Genetic deletion of osteopontin in TRAMP mice skews prostate carcinogenesis from adenocarcinoma to aggressive human-like neuroendocrine cancers

Osteopontin (OPN) is a secreted glycoprotein, that belongs to the non-structural extracellular matrix (ECM), and its over expression in human prostate cancer has been associated with disease progression, androgen independence and metastatic ability. Nevertheless, the pathophysiology of OPN in prostate tumorigenesis has never been studied. We crossed TRansgenic Adenocarcinoma of the Mouse Prostate (TRAMP) mice with OPN deficient (OPN−/−) mice and followed tumor onset and progression in these double mutants. Ultrasound examination detected the early onset of a rapidly growing, homogeneous and spherical tumor in about 60% of OPN−/− TRAMP mice. Such neoplasms seldom occurred in parental TRAMP mice otherwise prone to adenocarcinomas and were characterized for being androgen receptor negative, highly proliferative and endowed with neuroendocrine (NE) features. Gene expression profiling showed up-regulation of genes involved in tumor progression, cell cycle and neuronal differentiation in OPN-deficient versus wild type TRAMP tumors. Down-regulated genes included key genes of TGFa pathway, including SMAD3 and Filamin, which were confirmed at the protein level. Furthermore, NE genes and particularly those characterizing early prostatic lesions of OPN-deficient mice were found to correlate with those of human prostate NE tumours. These data underscore a novel role of OPN in the early stages of prostate cancer growth, protecting against the development of aggressive NE tumors.


INTRODUCTION
Prostate carcinoma (PCa) is one of the most common tumors in developed countries and the second leading cause of cancer-related death in the USA [1]. PCa is a multifocal disease, which progresses from high-grade prostate intraepithelial neoplasia (PIN). Initially, the tumor responds to therapies that mainly consist in androgen ablation, but it eventually becomes androgen resistant due to alterations in androgen receptor expression or related signalling. The main drivers of PCa development and progression are still unknown and effective therapies for the androgen-resistant stage are not yet available.
Stromal microenvironment modifications occurring at the stage of precancerous PIN lesions are no longer considered bystanders. In this context, SIBLINGs (Small Integrin-Binding Ligand N-Linked Glycoproteins), a family of non-structural extracellular matrix (ECM) proteins, seem to favor tumor progression by hijacking their physiological functions in wound healing and tissue remodeling [2].
Osteopontin (OPN) is a secreted protein, belonging to SIBLINGs, present in body fluids and almost ubiquitous in tissues. OPN contributes to tissue homeostasis through the regulation of stem cell pools within specific niches [3]. OPN over-expression in human PCa has been associated with disease progression towards androgen independence and metastasis [4,5]. Therefore, at the late stages of PCa OPN targeted inactivation might be a promising strategy. However, the role of OPN at onset of PCa has not been investigated so far. To address this issue, we crossed OPN deficient (OPN −/− ) mice with the TRansgenic Adenocarcinoma of the Mouse Prostate (TRAMP) mice, which carry the SV40 large T antigen (Tag) under the control of the rat probasin regulatory element, selectively activated by androgens in the prostate epithelia. TRAMP mice invariantly develop PIN (8)(9)(10)(11)(12)(13)(14)(15)(16) weeks) that progresses to focal invasive adenocarcinoma (16-20 weeks) and then to moderately differentiated (24 weeks) and poorly differentiated (24-30 weeks) adenocarcinoma [6,7].
Although the reported over-expression of OPN in advanced disease let envisage tumor reduction in OPN −/− TRAMP mice, we unexpectedly observed the early onset of highly proliferative, anaplastic androgenindependent tumors, characterized by a neuroendocrine (NE) phenotype. The corresponding human NE carcinoma characterized by aggressive behaviour, poor response to standard treatments and dismal prognosis [8,9] accounts for nearly 10% of cases and seldom occurs in relatively young patients. Notably, areas of NE differentiation occur in advanced PCa patients after androgen ablation therapy [10,11]. Five to ten percent of TRAMP mice spontaneously develop tumors with the NE phenotype. The mechanisms driving their onset in OPN deficiency are here investigated [12].

OPN deletion enhances anaplastic tumor development in TRAMP mice
We followed tumor onset and progression in OPN −/− TRAMP (n = 14) and TRAMP (n = 16) mice between 12 and 30 weeks (wks) of age by ultrasound (US) echographic examination. US allows evaluating prostate morphology and dimensions, and the elastomeric response to the echographic probe pressure, highlighting density alterations. The scan performed in 3D mode allows inspection of every prostatic lobe for the presence of tumor before it becomes palpable. US and histologic analyses of the prostate revealed no differences between the two strains until mice were 15 wks old. Between 18 and 22 wks of age a large, well-defined spheroid tumor appeared in OPN −/− TRAMP ( Figure 1A and 1B) but not in age-matched TRAMP mice ( Figure 1A). These spheroid tumors were even larger than the heterogeneous and multifocal tumors, also infiltrating the seminal vesicles, that characterize 30 wks old TRAMP mice ( Figure 1A). Such growth differences were confirmed by weighting the genitourinary organs after necropsy at 30 wks of age or following mice survival ( Figure 1C).
Histopathological analysis was performed on cohorts of mice sacrificed at different time points (18 to 20 wks, 20 to 30 wks, > 30 wks). Anaplastic lesions ( Figure  1D, right) were present in roughly 60% of OPN −/− TRAMP mice in all cohorts, but only in few TRAMP mice and only after the 20th week of age (Table 1). Conversely, multifocal adenocarcinoma lesions ( Figure 1D, left) were present in the majority of TRAMP mice at all time points ( Table 1). As expected, positivity for OPN was found in both epithelial and stromal cells in TRAMP prostates ( Figure 1E).
To compare the anaplastic lesions from the two strains, tumors were collected at 30 wks of age for histopathology and immunohistochemistry. At this age, both strains develop poorly differentiated lesions. However, tumors formed in OPN −/− TRAMP mice were characterized by marked anaplasia and higher expression of Ki67, N-cadherin and laminin ( Figure 2).
These data suggest that the lack of OPN at the beginning of the transformation process skews the nascent tumor towards a more aggressive, undifferentiated phenotype.

Absence of OPN promotes androgenindependent tumor growth in the early stages of prostate carcinogenesis
Androgen receptor was expressed by normal and transformed glandular prostatic tissue of TRAMP mice ( Figure 3A) as well as by normal epithelial cells of OPN −/− TRAMP prostates, but not by tumors of OPN −/− TRAMP mice ( Figure 3B). This finding suggested that in OPN −/− TRAMP mice tumors could rise as androgen independent.
Androgen ablation is a standard treatment for patients with advanced PCa (www.cancer.gov/types/ prostate/hp/prostate-treatment-pdq#section/all). Unfortunately these patients frequently develop androgen resistance and ineffective treatment options remain. This setting can be modeled in TRAMP mice by castration [13]. We investigated whether the lack of OPN affects mice susceptibility to hormone withdrawal in early phases of tumor development. TRAMP and OPN −/− TRAMP mice (both n = 11) were castrated before sexual maturation (6 wks) and tumor growth was monitored by US examination. Among castrated TRAMP mice, 9 showed prostate involution, 1 developed adenocarcinoma and the remaining one developed an anaplastic spheroid tumor with incidence similar to that of non-castrated TRAMP mice (12%; Figure 3C). Castrated OPN −/− TRAMP mice did not develop any evident adenocarcinoma, whereas 8 showed anaplastic tumors (73%, versus 60% of non-castrated littermates). Prostate involution occurred in the remaining 3 (27%) castrated mice ( Figure 3C). Castration after sexual maturation (12 wks) did not change the    incidence of anaplastic tumors in OPN −/− TRAMP mice, but increased up to 25% the rate of anaplastic tumors in TRAMP mice ( Figure 3C) as expected [13,14].
In TRAMP mice the expression of Tag is controlled by the rat Probasin promoter [6] which is regulated by androgens [15]. That means that Tag expression in TRAMP mice reflects androgen responsiveness of epithelial prostate cells. Immunohistochemistry revealed Tag expression by luminal cells of intact tumorigenic prostate glands of TRAMP mice ( Figure 3D) but not by cells of involute prostates ( Figure 3D). As already seen by others [13,16], Tag expression was maintained in cells of anaplastic lesions occurring in castrated TRAMP mice ( Figure 3F). Interestingly, Tag was expressed at high levels on anaplastic tumors of OPN −/− TRAMP mice regardless castration ( Figure 3G-3I). These results confirmed that the anaplastic tumors arising in OPN −/− TRAMP mice developed as androgen-independent.
Synaptophysin staining confirmed the NE phenotype of anaplastic tumors [12] developed in both non-castrated ( Figure 3L) and castrated ( Figure 3M) OPN −/− TRAMP mice. Synaptophysin did not stain the epithelia of OPN −/− TRAMP prostates regressed because of castration ( Figure 3N). Therefore, thereafter in this manuscript tumors from OPN −/− TRAMP mice will be referred as NE tumors.
To test the possible OPN source either epithelial or hematopoietic that is relevant in this experimental model, bone marrow (BM) transplantation was performed to obtain chimeras lacking or not OPN in the prostatic epithelium or in the radio-resistant stroma cell components. In detail, BM from OPN-competent wild-type (WT) mice was transplanted into 6 wks old OPN −/− TRAMP mice as well as BM, either from OPN −/− or WT mice, was transplanted into TRAMP mice. Mice were then sacrificed at 30 wks of age for histopathologic analysis of prostates. NE tumor frequency remained determined by the host genotype regardless that of donor BM, with high incidence in OPN −/− TRAMP and low incidence in TRAMP recipient chimeras ( Figure S1). These results suggest that OPN produced by the host radio-resistant cell component contrast the occurrence of NE tumors.

Basal origin of tumors in OPN −/− TRAMP mice
Whether prostatic NE cancers have a distinct cellular origin from adenocarcinomas or rather they originate from the peculiar differentiation of common precursors is still debated. As far as the TRAMP model is concerned, it is believed that NE tumors can arise from a subset of stem/ progenitor cells populating the basal cell layer of glandular ducts [12,17,18]. We thus tested the expression of the basal p63, [19][20][21] and the luminal CK8 [22] markers in association with the cell proliferation marker Ki67.
In prostate lobes of 15 wks old TRAMP mice, clusters of proliferating cells co-expressed Ki67 and CK8 ( Figure 4A) whereas p63 positive cells were Ki67 negative ( Figure 4C). On the contrary, in prostates of OPN −/− TRAMP mice luminal CK8 cells mostly lacked Ki-67 ( Figure 4B), which instead co-localized with p63 in the basal compartment ( Figure 4D). Notably, according to previous reports on prostate cancer patients [21], some cytoplasmic staining of p63 was detectable ( Figure 4C and 4D) whereas nuclear p63 localization was canonic in normal prostate of wild-type mice ( Figure S2).

Gene expression profile confirms the NE phenotype of OPN −/− TRAMP tumors
To obtain a broad view of the transcriptional features distinctive of NE and adenocarcinoma lesions, we compared gene expression profiles of prostate samples collected from 18 weeks old OPN −/− TRAMP mice bearing macroscopic NE lesions (OPT-tm) with those from 30 wks old TRAMP mice (T30) representative of adenocarcinomas (ADC). A detailed description of samples used from microarray comparison is reported in Suppl. Table 1, and a detailed flow of samples comparison is shown in Suppl. Figure 3 Unsupervised hierarchical clustering clearly distinguished gene expression profiles of TRAMP mice with adenocarcinoma from OPN −/− TRAMP mice with NE tumor ( Figure S4). We found 2536 differentially expressed genes between the two classes (absolute fold change ≥ 2 and FDR < 5%; Heatmap in Figure 5A). To identify biological processes distinguishing the two histotypes we performed Gene Set Enrichment Analysis (GSEA) using gene sets compiled from canonical pathway and Gene Ontology databases. Significant gene sets (FDR < 5%, p-value < 0.001) were visualized as interaction networks using the Enrichment map plugin in Cytoscape ( Figure S5). NE tumors were characterized by the activation of gene sets involved in cell cycle and proliferation, transcription and RNA-processing, and neural system. These findings were confirmed by functional analysis of the differentially expressed genes using Ingenuity Pathway Analysis (IPA) software. In OPT-tm, among genes involved in cell cycle, E2f2 and Ccne2 were the most up-regulated, while Ccnd1, whose up-regulation has been reported in TRAMP mice [23,24], was the most down-regulated. These results were confirmed by RT-qPCR analysis ( Figure 5B) and were concordant with the high cellular proliferative rate observed both macroscopically and histologically in OPN −/− TRAMP tumors. The most up-regulated genes in OPT-tm samples were those involved in neurological diseases and in neuronal system development, including L-Dopa Decarboxylase (Ddc, fold change = 172.2) and synaptophisin (Syp, fold change = 31.3), confirming the NE phenotype of anaplastic tumors arising in OPN −/− TRAMP mice. Upregulation at protein level of DDC and SYP in NE tumors was validated by western blot ( Figure 5C).
IPA predicted Tgfb1 to be an uprstream regulator of the differentially expressed genes in OPT-tm samples (p = 8.93e-37). Indeed 157 targets of Tgfb1 were listed as differentially expressed genes and 84 of them had an expression consistent with Tgfb1 pathway inhibition. This finding was supported by the significant downregulation of Tgfb1 in OPT-tm (FDR=0.005), although this gene was excluded from the list of differentially expressed genes because of a fold-change of -1.53. Moreover, Filamin A (FLNA), a key TGFβ activator [25], was among the down-regulated genes in our signature and western blot confirmed FLNA protein down regulation ( Figure 5C).
Tgfbr2 and Tgfbr3 transcripts were significantly down-regulated in OPT-tm, whereas Tgfbr1, which is responsible for SMADs phosphorylation and activation, was up-regulated. As protein levels of TGFbRI and TGFbRII were comparable between OPN −/− TRAMP NE tumors and TRAMP adenocarcinoma (not shown), we focused on proteins downstream the TGFβ pathway, particularly Smad2 and Smad3, which are crucial mediators of TGFβ activity in PCa [26]. Smad3 was down-regulated in GEP and in western blot in NE tumors from OPN −/− TRAMP mice ( Figure 5D). On the contrary, Smad2 transcript and protein, as well as phosphorylation of SMAD3 and SMAD, remained unchanged in OPN −/− TRAMP NE tumors versus TRAMP adenocarcinomas ( Figure 5D). These results suggested that in absence of OPN the TGFβ pathway and the cell cycle were down-regulated, likely through the activity of FLNA and SMAD3.

human NE prostate and can predict NE tumor outcome
We have the idea to test the possibility of defining a gene signature that could predict the onset of NE disease in PCa patients. Using the 2536 genes distinguishing NE from adenocarcinoma prostate tumors, we applied unsupervised hierarchical clustering to gene expression data from prostates of TRAMP mice of 18 wks of age (T18, n = 4) and from OPN −/− TRAMP mice of 15 (n = 5) or 18 (n = 9) wks of age (OPT15 and OPT18, respectively) all devoid of US-detectable lesions (Table S1 and Figure S3 describe samples and class comparisons). We found that early prostate samples fell into two distinct clusters (Figure 6a). The first cluster was composed of 5 samples, hereafter called Early-NE, whose gene expression pattern resembled that of NE OPT-tm. These samples were all from OPN −/− TRAMP mice (4 of 18 wks old and 1 of 15 wks), indicating that at the time of sacrifice they were likely bearing a developing NE tumor, well before any possible diagnosis. The second cluster contained 17 samples, hereafter called Early-ADC, whose gene expression profile was similar to that of TRAMP adenocarcinomas. All early samples from TRAMP (T18) and from OPN −/− TRAMP mice, without NE features, fell into this cluster. To identify early markers specific for NE tumor a comparison between Early-NE and Early-ADC GEPs was performed. We found that all the 184 differentially expressed genes (absolute fold change ≥ 2 and FDR < 5%), were upregulated in the Early-NE group ( Figure 6A). According to GSEA these genes are involved in cell cycle and nervous system development, but also in down-regulation of pro-inflammatory processes ( Figure  6B, Table S2 and S3). This list of 184 genes provides a transcriptional profile distinguishing NE tumors from adenocarcinoma at early stages of transformation.
We were interested in asking whether the transcriptional alterations observed in our mouse models could find correlation with human tumors counterpart. As gene expression datasets containing both NE and adenocarcinoma human PCa are not publicly available we used GSEA to test whether genes previously identified as up-or down-regulated in human prostate NE tumors versus prostate adenocarcinomas [27] were positively or negatively enriched in mouse NE lesions. Results showed that genes over-expressed in NE human tumors were also positively enriched in both OPT-tm and Early-NE samples ( Figure 6C, left, and Figure 6E, left, respectively). The core enrichment genes, accounting for the gene set's enrichment, were related to cancer, neuronal development, cell growth and cell cycle progression ( Figure 6D and 6F, and Table S4 and S5), confirming the results obtained with the murine NE signatures. Conversely, genes downregulated in NE patients were significantly negatively enriched in both OPT-tm and Early-NE samples while up-regulated in T30 (late adenocarcinoma) and Early-ADC samples ( Figure 6C, right, and Figure 6E, right, respectively). Taken together these results showed that gene expression profiles of tumors arising in OPN −/− TRAMP mice were similar to those of human NE prostatic cancers.

DISCUSSION
ECM confers proper architecture and function to tissues. It regulates connections and communication between cells and is responsible for the continued tissue remodeling in development or healing. Of great interest is the role of ECM in cancer, representing a tissue that does not heal [28], especially in premalignant to malignant transition [29].
OPN is an ECM non-structural protein, member of the SIBLINGs, endowed with pro-migratory and proangiogenic properties in transformed tissues [30] and considered a potential tumor marker [31]. In PCa its expression in primary tumors [4] and bone metastases [32] correlates with poor prognosis and its targeting has been considered a therapeutic option [33].
Despite the numerous studies, none has investigated the role of OPN at onset of PCa. The present work is the first to study the effect of OPN genetic deficiency in TRAMP, a mouse model mimicking human PCa. Considering its high expression in advanced human PCa we would have expected that OPN knockout in TRAMP mice would have delayed tumor onset or reduced tumor size. To our surprise, OPN −/− TRAMP mice displayed early onset and accelerated tumor development in comparison to TRAMP mice. These tumors developed as spherical masses, macroscopically detectable at 18 wks of age by echographic examination. Histopathology classified these tumors as undifferentiated and immunohistochemistry indicated the loss of AR, the retention of Tag and the gaining of synaptophysin. The same characteristics have also been described for NE tumors that seldom develop spontaneously in TRAMP mice unless stimulated by castration [12,18]. Castration in TRAMP mice has different outcomes depending on the age of mice at surgery. Before sexual maturation it causes prostate involution and prevents neoplastic transformation [13] whereas at late stages it promotes the development of androgen independent NE tumors; which express high levels of Tag [13,16,34]. In OPN −/− TRAMP mice castration at either 6 or 12 weeks had no effect on NE tumor development, which still occurred at high rate in an hormone independent way.
We also found that proliferating Ki67 positive cells in prostates of OPN −/− TRAMP mice express cytoplasmic p63, thus confirming that the basal prostate compartment is transforming in OPN −/− TRAMP mice. This is in line with the hypothesis that NE tumors arise www.impactjournals.com/oncotarget from basal stem-like cells in the TRAMP model [12]. The cytoplasmic staining of p63 ( Figure 4D) is in agreement with human PCa, carrying a translocation of p63 from the nucleus to the cytoplasm, an event that has been correlated with increased cell proliferation and dismal prognosis [21].
Clustering analysis performed on gene lists originated by prostate samples from OPN −/− TRAMP of 18 wks with US detectable NE tumors and by samples from 30 wks TRAMP with adenocarcinoma, proved the distinct nature of tumors developing early in mice deficient for OPN.
In silico analysis showed that genes promoting cell cycle were most up-regulated in OPN −/− TRAMP than TRAMP tumors. Among them CyclinE2 and E2f2 are inducible by estrogens [35], indicating their possible involvement in the development of androgen independent NE tumor [36]. CyclinD1, which promotes cell cycle progression in TRAMP adenocarcinomas [37][38][39], was down-regulated in NE tumors from OPN −/− TRAMP mice. These data are in line with the observation that up to 90% of human NE xenografts show loss of Cyclin D1 [40]. The similarity between murine and human NE tumors was corroborated by the concordant up-regulation of DDC, a biomarker for human NE PCa [41,42].
The key genes down-regulated in our signature belong to the TGFβ pathway and encode for cytoskeletal proteins. TGFβ mediates the activity of OPN in maintaining hematopoietic stem cells quiescence in the mesenchymal BM niche [43]. Not surprisingly, it might also maintain dormant the prostatic stem cell compartment [26]. The lack of OPN in the prostate stem cell niche may inactivate the TGFβ pathway and therefore favor the escape from its anti-proliferative control, towards NE tumors development.
In keratinocytes Filamin A (FLNA), a cytoskeletal SMAD-binding protein, binds SMAD2 for effective TFGβ receptor signaling; accordingly, FLNA-deficient melanoma cells have impaired TGFβ signaling [25]. In OPN −/− TRAMP NE tumors, FLNA was down-regulated at gene and protein levels. SMAD3 gene and protein were also down-regulated although the residual protein was still phosphorylated. SMAD3 is a key intermediated in the TGFβ pathway [44] and its depletion cannot be overcome by SMAD2 in sustaining TGFβ-mediated cell cycle arrest and growth inhibition [45]. Accordingly, a reduced expression of SMAD3 diminishes the tumorsuppressor function of the TGFβ pathway in a model of acute T-cell lymphoblastic leukemia [46]. Overall, our data support the conclusion that deregulation of TGFβ signaling is responsible for increased tumor proliferation and aggressiveness in the OPN −/− TRAMP model.
Nested class comparison between Early-NE vs Early-ADC tissues identified 184 genes that were all upregulated in prostates from OPN-/-TRAMP at the onset of the still undetectable NE tumors, and that might contain possible biomarkers for early diagnosis.
The validation of OPN −/− TRAMP as a reliable model for human NE prostate cancer came from the GSEA analysis on Beltran dataset of PCa patients [27]. Indeed, genes that are up-regulated in human NE prostate cancers were also enriched in our data set from OPN −/− TRAMP mice. Common were the genes belonging to canonical pathway related to cancer, neuronal development, cell growth and cell cycle progression.
Modification of the genetic background has been found to change the frequency of NE tumors in TRAMP mice [47], suggesting that polymorphic germline modifier genes may favor NE tumors in this model. Here, using the same C57BL/6 genetic background, we confined to OPN the genetic difference, underscoring a role of OPN in controlling prostate NE tumor outcome. Although this study is the first to demonstrate the role of OPN in prostate tumorigenesis, another SIBLINGs member, namely SPARC, has been shown to favor the occurrence of more aggressive cancers, of undefined histotype in TRAMP mice, if deleted [48]. Altogether these data support the idea that under the same oncogenic driver (SV40 large T antigen), the ECM composition can influence the phenotype and the aggressiveness of the upcoming tumor. Conversely, in advanced PCa patients both OPN and SPARC overexpression are associated with poor prognosis and metastatic phenotype [4,32,49], implying that the same ECM proteins can exert different functions at different tumor stages.

Immunofluorescence
Sections were incubated with primary antibody anti Ki67 o/n at 4°C, then washed in PBS and incubated for 1h with an Alexa-488 anti-rabbit antibody. Slides were then incubated with anti CK8 or anti p63 antibodies (Abcam) for 1h, washed and further incubated with Alexa-568 anti-mouse antibody. Sections were counter stained with DAPI and analyzed with a Microradiance 2000 (Bio-Rad Laboratories, Segrate, Italy) confocal microscope equipped with Ar (488 nm), HeNe (543 nm) and red laser diode (638 nm) lasers. Confocal images (512 × 512 pixels) were obtained using a x200, 0.5 NA Plan Fluor DIC or x600 1.4 NA oil immersion lens and analyzed using ImagePro 7.0.1 software.

RT-qPCR
Prostates were lysed with TRIzol (Invitrogen, Life Technologies, Monza, Italy) and RNA was purified by phenol/chloroform extraction followed by loading onto RNeasy MINI kit (Qiagen, Milan, Italy). On-column DNAse treatment was performed. RNA purity and yield was assessed using NanoDrop ND-100 Spectrophotometer (NanoDrop Technologies, Wilmington, DE).

Gene expression profiling
We collected prostates from 18 wks old TRAMP mice (T18, n = 4), with null or minimal disease, from 30 wks old TRAMP mice (T30, n = 4) in which adenocarcinoma was palpable and detectable by US, from OPN −/− TRAMP mice of 15 or 18 wks of age (OPT15, n = 5 and OPT18, n = 9, respectively) without any detectable lesion and from 18 wks old OPN −/− TRAMP mice with detectable NE tumor mass (OPT18tm, n = 3). A detailed description of samples used for microarray comparison is in Suppl. Table 1, and a detailed flow chart depicting how samples were compared is in Suppl. Figure 3 RNA was isolated using the guanidinium thiocyanate/cesium chloride gradient method [53]. RNA concentrations were measured with the NanoDrop ND-100 Spectrophotometer (NanoDrop Technologies) while RNA quality was assessed with the Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA, USA) using the RNA 6000 Nano kit (Agilent Technologies). The mean RIN value was 8.028 (SD = 0.65). RNA samples were processed for microarray hybridization by the Functional Genomics core facility at Fondazione IRCCS Istituto Nazionale dei Tumori, Milan. Briefly, 800 ng of total RNA was reverse transcribed, labeled with biotin and amplified overnight (14 hours) using the Illumina RNA TotalPrep Amplification kit (Ambion, Life Technologies, Grand Island, NY, USA) according to manufacturer's protocol. One μg of the biotinylated cRNA sample was mixed with the Hyb E1 hybridizatioin buffer containing 37.5% (w/w) formamide and then hybridized to array MouseRef-8 v2.0 Expression BeadChip (Illumina, Inc., San Diego, CA, USA) at 58°C overnight (18 hours). Array chips were washed with manufacturer's E1BC solution, stained with 1 μg/ml Cy3-streptavidine (Amersham Biosciences; GE Healthcare, Piscataway, NJ, USA) and scanned with Illumina BeadArray ™ Reader. www.impactjournals.com/oncotarget Microrarray data pre-processing Raw expression data were collected from scanned images using Illumina BeadStudio v3.3.8 and processed using the lumi package from Bioconductor [54]. Quality control was performed on raw and processed data by evaluation of array intensity distributions, distances between arrays, and by principal component analysis for the identification of possible outliers. All samples passed quality-control procedures. Raw data were log 2transformed, normalized with robust spline normalization and filtered, keeping only the probes with a detection p-value < 0.01 in at least one sample. Among probes mapping on the same gene the most detected one was chosen. All array data have been deposited in NCBI's Gene Expression Omnibus and are accessible through GEO Series accession number GSE69903.

Microarray data analysis
Differentially expressed genes between classes were identified using the limma package [55]. P-values were corrected for multiple testing using the Benjamini-Hochberg false discovery rate method. An FDR < 0.05 and an absolute fold change ≥ 2 were used as criteria for selection of differentially expressed genes. Agglomerative hierarchical clustering was performed using average linkage and 1-Pearson's correlation coefficient as distance measure. Gene Set Enrichment Analysis (GSEA [56]) was applied to identify biological processes enriched in the different groups. Gene sets were retrieved from the c2 canonical pathways and c5 biological process collections from the MSigDB database (http://www.broadinstitute. org/gsea/msigdb/index.jsp). Significantly enriched gene sets (p < 0.001 and FDR< 0.05) were visualized using Cytoscape v2.8.3 and the Enrichment Map plugin [57]. Functional analyses of gene lists were generated through the use of IPA(QIAGEN, www.qiagen.com/ingenuity).

Statistical analysis
We used GraphPad Prism software (GraphPad Software, La Jolla, CA, USA). Student's t-test was used for comparison of quantitative variables. Association between categorical variables was carried out using Fisher's exact test. Log-Rank test was applied for survival curves. All statistical tests were two-sided and a p-value <0.05 was considered significant.