Pathway, in silico and tissue-specific expression quantitative analyses of oesophageal squamous cell carcinoma genome-wide association studies data

Abstract

Background: Oesophageal cancer is the fourth leading cause of cancer death in China where essentially all cases are histologically oesophageal squamous cell carcinoma (ESCC). Agnostic pathway-based analyses of genome-wide association study (GWAS) data combined with tissue-specific expression quantitative trait loci (eQTL) analysis and publicly available functional data can identify biological pathways and/or genes enriched with functionally-relevant disease-associated variants.

Method : We used the adaptive multilocus joint test to analyse 1827 pathways containing 6060 genes using GWAS data from 1942 ESCC cases and 2111 controls with Chinese ancestry. We examined the function of risk alleles using in silico and eQTL analyses in oesophageal tissues.

Results: Associations with ESCC risk were observed for 36 pathways predominantly involved in apoptosis, cell cycle regulation and DNA repair and containing known GWAS-associated genes. After excluding genes with previous GWAS signals, candidate pathways (and genes) for ESCC risk included taste transduction (KEGG_hsa04742; TAS2R13, TAS2R42, TAS2R14, TAS2R46,TAS2R50 ), long-patch base excision repair (Reactome_pid; POLD2 ) and the metabolics pathway (KEGG_hsa01100; MTAP , GAPDH , DCTD , POLD2, AMDHD1 ). We identified and validated CASP8 rs13016963 and IDH2 rs11630814 as eQTLs, and CASP8 rs3769823 and IDH2 rs4561444 as the potential functional variants in high-linkage disequilibrium with these single nucleotide polymorphisms (SNPs), respectively. Further, IDH2 mRNA levels were down-regulated in ESCC (tumour:normal-fold change = 0.69, P  = 6.75E-14).

Conclusion: Agnostic pathway-based analyses and integration of multiple types of functional data provide new evidence for the contribution of genes in taste transduction and metabolism to ESCC susceptibility, and for the functionality of both established and new ESCC risk-related SNPs.

Key Messages

• We carried out an agnostic evaluation of variants in functionally-related genes in established biological pathways and examined tissue-specific expression quantitative trait loci (eQTLs) to provide evidence of functionality and identify true risk variants.

• From 1827 biological pathways and 6600 genes, we found new evidence linking genes in bitter taste transduction, long-patch base excision repair and metabolics pathways with risk of ESCC in Han Chinese.

• We identified eQTL relations for CASP8 (a gene previously-associated with ESCC) in apoptotic pathways, and IDH2 (a gene not previously-associated with ESCC) in the metabolics pathway.

• Although rs13016963 is the strongest CASP8 SNP in GWAS and in this study, eQTL analyses indicated that the highly correlated rs3769823 is more functionally relevant, as it locates to a secondary CASP8 promoter and has the potential to interact with the transcriptional regulator HDAC2.

• Further, whereas rs11630814 was the top SNP in IDH2 , the correlated rs4561444, eQTL, locates to an active enhancer in oesophageal tissue, whereby changes in IDH2 expression/production may attenuate the neutralization of reactive oxygen species and TET2-dependent demethylation of DNA.

• The identification of predisposing genetic factors in genes and pathways as well as functional variants associated with ESCC development may ultimately lead to new diagnostic, prognostic and therapeutic strategies in high-risk populations.

• The possible involvement of HDAC2 in the transcriptional regulation of CASP8 , and IDH2 in the 5-hydroxymethylcytosine generating pathway, suggest potential for targeted epigenetic therapies.

Introduction

Oesophageal cancer (EC) causes more than 400 000 deaths each year and is the sixth leading cause of cancer death worldwide. ^1,2 Morbidity and mortality rates for EC in China are particularly high in northern areas including Shanxi and Henan Provinces where essentially all cases of oesophageal cancer are oesophageal squamous cell carcinoma (ESCC) as opposed to adenocarcinoma, which has become the most common EC histology in the Western world. ³ While smoking tobacco and drinking alcoholic beverages account for nearly 90% of ESCC cases in Western countries including the USA, ^4–6 these exposures explain very little of the disease in high-risk populations in China. ^7–9 Low levels of vitamins and minerals, consumption of pickled food and exposure to nitrosamines ^3,7,8 have been investigated but not convincingly linked to the high incidence, suggesting that these environmental risk factors alone may not be responsible for ESCC. Several lines of evidence for a genetic influence in these high-incidence regions exist, including associations with family history, segregation of disease within families and results from genome-wide association studies (GWAS). ^10–14

GWAS aims to detect single nucleotide polymorphisms (SNPs) associated with trait variation. We and others have identified a number of genetic loci linked to risk of ESCC from GWAS. ^13–18 However, mechanisms that translate these variants into cancer risk have not been delineated. Also, due to the large number of tests, standard analysis techniques impose highly stringent significance thresholds, likely missing many truly associated loci and leaving much of the genetic variation in the trait unexplained. Pathway-based analyses may allow the detection of associations missed by standard single-marker approaches. Such an approach acknowledges the complexity of carcinogenesis, where the expression products of numerous genes interact together and temporally, to form malignant cells. Pathway analysis jointly considers variants in the functionally-related genes of established biological pathways, potentially offering insight into the polygenic basis of cancer susceptibility. ¹⁹ Notably, pathway-based GWAS analyses have provided novel insights into the aetiology of cancers of the breast, pancreas, lung, and melanoma. ^20–23 We recently analysed GWAS data focusing on candidate gene pathways including DNA repair and epidermal growth factor receptor (EGFR) signalling. ^24,25 Although these data support the usefulness of this approach, previous analyses were limited by the number of pathways/genes examined and preconceived notions of which pathways may be important. A comprehensive agnostic analysis of all known pathways may uncover novel genes that contribute to ESCC susceptibility. Moreover, recent studies have profiled mRNA transcripts and genotyped SNPs, resulting in the detection of SNPs associated with gene expression [termed expression quantitative trait loci (eQTLs)] in disease-relevant tissues that can be further considered to aid in sorting out functional relevance. ²⁶ In the current study we conducted pathway, eQTL and in silico functional analyses using data from previously published ESCC GWAS ¹³ and expression studies ²⁷ coupled with publicly-available data from the NIH Roadmap Epigenomics, the Encyclopedia of DNA Elements (ENCODE) and the NIH Genotype-Tissue Expression (GTEx) projects.

Methods

Study population

This study reports agnostic analyses of data from a GWAS of ESCC conducted in ethnic Chinese; full details have been described elsewhere. ¹³ Briefly, participants were drawn from two studies, the Shanxi Upper Gastrointestinal Cancer Genetics Project (Shanxi) and the Linxian Nutrition Intervention Trial (NIT), a prospective cohort study.

Genotyping, quality control, and exclusions

Generated GWAS data ¹³ are available upon request from the NIH Data Access Committee [ http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000361.v1.p1 ]. An overall subject completion rate of 85% was applied to cases and controls in the combined population for all assays analysed. We excluded SNPs with <98% completion and <95% concordance, a minor allele frequelncy (MAF) <5% and a Hardy–Weinberg equilibrium P -value < 1.0E-6.

Pathway construction

The overall pathway data source was built and updated at DCEG-NCI in 2011 (available upon request). Pathway data were retrieved from five resources or pathway catalogues: BioCarta [ http://www.biocarta.com/genes/index.asp ]; HumanCyc [ http://humancyc.org/ ]; the Kyoto Encyclopedia of Genes and Genomes database [KEGG] [ http://www.genome.jp/kegg/ ]; the NCI-Nature Curated database [NCI-PID] [ http://pid.nci.nih.gov/download.shtml ]; and Reactome [ http://www.reactome.org/ ]. The Reactome database contains both Reactome_cpath [systems biology] and Reactome_pid designated pathways. The five sources contained a total of 2814 pathways. To exclude small functional categories, we only analysed pathways with five or more genes. Therefore, the total number of pathways analysed was 1827. Autosomal SNPs were mapped within a genomic region encompassing 20 Kb 5′ upstream and 10 Kb 3′ downstream of a coding gene contained in the pathway. For a number of large pathways such as the KEGG_hsa01100 metabolics pathway, which contains 972 genes, we also identified sub-pathways and/or specific gene sets using Pathway Studio 9.0 (Ariadne Genomic, Rockville, MD) [ http://www.ariadnegenomics.com/products/pathway-studio/ ] and Ingenuity Pathway Analysis (IPA) (Ingenuity®Systems at: http://www.ingenuity.com ).

Statistical analyses

To investigate variation in the genes of each pathway, we conducted individual SNP-, gene- and pathway-based analyses. SNP-based analyses were tested under the additive model, and odds ratios (ORs) and 95% confidence intervals (CIs) were calculated using unconditional logistic regression with adjustment for age (10-year categories), sex and study. We adjusted for population stratification using the top five eigenvectors. All P -values for SNPs are nominal except where otherwise specified. SNP-based analyses were performed using R program language [ http://www.r-project.org/ ].

For the gene-based analyses, the associations between candidate genes and ESCC were assessed using the AdaJoint method [ http://dceg.cancer.gov/tools/analysis/adajoint/ ] as previously described. ²⁸ Pathway analyses were based on the adaptive rank truncated product (ARTP) ²⁹ method which combines P -values from the AdaJoint method across all relevant genes. Statistical associations for the gene- and pathway-based tests were determined through use of 10 million resampling steps as previously described. ²⁹ The ARTP method adjusts for the size of genes (and number of SNPs in these genes) and pathways automatically through a resampling procedure. The final P -value is computed from the resampled null distribution. We further used Bonferroni thresholds for multiple testing for genes ( P  < 8.25E-06) and pathways ( P  < 2.74E-05). Since GWAS SNPs in specific genes have previously been associated with ESCC risk in this study population, we also excluded these signals and genes and then repeated our pathway-based analyses for all 1827 pathways.

In silico functional and eQTL annotations

We used custom tracks on the UCSC Genome browser [ http://genome.ucsc.edu ] to screen Roadmap Epigenomics [ http://www.roadmapepigenomics.org/ ] and ENCODE ³⁰ data containing the SNP regions of interest for evidence of regulatory relevance ^31,32 in oesophageal tissues. We also used the online tools HaploReg v2 [ http://www.broadinstitute.org/mammals/haploreg/haploreg.php ] and RegulomeDB [ http://regulome.stanford.edu ] as complementary analyses and to confirm the location of each SNP in relation to protein-coding and/or non-coding RNA (ncRNA) genes. To evaluate the potential of a particular SNP to affect oesophageal mRNA expression, genotypes were correlated with gene mRNA levels in adjacent histologically normal squamous oesophageal tissues from 100 ESCC cases with available GWAS data. ²⁷ Paired tumour and normal oesophageal mucosal tissues distant to tumour(s) were collected from Shanxi cases during surgery and all cases were pathologically confirmed as ESCC. Spearman rank correlations were used for eQTLs testing and only correlations surviving Bonferroni thresholds for multiple testing were considered. Array probes for genes of interest were identified using GeneAnnot annotation of Affymetrix probe sets from HG-U133A [ http://genecards.weizmann.ac.il/geneannot/index.shtml ]. We used data from the NIH GTEx project and portal (dbGaP release phs000424. v4 .p1, December 2014) for external validation and to evaluate tissue specificity of the genetic influence [ http://www.gtexportal.org/home/ ]. The Affymetrix U133A expression array does not contain ncRNA probes, so we used available GTEx data to evaluate annotated risk SNPs as potential ncRNA eQTLs.

We also compared expression levels in ESCC tumour tissue with that of matched normal squamous oesophageal tissue for 133 cases using Affymetrix U133A array (GEO Accession number GSE23400), as described previously. ²⁷ We used linear regression to evaluate the gene expression of 22 277 probes in ESCC tumour compared with normal squamous oesophagus. Fold-change (fc) was defined as 2β, where log ₂ expression = α + β × tumour status. P -values were adjusted for multiple comparisons using Bonferroni correction and were considered associated only if they reached threshold of P  < 2.24E-6. All statistical analyses were conducted using R [ http://www.r-project.org/ ].

Results

Population characteristics

A total of 1942 ESCC cases and 2111 controls from the combined studies were analysed. Demographic and risk factor information for both individual studies and the combined population are shown in Supplementary Table 1 (available as Supplementary data at IJE online).

Pathway-based analyses

In our GWAS data, we identified a total of 82 772 unique SNPs in 6060 genes [Reference Sequence (RefSeq) hg 19] contained within 1827 pathways. The distribution of the pathways by source is summarized in Supplementary Figure 1 (available as Supplementary data at IJE online). Variation between databases was evident even in pathways that represented the same biological processes. For example, the base-excision repair pathway included 35 genes in KEGG but only 18 genes in Reactome_cpath; likewise, the tumour necrosis factor (TNF) receptor signalling pathway included 36 genes in NCI-PID but only 7 genes in Reactome_pid.

Of the 1827 examined pathways, 289 candidate pathways at a P  < 0.05 level were identified. Of these 289, 157 came from a single source pathway and 132 were found in more than one source pathway. For example, we observed associations for apoptotic pathways from Reactome_pid (7 genes, P  = 1.00E-07), BioCarta (22 genes, P  = 3.50E-06) and Reactome_cpath (11 genes, P  = 4.80E-06). Supplementary Table 2 (available as Supplementary data at IJE online) shows the top 50 pathways (i.e. those with the lowest P -values) and their associated genes. The 36 pathways that survived Bonferroni correction (0.05/1827 =  P  < 2.74E-05) for multiple comparisons included independent and overlapping pathways for p53 signalling, apoptosis, cell cycle regulation, co-regulation of androgen receptor activity, pathways in cancer and base excision repair ( Figure 1 ).

Figure 1.

Thirty six agnostic pathways associated with ESCC risk and ranked by the smallest (top) to the largest pathway P -value. Pathway source and name, together with the total number of genes in each pathway in parentheses, are indicated on the Y axis, and the proportion of genes in each pathway with a P -value < 0.05 is on the X axis. The generic pathway function or grouping for a specific pathway is noted on the axis to the right of the bar chart, which is colour-coded: red, p53 pathway; yellow, apoptosis; orange, cell cycle control and apoptosis; green, cell cycle control; turquoise, co-regulation of androgen receptor activity; pink, DNA damage response; purple, pathways in cancer; grey, control of translation; and blue, DNA repair.

Gene-based analyses identified 365 ESCC candidate genes ( P  < 0.05) ( Figure 2 ) that were involved in 202 different pathways (KEGG, 86; REACTOME_cpath, 45; NCI-PID, 39; BioCarta_pid, 19; and HumanCyc, 13). Five genes survived correction for multiple comparisons. CASP8 (1.0E-07) had the smallest P -value and was included in 22 separate pathways. Other associated gene P -values were observed for: CHEK2 (8.0E-07), involved in 9 pathways; CDKN2B (3.0E-06), involved in 7 pathways; PLCE1 (4.0E-06), involved in 8 pathways; and XBP1 (7.0E-06), involved in 1 pathway. Figure 2 illustrates the distribution and association of all 6060 analysed genes. Supplementary Table 3 (available as Supplementary data at IJE online) shows SNP associations in genes in the top 50 ESCC candidate pathways including the 36 ESCC-associated that survived correction for multiple comparisons. Of the top 50 pathways, 12 involved cell death or apoptosis, 7 involved base excision repair and 5 involved cell cycle processes. In general, we observed a large overlap in the function of associated pathways, as well as overlap in the identity of top-ranked gene associations within them ( Supplementary Table 3 , available as Supplementary data at IJE online).

Figure 2.

Plot showing the distribution and association of 6060 autosomal candidate genes in 1827 pathways. Examples of ESCC-associated genes and genes containing previously identified GWAS hits are labelled.

Pathway-based analyses after exclusion of genes with known ESCC-association signals

The overlap in pathway function as well as the identity of the gene(s) with the smallest P -values (e.g. CASP8 , PLCE1 , CHEK2 and CDKN2B ) in the 36 ESCC-associated pathways suggested that these pathway results might be driven by previously identified associations signals for ESCC. We therefore excluded all genes containing a main effect GWAS signal based on the recently published joint analysis ¹⁸ as well as ESCC-associated signals at the 9p21 region. ^18,33 Because a SNP may exert a regulatory effect on other neighboring genes ( in cis ), we excluded additional genes that overlapped (including 20 Kb up from transcriptional start site (TSS) and 10 Kb down from 3’ untranslated region (UTR) regions) with the cited/candidate gene for the association signal ( Supplementary Table 4 , available as Supplementary data at IJE online). After excluding these genes ( Supplementary Table 4 ), candidate pathways identified included taste transduction, amino acid metabolism and catabolism, oxidative decarboxylation, glycolysis, nucleotide metabolism, DNA synthesis and repair and RNA Pol II transcription pathways ( Supplementary Table 5 , available as Supplementary data at IJE online). Supplementary Table 5 shows the 50 pathways with the lowest P -values after exclusion of genes ( Supplementary Table 4 ). Of note, the top three pathways included the taste transduction pathway (KEGG_hsa04742), the removal of DNA patch containing abasic residue (Reactome_pid) [now referred to as long-patch base excision repair (BER) (Reactome_pid)] and the metabolics pathway (KEGG_hsa01100) ( Figures 3–5 ). Genes with the smallest P -values ( P  < 0.001) in these pathways were TAS2R13, TAS2R42, TAS2R14, TAS2R46 and TAS2R50 ( Figure 3 ), and MTAP , GAPDH , DCTD , POLD2 and AMDHD1 ( Figures 4 and 5 ). The top SNPs in the TAS2R13 , TAS2R42 , TAS2R46 and TAS2R50 genes of the taste transduction pathway were in high linkage disequilibrium (LD) (r ² > 0.9) ( Supplementary Figure 2 , available as Supplementary data at IJE online), suggesting a single signal in this region may be linked with increased ESCC risk.

Figure 3.

Candidate ESCC genes in the taste transduction pathway (KEGG_hsa04742) after exclusion of genes with known association signals. Top candidate genes ( P  < 0.001, circled in red) included TAS2R13, TAS2R42, TAS2R14, TAS2R46 and TAS2R50 . For all gene P -values, see Supplementary Table 8 (available as Supplementary data at IJE online).

Figure 4.

Candidate ESCC genes in the long-patch base excision repair pathway (Reactome_pid) after exclusion of genes with known association signals. The top candidate gene with the smallest P -value ( P  < 0.001, circled in red) is POLD2 . For all gene P -values, see Supplementary Table 8 (available as Supplementary data at IJE online).

Figure 5.

Candidate ESCC genes in the metabolics pathway (KEGG_hsa01100) after exclusion of genes with known association signals. Top candidate genes with the smallest P -values ( P  < 0.001, circled in red) include MTAP , GAPDH , DCTD , POLD2 and AMDHD1 . For all gene P -values, see Supplementary Table 8 (available as Supplementary data at IJE online). PLCE1 was not included in the pathway analysis and is labelled in the plot only as a reference to indicate that this pathway normally contains this ESCC-associated gene.

Subpathway-based analyses of the metabolics pathway

The metabolics pathway (KEGG_hsa01100) was the largest pathway among the five data sources, and included a total of 972 autosomal genes involved in many catabolic and anabolic reactions ( Supplementary Table 6 , available as Supplementary data at IJE online). Within this pathway, we identified 24 subpathways using Pathway Studio 9.0, and 239 subpathways using IPA. We observed 15 candidate subpathways ( P  < 0.05) within the parent metabolic pathway ( Supplementary Table 6 ) that were enriched with genes involved in the metabolism of specific amino acids (i.e. histidine, valine and tryptophan), glucose, nucleotides, triacylglycerol and D -myo-inositol-trisphosphate, as well as leukotriene and IL-12 signalling subpathways ( P  < 0.05).

In silic o functional annotation of SNPs in genes in ESCC pathways

Based on Roadmap Epigenomics and ENCODE data for normal oesophageal (and other) tissue/cells, SNPs in each of the 42 top genes in the 36 agnostic ESCC-associated pathways were mapped to potential DNA regulatory regions [e.g. TSS, enhancers and DNaseI /open chromatin sites)( Supplementary Table 7 , available as Supplementary data at IJE online). The SNPs rs1063192 ( CDKN2A/CDKN2B ) and rs4913355 ( DYRK2 ) were located proximal to ncRNA genes, and rs13928 ( POLD2) resulted in a missense change in the overlapping AEBP1 gene ( Supplementary Table 7 ). In addition, SNPs in genes in the top three ESCC candidate pathways (i.e. taste transduction, long-patch BER and the metabolics pathway) after exclusion of genes with known ESCC association signals were also mapped ( Supplementary Table 8 , available as Supplementary data at IJE online). The TAS2R rs10772380 (RP11-144O23.18) and rs10875871 (RP11-579D7.1), as well as MTAP rs2764736 ( CDKN2B-AS1 ), were located proximal to ncRNA genes ( Supplementary Table 8 ). The SNPs rs3897926 ( PRHOXNB ) and rs2274976 ( MTHFR ) resulted in missense changes, whereas rs619381 ( TAS2R9 ), rs1129649 ( GNB3 ) and rs2305030 ( ITPKA ) resulted in missense changes in the overlapping TAS2R7 , LEPREL2 and LTK genes, respectively.

eQTL analysis of ESCC-SNPs in normal oesophageal tissue

To identify eQTLs, we examined correlations between top SNPs (and SNPs in LD) and gene mRNA levels using gene expression data from adjacent normal tissue from 100 ESCC cases. We further validated and examined the tissue specificity of any eQTLs identified using the GTEx portal. The differential expression of genes with potential eQTLs was also examined in matched tumour-normal tissues from 133 cases. We observed correlations between genotype and mRNA probe levels of CASP8 (rs13016963), TP73 (rs3765705) and SNURF (rs220030) genes in the top 36 agnostic pathways ( Supplementary Table 9 , available as Supplementary data at IJE online). In particular, we observed a negative correlation for variant A of the ESCC risk SNP rs13016963 (ESCC OR = 1.26, 95% CI = 1.14-1.39, P  = 4.36E-06) and CASP8 mRNA levels using two independent probes 213373_s_at ( rho  = −0.36, P  = 1.9E-04) and 207686_s_at ( rho  = -0.20, P  = 4.7E-02) ( Figure 6 A and Supplementary Table 9 ). The former remained associated after Bonfferoni correction for multiple comparisons (0.05/85 tests, P  < 5.9E-04). We did not observe a strong decrease in CASP8 mRNA in ESCC tumour compared with normal tissues (data not shown). For the overall pathway analysis (i.e. before GWAS gene exclusions), genes such as PLCE1, ATP1B2, HEATR3, PDE4D, RUNXI, ST6GAL1 and TMEM173 with known ESCC-associated signals were not contained in our top 36 pathways, but were in our top 50 pathways. Therefore, we have included functional annotations and eQTL results for these ESCC-association signals and others in Supplementary Table 10 (available as Supplementary data at IJE online) for reference. Of note, known GWAS signals HEATR3 rs4785204 and TMEM173 rs7447927 were identified as eQTLs. Further, TMEM173 rs7447927 also located to a promoter region in oesophageal tissue.

Figure 6.

A. Expression quantitative trait loci (eQTL) analysis between rs13016963 (risk allele = A) and rs3769823 (risk allele = T) genotypes and CASP8 mRNA expression in normal oesophageal squamous tissues from 100 ESCC cases. CASP8 mRNA levels were assessed using two independent Affymetrix_U133A probes: 213373_s_at and 207686_s_at from the Affymetrix U133A platform. Expression using 213373_s_at is shown here. B. Genome Browser [ http://genome.ucsc.edu/ ] image of CASP8 gene region on human assembly hg19 based on NIH Roadmap Epigenomics data [ http://www.roadmapepigenomics.org/ ] and ENCODE data. ^31,32 The position of the Affymetrix_U133A probes 207686_s_at and 213373_s_at are indicated. RefSeq profile of CASP8 and ALS2CR12 mRNA transcripts. Regulatory domains [chromatin state segmentation using a hidden Markov Model (ChromHMM)] and core histone marks: crimson, flanking TSS; red, active transcriptional start site (TSS); dark green: transcription elongation/transition; yellow green: transcription enhancer-like; orange, active enhancer. DNA methylation profile of normal esophageal tissue (NIH Roadmap) using bisulphite sequencing (BS). SNP locations (dbSNP 141) for CASP8 SNPs rs3769823, rs10931926, rs13016963, rs9288318 and rs10201587. CASP8 rs3769823 locates to a secondary promoter (red) and rs10931936 locates to an enhancer downstream of ALS2CR12 (orange). For clarity, all SNPs contained within this region (dbSNP) are not shown. DNase I: open chromatin DNase I hypersensitivity clusters in 125 cell types from ENCODE [ https://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg19&g=wgEncodeAwgDnaseUniform ]. TF (transcription factor): ChIP-seq from ENCODE with Factorbook Motifs (green-coloured bars in black rectangles). Details of TFs and cell lines tested can be found at: [ https://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg19&g=wgEncodeAwgTfbsUniform ]. For clarity, only proteins interacting with the DNA region of interest are shown. Sources and acknowledgements for UCSC genome browser data ³² and extracted tracks can be found at: [ http://genome.ucsc.edu/goldenPath/credits.html#human_credits ].

We did not observe eQTLs for TAS2R9 , TAS2R13, TAS2R14, TAS2R42, TAS2R46 , TAS2R50, ADCY6, GNB3 or CACNA1A from the taste transduction pathway (KEGG_hsa04742) nor POLD2, FEN , or POLD1 from the long-patch BER (Reactome_pid) pathway ( Supplementary Table 11 , available as Supplementary data at IJE online). The SNP rs174556 in FEN1 from the long-patch BER pathway also maps to the intron of FADS1 , and this SNP showed a negative influence on FADS1 mRNA levels in normal oesophageal tissue ( Supplementary Table 11 ).

In contrast, we observed a number of correlations between mRNA levels and SNP genotypes of IDH2 , GALNT3 , ACADSB and DGKQ genes from the KEGG_hsa01100 metabolics pathway ( Supplementary Table 11 ). In particular, we observed a negative correlation for variant G of rs11630814 (ESCC OR = 1.16, 95% CI = 1.05-1.16, P  = 4.1E-03) with IDH2 mRNA levels using two independent probes 210046_s_at ( rho  = -0.39, P  = 8.27E-05) and 210045_at ( rho  = -0.23, P  = 2.45E-02) ( Figure 7 A and Supplementary Table 11 ); the former remained associated after Bonferroni correction for multiple comparisons (0.05/101 tests, P  < 5.0E-05). Further, IDH2 mRNA was down-regulated using both of these probes in ESCC vs normal oesophageal tissues from 133 cases ( Supplementary Table 12 ). We further validated the CASP8 and IDH2 eQTLs using the GTEx portal ( Supplementary Tables 12 and Supplementary Data , available as Supplementary data at IJE online). Lastly, 10 candidate ESCC SNPs were annotated to nine different long ncRNA (lncRNA) genes; however, no SNPs were shown to be lncRNA eQTLs in oesophageal mucosa tissue using GTEx data (data not shown).

Figure 7.

A. Expression quantitative trait loci (eQTL) analysis between rs11630814 (risk allele = G) and rs456144 (risk allele = T) genotypes and IDH2 mRNA expression in normal oesophageal squamous tissues from 100 ESCC cases (one individual had no genotype data for rs11630814). IDH2 mRNA levels were assessed using two independent Affymetrix_U133A probes; 210046_s_at and 210045_at from the Affymetrix U133A platform and data are shown for 210046_s_at. B. Genome Browser [ http://genome.ucsc.edu/ ] image of IDH2 gene region on human assembly hg19 based on NIH Roadmap Epigenomics data [ http://www.roadmapepigenomics.org/ ] and ENCODE data. ^31,32 The position of Affymetrix_U133A probes 210045_at and 210046_s_at, which detect variant mRNAs 1 and 2 of IDH2 , are indicated. RefSeq profile of IDH2 mRNA transcripts. Regulatory domains [chromatin state segmentation using a hidden Markov Model (ChromHMM)] and core histone marks: crimson, flanking TSS; red, active transcriptional start site (TSS); dark green: transcription elongation/transition; yellow green: transcription enhancer-like; orange, active-to-weak enhancer. The promoter CpG island of IDH2 is highlighted by a green-filled box. DNA methylation profile of normal oesophageal tissue (NIH Roadmap) using bisulphite sequencing (BS). IDH2 rs4561444, which locates to an active enhancer (orange) upstream of the IDH2 promoter (red) is highlighted by a pink-coloured box. For clarity, not all SNPs contained in this region (dbSNP) are shown. DNaseI: open chromatin DNase I hypersensitivity clusters in 125 cell types from ENCODE [ https://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg19&g=wgEncodeAwgDnaseUniform ]. TF (transcription factor): ChIP-seq from ENCODE with Factorbook Motifs (green-coloured bars in black rectangles). Only proteins interacting with the DNA region of interest are shown. Details of TFs and cell lines tested can be found at: [ https://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg19&g=wgEncodeAwgTfbsUniform ]. RNA-seq analysis of normal oesophageal tissue (NIH Roadmap data). Sources and acknowledgements for UCSC genome browser data ³² and extracted tracks can be found at: [ http://genome.ucsc.edu/goldenPath/credits.html#human_credits ].

CASP8 rs13016963 and IDH2 rs11630814 eQTLs are linked to the underlying functional variant

In silico functional analysis indicated that the DNA regions containing CASP8 rs13016963 and IDH2 rs11630814 did not map to potential regulatory regions in normal oesophageal tissues ( Supplementary Tables 7 and Supplementary Data ). To examine whether rs13016963 and rs11630814 might be proxies for the functional SNP, we used controls from our own population to compare LD structure, and mapped all correlated SNPs (r ² > 0.5) to each identified eQTL. CASP8 rs13016963 was previously shown to exist in a single risk haplotype with rs3769823, rs10931936, rs9288318 and rs10201587 ¹⁴ and these SNPs were also genotyped in the current study ( Supplementary Table 13 ). We observed that the DNA region containing rs3769823 (missense K14R) mapped to an active promoter for the CASP8 mRNA transcript G ( ENST00000358485 ) and a non-coding transcript in normal oesophageal tissues ( Figure 6 B). According to PolyPhen, K14R is benign, suggesting it has little impact on the structure and function of the CASP8 protein isoform, but the variant T of rs3769823 can alter a number of DNA motifs including STAT3, and the overlapping DNA region interacts with a number of proteins including YY1 and the HDAC2 complex (Haploreg v2). CASP8 rs10931936 also mapped to a strong enhancer downstream of ALS2CR12 ( Figure 6 B); however, oesophageal tissues appeared to express little ALS2CR12 mRNA ( Figure 6 B). The remaining SNPs in the haplotype (including the top association signal rs13016963) did not map to potential promoter or enhancer regulatory regions ( Figure 6 B). IDH2 rs11630814 was in LD (r ² = 0.67) only with rs4561444 in this study population ( Supplementary Table 12 and Figure 7 A). The DNA region containing rs4561444 maps to an active enhancer upstream of the IDH2 promoter ( Figure 7 B). The variant T SNP can alter the DNA binding motif of Zfx (a key activator of IDH2 ) as well as AP-1 and BDP1 (Haploreg v2). We also observed correlations between the genotypes of the LD SNP(s) with CASP8 and IDH2 mRNA levels, respectively in our own data and in GTEx ( Supplementary Tables 12 and Supplementary Data ). An evaluation of tissue-specificity for these eQTLs using GTEx data is presented in Supplementary Table 14 (available as Supplementary data at IJE online).

Discussion

In this study, we examined the cumulative variation of polymorphisms in multiple genes acting in functional pathways in order to provide a complementary approach to the more commonly conducted single SNP association studies to understand genetic susceptibility to ESCC in a high-risk Chinese population. We further leveraged information from gene expression data from the same population as well as data from in silico sources for oesophageal tissues to identify eQTLs and characterize SNP functionality in genes in specific pathways associated with ESCC.

Initial pathway analysis of all GWAS data found 36 pathways highly associated with ESCC risk, whose functions were predominantly apoptosis, cell cycle regulation and DNA repair. These pathways were enriched with genes containing previously identified GWAS association signals. By examining the top SNPs in genes in these pathways, we identified and validated CASP8 rs13016963 as an eQTL with a negative effect on CASP8 mRNA levels in oesophageal tissues. We found no evidence of a potential regulatory function for rs13016963, but we did observe that the closely correlated SNP rs3769823 mapped to a secondary promoter region for a specific CASP8 mRNA in normal oesophageal tissues, whose activity has recently been confirmed in oesophageal cells in vitro . ³⁴ The G mRNA transcript of CASP8 encodes the initiator proCASP8L isoform, which is recruited to the BAP31 complex in the endoplasmic reticulum in response to apoptotic signalling in vitro . ³⁵CASP8 rs13016963 (A) and rs3769823 (T) variants are associated with increased risk of ESCC in our study population, suggesting that low basal expression/activation of proCASP8L and reduced downstream activation of caspases and cell death may be involved in susceptibility. Interestingly, the rs3769823 (T) containing region can interact with the HDAC2 complex (Haploreg v2), which can repress transcription and is an important target for inhibition in cancer therapies. Our data also suggests that CASP8 rs13016963 and/or rs3769823 can reduce CASP8 mRNA expression in a number of other tissues, including skin. Furthermore, rs13016963 has been associated with risk of melanoma in GWAS ³⁶ and rs3769823 also maps to a secondary promoter region in normal melanocytes from the NIH Roadmap Epigenomics project (data not shown). Testing the expression of CASP8 mRNA transcripts as well as the apoptotic function of proCASP8L in normal oesophageal tissues, and implications for ESCC prevention or therapy, remain interesting biological questions for the future.

While this paper was being prepared, Yang et al . ³⁷ published a brief report on an ESCC pathway analysis (based on missense SNPs and SNPs located to gene coding regions) which used Gene Ontology (GO) as the main pathway data source. In agreement with our current results, Yang et al . ³⁷ reported that CASP8 rs376982 and not rs13016963 was the more likely causal ESCC SNP, but the authors did not describe the regulatory mechanism linking CASP8 rs376982 to ESCC susceptibility or the associated eQTL influence on CASP8 mRNA expression in oesophageal tissues that we describe here. We do not observe any potential for the other proposed candidate causal SNPs described ³⁷ to influence gene expression or to act as regulatory SNPs at the genomic level in oesophageal tissues, using in silico and publicly available online resources ( Supplementary Table 15 , available as Supplementary data at IJE online).

By excluding genes containing previously identified ESCC association signals, we also identified a number of novel candidate pathways, the top three of which were taste transduction (KEGG_hsa04742), long-patch BER (Reactome_pid) and the metabolics pathways (KEGG_hsa01100). In the taste transduction pathway, the genes predominantly responsible for the observed pathway effect were the bitter taste receptor ( TAS2R ) genes ( TAS2R9 , TAS2R13, TAS2R14, TAS2R42, TAS2R46 and TAS2R50 ), ADCY6, GNB3 and CACNA1A. In humans, 25 TAS2R genes are present in gene clusters on chromosomes 5, 7 and 12. ³⁸ They encode chemosensory receptors that discriminate bitter taste and are activated by hundreds of chemical molecules including alcohol, tobacco and pesticides. ^38–41 Moreover, TAS2Rs are expressed not only in the oral cavity ⁴² but also in cells in the gastrointestinal (GI) tract, lungs and breast. ^43–46 The capability to discriminate bitter taste has evolved as a central warning signal against the ingestion of possible toxic substances and as an important system to start hormonal and/or neural pathways leading to the regulation of caloric intake, insulin secretion and metabolism. ⁴⁶ Specific TAS2Rs are also active and differentially expressed in mammary epithelial cells and cancer cells in vitro , suggesting tissue-specificity for the receptors and dysregulation in the development and/or progression of cancer. ⁴⁵

Only two candidate gene studies have examined the association between variants in TAS2R genes and cancer; both studies investigated colorectal cancer risk in Caucasian populations and observed null associations. ^47,48 No study to date has examined associations at the gene or pathway level for taste transduction and cancer risk. Epigenomic data from the NIH Roadmap project showed that TAS2R13 rs10772380 maps to a histone activation mark and TAS2R9 rs619381 maps to a potential repressed polycomb region in normal oesophageal tissue; however, we found no evidence that either variant influenced mRNA expression, nor did genotypic variation in TAS2R14 . The identification of the target gene(s) of a potential regulatory variant is, however, complicated by the fact that regulatory regions may lie at great distances from the genes they actually control. ⁴⁹ No mRNA probe data were available for TAS2R42, TAS2R46 or TAS2R50 .

A few studies have reported associations between specific TAS2R SNPs with alcohol ^39,50 and tobacco use ^50,51 in Caucasian and African American populations. Also, it is well known that diets high in vegetables and fruits confer a degree of protection against cancer, including ESCC in our high-risk population. ^9,52 We initially hypothesized that genetic variation in this pathway might be linked with ESCC susceptibility due to taste perception and subsequent intake of fruits and vegetables, and/or with alcohol and tobacco use. However, we found no association between the top SNPs in TAS2R genes with intake of fruits, vegetables, mouldy food, pickled vegetables/vegetable juice, hot scalding food, alcohol intake or tobacco use in Shanxi cases and controls (data not shown). This result could be due to sample size and imprecision in our questionnaire data. TAS2Rs may also play a vital role in transmitting an extracellular signal into an intracellular response in oesophageal cells. The importance of these receptors in relation to signalling and disease susceptibility as well as therapeutic relevance is underscored by their expression and functionality in extra-oral tissues, ^44,45,53 and their ability to cause bronchodilation in the lungs and influence host susceptibility to infection. ^44,53 However, such a role for TAS2Rs expressed in the oesophagus remains to be determined by future studies. Last, the other genes (i.e. ADCY6, GNB3 and CACNA1A) in taste transduction pathway encode proteins that function in other biological pathways (for description see Supplementary Table 16 , available as Supplementary data at IJE online). We found no evidence that the top SNPs in ADCY6 , GNB3 or CACNA1A correlated with probe mRNA levels in normal oesophageal tissues from our population.

The second candidate ESCC pathway in this study was the long-patch BER pathway. This form of BER may be important to correct oxidative damage, ⁵⁴ but the rules that govern ‘long-patch’ over ‘short-patch’ repair are not yet understood. Three genes including POLD2 (also important in the metabolics pathway), FEN1 and POLD1 genes were linked with risk of ESCC. POLD1 and POLD2 encode separate catalytic subunits of the DNA polymerase delta (POLδ) critical for polymerase and exonuclease activity as well as cofactor interaction, respectively. The long-patch repair pathway involves the replacement of two to eight nucleotides including the damaged base, and can be evoked when the structure of the terminal sugar phosphate is such that it cannot be cleaved through the apurinic/apyrimidinic (AP) lyase activity of DNA polymerase beta (POLβ). Under these circumstances, POLδ mediates the synthesis of additional residues resulting in the displacement of the DNA flap containing the abasic site and 3' flanking residues. The flap structure is recognized and cleaved by FEN1 and the replacement residues are then ligated by the DNA ligase1. ⁵⁵ Using a candidate approach, we previously reported an association between BER (pathway containing 27 genes) and ESCC risk. Genes associated with ESCC in this candidate BER pathway included FEN1, SMUG1 and TGD . ²⁴ We did not observe a correlation between rs13928 and POLD2 mRNA. Thus, the present study extends our and others previous BER-association results for ESCC ^24,37 and provides evidence of a potential role for genetic variation in genes involved in long-patch BER in ESCC susceptibility.

The third pathway linked to ESCC risk after GWAS exclusions was the metabolic pathway, the largest gene-containing pathway from all five data sources. A total of 69 genes in the metabolics pathway (KEGG_hsa01100) had P -values < 0.05, but the top candidate ESCC genes were MTAP , GAPDH , DCTD , POLD2 and AMDHD1 . MTAP and AMDHD1 gene products are involved in methionine and histidine metabolism, and the enzyme GAPDH is essential for glycolysis and gluconeogenesis. POLD2 is essential for DNA repair and replication, and the enzyme DCTD is required for the biosynthesis of the pyrimidine nucleotides. We further observed enrichment of variants in a number of subpathways within the metabolics pathway which were also involved in the metabolism of: histidine, valine and tryptophan; glucose; nucleotides; and triacylglycerol and D -myo-inositol-triphosphate, as well as leukotriene and IL-12 signalling.

A number of recent studies have used high-performance liquid chromatography and nuclear magnetic resonance methods to examine metabolic profiles of EC in tissue, plasma, serum and urine. ^56–61 Collectively, the most prominent EC-related metabolic pathways identified among dysregulated biomarker metabolites included those from glycolysis, gluconeogenesis, amino acid metabolism (of tryptophan, valine, cysteine and histidine) and lipid metabolism. ^56,57 Two studies specific to ESCC, including one conducted in a Chinese population, reported these same dysregulated metabolites in plasma and ESCC tissue compared with control samples, as well as additional biomarkers including phospholipids, fatty acids, pyrimidine nucleotides and energy metabolites such as AMP. ^59,61 We observed relations between genotypes of the top ESCC risk SNPs in GALNT3, ACADSB and DGKQ , and mRNA levels in histologically normal oesophageal tissue; however, these correlations did not survive correction for multiple comparisons. Moreover, we identified and validated IDH2 rs11630814 in this pathway as an eQTL with a negative effect on IDH2 mRNA expression. In addition, IDH2 mRNA was down-regulated in ESCC tumours compared with normal oesophagus in cases from our population. IDH2 rs11630814 did not map to a regulatory region in oesophageal tissues, but it is in close LD with rs4561444, which maps to a strong enhancer upstream of the IDH2 promoter and has an even greater negative effect on IDH2 expression.

IDH2 is a homodimeric mitochondrial enzyme which converts isocitrate to α-ketoglutarate (α-KG) and in the process produces the antioxidant nicotinamide adenine dinucleotide phosphate (NADPH). ⁶² Mitochondria require a steady supply of NADPH to support antioxidative defences, lipogenesis, redox balance and mitochondrial DNA maintenance. The production of NADPH has been shown to be modulated in cancers, and somatic monoallelic mutations that result in reduced expression and activity of IDH2 have been described in lymphoid and thyroid tumours. ^62,63 Thus, we propose that individuals heterozygous or homozygous for the G variant of rs11630814 or the T variant of rs4561444 may have lower basal expression and production of IDH2 and thus a reduced ability to neutralize reactive oxygen species. Oxidative DNA damage attributable to disturbances in DNA repair systems, as well as mitochondrial genome instability and copy number, is closely related to ESCC carcinogenesis and its progression. ^64,65 Moreover, TET2 is an αKG-utilizing enzyme that hydroxylates 5-methylcytosine to 5-hydroxymethylcytosine (5-hmC), an important step in the demethylation of DNA. Recently, loss of 5-hmC in melanoma caused (at least in part) by the decreased expression of IDH2 and TET2 has been reported. ⁶⁶ Using GTEx data from other normal tissues, we observed reduced IDH2 mRNA levels with rs11630814 and rs4561444 risk variants specifically in skin tissues (GTEx). Thus, decreased IDH2 expression and reduced production of αKG in oesophageal tissue may also attenuate 5-hmC production and TET2 -dependent demethylation of genomic DNA.

This study had several strengths and limitations. Currently, there is no gold-standard reference to define a pathway representing a biological process. As a result, the gene content of pathways representing the same biological process vary substantially across different databases. We minimized this effect by selecting pathways from five commonly used resources and evaluated the largest number of pathways to date for this analysis of GWAS and cancer. Our examination of 1827 pathways associated with ESCC risk is a strength, in addition to our comprehensive assessment of over 6060 candidate genes within these individual pathways. Examination of many pathways and genes does, however, create concerns about redundancy as well as multiple testing. A further strength of this study is the integration of gene expression data with trait-associated GWAS SNPs to explore the potential functional relevance of our findings and enhance our understanding of mechanisms involved in the aetiology of ESCC. The large number of cases studied also allowed us to assess risks with reasonable power. However, despite the large size of our study, further studies are needed to replicate these findings. Finally, the generalizability of our findings to other ethnic populations remains to be determined.

In conclusion, this study is the largest and most comprehensive agnostic attempt to date to use GWAS data to investigate pathways in cancer. In addition to affirming the roles of cell cycle, DNA repair and apoptotic pathways involving known ESCC-related genes, our results suggest that genetic alterations in taste transduction, long-patch base excision repair and specific metabolic pathways may also contribute to ESCC susceptibility. Links to these pathways include genetic variants in candidate genes involved in the bitter taste transduction, DNA repair synthesis, nucleotide metabolism, amino acid metabolism, lipid metabolism, glycolysis and gluconeogenesis, and energy metabolism. These gene variants may alter protein signalling and/or interactions with other proteins, and we provide evidence that some of their strongly correlated regulatory SNPs can alter gene expression which may lead to changes in oesophageal cell phenotype and functions important for cancer susceptibility. Further investigation into the association of these pathways and genes with risk of ESCC is warranted.

Funding

This research was funded by the Intramural Research Program of the National Institutes of Health, National Cancer Institute, Division of Cancer Epidemiology and Genetics.

Conflict of interest : None declared.