Gene set enrichment analysis and meta-analysis to identify six key genes regulating and controlling the prognosis of esophageal squamous cell carcinoma

Abstract

Background

Esophageal squamous cell carcinoma (ESCC) is a common malignancy with high mortality. Because of the lack of clarity in the relevant genes and mechanisms involved, and the current difficulty for oncotherapy in providing therapeutic solutions, there is an urgent need to study this matter. While gene probe studies have been used to select the most virulent genes and pathways, paucity of case controls during gene screening and lack of conclusive results to expound the etiology and pathogenesis of the disease, have reduced study reliability.

Methods

We chose six datasets from independent studies in the Gene Expression Omnibus (GEO) database and used gene set enrichment analysis and meta-analysis to select key genes and pathways.

Results

We found four down-regulated and four up-regulated pathways through gene set enrichment analysis, and 406 differential genes through meta-analysis. Based on The Cancer Genome Atlas (TCGA), 995 differentially expressed genes were screened out. Comparing the 406 gene set with the 995 gene set, we found 19 common genes, of which 6 had a common pathway and were screened out as key genes regulating and controlling the prognosis of ESCC.

Conclusions

Among the 19 genes, we found three genes that affect the chemotherapy of ESCC: BUB1B, BUB1, and TTK. Another three genes NDC1, NUP107, and NUP155 on the RNA transport pathway were also found. Altogether, these six genes are not only crucial in the development of ESCC, but also determine the prognosis of patients. The key genes and pathways identified in the present study will be used for the next stage in our study, which will involve gene elimination and other experimentation methods.

Introduction

Originating from the esophageal mucosa or gland, esophageal squamous cell carcinoma (ESCC), which is a predominant type of esophageal carcinoma, is malignant and aggressive with typically poor prognosis (1). According to the data provided by the International Agency for Research on Cancer, of the incidence of 27 cancers in 184 countries in 2012, we found that the crude rate of esophageal carcinoma worldwide is 6.5/100,000 while the incidence of China is 16.4/100,000; thus, China’s incidence of esophageal carcinoma far exceeds the global average. Furthermore, squamous cell carcinoma accounts for about 90% of the global esophageal carcinoma, while the 5-year survival rate of ESCC patients, although diagnostic methods and treatments have improved over the last few years, remains generally poor (2).

Gene chips have been widely used in cancer research, and the analysis of genome-wide mRNA expression chips—a kind of gene chips can help identify the disease-related genes, and provide an important theoretical basis for the pathogenesis of ESCC. For instance, Subramanian used a method of gene set enrichment analysis (GSEA) (3,4) to reveal significant differences in the expression between normal people’s and patients’ samples. GSEA, in contrast with other analytical methods, shows its distinction in gene detection by testing groups, rather than individual genes. In order to provide a better personalized therapy for ESCC patients, we upgraded the analytical method to construct the disease-related gene regulatory network by combining GSEA with meta-analysis. These two methods were utilized to select significant genes for gene ontology (GO) annotation.

Methods

Datasets selection

We systematically for x using the key phrase “Esophageal Carcinoma” in the Gene Expression Omnibus (GEO) (5) with a subject limit of expression profiled by an array and the species limit of humans 115 identified datasets were found up to the date of September 1st, 2017. Any dataset that met the following standards was selected for inclusion: (I) datasets were about genome-wide RNA expression; (II) datasets provided a comparison between the patient and control; (III) datasets contained more than 3 samples; (IV) the samples were from esophageal tissue; (V) datasets compared ESCC to normal controls from the same patients with ESCC. Finally, there were six gene expression datasets that met the selection criteria (Table 1).

Table 1. Characteristics of datasets selected in the studies.

GEO accession	Contributor	Submission date	Region	Experimental design	Chip	Organism	Probes	Disease	Normal
GSE17351	Nakagawa H	1-Sep-09	USA	Paired, tissues	GPL570 [HG-U133_Plus_2]	Homo sapiens	35166	5	5
GSE20347	Clifford RJ	15-Mar-11	USA	Paired, tissues	GPL571 [HG-U133A_2]	Homo sapiens	22277	17	17
GSE23400	Su H	1-Sep-10	USA	Paired, tissues	GPL96 [HG-U133A]	Homo sapiens	22283	53	53
GSE38129	Hu N	30-Dec-15	USA	Paired, tissues	GPL571 [HG-U133A_2]	Homo sapiens	22277	30	30
GSE77861	Erkizan HV	15-Aug-17	USA	Paired, tissues	GPL570 [HG-U133_Plus_2]	Homo sapiens	54675	7	7
GSE100942	Zhang X	15-Jul-17	Hong Kong	Paired, tissues	GPL570 [HG-U133_Plus_2]	Homo sapiens	54675	4	4

GEO, Gene Expression Omnibus.

Significant gene detecting through gene set enrichment analysis

Software packages developed in the version 2.10.1 of Bioconductor (6) were applied for standardized preprocessing. The Robust Multichip Averaging (7,8) algorithm in the affy conductor package was used for each Affymetrix raw dataset, to calculate the background adjusted, normalized and log2 probe-set intensities. The measure of variability was within the interquartile range (IQR), and a cut-off was set up to remove IQR values under 0.5 for all of the remaining genes. If one gene was targeted for multiple probe sets, we retained the probe set with the largest variability. Pathway analysis of each dataset was performed independently. GSEA was performed using the category version 2.10.1 package. Gene sets contained more than ten genes were retained, Student’s t-test statistical score was implemented for each pathway and also the mean of the genes was calculated. Additionally, a permutation test was performed 1,000 times to obtain the P value of each pathway.

Meta-analysis

Six datasets were analyzed by independent sample using the Student’s t-test method. Meta-analysis was carried out in SAS 9.42. We calculated the chi-square values of the remaining genes of each database above, and retained the differently expressed genes with chi-square values under 0.05. The selection process was based on the following formula (9)

The Cancer Genome Atlas (TCGA)

We downloaded the clinical data and expression profiles of esophageal carcinoma into TCGA database, and removed the esophageal adenocarcinoma group and the control group data. Then, we analyzed the survival data package of R language, used cox-regression analysis, and adjusted P<0.05 so that we could acquire the significant differential genes and obtain their Kaplan-Meier (K-M) survival curves. These genes, together with the genes found in the meta-analysis above, were then used to screen the most differentially expressed genes and obtain significant pathways.

Gene annotation

While searching the intersection of the 406 genes that were obtained from meta-analysis with the 995 genes obtained from TCGA, 19 genes had statistical differences that could affect survival prognosis, and were consequently screened out. We only selected the genes that have been mapped to any explicit KEGG (10) pathway for the sake of having a better insight of the results from the GSEA and meta-analysis. The consistent genes were annotated by the software, Blast2go. The 19 significant genes were used to obtain the pathways of the KEGG from DAVID Bioinformatics Resources 6.7 and annotated by Blast2go. Finally, we put significant genes into the Su Esophagus 2 Oncomine database (www.oncomine.org) for validation.

Results

Common significant pathways were obtained from six esophageal carcinoma tissue datasets by GSEA. According to the inclusion criteria, there were six datasets containing 116 esophageal carcinoma tissues and 116 normal tissues screened out. They are GSE17351, GSE20347, GSE23400, GSE38129, GSE77861 and GSE100942. More details about the datasets are shown on Table 1. The GSEA method was used on each dataset to identify significantly altered genes and significant common pathways. Tissues used to extract the total RNA were matched by pairs from esophageal carcinoma, and normal tissue that was adjacent to esophageal carcinoma. Genomic profiles were matched in pairs in samples above, which reduced the influence of the multiple factors on GSEA and meta-analysis with the purpose of ensuring the reliability of the conclusions obtained. After performing the GSEA conclusion, we identified four up-regulation pathways as well as four down-regulation pathways from six groups of datasets in the path comparison. The four up-regulation pathways were aminoacyl-tRNA biosynthesis, proteasome, ribosome biogenesis in eukaryotes and homologous recombination. The four down-regulation pathways were histidine metabolism, arachidonic acid metabolism, fatty acid degradation and valine, leucine and isoleucine degradation (Table 2). Then, we used the volcano plot method (Figure 1), to roughly screen the genes for the first time.

Table 2. Summary of six datasets using GSEA.

Up-regulation pathway

Aminoacyl-tRNA biosynthesis

Proteasome

Ribosome biogenesis in eukaryotes

Homologous recombination

Low-regulation pathway

Histidine metabolism

Arachidonic acid metabolism

Fatty acid degradation

Valine, leucine and isoleucine degradation

GSEA, gene set enrichment analysis.

Figure 1.

Volcano plot of six datasets.

Conclusion of the meta-analysis

Due to the differences between experimental platforms and the different choices of samples, standardization methods and analysis methods, it’s not surprising that different experimental platforms yield differing data. Taking the results of multiple independent studies on the same subject as the object of the research and based on strict design, the meta-analysis uses appropriate statistical methods to comprehensively analyze multiple research results that are both objective and quantitative. The advantage of meta-analysis is that by using this method, we can improve the reliability of conclusions and reduce the inconsistencies in the results of study by increasing the sample content. With an independent sample, Student’s t-test, six datasets were analyzed, and we obtained the P value of each gene, retaining only the differently expressed genes with chi-square values under 0.05. We inserted the genes’ names into the software SAS 9.42 for statistical analysis, and used the selected meta-formula for integration analysis. Thus, a total of 406 genes were screened out (P<0.05).

Conclusion of the TCGA

We downloaded the clinical data and expression profiles of esophageal tumor into TCGA database, removed the esophageal adenocarcinoma group and the control group data, used the survival data package of R language, used cox-regression analysis, and adjusted P’ value <0.05 (P’ value stands for the P value in TCGA) so that a total of 995 significant differential genes and their K-M survival curves were obtained. Although these genes are all confirmed to be related to prognosis, they have different effects on the prognosis of patients. To improve the patient’s prognosis, we compared these genes with those found in the meta-analysis mentioned above, allowing us to identify the genes that play an important role in the development of ESCC, as well as the ones that are closely related to the prognosis. By studying the pathways where these genes are located, we can better understand the mechanism of ESCC and provide a targeted research direction for the future improvement of patient prognosis.

Gene annotation for key genes

By taking the intersection of the 406 meta-analysis and the 995 TCGA databases, we selected a total of 19 common genes to be screened from the meta-analysis and TCGA databases. The 19 genes had statistical differences and could affect survival prognosis. The names, P’ and P value of the 19 common genes are shown in Table 3. In addition, we drew the Venn Diagram (Figure 2) to visually display the process of taking the intersection. Furthermore, we downloaded the gene network map of 19 key genes from the String database (Figure 3) to better demonstrate the relationship among these genes. As for these 19 genes, we performed a GO annotation (Figure 4) and looked for their common KEGG pathways by DAVID. We utilized Blast2go for gene annotation, and the result was that 19 genes were divided into three parts named the Cellular Component, the Molecular Function and the Biological Process. We found two common KEGG pathways from DAVID Bioinformatics Resources 6.7, which were the Cell cycle and the RNA transport pathways, with each pathway having three key genes. These six key genes’ K-M survival curves were obtained (Figure 5). The cell cycle pathway contained the following three genes (Figure 6): BUB1B (BUB1 mitotic checkpoint serine/threonine kinase B), BUB1 (BUB1 mitotic checkpoint serine/threonine kinase), and TTK (TTK protein kinase). The RNA transport pathway contained the following three genes (Figure 7): NDC1 (NDC1 transmembrane nucleoporin), NUP107 (nucleoporin 107), and NUP155 (nucleoporin 155).

Table 3. Common genes screened from the meta-analysis and TCGA databases.

Gene name	Location	P value	P’ value	Type
NUP155	5p13.2	1.22E-15	0.018650	Up
ASAP2	2p25.1; 2p24	0.001135	0.030007	Up
RAD54B	8q22.1	6.98E-13	0.020964	Up
SH3GLB2	9q34.11	0.00E+00	0.016012	Down
CLDN10	13q32.1	0.002535	6.15E-04	Down
ACPP	3q22.1	2.31E-13	0.038153	Down
ABCA12	2q35	0.038204	0.040447	Up
CHMP2A	19q13.43	1.71E-09	0.010392	Down
DNMT1	19p13.2	2.22E-16	0.026660	Up
VARS	6p21.33	0.034115	0.021041	Up
EXTL2	1p21.2	0.033779	0.029618	Up
MSH2	2p21-p16.3	1.19E-06	0.043425	Up
BUB1B	15q15.1	5.55E-16	0.031915	Up
SNRNP40	1p35.2	0.005732	0.048899	Up
IRF9	14q12	7.23E-05	0.040779	Up
TTK	6q14.1	1.94E-13	0.010229	Up
NDC1	1p32.3	9.18E-09	0.031411	Up
BUB1	2q13	0.00E+00	0.016732	Up
NUP107	12q15	0.00E+00	0.006171	Up

P’ value, P value in TCGA. TCGA, The Cancer Genome Atlas.

Figure 2.

Venn chart of 19 common genes.

Figure 3.

A gene network map of 19 key genes from the String database.

Figure 4.

GO annotation of 19 common genes. GO, gene ontology.

Figure 5.

The K-M survival curves of six key genes. K-M, Kaplan-Meier.

Figure 6.

Cell cycle pathway (the chart is from the KEGG database, gene symbolized by ★ and correlation P values can be found in Table 3).

Figure 7.

RNA transport pathway (the chart is from the KEGG database, gene symbolized by ★ and correlation P values can be found in Table 3).

Finally, we put these six key genes into the Su Esophagus 2 database in Oncomine for validation. The six key genes were found to be highly expressed, with statistical difference (Figure 8).

Figure 8.

The boxplots of 6 key genes from Su Esophagus 2 database in Oncomine.

Discussion

Currently, analysis of gene chip data is an important part of the study of tumor-related genes. The international esophageal carcinoma situation is severe, and the lack of esophageal cancer gene chip data analysis makes the pathogenesis of esophageal a poorly understood phenomenon. We combined GSEA and meta-analysis to analyze six datasets and compared them with TCGA database intending to find important genes and pathways that affect the development and prognosis of ESCC, as well as to reveal the role of these genes in the pathogenesis of esophageal carcinoma.

It’s difficult for researchers to gain enough samples of a disease, so many studies only focus on a small sample that are available. If analysis is only for a single experimental result, and limited to a single gene, a lot of useful information will be missed. Moreover, restricted by sample size, the Student’s t-test of the gene chip has a certain limitation, which leads to the estimation of untrustworthy variants, resulting in higher false positives, and the ignoring of different levels of expression from different samples (11). Gene Set Enrichment Analysis (GSEA) is a computational method that assesses whether a priori defined set of genes shows statistically significant and concordant differences between two biological states, and determines the presence or absence of a common expression by analyzing data of two different biological states (e.g., normal and cancerous) to infer the genes or pathways associated with the disease (4). We performed a meta-analysis with the version 9.42 SAS software, limited P<0.05, to find 406 differently expressed genes. In contrast to only comparing six gene sets to get common genes, which the gene sets have at the same time, meta-analysis avoids the following drawbacks: (I) since the sample size is too small, genes that are not common to six gene sets but are still important may be missed; (II) a simple comparison does not restrict the P value, which may result in statistical bias.

Based on the consideration of the following two reasons, we did not use 406 differently expressed genes from the meta-analysis as the final result: (I) the number and range of these genes are too big or wide, which makes it laborious to go through them in detail; (II) by meta-analysis alone, it is impossible to figure out whether the role the 406 genes play in the development of ESCC is a key gene or just a biomarker. In TCGA database, a total of 995 significant differential genes and their K-M survival curves were obtained. While searching the intersection of the 995 genes and the 406 genes that were obtained through meta-analysis, 19 genes that had statistical differences and could affect survival prognosis were screened out. The 19 genes were put into DAVID for GO annotation and for finding the common KEGG pathway. Unfortunately, the two pathways did not overlap with the eight pathways found by the GSEA. However, these two pathways and their genes may be closely related to ESCC. The two common pathways were the cell cycle and RNA transport, each pathway having three related genes. The three genes on cell cycle were BUB1B (BUB1 mitotic checkpoint serine/threonine kinase B), BUB1 (BUB1 mitotic checkpoint serine/threonine kinase), and TTK (TTK protein kinase). The three genes on RNA transport were NDC1 (NDC1 transmembrane nucleoporin), NUP107 (nucleoporin 107), and NUP155 (nucleoporin 155). The respective discussion of the genes of each pathway follows below.

Interestingly, BUB1B and BUB1 have a common effect in the cell cycle. In a variety of cancer tissues including ESCC, the expression of BUB1B was significantly higher than in adjacent normal tissues. Tumor BUB1B was significantly reduced after chemoradiotherapy. In ESCC-related pharmacological studies, high levels of BUB1B are less sensitive to the parenteral drugs, paclitaxel and nocodazole (12). For the second chemoradiotherapy of recurrent and metastatic esophageal cancer, the potential efficacy of taxanes is reduced, due to the detection of BUB1B (13). As an important molecule in the formation of mitotic spindles, the application value of BUB1B as a drug target or biomarker in the diagnosis and treatment of hepatocellular carcinoma and primitive neuroectodermal tumors is constantly being explored (14,15). Clinically, studies have found that BUB1B expression is associated with a poor prognosis in patients with glioblastoma multiforme (GBM). BUB1B promotes tumor proliferation and induces radioimmunoassay of GBM, and BUB1B may confer to aggressive and effective drugs. This reaction provides a predictive marker. The BUB1BR/S classification of GBM tumors can predict the clinical course and sensitivity to drug treatment (16,17). The ZWINT gene is highly expressed in a variety of cancers, including esophageal cancer. Studies have shown that BUB1B and BUB1 may be important components of the ZWINT mitotic checkpoint for lung cancer (18). Although there is no clear application of BUB1B as a target for clinical diagnosis or treatment, the above studies have pointed out the direction for the application of BUB1B in ESCC.

During apoptosis, BUB1 is cleaved and altered expression is associated with treatment failure, and death in a variety of cancer patients (19). BUB1 is carcinogenic; not only does it regulate the cell cycle, but also may be involved in cytoskeletal control, and Aurora B is a key target for overexpression of Bub1-driven aneuploidy and tumorigenesis (20). In pharmacological research, dipyridamole can increase the concentration of various anticancer drugs represented by 5-fluorouracil in cancer cells, thereby improving the efficacy of the treatment of cancer. BUB1 may then be a molecular target for the action of dipyridamole (21).

Many studies have shown that TTK can affect the development of breast cancer (22), melanin (23), pancreatic ductal adenocarcinoma (PDAC) (24), hepatocellular carcinoma (25) and other cancers. TTK protein kinase is up-regulated in ESCC, which may be involved in the tumor progression and/or represent ESCC-specific properties. In the clinical trials of different drugs for ESCC treatment, the expression of the TTK antigen was observed, which is an important indicator for observing the clinical response of patients to drugs, and also as a cancer vaccination for ESCC in the experiment (26,27). The above evidence provides direction for an individualized treatment of ESCC (28).

Among the RNA transport pathways, the three genes are essential components of the nuclear pore complex. These three genes are rarely used in ESCC, but have a certain role in many other diseases. NDC1-mediated ALADIN localization of the nuclear pore complexes is critical for selective nuclear protein introduction (29). The research on NDC1 and Nup107 is mostly about the effects of these two genes on sperm formation and infertility (30,31). The central domain of Nup107 interferes with Apaf-1 nuclear translocation during genotoxic stress, leading to Chk-1 activation, and a significant reduction in cell cycle arrest (32). In ovarian cancer, Nup107 is associated with platinum sensitivity. The single nucleotide variant (SNV) is significantly associated with platinum resistance, and can be used to clinically predict patient response to drugs (33). NUP107 is considered to be a candidate gene for the detection of nephrotic syndrome and families of developmental delays (34,35). Based on high-throughput techniques, NUP107 is a potential molecular marker for early diagnosis of PDAC (36). NUP155 is specific in many tissues, but its specificity in esophageal cancer tissues has not been tested. As a nuclear pore complex protein, NUP155 has been identified as a clinical driver of atrial fibrillation (37,38) and NUP155 may be a new potential drug to target NUP214-ABL1-positive T-cell acute lymphoblastic leukemia (39). Some studies have explored the role that the three genes play on RNA transport pathways; however, little of this research explains the relationship between these genes and cancer, let alone their relationship with ESCC.

In summary, there are a few studies that illuminate the relationship between these genes and esophageal cancer. Currently, we only have information related to genes and pathways, extracted through data analysis, rather than experimental study. We plan to experimentally verify our conjecture about the mechanism of action and the specific functions of these genes, and discuss them in depth, so that we can further reveal their important role in ESCC and provide a theoretical basis for clinical prevention, diagnosis, and individualized treatment of ESCC.