Background

Nasopharyngeal carcinoma (NPC) can be found most frequently in southern China and southeast Asian countries. According to the global cancer statistics of 2018, it was estimated that there were 129,079 new NPC cases and 72,987 deaths that year [1]. Despite improved treatment for NPC patients, especially concurrent chemoradiation therapy combined with adjuvant chemotherapy, those patients with recurrence and metastasis still have worse overall survival [2]. Tumor-node-metastasis staging system can provide an effective prognostic prediction for NPC patients [3]. However, this factor fails to be accurate enough for evaluating the survival outcomes.

It is well documented that the number of biomarkers might predict the prognosis of NPC, such as the EBV DNA levels, lactate dehydrogenase (LDH), SOX2, NGX6, FMNL3 and ANXA2 [4,5,6,7,8,9]. However, the molecular mechanism of NPC has not been completely established. Therefore, the identification of novel biomarkers that reflect tumors heterogeneity is urgently needed to predict outcomes and tailor treatment for NPC patients.

It is noted that bioinformatics analysis is commonly used to identify candidate biomarkers from high-throughput data. Weighted gene co-expression network analysis (WGCNA) [10], which is mainly applied to filtrate hub genes from large-scale data sets, has been successfully applied in several human tumors [11,12,13,14]. WGCNA had been used to analyze the NPC data sets and identify some candidate diagnostic genes [15, 16], which have limited application in the predication of prognosis in NPC patient.

In this study, we performed WGCNA to identify hub genes and validate these hub genes using three data sets (GSE68799, GSE12452 and GSE53819). Another one data set GSE102349 that contains follow-up information was used for univariate Cox regression analysis to screen hub genes with prognostic values. Numerous reports have testified that tumor hypoxia leads to poor prognosis [17,18,19]. We hypothesized that one prognostic hub gene (IGSF9) might be involved in hypoxia in NPC. Hence, Pearson’s correlation analysis was conducted and the result showed that expression of IGSF9 was related to the scores of three hypoxia signatures. Furthermore, functional studies confirmed that the elevated expression of IGSF9 induced by hypoxia increased NPC cells proliferation, invasion and migration in vitro.

Materials and methods

Data download and pre-processing

Four NPC micro-array data sets (Additional file 1: Table S1) were downloaded from the well-known gene expression omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Since the numerical distribution of raw data value of four data sets is too wide to be used for further analysis, we formulated each raw data value in log2 (X + 1), where X is the raw data value, followed by quantiles of them. After normalization, expression values of each gene of data set GSE68799 were ranked from large to small by variance analysis report (VAR) value. The expression profile of data set GSE68799 which contains the top 6000 genes was used for WGCNA.

Co-expression network construction and identification of key modules and hub genes

The processed matrix of the data set GSE68799 was exploited to construct an unsigned gene co-expression network by WGCNA. Genes were clustered by using the hierarchical clustering method according to the topological overlap measure (TOM). In accordance with the criterion of hybrid dynamic tree, each gene module should contain no less than 50 genes. In line with the dissimilarity of module eigengenes (MEs), the height 0.25 was set as a cut off line to combine some modules. In the process of generating the network, 1000 genes were randomly selected every time. Then, modules with the strongest correlation between MEs in each module and clinical traits were identified as key modules. In a similar way, genes with higher module membership (MM) and higher gene significance (GS) were considered as hub genes, filtered criteria was set as MM > 0.8 and GS >|0.7| [14].

Validation of hub genes

Differentially expressed genes (DEGs) between NPC group and non-NPC group of three data sets (GSE68799, GSE12452 and GSE53819) were screened by the “limma” R package. Cut-off value was determined by adjusted P < 0.05. Validated hub genes were obtained by overlapping hub genes and common DEGs. Comparison analysis for expression of validated hub genes was conducted in benign and malignant groups of three data sets (GSE68799, GSE12452 and GSE53819), and the same analysis for expression of validated hub genes was performed in different clinical stages of the data set GSE102349.

Identification and functional analysis of prognostic hub genes

The profile contains 23,706 genes from the data set GSE102349 was exploited for univariate Cox regression analysis by the “survival” R package. Cut-off value was set as hazard ratio (HR) >|1|, and adjusted P < 0.05. If one validated hub gene was significantly associated with progression-free survival (PFS), it could be identified as a prognostic hub gene. Functional analysis of them was proceeded with gene set enrichment analysis (GSEA) (https://www.gsea-msigdb.org/gsea/index.jsp) [20]. This analysis was performed on the version of MSigDB.v7.0, and each analysis would conduct on 1000 times gene set substitution. Significant pathways were identified by follow criterion: absolute value of normalized enrichment score (NES) > 1, P < 0.05.

Correlation of hypoxia signatures and prognostic hub genes

To explored whether prognostic hub genes are related to hypoxia signatures in NPC, we focused on three previously reported hypoxia signatures (99-gene, 26-gene and 15-gene) [21,22,23], which have been regarded as effective prognostic tools in head and neck cancer. The data set GSE102349 was used to summarize the scores of hypoxia signatures. For each gene of hypoxia signatures, each sample in the data set was given a primary score of either + 1 or − 1 as the following criteria [24]: If the expression of that gene of one sample is greater than the median value of that gene of all samples in the data set, this sample would be marked as + 1 score. If not, the sample would be marked as − 1 score. For each sample, the summed value was calculated from the + 1′s and − 1′s of total genes of hypoxia signatures, which was the scores of hypoxia signatures for that sample in data set GSE102349. After that, the scores of hypoxia signatures were exploited for survival analysis and correlation analysis between the scores and the expression of prognostic hub genes and parallel analysis was performed in P4HA1 which was one of the common genes from three signatures.

Cell culture and siRNA transfection

CNE-2 and 5-8F cells were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS, Gibco BRL, NY, USA) and incubated in a humidified atmosphere at 37 °C with 5% CO2. For hypoxia exposure, they were incubated in a anoxic chamber at 37 °C with 5% CO2 and 1% O2 for 24 h. Two siRNA oligonucleotides (siIGSF9#1 siIGSF9#2) and negative control were purchased from Sangon Biotech (Shanghai, China) and LipofectamineTM 2000 transfection reagent were from Invitrogen (Camarillo, USA).

Quantitative real time polymerase chain reaction (qRT-PCR)

Total RNA was extracted from cells using TRIzol reagent (Invitrogen) based on the product specification. qRT-PCR was executed using SYBR green qPCR Master Mix (Takara Bio, Otsu, Japan). Comparative CT method was used for quantitative analysis and GAPDH was served as the internal reference. Primer sequences as follow: IGSF9 sense 5′-TCTGTGGACGAGAACTATGAGTG-3′, antisense 5′-AGCCGAGCCCTAGTAGCAT-3′; P4HA1 sense 5′-GGGACAAGCCTCGTATT-3′, antisense 5′-TTGGGTTTGAAATGGTG-3′; GAPDH sense 5′-GACTCATGACCACAGTCCATGC-3′, antisense 5′-AGAGGCAGGGATGATGTTCTG-3′.

Western blotting

Proteins from cells were lysed using RIPA lysate (Beyotime, Shanghai, China), and quantified using Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, MA, USA). Protein lysates were separated by 10% sodium dodecyl sulfate containing polyacrylamide gels and transferred onto a polyvinylidene difluoride membrane (Millipore, MA, USA) by electrophoresis. GAPDH antibody (1:5000, Abcam, MA, USA), IGSF9 antibody (1:1000, Ybscience, Shanghai, China) and P4HA1 antibody (1:1000, Invitrogen) were incubated at 4℃ overnight. Horseradish Both peroxidase-labeled anti-rabbit (1:8000, Santa Cruz, CA, USA) and anti-mouse (1:8000, Santa Cruz) were used at 4 °C for 2 h. The bands on the membrane were visualized by enhanced chemiluminescence (Thermo Fisher Scientific).

Cell proliferation, migration and invasion assays

Cell proliferation was estimate using cell counting kit-8 (CCK8) assay. 1000 cells were seeded in per well of 96-well plates. The 450 nm absorbance of each well was detected using CCK8 reagent (Dojindo, Kumamoto, Japan) incubation for 0, 24, 48, 72 and 96 h. Cell migration was evaluated using wound-healing assays. An artificial scratch was conducted on fused monolayer of adherent cells. 48 h later, photographs were got from an inverted microscope (Olympus,Tokyo, Japan). For cells invasion assay, cells were seeded on the matrigel (Corning, NY, USA) of the upper layer of transwell chamber without serum medium. Then, FBS was added into the lower layer of the transwell chamber. After 48 h, non-invasive cells and matrigel were softly removed, while invasive cells that pass through matrigel into the lower layer were colored with crystal violet, counted and imaged.

Statistical analysis

Data were expressed as mean ± standard deviation (SD) and the difference between two groups was estimated using two-tailed Students’ t-test. Survival difference was evaluated using Log-rank test. Pearson’s correlation analysis was used to calculate the correlation between two continuous variables. P < 0.05 was defined to be statistically significant.

Results

Identification of two key modules and 79 hub genes

Expression value of all genes of four NPC data sets were normalized by log2 (X + 1) and quantiles (Additional file 1: Fig. S1). The data profile of the top 6,000 genes collated by the VAR values from the data set GSE68799 was used for WGCNA, and 46 samples of this data set could be divided into two clusters (Fig. 1a). Four was recognized as a soft-thresholding power value to create unsigned co-expression network (Fig. 1b, c), and the network contains 22 independent modules (Fig. 1d, f), which were split into two clusters (Fig. 1e). The blue module contains 869 genes, while the dark red module contains 56 genes (Additional file 1: Table S2). These two modules were identified as key modules (Fig. 2a-c). 74 hub genes and five hub genes were found in the blue and dark red module respectively (Additional file 1: Fig. S2).

Fig. 1
figure 1

Construction of a co-expression network. a Hierarchical clustering was used in the data set GSE68799, which included 42 NPC samples and four non-NPC samples. Matrix of the top 6,000 genes from the data set GSE68799 in order of their VAR values was used for WGCNA. Different colors represent different kinds of clinical traits (Group of benign and malignant, WHO histology, Age and Gender). b, c Scale-free fit index (b) and mean connectivity (c) for various soft-thresholding powers. d Dendrogram clustered according to a dissimilarity measure (1-TOM). Different branches represent different genes, and different colors below represent different modules. e Plot of hierarchical clustering of the module. f Heat-map plotted according to adjacencies of the network

Full size image
Fig. 2
figure 2

Identification of two key modules and 79 hub genes. a Heat-map of the correlation between modules and clinical traits (Group of benign and malignant, WHO histology, Age and Gender) of NPC. Each cell corresponds the correlation. Color legend on the right of the heat-map. Modules framed in red (blue module and dark red module) were considered as key modules. b, c Scatter plots of MEs in the dark red module (b) and the blue module (c)

Full size image

Validation of 79 hub genes

To verify the 79 hub genes, we screened DEGs between NPC samples and non-NPC samples of three data sets (GSE68799, GSE12452 and GSE53819). 220 up-regulated common DEGs (Fig. 3a) and 250 down- regulated common DEGs (Fig. 3c) were found. Seven of 74 hub genes of the blue module were significantly up-regulated (Fig. 3b) and (Additional file 1: Fig. S3a) in the NPC patients of three data sets, while none of hub genes of the dark red module was significantly down-regulated (Fig. 3d). In addition, expression of four genes (IGSF9, AGRN, STAP2 and PKP4) was significantly up-regulated in advanced NPC patients of the data set GSE102349 (Additional file 1: Fig. S3b), but expression of other three genes (TRIP6, EPCAM and IRF6) was not.

Fig. 3
figure 3

Validation of 79 hub genes. a Venn diagram analysis among up-regulated DEGs from three NPC data sets (GSE68799, GSE12452 and GSE53819). b Intersection genes of hub genes of the blue module and up-regulated common DEGs. c Venn diagram analysis among down-regulated DEGs from three data sets. d Intersection genes of hub genes of the dark red module and down-regulated common DEGs

Full size image

Identification and functional analysis of a prognostic hub gene (IGSF9)

Univariate Cox regression analysis was carried out in the data set GSE102349 using survival package in R. IGSF9 included in seven validated hub genes was identified as a prognostic hub gene (Fig. 4a, b), and an up regulation of IGSF9 could lead to poor PFS (Fig. 4c). Moreover, functional analysis results showed that IGSF9 may promote metastasis of NPC cells through Akt signaling pathway, including XU AKT1 TARGETS 6HR, LEE METASTASIS AND RNA PROCESSING UP and MTOR UP. N4.V1 UP (Fig. 4d, e).

Fig. 4
figure 4

Survival and functional analysis of IGSF9. a Venn diagram analysis between validated hub genes of the blue module and genes related with PFS from GSE102349. b Univariate Cox regression analysis result of seven validated hub genes. Only IGSF9 was identified as a prognostic hub gene. c Kaplan–Meier survival plots of IGSF9, which is significantly associated with the PFS for NPC patients. Patients were grouped by the trichotomization (Q1 vs Q4) value of seven genes. d GSEA results of IGSF9 using the data set GSE102349. e Three significant metastatic pathways involved IGSF9. Absolute value of NES > 1, P < 0.05 were considered as statistical significance

Full size image

Correlation of the scores of hypoxia signatures and the expression of IGSF9

The scores of three hypoxia signatures (99-gene, 26-gene and 15-gene) were calculated for survival and correlation analysis in the data set GSE102349. Patients with high scores of these hypoxia signatures had unfavorable PFS (Fig. 5a), and a parallel result was found in P4HA1 included in four common genes (Fig. 5d, e); in other words, an over expression of P4HA1 was associated with a poor PFS. Moreover, patients with high scores of these hypoxia signatures owned an over expression of IGSF9 (Fig. 5b, c), which could also be seen in patients with high expression of P4HA1 (Fig. 5f, g).

Fig. 5
figure 5

Correlation of hypoxia signatures and the expression of IGSF9. a Kaplan–Meier survival plots of three hypoxia signatures, these three signatures were related with prognosis in NPC. b Correlation of the scores of three hypoxia signatures and the expression of IGSF9. c Comparison results of the expression level of IGSF9 between high and low groups according to the scores of three hypoxia signatures from the data set GSE102349. Patients used in above analysis (a–c) grouped by the median score of those signatures. d Venn diagram of genes of three hypoxia signatures.e Kaplan–Meier survival plots of P4HA1, which is significantly associated with PFS in NPC. Patients were grouped by the trichotomization (Q1 vs Q4) of P4HA1. f Correlation between  the expression level of P4HA1and IGSF9. g Comparison of the expression level of IGSF9 between low-exp and high-exp group of P4HA1 of the data set GSE102349. Patients used in above analysis (f, g) were grouped by the median value of P4HA1. *P < 0.05, **P < 0.01, ****P < 0.0001

Full size image

In vitro functional analysis of IGSF9

To investigate the effect of IGSF9 on the aggressive phenotype, we transfected two different IGSF9 siRNAs (siIGSF9#1 and siIGSF9#2) into CNE-2 and 5-8F cells to down-regulate IGSF9 (Fig. 6a) and performed CCK8, wound-healing and invasion assays to evaluate whether regulation of IGSF9 expression affects the proliferation, invasion and migration of NPC cells. The results showed a significant inhibition of proliferation, invasion and migration in the cells transfected with IGSF9 siRNAs compared with the control cells (Fig. 6b–d). These data indicated that IGSF9 could promote NPC cell malignant biological behaviors that might be associated with poor prognosis. Next, we tried to verify the interaction between the expression of IGSF9 and hypoxia. qRT-PCR and Western blotting were used to measure the expression of IGSF9 and P4HA1 in CNE-2 and 5-8F cells exposed to hypoxia. We observed that expression levels of IGSF9 and P4HA1 were significantly increased in NPC cells exposed to anoxic conditions (Fig. 6e, f). There was an evidence that expression of IGSF9 might be induced by hypoxia, which was consistent with the result from our bioinformatics analysis.

Fig. 6
figure 6

In vitro functional analysis of IGSF9. a CNE2 and 5-8F cells were transfected with siIGSF9#1 and siIGSF9#2 and subjected to western blotting for IGSF9. b Proliferative potential of two cell lines with interference of two siRNAs were analyzed by CCK8 assay. c, d NPC cells were treated with siRNA (siIGSF9#1) and subjected to transwell assay (c) and wound healing assay (d). e, f qRT-PCR (e) and western blotting (f) in CNE2 and 5-8F cells cultured under normoxia or hypoxia. All data are displayed as mean ± SD, *P < 0.05, **P < 0.01, ***P < 0.001

Full size image

Discussion

To identify prognostic hub genes of high throughput data sets, WGCNA and univariate Cox regression analysis are very powerful tools. There are two other studies [15, 16] aiming at identifying candidate genes by WGCNA. Besides this way, our research added univariate Cox regression analysis via the mRNA-seq data set GSE102349 that contains follow-up information. By these two tools, IGSF9 was identified as a prognostic hub gene in NPC.

IGSF9 was highly expressed in NPC patients, and it was an appropriate prognostic gene. IGSF9 belongs to the immunoglobulin superfamily [25], which plays a key role in inhibiting synaptic development by regulating calmodulin-like activity [26]. Calmodulin is involved in tumor metastasis, and our GSEA result also shows that IGSF9 may cause metastasis of NPC cells through Akt signaling pathway. Meanwhile, we verified that IGSF9 promoted the proliferation, migration and invasion of NPC cells in vitro. To date, there is one published report [27] on over-expression of IGSF9 associated with poor PFS in endometrial cancer tissue. Similar result was found in NPC tissue in our research. Thus, IGSF9 is likely to be a prognostic gene which would promote the invasion and metastasis of NPC cells.

Our study revealed that three hypoxia signatures (99-gene, 26-gene and 15-gene) have prognostic value in NPC patients using the data set GSE102349, though previous evidences have proved that those three hypoxia signatures can successfully discriminate the prognosis of patients with head and neck cancer [21,22,23, 28]. Interestingly, one research suggested that those three hypoxia signatures might be replaced by P4HA1 [28] and another study also confirmed that P4HA1 was a significantly prognostic factor for oral squamous cell carcinoma patients [29]. We also got a similar result in NPC patients. It is clear that our research provides valuable support for consideration of three hypoxia signatures and one single-gene (P4HA1) hypoxia signature as new prognostic tools like IGSF9 in NPC patients. Moreover, we found the expression of IGSFP was positively associated with both the scores of three hypoxia signatures and the expression of P4HA1 utilizing data set GSE102349, and the expression of IGSFP was significantly increased in NPC cells exposed to anoxic microenvironment. These data revealed that the expression of IGSF9 in NPC cells might be affected by hypoxia.

Taken together, we identified IGSF9 as a prognostic hub gene by WGCNA and univariate Cox regression analysis and found the expression of IGSF9 might be induced by hypoxia. We then performed cytology experiments and verified the role of IGSF9 in NPC cells. However, the investigation of deeper mechanism and in vivo experiment of IGSF9 needs to be supplemented further.

Conclusion

Our study suggests that the elevated expression of IGSF9 will be result in a poor PFS and involved in hypoxia in NPC, implying that IGSF9 might be expected to be a new prognostic indicator of NPC in the future.