XPO1 Expression Is a Poor-Prognosis Marker in Pancreatic Adenocarcinoma

Pancreatic adenocarcinoma (PAC) is one of the most aggressive human cancers and new systemic therapies are urgently needed. Exportin-1 (XPO1), which is a member of the importin-β superfamily of karyopherins, is the major exporter of many tumor suppressor proteins that are involved in the progression of PAC. Promising pre-clinical data using XPO1 inhibitors have been reported in PAC, but very few data are available regarding XPO1 expression in clinical samples. Retrospectively, we analyzed XPO1 mRNA expression in 741 pancreatic samples, including 95 normal, 73 metastatic and 573 primary cancers samples, and searched for correlations with clinicopathological and molecular data, including overall survival. XPO1 expression was heterogeneous across the samples, higher in metastatic samples than in the primary tumors, and higher in primaries than in the normal samples. “XPO1-high” tumors were associated with positive pathological lymph node status and aggressive molecular subtypes. They were also associated with shorter overall survival in both uni- and multivariate analyses. Supervised analysis between the “XPO1-high” and “XPO1-low” tumors identified a robust 268-gene signature, whereby ontology analysis suggested increased XPO1 activity in the “XPO1-high” tumors. XPO1 expression refines the prognostication in PAC and higher expression exists in secondary versus primary tumors, which supports the development of XPO1 inhibitors in this so-lethal disease.


Introduction
Pancreatic adenocarcinoma (PAC) is a major public health problem worldwide with the highest mortality rate of all human cancers and a rising incidence [1,2]. Complete surgical tumor resection followed by chemotherapy is the only curative treatment available, but less than 20% of patients are eligible for surgery at diagnosis [3,4]. In the case of inoperable or metastatic form, the median survival is six months and the long-term survival is null. The improvements in radiotherapy and systemic treatments during the past 20 years have achieved limited impact. The few chemotherapeutic agents that are efficient against PAC include gemcitabine with or without nab-paclitaxel and the FOLFIRINOX regimen that combines 5-FU, leucovorin, oxaliplatin, and irinotecan. The survival benefit is modest, making the development of novel drugs crucial. Few molecular alterations, such as KRAS, TP53, SMAD4, CDKN2A, BRCA2, and ARID1A mutations and GATA6 amplification have been identified in PAC [5][6][7][8][9][10][11][12], but most of them remain non-druggable. Furthermore, many other tumor suppressors are

Gene Expression Data Sets
We gathered clinicopathological and gene expression data of clinical pancreatic carcinoma samples from ten publicly available data sets [54][55][56][57][58][59][60][61][62][63], which comprise at least one probe set representing XPO1 (Supplementary Table S1). Data were collected from the National Center for Biotechnology Information (NCBI)/Genbank GEO, ArrayExpress, and TCGA databases. The samples were profiled using whole-genome DNA microarrays (Affymetrix, Agilent) and RNASeq (Illumina). The pooled data set contained 1052 samples, including 573 primary PAC samples, 73 metastatic samples, and 95 normal pancreatic samples. Our institutional board approved the study. A total of 573 PAC samples that were informative for overall survival were included in the present analysis.

Gene Expression Data Analysis
Data analysis required pre-analytic processing. First, we separately normalized each DNA microarray-based data set by using quantile normalization for the available processed Agilent data, and Robust Multichip Average (RMA) with the non-parametric quantile algorithm for the raw Affymetrix data sets. Normalization was done in R using Bioconductor and the associated packages. Subsequently, we mapped hybridization probes across the different technological platforms present. We used SOURCE (http://smd.stanford.edu/cgi-bin/source/sourceSearch) and EntrezGene (Homo sapiens gene information db, release from 04/27/2017), ftp://ftp.ncbi.nlm.nih.gov/gene/) to retrieve and update the Agilent annotations, and NetAffx Annotation files (www.affymetrix.com; release from 01/12/2008) for the Affymetrix annotations. The probes were then mapped according to their EntrezGeneID and, when multiple probes represented the same GeneID, we retained the one with the highest variance in a particular dataset. For the RNA-seq data, we used the available normalized RNASeq data that we log 2 -transformed. Subsequently, we corrected the ten studies for batch effects using z-score normalization. Briefly, for each separate XPO1 expression value in each study, subtracting the mean of the gene in that dataset divided by its standard deviation, mean, and standard deviation only being measured on primary cancer samples transformed the value. XPO1 expression in tumors was measured as discrete value after comparison with mean expression in the 573 primary tumors: high expression was defined by value > mean and low expression by value ≤ mean.
We separately applied different multigene classifiers to each sample in each data set: the subtype classifiers that were reported by Bailey [61], Collisson [58], and Moffitt [60], and the 25-gene prognostic signature that we recently developed [64]. Finally, to explore the biological pathways linked to XPO1 expression in pancreatic cancer more-in-depth, we applied a supervised analysis by using the largest data set (TCGA: 150 samples) as a learning set, and the other data sets as independent validation sets (423 samples). In the learning set, we compared the whole-genome expression profiles between tumors with (N = 76) versus without (N = 74) high XPO1 expression using a moderated t-test with an empirical Bayes statistic [65] being included in the limma R packages. False discovery rate (FDR) [66] was applied to correct the multiple-testing hypothesis and the following thresholds: p < 1.0 × 10 −5 and q < 1.0 × 10 −5 defined significant genes. Ontology analysis of the resulting gene list was based on the GO biological processes of the Database for Annotation, Visualization and Integrated Discovery (DAVID; david.abcc.ncifcrf.gov/). We verified the robustness of the resulting gene list in the validation set (295 tumors with and 283 without high XPO1 expression) by computing a metagene-based prediction score defined by the difference between the "metagene XPO1-high" (mean expression of all genes upregulated in the "XPO1-high" class) and the "metagene XPO1-low" (mean expression of all genes upregulated in the "XPO1-low" class) for each tumor. This score was then compared between the "XPO1-high" and "XPO1-low" samples.

Statistical Analysis
The t-test or the Fisher's exact test, when appropriate, were used to analyze the correlations between XPO1-based tumor classes and clinicopathological features. Overall survival (OS) was calculated from the date of diagnosis to the date of death from pancreatic cancer. Follow-up was measured from the date of diagnosis to the date of last news for event-free patients. The survivals were calculated using the Kaplan-Meier method and the curves were compared with the log-rank test. Univariate and multivariate survival analyses were done using Cox regression analysis (Wald test). The variables tested in univariate analyses included patients' age at time of diagnosis (continuous value), sex, American Joint Committee on Cancer (AJCC) stage (4, 3, 2 vs. 1), pathological features including pathological type, tumor grade (3, 2 vs. 1), tumor size (T4, T2, T3 vs. T1), regional lymph node status (positive vs. negative), and XPO1 expression ("high" vs. "low"). Variables with a p-value < 0.05 were tested in multivariate analysis. All of the statistical tests were two-sided at the 5% level of significance. Statistical analysis was done using the survival package (version 2.30) in the R software (version 2.15.2; http://www.cran.r-project.org/). We followed the reporting REcommendations for tumor MARKer prognostic studies (REMARK criteria) [67]. Table 1 summarizes the analyzed XPO1 mRNA expression in 573 clinical primary PAC samples. Their clinicopathological characteristics are summarized. Briefly, most of patients were more than 60 year-old and 53% were male. Most of the tumors were ductal type (93%), grade 2 (57%), and they were classified as AJCC stage II (85%); most of them were pT3 tumors (78%) and most had at least one lymph node involved (70%). All but one had been initially treated by surgery. None of them had received neoadjuvant chemotherapy or radiotherapy. All of the molecular subtypes were represented with more frequent squamous Bailey's subtype (36%), more frequent classical Collisson's subtype (45%), more frequent classical Moffitt's tumor subtype (60%), and more frequent activated Moffitt's stroma subtype (59%). Table 1. Clinico-pathological and molecular characteristics of 573 primary pancreatic adenocarcinoma (PAC) samples in the whole population and in each exportin-1 (XPO1)-based group.

XPO1 Expression and Clinicopathological Features
The XPO1 expression was variable and different between the normal tissue samples, the primary tumors, and the metastatic samples ( Figure 1A), with an increasing gradient from normal samples to primary cancer samples (p = 4.88 × 10 −18 , Student t-test), and from primary cancer samples to metastatic samples (p = 1.22 × 10 −21 , Student t-test). We defined two classes of primary cancer samples that were based upon XPO1 expression in tumors when compared with mean expression in normal pancreatic samples: the "XPO1-high" class (N = 298; 52%) and the "XPO-low" class (N = 275, 48%). We then searched for correlations between these two classes and the clinicopathological and molecular features (Table 1). There was no correlation with patient's age and sex, AJCC stage, and pathological type, grade, and tumor size. By contrast, correlations (Fisher's exact test) existed with the pathological lymph node status (pN) and the molecular subtypes ( Figure 1B (Table 1). There was no correlation with patient's age and sex, AJCC stage, and pathological type, grade, and tumor size. By contrast, correlations (Fisher's exact test) existed with the pathological lymph node status (pN) and the molecular subtypes ( Figure 1B-E). The "XPO1-high" tumors were enriched in pN-positive tumors (p = 1.97 × 10 −2 ) and in squamous Bailey's subtype (p = 3.20 × 10 −5 ), in classical and quasi-mesenchymal Collisson's subtypes (p = 9.07 × 10 −4 ), in basal-like Moffitt's tumor subtype (p = 4.55 × 10 −4 ), and in activated Moffitt's stroma subtype (p = 4.69 × 10 −4 ).

XPO1 Expression and Associated Biological Processes
To explore the biological alterations that are associated with the XPO1 expression status, we applied supervised analysis to the TCGA data set (N = 150). We identified 268 genes that were differentially expressed between the tumors with (N = 76) versus without (N = 74) XPO1 upregulation, including 191 genes that were upregulated and 77 genes that were downregulated in the "XPO1-high" samples (Supplementary Table S3). Ontology analysis (Supplementary Table S4) revealed the strong involvement of genes that were overexpressed in the "XPO1-high" tumors in cell cycle, nuclear division, DNA repair, signal transduction, chromosome segregation, DNA replication, and RNA processing. Ontologies that were associated with the genes underexpressed in the "XPO1-high" tumors were fewer and mainly related to metabolism and development. The robustness of this gene signature was verified in the learning set, and more importantly confirmed in the independent validation set by using a metagene-based prediction score (Supplementary Figure S1A,B).

Discussion
The need for new therapeutic and/or prognostic targets is crucial in PAC. We have analyzed XPO1 mRNA expression in 573 clinical PAC samples because of the promising therapeutic value of XPO1 inhibitors in oncology and the paucity of data in the literature: high expression was associated with shorter OS in multivariate analysis. To our knowledge, this is by far the largest study analyzing XPO1 expression in PAC.
Our analysis was based on mRNA expression rather than protein expression as measured using immunohistochemistry (IHC) for several reasons: i) avoiding the limitations of IHC with different non-standardized protocols for XPO1; ii) working on an available large series of clinical samples; and, iii) searching for associations with expression of other genes on a whole-genome scale. When compared to normal pancreatic tissue, XPO1 expression was higher in primary PAC. Of note, it was also higher in secondary tumors as compared to primary tumors. Expression was heterogeneous between samples in our series of 573 operated primary PAC. This range of expression values allowed for searching for correlations with clinicopathological features. Correlations existed with the pathological lymph node status (pN) and aggressive molecular subtypes, squamous Bailey's subtype, quasi-mesenchymal Collisson's subtype, basal-like Moffitt's tumor subtype, and activated Moffitt's stroma subtype. Such an association with adverse prognostic features was confirmed in univariate analysis with shorter metastasis-free survival (MFS) in the "XPO1-high" class. However, interestingly, such unfavorable prognostic value remained significant in multivariate analysis, suggesting independence. Our analysis was based on discrete values while using the mean expression level in normal tissues as cut-off, but a similar correlation was found when XPO1 expression was analyzed as continuous values.
It is not surprising to find frequent high expression in PAC and association with poor prognosis given the XPO1 function of inactivation of TSPs. This has already been reported in several cancers [18][19][20][21][22][23][24][25][26][27]68,69]. Regarding PAC, to our knowledge, only three studies have described XPO1 protein expression in clinical samples using Western blot and IHC with different antibodies [23,52,53]. The first one, which was published in 2009 [23], included 69 primary pancreatic cancer samples and 10 normal tissues that were tested using Western blot. Increased XPO1 expression was shown in pancreatic cancer, and high expression was associated with increased serum levels of CEA and CA19-9, with tumor size, lymphadenopathy, and liver metastasis, and with shorter progression-free survival (PFS) and OS in uni-and multivariate analyses. The second study [53] concerned 91 pancreatic cancer tissues and 70 non-malignant pancreatic samples and it showed higher expression in cancer samples than in the matched normal control, but no correlation with the clinicopathological features and survival tested. In the last study [52], IHC was applied to a tissue microarray comprising 76 primary cancer samples: XPO1 was expressed in 86% of pancreatic cancers, and increased expression was correlated with both survivin expression and increased proliferative activity; no correlation with clinicopathological features and survival was searched.
Our analysis of genes upregulated in XPO1-high tumors identified several ontologies that were related to cell proliferation, such as cell cycle, nuclear division, chromosome segregation, and DNA replication; other ontologies, such as DNA repair, signal transduction, or RNA processing also agreed with the multiple protein targets of XPO1 exported from the cell nucleus to the cytoplasm in eukaryotic cells [13,16], and with recent publications showing that selinexor, which is an XPO1 inhibitor, reduces the expression of DNA damage repair proteins [51], and showing the correlation of XPO1 expression with the proliferative activity [52]. These correlations explain, at least in part, the oncogenic effect of XPO1 and the association with tumor stage (higher in metastatic samples than in primary tumors, and higher in primary tumors than in normal tissue), with shorter survival and with resistance to cytotoxic therapies. Importantly, they also provide indication that increased XPO1 expression in PAC is likely associated with an increase in its biological activity.

Conclusions
In conclusion, we showed that XPO1 mRNA expression is heterogeneous in PAC and is associated with progression stage and shorter survival independently from the other prognostic features. The strength of our study lies in the size of the series (the largest series of tumors reported to date regarding analysis of XPO1 expression), the biological and clinical relevance of XPO1 expression, and its independent prognostic value. The limitations include its retrospective nature and associated biases, such as the lack of available information regarding the delivery or not of adjuvant chemotherapy for most of cases. No patient had received neoadjuvant chemotherapy, impeding the search for an eventual correlation of XPO1 expression with response to chemotherapy. Obviously, an analysis of larger series, retrospective, then prospective, is warranted to confirm our observation. Functional studies need to be conducted with fresh patient samples as well as retroactive meta-data analysis to link mRNA levels to actual protein expression and patient outcomes. If such a prognostic value is confirmed, XPO1 expression might refine the prognostication and improve our ability to tailor adjuvant chemotherapy. However, more importantly, and given this unfavorable prognostic value and the likely association with increased XPO1 biological activity, patients with high level of XPO1 expression would warrant a more aggressive treatment plan, which should include SINE compounds that are associated with classical drugs, notably the DNA-damaging agents. In the metastatic setting, clinical trials are ongoing, and it will be important to test whether XPO1 mRNA expression can predict the clinical response to SINE compounds.
Supplementary Materials: The following are available online at http://www.mdpi.com/2077-0383/8/5/596/s1, Table S1: List of pancreatic cancer data sets included in the analysis, Table S2: Uni-and multivariate prognostic analyses for OS, Table S3: List of 268 genes differentially expressed between the "XPO1-high" PAC samples (N = 76) and the "XPO1-low" PAC samples (N = 74), Table S4: Ontology analysis of the 268 genes differentially expressed between the "XPO1-high" PAC samples (N = 76) and the "XPO1-low" PAC samples (N = 74), Figure S1: Supervised analysis of gene expression profiles between the XPO1-high" primary PAC samples and the "XPO1-low" primary PAC samples. (A) Classification of 150 samples from the learning set (TCGA) using the 268-gene expression signature. Top panel: indicates the XPO1 expression status of tumors (black, "XPO1-high"; white, "XPO1-low"). Middle panel: matrix of gene expression levels. Each row represents a gene and each column represents a sample. Expression levels are depicted according to the color scale shown below. Genes are ordered from top to bottom by their decreasing t-statistic. Tumor samples are ordered from left to right according to the metagene-based prediction score (middle panel). The solid orange line indicates the threshold 0 that separates the two classes of samples, "XPO1-high-like class" (at the left of the line) and "XPO1-low-like class" (right to the line). Bottom panel: Box-and-whisker plot of the metagene based prediction score in the "XPO1-high" samples compared to the "XPO1-low" samples. The p-value is for the Student's t-test assessing the difference of the prediction score between the observed XPO1 classes, which is, as expected, very significant in the learning set.