Molecular Classification Of Nonkeratining Nasopharyngeal Carcinoma

BackgroundAccording to the International Agency for Research on Cancer, there were84,400cases of nasopharyngeal carcinoma (NPC), and51,600deaths from it, in2008. NPC is mainly a non-keratinizing, squamous cell carcinoma (NK-NPC). It mainly afflicts middle aged men and is a common cancer among Chinese, Greenland Eskimos, and North Africans. The major etiological factors for NK-NPC are genetic, environmental, and viral factors. Strong association with Epstein-Barr virus (EBV) is a unique feature of NK-NPC, in contrast to other head and neck cancer, and epithelial malignancy in general. Currently, standard treatment of NK-NPC consists of concurrent chemo-radiotherapy. Although a majority of patients could be cured by this treatment approach, distant metastasis remains a formidable challenge, and many patients still suffer from long-term toxicity. There is a great need for further understanding of the heterogeneity of NK-NPC, which could aid in the development of more effective and safe therapies, and push the boundaries of personalized treatment further. Molecular classification holds promise by providing further insight into this disease. However, whether NK-NPC tumors could be further classified into molecular subgroups has been disputed. Sengupta et al. have failed to classify NK-NPC tumors into any specific subgroups based on genome-wide human gene expression profiles analysis of31NK-NPC tumors. While Wang et al. have suggested that at least two molecular subtypes existed in NK-NPC based on human gene expression profiles analysis of24differentiated-type NK-NPC tissues. In this study herein, we have confirmed that at least two molecular subtypes exist in NK-NPC by investigating three microarray datasets (including dataset of Sengupta et al.). We further delineated biological characters, origin of subtypes, molecular networks, and targeted drugs for each subtype. Our work is a first step in improvement of diagnosis and treatment of NK-NPC.Methods1. Microarray datasets Raw microarray data was downloaded from GEO including GSE1245and GSE13597. A dataset of25NPC samples and8normal nasopharynx tissues was generated by our institute in a previous study.2. Cluster analysis Three methods of cluster analysis were applied including unsupervised hierarchical clustering, K-means clustering and self-organized map (SOM).hierarchical clustering was performed in Cluster3.0while the other two analysis were carried out by software Gene Pattern.3. Identification of differentially expressed genes Significance Analysis of Microarray (SAM) was performed to identify significantly changed genes by statistical software R. Each subtype was also compared with normal nasopharynx tissues separately.4. Gene Set Enrichment Analysis Gene Set Enrichment Analysis (GSEA) was performed to identified dysregulated biological processes between NPC subtypes. Thousands of curated gene expression signatures were used for the analysis. 5. Dataset GSE12452was used as training set while GSE13597as testing set. The weightedVotingXValidation module in Gene Pattern platform was used to identify gene predictors.50genes were chosen in each round of cross-validation. Genes that were picked in every round were retained in the predictor. WeightedVoting was performed to predict class label of samples in GSE13597.6. Construction of Transcriptional regulatory network transcriptional regulatory network was constructed by software ARACNe. Besides GSE12452and GSE13597, another datasets of22arrays were included in this analysis. A list known transcriptional factors was downloaded from animal TFDB. Only genes with a sample standard deviation larger than0.5were retained for the ARACNe analysis. Fisher’s exact test was applied to determine the significance of overlap between NPC subtype gene signature and regulon of TFs.7. Construction of protein-protein interaction network Protein-protein interaction (PPI) networks for each subtype of NK-NPCGES genes were constructed using a STRING9.0database [17]. Only experimentally validated PPIs were used to construct networks (confidence score>0.400). Each edge of the resultant networks represents a PPI. The edge number of each network was counted. To determine the statistical significance of each network, we randomly selected the same number of genes as the number of each subtype NK-NPCGES genes from the common genes of GSE12452and GSE135971,000times. Each time, the PPI network of randomly selected genes was constructed using the STRING database with the same parameters. The edge number of the result networks was counted. A p value was calculated as the proportion of edge numbers of random networks larger than that of the real networks. 8. Competing endogenous RNA network MicroRNA-target interactions were downloaded from database StarBase. According to Pandolfi et al, a pair of ceRNAs was defined as two mRNAs that shared at least seven common nonredundant microRNA recognition elements. For genes corresponding to multiple refseq mRNA transcripts, we chose the transcript with the most targeting microRNA.9. Drug prediction On-line tool Connectivity map was used to predict chemical compounds that could be used to treat NPC tumors, especially type II. EntrezGene IDs of signature genes were transformed into Affymetrix Probeset identifiers before submitting to Connectivity map.ResultNK-NPC tumors were clearly grouped into two classes, in which one class tumor was grouped together with NP tissues. NK-NPC subclasses were independent from TNM stages. Type I NK-NPC (NK-NPC_I) had low EBV load, while Type II NK-NPC (NK-NPC_II) had high EBV load. Gene Set Enrichment Analysis (GSEA) showed that NK-NPC_I may be associated with normal epithelium-like, enhanced immune response, and good survival; while NK-NPC_II with escape of immunosurveillance, cell proliferation, cancer stem cells (CSCs) enrichment, high metastasis, and poor survival. NK-NPC_II may develop from NK-NPC_I, mainly caused by increase of EBV load.29genes were selected as the predictor of NPC subtype.Top four transcriptional regulators for NPC I were SP140,IRF8,KLF2and STAT4which regulated69%NPC I signature genes. Top six transcriptional regulators were KLF2, DNAJC2, FOXM1HMGB2, ZNF146and MYC which regulated70%NPC II signature genes. 158/327NPC II signature genes participated in249protein-protein interactions while117/351NPC I signature genes participated in222protein-protein interactions. In1000random simulations,42/327random gene formed30protein-protein interactions in avergae while43/351random genes formed31protein-protein interctions in avergae. CDK1was the hub genes in NPC-II PPI network with25interactions with other genes.A undirected ceRNA network was constructed which was formed by1506genes and48338interactions.22signature genes of type II NPC participated in38interactions and among them Weel interacted with20partners.A known dopamine receptor antagonist Thioridazine was predicted to treat type II NPC by Connectivity map analysis.ConclusionBy analysis of genome-wide expression profile of nasopharyngeal carcinoma and normal nasopharynx tissues, and QRT-PCR data of EBV encoded mRNA, we found that non-keratinizing nasopharyngeal carcinoma could be divided into at least two molecular subtypes. Type I NPC had low EBV load while type II NPC had high EBV load. According to Gene Set Enrichment Analysis, Type I NPC may be associated with a normal epithelial-like phenotype, enhanced immune response and better survival while type II NPC could be associated with escape of immunosurveillance, cell proliferation, cancer stem cells enrichment, high risk for metastasis and worse survival. Type II NPC may develop from type I mainly caused by increase of EBV load. This may explain why serological EBV antibody/DNA titers are now the unique biomarkers for early diagnosis of NK-NPC and complementing TNM staging prognostication in NK-NPC.FOXM1, CDK1and WEE1were found to be the major regulator/activator in type II NPC transcriptional regulatory network, protein-protein network and competing endogenous RNA networkThe major limitations of our work are as following:(i) there is no clinical outcome information for our microarray data, as such, we cannot establish a clinical standard to differentiate the two NPC subtypes;(ii) the microarray sample size is still small; and (iii) experiment validation for the major nodes of each molecular network was not done.