (CAFs) play an important role in tumour development and progression. In this study we investigated the functional role of CAFs in oesophageal adenocarcinoma (EAC). We used immunochemistry to analyse a cohort of 183 EAC patients for CAF markers related to disease mortality. We characterized CAFs and normal oesophageal fibroblasts (NOFs) using western blotting, immunofluorescence and gel contraction. Transwell assays, 3D organotypic culture and xenograft models were used to examine the effects on EAC cell function and to dissect molecular mechanisms regulating invasion. Most EACs (93%) contained CAFs with a myofibroblastic (α-SMA-positive) phenotype, which correlated significantly with poor survival [p = 0.016; HR 7. 1 (1.7–29.4)]. Primary CAFs isolated from EACs have a contractile, myofibroblastic phenotype and promote EAC cell invasion in vitro (Transwell assays, p ≤ 0.05; organotypic culture, p < 0.001) and in vivo (p ≤ 0.05). In vitro, this pro-invasive effect is modulated through the matricellular protein periostin. Periostin is secreted by CAFs and acts as a ligand for EAC cell integrins αvβ3 and αvβ5, promoting activation of the PI3kinase–Akt pathway. In patient samples, periostin expression at the tumour cell–stromal interface correlates with poor overall and disease-free survival. Our study highlights the importance of the tumour stroma in EAC progression. Paracrine interaction between CAF-secreted periostin and EAC-expressed integrins results in PI3 kinase–Akt activation and increased tumour cell invasion. Most EACs contain a myofibroblastic CAF-rich stroma; this may explain the aggressive, highly infiltrative nature of the disease, and suggests that stromal targeting may produce therapeutic benefit in EAC patients.


Introduction
Oesophageal cancer is associated with extremely poor overall survival; 60-70% of patients present with late-stage disease, too advanced for treatment with curative intent, and in those patients who are suitable for multimodal therapy, 5 year survival is only 35% [1]. This is partly due to the mechanically compliant oesophageal anatomy, which allows symptomless tumour expansion [2]. Additionally, tumour invasion and metastasis are facilitated by the absence of an outer serosal layer and the presence of a rich lymphatic plexus. Depth of invasion and nodal metastasis are both prognostic, and improving survival rates will require an increased understanding of the molecular mechanisms regulating local and regional tumour spread.
Although most cancer-related research has focused on tumour cells [3][4][5], accumulating evidence suggests that 'normal' cells within the tumour stroma (including fibroblasts, endothelial and immune cells) play a major role in tumour development and progression [6]. Stromal features are prognostic in many tumour types [7][8][9][10][11], and expression of stromal genes has been shown to be associated with poor outcome in oesophageal cancer [12].
Fibroblasts are probably the most abundant stromal cells in most cancers, and are recognised to play a role in the development and progression of a range of epithelial tumours [13]. Although fibroblasts within cancers comprise a phenotypically heterogeneous population, and the terminology is confusing [peritumour fibroblasts, cancer-associated fibroblasts (CAFs), 'activated' fibroblasts], 'activated' CAFs are commonly described as having a myofibroblastic phenotype; ie a secretory and contractile cell which expresses α-smooth muscle actin (α-SMA) [14,15]. These cells are responsible for the stromal desmoplasia observed in a number of solid cancers [16][17][18] and are associated with poor prognosis in several carcinoma types, including colorectal [10], breast [19], ovarian [20] and head and neck cancers [8]. CAFs regulate a number of tumour-promoting functions, including invasion [21] and angiogenesis [22], and may also affect tumour cell function by remodelling and generating tissue tension [23]. CAFs are predominantly derived from local fibroblasts, but it is clear that other cell types, including stellate cells, pericytes and circulating mesenchymal stem cells, can undergo myofibroblast transdifferentiation [24]. Irrespective of the progenitor cell, TGFβ1 signalling, in conjunction with increased mechanical resistance of the extracellular matrix (ECM), is central to the transdifferentiation process [25,26]. Notably, the stromal signature identified by Saadi and colleagues [12] in oesophageal cancer patients contained a predominance of inflammation and TGFβ-related genes.
To date, little is known regarding the role of the tumour microenvironment, and in particular CAFs, in the development and progression of EAC [27,28]. In this study we investigated the prognostic and functional role of CAFs in EAC and provide new insights into the role of periostin in these processes.

Tissue and cell collection, maintenance and RNAi
Tissue was collected and stored with ethical agreement and informed consent; 09/H0504/66. Fibroblasts were extracted from normal oesophagus and oesophageal adenocarcinoma and subcultured as previously described [29]. siRNA-mediated silencing of periostin was carried out using INTERFERin transfection reagent and two commercially available siRNA sequences (for full details, see supplementary material, Supplementary materials and methods).

Recombinant periostin
Periostin isoform 2 was subcloned and subsequently purified from Hek293F cells (for full details, see supplementary material, Supplementary materials and methods).

Western blotting
Cells (FLO-1 or primary fibroblasts) were treated under different conditions including ± fibroblast conditioned medium, ± recombinant periostin, ± TGFβ1, ± TGFβR1 kinase inhibitor IV (Calbiochem), ± LY294002 (Sigma) and ± αvβ3 and αvβ5 integrin antibodies (R&D Systems). Cells were treated with either TGFβ1 or integrin receptor inhibitors for 1 h before treatment with recombinant proteins (TGFβ1 and periostin) or clarified conditioned medium (for full details of antibodies used, experimental conditions and western blotting, see supplementary material, Supplementary materials and methods).

Organotypic cultures
Organotypic cultures were carried out as previously described [29,32]. NOFs or CAFs were co-cultured with either OE33 or FLO-1 cells for 14 days. Fibroblasts were pretreated with POSTN siRNA sequence 1 (Ambion) before embedding in organotypic cultures, and matured for 10 days. Organotypics treated with 10 μM LY294002 were matured for 14 days. Invasion was quantified and compared as previously described [33].

Transwell invasion
Transwell invasion assays were carried out as previously described [32]: 24 h-conditioned medium from NOFs or CAFs (plated in identical numbers) ± POSTN siRNA pre-treatment, ± TGFβ1 pre-treatment. Addition of 5 μg/ml recombinant POSTN was used as the chemoattractant in the bottom chamber where indicated. Blocking antibodies for αvβ3 and αvβ5 integrins (10 μg/ml; R&D Systems) and PI3K inhibitor LY294002 (10 μM; Sigma) were added to the top chamber in combination with FLO-1 or OE33 cells for the duration of the experiment (72 h). All experiments were repeated three times, with four replicates/experiment. Normalisation was performed to the mean value of replicate 1 for all experimental conditions and data expressed as percentage invasion compared to the appropriate control (ie negative control siRNA-or vehicle-treated cells).

Data acquisition and processing for network analysis
Weighted gene correlation network analysis (WGCNA) [34,35] was applied to microarray datasets generated on Affymetrix HG-U133A GeneChips from two independent studies (for full details of this analysis, see supplementary material, Supplementary materials and methods).

Statistical analysis
Statistical analysis was performed with SPSS ® v 19 (SPSS, Chicago, IL, USA). Survival of patients was plotted using the Kaplan-Meier method and analysed using the log-rank test, with patients censored at last follow-up. Multivariate Cox logistic regression was used to assess the relationship between α-SMA and standard pathological variables after resection. Kruskal-Wallis and Mann-Whitney U-and t-tests were used to compare groups, as appropriate; p < 0.05 was considered statistically significant for all tests (*p < 0.05, **p < 0.01).

Results
High α-SMA expression is common in EAC and correlates with poor overall survival following oesophageal resection We used immunochemistry to examine the expression of α-SMA in the stroma of tumours from 183 patients following oesophageal resection (median follow-up = 4.5 years; 81 cancer-related deaths). Moderate/high expression of α-SMA (α-SMA-positive) was found in the majority of cases (n = 171). In 12 cases an α-SMA-negative stroma was observed, and this group of patients had a significant overall survival advantage [ Figure 1A; mean survival α-SMA negative = 78.66 months (median not yet reached) versus α-SMA-positive = 48.14 months; p = 0.016 (median 39 months; 95% CI 29.2-48.8)]. The α-SMA-negative group was further analysed to exclude the possibility that it was biased towards early-stage disease; this was not the case. The pathological stages according to the AJCC (7th edn) classification of EAC [36] for the patients in this group were: stage 0, one patient; stage IA, three patients; stage IIA, one patient; stage IIB, two patients; stage IIIA, two patients; stage IIIB, one patient; and stage IIIC, two patients. When α-SMA expression was compared against other commonly used indicators of prognosis following surgery (TNM stage [36,37] and resection clearance: R status) in a Cox regression model, α-SMA expression was the most predictive factor analysed: hazard ratio (HR) 7.1 (1.7-29.4); p = 0.007 (Table 1). Full details of the patient demographics are shown in Table S1 (see supplementary material).

Ex vivo analysis of primary oesophageal CAFs reveals a myofibroblast phenotype that can be induced by TGFβ
CAFs were cultured directly from tumour tissue and NOFs from macroscopically and microscopically normal areas of mucosa at the proximal resection margin. NOFs and CAFs were characterised using cell-type specific markers, confirming a fibroblastic phenotype [vimentin-positive, cytokeratin-negative, CD31-negative ( Figure 1B; see also supplementary material, Figure S1A) and desmin-negative (data not shown)]. CAFs expressed high levels of α-SMA, confirming a myofibroblast phenotype ( Figure 1B). In an additional panel of primary fibroblasts, seven of eight CAFs expressed high levels of α-SMA compared with low or absent α-SMA levels in NOFs, in keeping with the immunohistochemical analysis of paraffin-embedded tissues (see supplementary material, Figure S1B). α-SMA stress fibre formation was observed in CAFs but not NOFs ( Figure 1C). In gel contraction assays there was a significant reduction in gel size and weight in the presence of CAFs compared to NOFs (weight, p < 0.01), demonstrating the ability of CAFs to contract an in vitro ECM ( Figure 1D).
Gene ontology analysis of the stromal signature in the progression of pre-invasive to invasive disease in EAC has previously identified TGFβ signalling as a critical pathway in this transition [12]. Treatment of NOFs with TGFβ1 for 72 h resulted in high levels of α-SMA expression, stress fibre formation and increased gel contraction (gel weight NOFs 0.21 g versus NOFs + TGFβ1 0.15 g; p < 0.01), confirming myofibroblast transdifferentiation ( Figure 1E). In addition, exposure of NOFs to conditioned medium from FLO-1 cells for 72 h resulted in a TGFβ1-dependent up-regulation of α-SMA (see supplementary material, Figure S1C).

CAFs promote EAC cell invasion in vitro
To determine whether CAFs secrete factors that promote EAC cell invasion, Transwell invasion assays were performed using EAC cell lines FLO-1 and OE33. Conditioned medium from NOFs or CAFs was used as a CAFs-conditioned medium promoted a > two-fold increase in invasion of both EAC cell lines (p < 0.05; Figure 2A). The proliferation of FLO-1 cells exposed to either NOF-or CAF-conditioned medium was assessed by MTS assay, and no difference was observed (see supplementary material, Figure S2B). Using organotypic cultures as a more physiologically relevant model of tumour cell invasion [29], we further explored the role of EAC CAFs in this process. Invasion in these models was quantified and compared, taking into account the average depth of tumour invasion and the number and area of invading tumour islands [33]. Minimal invasion was seen in NOFs-containing gels, but in the presence of CAFs, EAC cell invasion was significantly increased for both tumour cell lines [mean total area of invasion, arbitrary units (n = 3), FLO-1 + NOFs 2670.3 versus FLO-1 + CAFs 3862.3; p < 0.001: OE33 + NOFs 501 versus OE33 + CAFs 6238.3; p < 0.001: Figure 2B; see also supplementary material, Figure S2A].

CAFs promote tumour growth in vivo
The ability of CAFs to support tumour growth in vivo was assessed using a xenograft model. The cell line FLO-1 was non-tumourigenic in this model and therefore experiments were performed using OE33 cells. Xenografts from OE33 cells injected subcutaneously, either alone (n = 15) or mixed with NOFs (n = 9) or CAFs (n = 9), did not grow uniformly and therefore the mice were culled when the tumour volume had reached 500 mm 3 , and groups were compared by Kaplan-Meier analysis. Mice receiving co-injections of OE33 cells and CAFs were culled significantly earlier (median survival OE33 + CAFs = 100 days versus OE33 + NOFs = 163.5 days; p = 0.04; Figure 2C). Immunostaining for α-SMA confirmed the retained presence of CAFs in these tumours surrounding cytokeratin-positive tumour cells ( Figure 2D).

Periostin is secreted by CAFs and promotes tumour cell invasion
Saadi and colleagues [12] recently described a stromal gene expression signature that is predictive of outcome in EAC. To complement this differential expression study, we applied WGCNA [38] to a publicly available EAC microarray dataset [39]. Seven modules of highly co-expressed genes were identified by unsupervised hierarchical clustering on the basis of topological overlap (TO) and labelled by colour ( Table 2). The co-expression network was strongly preserved in an independent dataset [40]. ToppGene analysis indicated that the pink module was highly enriched for genes associated with fibroblasts (GSM777043; Bonferroni adjusted p = 3.1 × 10 −83 ). Concomitant with this finding, the top Gene Ontology (GO) category for the pink module was 'extracellular matrix organisation' ( Table 2); among the enriched  The Zsummary is a composite statistic of density and connectivity preservation for the module: < 2, no preservation; 2-10, weak to moderate preservation; > 10, strong evidence of preservation.
genes were α-SMA, fibrillin 1 and periostin. Unsupervised hierarchical clustering of the ECM module genes differentiated between normal oesophageal epithelium and EAC ( Figure 3A). These observations are consistent with the ECM module representing genes expressed by a CAFs-rich stroma.
The relative importance of each gene within its respective module can be estimated by a module membership measure (KME) [38]. Periostin had a high KME value (0.87), denoting centrality and importance within the ECM module; a full list of KME values is provided in Table S2 (see supplementary material). Of the top 86 nearest neighbours of periostin (TO threshold ≥ 0.15; Figure 3B), 52 of the genes were annotated for the GO categories ECM organization (adjP = 4.1 × 10 −27 ), biological adhesion (adjP = 7.0 × 10 −17 ) and locomotion (adjP = 3.2 × 10 −8 ).
We therefore examined whether the tumourpromoting effects of CAFs are modulated through periostin. Conditioned medium from CAFs contained higher levels of periostin compared to NOFs by western blot and immunocytochemistry ( Figure 3C). siRNA-silencing of POSTN in CAFs suppressed periostin expression and resulted in a > 50% suppression of tumour cell invasion in Transwell assays (p < 0.05; Figure 3D). POSTN silencing in CAFs also resulted in complete loss of invasion by EAC cells in organotypic culture (mean total area of invasion, arbitrary units; n = 3; control siRNA 1187.7 versus POSTN siRNA 3581; p < 0.001; Figure 3E). To further confirm that the inhibition of tumour cell invasion was directly modulated through reduced periostin expression, we immunodepleted periostin from CAFs-conditioned medium and observed a similar reduction in invasion (p < 0.05). This was restored by addition of recombinant periostin (Figure 3 F). Fibroblast proliferation after periostin silencing was not significantly different to control siRNA-treated cells (data not shown).
We had previously found that TGFβ1 treatment of NOFs generated a myofibroblastic phenotype ( Figure 1D, E). TGFβ1-treated NOFs also up-regulated POSTN ( Figure 3G) and promoted EAC invasion (p < 0.01). Periostin silencing did not affect the up-regulation of α-SMA in TGFβ1-treated NOFs (see supplementary material, Figure S3F). The increase in EAC invasion was again periostin-dependent and abrogated by POSTN down-regulation. Furthermore, the addition of recombinant periostin to NOF-conditioned medium significantly increased EAC cell invasion in Transwell assays to levels comparable to CAF-conditioned medium ( Figure 3H).

Periostin treatment of EAC cells results in integrinand PI3K-dependent activation of Akt
We examined potential periostin-dependent downstream signalling pathways in EAC cells and found Akt to be consistently phosphorylated following periostin treatment. PI3K-Akt signalling has been reported to be a critical regulator of cell motility in several cancer types, including EAC, and we explored this response in more detail. Recombinant periostin was added to tumour cells and phosphorylation of Akt was observed at 30 min post-stimulation, which persisted for at least 120 min ( Figure 4A). Periostin is a ligand for several integrins, including αvβ3 and αvβ5, and so we examined whether periostin-dependent Akt activation is modulated through integrin binding. After confirming the presence of αvβ3 and αvβ5 on the surface of EAC cells by western blot and flow cytometry (see supplementary material, Figure S3A, B), we found that phosphorylation of Akt in response to periostin was suppressed by blocking antibodies directed against αvβ3 or αvβ5 ( Figure 4B). The suppression of Akt phosphorylation in periostin-treated cells, in the presence of the PI3K inhibitor LY294002 confirmed that Akt phosphorylation was PI3K-dependent ( Figure 4C). Similar effects on Akt activation were observed in FLO-1 cells after 30 min treatment with CAFs-conditioned medium (see supplementary material, Figure S3E).

Periostin-dependent invasion is suppressed through inhibition of αvβ3, αvβ5 and PI3K
We repeated Transwell invasion assays to examine the effect of inhibiting αvβ3/αvβ5-PI3K-Akt signalling in EAC cells. NOF-conditioned medium supplemented with recombinant POSTN or CAF-conditioned medium promoted EAC invasion (>3.5-fold; Figure 4D) and this was inhibited by the addition of αvβ3-or αvβ5-blocking antibodies, or a PI3K inhibitor, LY294002, to the tumour cells. Invasion of EAC cells in organotypic culture containing CAFs was significantly reduced in the presence of LY294002 (mean total area of invasion, arbitrary units (n = 3), FLO-1 + CAFs 11961 versus FLO-1 + CAFs + LY294002 338.7; p < 0.001; Figure 4E). Proliferation of FLO-1 cells was not affected by treatment with LY294002 in organotypic assays, shown by Ki67 staining (Figure 4E), percentage of Ki67-positive cells in 10 high-power fields (vehicle = 87 ± 5, LY294002 = 89 ± 3). This was confirmed by MTS assays across a range of concentrations (see supplementary material, Figure S3C). An increase in caspase-8 expression (not cleavage) was observed (see supplementary material, Figure S3D). To confirm Akt activation in human EAC tumours, we performed immunohistochemistry (IHC) and observed pAkt in tumour cells, with the strongest staining adjacent to CAFs (Figure 4 F). High stromal periostin expression was observed in the mouse xenograft model containing OE33 and CAFs, and activation of Akt was confirmed in this model by high pAkt staining in adjacent tumour cells ( Figure 4G).

Periostin expression predicts survival after resection for EAC
Having established the importance of periostin to CAFs-induced EAC cell invasion in vitro, we examined our EAC patient cohort to confirm the presence of stromal periostin by IHC. Similar to α-SMA expression, moderate/high expression of periostin was observed in the majority of cases (n = 166; periostin-positive), with periostin localised to the cancer cell-stromal interface ( Figure 5C, D), corresponding to areas of high α-SMA expression ( Figure 5D) in keeping with a myofibroblastic CAFs origin. Periostin expression was limited to blood vessel walls in normal oesophageal mucosa ( Figure 5C).
In 17 cases there was no/minimal periostin staining (periostin-negative) and this group of patients had a significant overall survival advantage when compared with periostin-positive patients [ Figure 5A; mean survival periostin-negative = 76.45 months (median not yet reached) versus periostin moderate/high = 46.80 months

Discussion
The traditional model of EAC development describes progressive cellular dysplasia and accumulation of mutations in response to persistent acid reflux and inflammation [41]. Whole-genome and exome-sequencing studies are beginning to define the complexity of EAC, but suggest that the mutational burden may already be established in non-dysplastic BE [42][43][44], meaning that the drivers for disease progression lie beyond gene mutations in the epithelial cells and that the stromal tissues play a critical role in disease development and progression.
However, very little is known about the role of the tumour microenvironment/stroma in oesophageal cancer [27,28]. CAFs are a heterogeneous cell type, and no single marker reliably identifies all CAFs. They are most commonly identified by α-SMA expression, indicative of an 'activated' myofibroblast-like phenotype. We found that the vast majority of patients with EAC have tumours with high or moderate levels of stromal α-SMA; this predicts poor survival and may account, in part, for the aggressive nature of the disease. Retrospective studies have intrinsic limitations and the uneven distribution of SMA expression in our cohort raises the possibility of a type 1 error; however, this may reflect the importance of the CAFs phenotype in EAC, which is clearly supported by the in vitro and in vivo data. EAC CAFs cultured from primary tumours have an α-SMA-positive, contractile, myofibroblastic phenotype; they express a range of mesenchymal and fibroblast-specific markers, and promote EAC invasion in vitro and tumour growth in vivo. We provide evidence of a potential mechanism by showing that the invasion-promoting effect of CAFs is modulated, in part, through periostin, a secreted matricellular protein that plays an important role in tissue remodelling and collagen fibrillogenesis, interacting with ECM proteins such as fibronectin, tenascin C and collagen V, and with cell surface receptors, most notably integrins, including αvβ3, αvβ5 and α6β4 [45][46][47]. We found that CAFs express high periostin levels in vitro and in vivo, and activate tumour cell PI3K-Akt signalling through integrin binding. Similar to α-SMA, periostin expression also predicts poor outcome in EAC patients.
A stromal gene-expression signature containing periostin has previously been shown to predict outcome in EAC [12]. However, most cancer studies have found periostin to be of tumour cell origin, with over-expression documented in a range of tumours, including lung, breast, ovary and pancreas [48]. Its functional role in cancer is not clear, although expression correlates with poor outcome in several cancers, suggesting a tumour-promoting effect, and, consistent with this, periostin has been reported to modulate a number of the hallmarks of malignancy [48]. Metastasising breast cancer cells induce periostin secretion in the cancer stem cell niche and require continued stromal periostin expression for cancer stem cell maintenance [49]. Importantly, disruption of the interaction between periostin and the integrin receptors αvβ3 and αvβ5 by periostin-binding DNA aptamers blocked signal transduction and reduced primary tumour growth and metastasis in an orthotopic breast cancer mouse model [50], suggesting that targeting periostin has therapeutic potential. Periostin is involved in remodelling of the ECM to support tumour development, invasion and metastasis [48]. Periostin has also been identified as a key component of an invasive signature in a three-dimensional (3D) organotypic model of oesophageal squamous cell carcinoma [51]. We have confirmed the importance of periostin in EAC biology. Applying WGCNA to publicly available EAC microarray datasets has identified periostin to have a 'nodal' position within ECM gene expression in EAC, with close associations with genes involved in cancer cell invasion, adhesion and locomotion.
Periostin contains a FAS1 domain that allows binding of αv-integrins and glycosaminoglycans in vivo. Interaction with these cellular receptors can lead to diverse downstream signalling effects in a context-dependent manner (NF-κB, STAT3, PI3K-Akt and FAK), controlling the expression of many genes (including α-SMA, collagen and fibronectin) [52]. We found that periostin treatment of EAC cells results in integrin-and PI3K-dependent activation of Akt, and furthermore that periostin-dependent EAC cell invasion is suppressed through inhibition of αvβ3, αvβ5 and PI3K. These findings suggest not only that periostin is very likely to be important for EAC cell invasion in vivo, but also that this is mediated by integrin-PI3K-Akt signalling. The PI3K signalling pathway integrates intra-and extracellular signals to modulate a multitude of downstream responses, effected by multiple substrates [53]. PI3K activation in response to integrin signalling is well documented in cancer, and there are some reports of Akt activation leading to tumour cell invasion in response to integrin-mediated periostin signalling in epithelial tumours, eg integrin α5β1 in cholangiocarcinoma [47]. Canonical PI3K-Akt signalling is frequently dysregulated in cancer and can determine cell growth, proliferation, angiogenesis, migration and invasion [54]. The fact that periostin has been implicated in many of the same processes suggests that the PI3K-Akt pathway may act as a node for periostin signalling in cancer. Our findings suggest that in EAC it may be possible to target periostin-intergrin-PI3K-Akt signalling at multiple levels with new and existing therapeutics. Further work is needed to confirm this hypothesis. For instance, it will be important to know whether periostin is prognostic in pretreatment biopsies from EAC, and whether or not the presence of periostin predicts response to neoadjuvant chemotherapy.
In conclusion, our study highlights the importance of the tumour stroma in EAC progression. The majority of EACs contain a prominent myofibroblastic (α-SMA-positive) CAFs-rich microenvironment, and stromal α-SMA expression is more significantly prognostic than conventional histopathological criteria. CAFs isolated from EAC have a functional myofibroblastic phenotype, and promote tumour cell invasion in vitro and growth in vivo, signalling to EAC cells via secretion of the ECM protein, periostin. Disruption of periostin signalling via integrin receptors and the PI3K-Akt pathway leads to abrogation of cancer cell invasion. This opens the possibility of targeting CAFs and their paracrine signals as treatment strategies in oesophageal cancer. design, acquisition of data, analysis and interpretation of data and statistical analysis; ST carried out the microarray data analysis; TJU, JNP and GT obtained funding for the work and were responsible for administrative, technical and material support and study supervision; TJU, AH and GT drafted the manuscript; and all authors undertook critical revision of the manuscript for important intellectual content. TJU and GT take overall responsibility for the scientific content of the manuscript.

SUPPLEMENTARY MATERIAL ON THE INTERNET
The following supplementary material may be found in the online version of this article: Supplementary materials and methods Table S1. Clinicopathological characteristics of patient cohort for immunohistochemistry Table S2. KME values for all of the proteins clustered in the ECM module