Common genetic variants in epigenetic machinery genes and risk of upper gastrointestinal cancers

Abstract

Background: Populations in north central China are at high risk for oesophageal squamous cell carcinoma (ESCC) and gastric cancer (GC), and genetic variation in epigenetic machinery genes and pathways may contribute to this risk.

Methods: We used the adaptive multilocus joint test to analyse 192 epigenetic genes involved in chromatin remodelling, DNA methylation and microRNA biosynthesis in 1942 ESCC and 1758 GC cases [1126 cardia (GCA) and 632 non-cardia adenocarcinoma (GNCA)] and 2111 controls with Chinese ancestry. We examined potential function of risk alleles using in silico and expression quantitative trait loci (eQTLs) analyses.

Results: Suggestive pathway-based associations were observed for the overall epigenetic (P-value^PATH = 0.034) and chromatin remodelling (P-value^PATH = 0.039) pathways with risk of GCA, but not GC, GNCA or ESCC. Overall, 37 different epigenetic machinery genes were associated with risk of one or more upper gastrointestinal (UGI) cancer sites (P-value^GENE< 0.05), including 14 chromatin remodelling genes whose products are involved in the regulation of HOX genes. We identified a gastric eQTL (rs12724079; rho = 0.37; P = 0.0006) which regulates mRNA expression of ASH1L. Several suggestive eQTLs were also found in oesophageal (rs10898459 in EED), gastric cardia (rs7157322 in DICER1; rs8179271 in ASH1L), and gastric non-cardia (rs1790733 in PPP1CA) tissues.

Conclusions: Results of our analyses provide limited but suggestive evidence for a role of epigenetic gene variation in the aetiology of UGI cancer.

Introduction

Gastric cancer (GC) is the third and oesophageal cancer (EC) is the sixth leading causes of cancer mortality in the world.¹ Despite marked declines in recent decades, around half of the total GC and EC deaths in the world occur in China.¹ Populations in north central China are at particularly high risk for both of these cancers. GCs here occur primarily in the cardia, and essentially all ECs are oesophageal squamous cell carcinomas (ESCC).² Although Helicobacter pylori (H. pylori) infection is causally associated with GC and environmental risk factors (e.g. cigarette smoking and alcohol drinking) are recognized aetiological factors for ESCC in the West,^3,4 the attributable fractions for these factors vary substantially by geographical region and ethnicity.^5,6 Familial aggregation studies⁷ and germline mutations among familial cases provide primary evidence of genetic predisposition for upper gastrointestinal (UGI) cancers,⁸ and genome-wide association studies (GWAS) have identified around 20 genetic loci associated with risk of ESCC^9–15 or GC^11,16,17 in East Asians.

It is now well recognized that epigenetic regulation of gene expression by DNA methylation, chromatin remodelling and histone modifications, and non-coding RNAs is opening new and exciting insights into novel molecular mechanisms in carcinogenesis. Furthermore, there is growing evidence to suggest intricate interplay among epigenetic regulators and mechanisms. For example, more than 100 microRNAs have been reported to be affected by epigenetic aberrations through DNA methylation and/or histone modifications.^18,19 Also, changes in DNA methylation occur in the presence of histone modifications and chromatin remodelling,²⁰ which can either activate or suppress gene expression including microRNA expression in a tissue-specific manner and in different types of cancers.^20–22 Collectively, these observations suggest that an intricate interplay exists across epigenetic machinery genes or mechanisms.

Previous studies related germline variants in DNA methyltransferase and microRNA biosynthesis genes to risk of a number of different cancers, including gastric, lung, liver, bladder, oesophageal, kidney, and head and neck cancers.^23–33 However, to date, no comprehensive examination of the relation of genetic variants in established epigenetic machinery genes to risk of UGI cancer has been performed, nor has a mechanism that translates these variants into cancer risk been delineated. Therefore, an examination of genetic variants from a comprehensive list of epigenetic machinery genes may uncover novel genes and pathways that contribute to UGI cancer susceptibility.

In the present study we used genome-wide association study (GWAS) data¹¹ to examine the relation of variation in epigenetic machinery genes and pathways to risk of ESCC and GC. We carried out bioinformatics analyses of SNP-containing regions using available ENCODE and NIH Roadmap Epigenomics project data. Further, we integrated gene expression data from normal tissues with genotype data from our GWAS participants, and used data from the NIH Genotype-Tissue Expression project (GTEx) portal to explore the functional relevance of our findings through expression quantitative trait loci (eQTLs).

Methods

Study population

This study reports analyses of data from a GWAS of ESCC and GC conducted among ethnic Chinese; full details have been described elsewhere.¹¹ Briefly, participants were drawn from two studies/projects, the Shanxi UGI Genetics Project (Shanxi) and the Nutrition Intervention Trials (NIT). The Shanxi Project was conducted between 1997 and 2007 and had several components including a case-control study and a case-only study (tumour/nontumour study). Case-control and case-only studies were conducted contemporaneously, and the catchment area, enrolment process, information gathering and biological specimen collections were essentially identical. All cases were incident, histologically confirmed ESCC and GC cases who underwent surgical resection at the Shanxi Cancer Hospital/Institute in Taiyuan. Cases were individually matched to controls on age (±5 years), sex and neighbourhood for the case-control portion.³⁴ The NITs were randomized, double-blind, placebo-controlled trials initiated in the mid 1980s that tested the effect of multiple vitamin and mineral combinations on the incidence and mortality of UGI cancers. NIT subjects included in this GWAS were drawn from the cohort of trial participants who participated in a 1999/2000 blood survey which collected blood for DNA. All incident and histologically confirmed ESCC and GC cases identified from blood survey participants followed through to the end of 2007 were included in this GWAS analysis. Controls were selected as a case-cohort from an age- and sex-stratified randomly sampled sub-cohort.³⁵ Both projects in this GWAS collected self-reported risk factor information from interviews, obtained written informed consent from all participants and were approved by institutional review boards. The NCI Special Studies Institutional Review Board also approved the overall GWAS.

Genotyping, quality control, and exclusion

DNAs were genotyped as part of the GWAS at the Cancer Genomic Research laboratory of the National Cancer Institute's Division of Cancer Epidemiology and Genetics as previously described.¹¹ After the initial report, additional subjects from the same contributing studies were scanned on the same platform at the same facility. Both the initial and additional scan data underwent similar processing and quality control metrics. We excluded SNPs with minor allele frequency (MAF) < 5%, < 98% completion and < 95% concordance and a Hardy–Weinberg equilibrium P-value < 1E-6. Linkage disequilibrium (LD) was measured by pairwise r² among controls using Haploview [http://www.broad.mit.edu/mpg/haploview/]. We excluded subjects whose SNP completion rates were < 94%, who had abnormal mean heterozygosity values (> 30 or < 25%), were gender discordant, or were an unexpected duplicate pair. Following all subject exclusions, data on a total of 5811 subjects were included in this analysis, including 1942 ESCCs, 1758 GCs (1126 cardia [GCA] and 632 non-cardia adenocarinoma [GNCA]) and 2111 controls. Data 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].

Gene and SNP selection in epigenetic pathway(s)

To obtain as comprehensive a list of epigenetic machinery genes as possible, we identified from the literature a priori genes which encode proteins associated with known epigenetic mechanisms (DNA methylation, histone modifications, chromatin remodelling and microRNA biosynthesis; collectively referred to here as epigenetic pathway genes).^36–39 We searched PubMed for epigenetic genes and associated mechanisms/pathways using the terms: epigenetics, epigenetic modifications, epigenetic enzymes, epigenetic machinery genes, genes encoding epigenetic regulator, DNA methylation machinery, microRNA biosynthesis, histone modifications, chromatin enzymes, chromatin remodelling.

We extended the search and confirmation of gene data and associated pathways to 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 [http://www.genome.jp/kegg/], the NCI-Nature Curated database [http://pid.nci.nih.gov/download.shtml] and Reactome [http://www.reactome.org/]. We identified 198 autosomal epigenetic pathway genes and mapped SNPs including regions located 20 Kb upstream and 10 kb downstream of each coding gene (Human Genome Build 36 Hg18). We excluded two genes with no SNP data (TARBP2P and PCGF1), one gene which contained SNPs with MAFs < 5% (HDAC1) and four genes for which the proportion of subjects with at least one missing genotype (gene maximum missing rates) was higher than 10% [HDAC9, KDM4C and SMYE3 in all cancers; MBD5 in ESCC, GC and GCA]. After this, a total of 192 genes were included in the epigenetic pathway (Supplementary Table 1, available as Supplementary data at IJE online). To allow subpathway-based analyses, we further subdivided the genes into groups based on their gene regulation mechanism: chromatin remodelling (including histone modification, 142 genes), DNA methylation (15 genes) and microRNA biosynthesis (35 genes) pathways. After quality control processes, we included a total of 2181 unique SNPs in all epigenetic pathway genes for ESCC, 2181 SNPs for GC, 2173 SNPs for GCA and 2225 SNPs GNCA in the final analysis.

Statistical analyses

We carried out individual SNP-, gene- and pathway-based analyses for ESCC and GC overall in addition to the GC subsites of GCA and GNCA. SNP-based analyses were tested under the additive model, and per-allele odds ratios (ORs), 95% confidence intervals (CIs) and corresponding P-values (called P-value^SNP) were calculated using unconditional logistic regression with adjustment for age (10-year categories), sex, study (Shanxi or NIT) and the top 5 eigenvectors (to control for population stratification). After excluding SNPs with LD r²≥ 0.80 in controls, our Bonferroni-corrected threshold for SNP analyses was P = 3.1E-05 (the number of independent signals varied from 1563 to 1593 for ESCC, GC, GCA and GNCA).

Gene-based analyses were conducted using the AdaJoint method [http://dceg.cancer.gov/tools/analysis/adajoint]. The test statistic for gene-based analyses (P-value^GENE) accumulated the evidence of association from the best single SNP and the best pair of SNPs within the candidate gene.

The test statistics for pathway-level analyses (P-values^PATH) were based on the adaptive rank truncated product (ARTP)⁴⁰ method which combined P-value^GENE from the AdaJoint method across all relevant genes in the pathway. The P-values for gene-based analyses and pathway-based analysis were determined through 10⁶ resampling steps.⁴¹ The resampling procedure was used to generate datasets under the null hypothesis while keeping the correlation among SNPs the same as in the observed dataset.

The Bonferroni-corrected significance threshold was P = 2.6E-04 (0.05/192 genes) for gene-based analyses and P = 0.013 (0.05/4 pathways) for pathway-level analyses.

mRNA expression data

Paired tumour (ESCC and GC) and histologically confirmed normal squamous oesophageal and gastric mucosa tissues were collected from a subset of Shanxi cases enrolled between 1998 and 2001. Cases enrolled had no prior therapy for their cancer and all underwent surgical resection at the time of hospitalization. Selection of subjects for tissue study was based solely on the availability of appropriate tissues for RNA testing [i.e. consecutive testing of cases with available frozen tissue, tumour samples that were predominantly (>50%) tumour, normal paired tissues that were dissected at least 2 cm distant regions from the tumours and RNA quality/quantity adequate for testing]. All tissues were snap-frozen in liquid nitrogen and stored at −130°C until required for RNA extraction. RNA extraction and mRNA microarray analyses have been described elsewhere.^42,43 Among the subjects with mRNA expression data available for paired tumour/normal tissues (133 ESCC cases, 62 GCA cases and 72 GNCA cases), 100 ESCC, 34 GCA and 56 GNCA cases also had GWAS germline scan data available to be used for eQTL analyses. Gene expression data are publically available from Gene Expression Omnibus [http://www.ncbi.nlm.nih.gov/geo/] with accession numbers of GSE2340021 for ESCC and GSE2927222 for GCA and GNCA.

Functional annotation of genes and SNPs

We used custom tracks on the UCSC Genome browser [http://genome.ucsc.edu] to screen NIH Roadmap and ENCODE data containing the SNP region for evidence of regulatory relevance⁴⁴ in oesophageal and gastric tissues/cells. We also used the online tools HaploReg [http://www.broadinstitute.org/mammals/haploreg/haploreg.php] and RegulomeDB [http://regulome.stanford.edu] as complementary analyses to confirm the location of each SNP in relation to protein-coding and/or non-coding RNA genes.

Expression quantitative trait loci (eQTLs) analysis

To assess the potential of a risk variant to regulate mRNA expression, we conducted eQTL analysis in normal tissues: 100 normal squamous oesophageal, 34 normal cardia, and 56 normal non-cardia tissues. Normal cardia and non-cardia were also combined to evaluate 90 normal gastric tissues. Candidate SNPs were selected as the SNP with the lowest P-value^SNP among all SNPs tested within a gene; genes evaluated were limited to those with a P-value^GENE<0.05. Based on available probes from the U133A Affymetrix array, we had 16, 20, 23 and 23 SNP:probe pairs(or tests) available in normal oesophageal, gastric, cardia and noncardia tissues, respectively (Supplementary Table 6, available as Supplementary data at IJE online). Correlation coefficients (rho) between SNPs and mRNA expression level were estimated using the nonparametric Spearman rank correlation test. For selected SNPs (i.e. rs12724079 and rs1879271) where fewer than five cases were homozygous for the minor allele, we combined heterozygote cases with minor allele homozygote cases and applied the dominant model to increase power. We used data from GTEx [http://www.gtexportal.org/home/] as an alternative measure for unavailable gene probes from our array (e.g. MDB3L1) and as a source for independent validation of potential eQTLs. GTEx eQTL results were based on RNA-sequencing data, and the relevant tissues available included normal oesophageal mucosa (n = 95), normal oesophageal muscularis (n = 91) and normal gastric (n = 76) tissues. Based on the number of SNPs tested as eQTLs, our Bonferroni-corrected significance threshold for eQTL testing was P = 0.002 (0.05/23 number of pairs).

Results

Population characteristics

A total of 1942 ESCC cases, 1758 GC (1126 GCA and 632 GNCA cases) and 2111 controls from the combined studies were analysed in this study. Demographic and risk factor information for both individual studies and the combined population are shown in Supplementary Table 2 (available as Supplementary data at IJE online).

Pathway-based analyses

Associations were observed for the overall epigenetic pathway and risk of GCA (P-value^PATH= 0.034) and for the chromatin remodelling subpathway with risk of GCA (P-value^PATH= 0.039) (Table 1). However, these associations did not reach the Bonferroni P-value threshold for multiple comparison adjustment.

Table 1.

Associations between epigenetic machinery pathways and the risk of UGI cancer

Pathway/Subpathway	No of genes	ESCC		GC		GCA		GNCA
		No of SNPs (total/unique)	P-value PATH	No of SNPs (total/unique)	P-value PATH	No of SNPs (total/unique)	P-value PATH	No of SNPs (total/unique)	P-value PATH
Epigenetic pathway	192	2181/2174	0.219	2181/2174	0.267	2173/2166	0.034	2225/2218	0.607
Chromatin remodelling	142	1731/1731	0.174	1728/1728	0.349	1724/1724	0.039	1727/1727	0.432
DNA methylation	15a	126/126a	0.441	124/124a	0.245	123/123a	0.081	172/172	0.899
microRNA biosynthesis	35	324/324	0.677	328/328	0.471	326/326	0.392	327/327	0.612

Abbreviations: ESCC, oesophageal squamous cell carcinoma; GC, gastric cancer; GCA, gastric cardia adenocarcinoma; GCNA, gastric non-cardia adenocarcinoma; UGI, upper gastrointestinal.

^aMBD5 was excluded due to gene maximum missing rates  >10%.

Note: Pathway-based p-values (P-value^PATH) are shown for the overall epigenetic pathway and three subpathways in each cancer.

Gene-based analyses

Overall, 37 genes were associated with at least one type of cancer risk (P-value^GENE< 0.05, Table 2): 11 with ESCC, 11 with GC, 14 with GCA and 12 with GNCA risk. Among these 37 genes, 14 (ASH1L, EED, EZH1, KAT2A, KAT6A, KAT8, KDM5A, PRMT1, PRMT7, RING1, SETD1A, SETD1B, SMARCC, and SMARCD1) were chromatin remodelling genes (Supplementary Table 7, available as Supplementary data at IJE online). The genes with the lowest P-values for each site were SETD1B in ESCC (P-value^GENE= 0.006), SUV420H1 in GC (P-value^GENE= 0.006), GSG2 in GCA (P-value^GENE= 0.003) and SMARCC1 in GNCA (P-value^GENE= 0.012). ASH1L was the only gene associated with risk for three cancer sites: overall GC (P-value^GENE= 0.010), GCA (P-value^GENE= 0.018) and GNCA risk (P-value^GENE= 0.019). None of the gene associations had P-values below the Bonferroni threshold following correction for multiple comparisons. Results for all 192 genes evaluated are shown in Supplementary Table 1 (available as Supplementary data at IJE online).

Table 2.

Top-ranked SNPs in genes (P-value^GENE< 0.05) associated with risk of UGI cancers

Oesophageal squamous cell carcinoma (ESCC)
Gene (n = 11)	Gene function	Genomic location	P-valueGENE	Top-ranked SNP	Major/minor	MAF Cases	MAF Controls	OR (95% CI)	P-valueSNP
SETD1B	Chromatin remodelling	12q24.31	0.006	rs2242259	C/T	0.39	0.35	1.14 (1.04–1.24)	0.006
CBX4	Chromatin remodelling	17q25.3	0.020	rs4889898	C/A	0.33	0.30	1.15 (1.05–1.27)	0.003
KAT5	Chromatin remodelling	11q13	0.021	rs1151500	C/T	0.13	0.11	1.21 (1.06–1.39)	0.006
ZGPAT	Chromatin remodelling	20q13.3	0.022	rs8957	C/A	0.34	0.36	0.93 (0.84–1.02)	0.104
BTD	Chromatin remodelling	3p25	0.025	rs2455823	G/A	0.50	0.47	1.14 (1.04–1.24)	0.005
EED	Chromatin remodelling	11q14.2-q22.3	0.026	rs10898459	C/T	0.47	0.50	0.88 (0.81–0.96)	0.005
SMARCD1	Chromatin remodelling	12q13-q14	0.027	rs836178	G/T	0.14	0.16	0.86 (0.76–0.97)	0.013
POLE3	Chromatin remodelling	9q33	0.030	rs8177812	C/T	0.22	0.20	1.17 (1.05–1.30)	0.005
AGO2	microRNA biosynthesis	8q24.3	0.031	rs6983924	A/G	0.50	0.47	1.16 (1.06–1.27)	0.001
RING1	Chromatin remodelling	6p21.3	0.046	rs213194	G/A	0.08	0.07	1.18 (1.00–1.40)	0.051
MBD3L1	DNA methylation	19p13.2	0.047	rs10412487	T/C	0.32	0.34	0.88 (0.80–0.97)	0.008

Gastric cancer (GC)
Gene (n = 11)	Gene function	Genomic location	P-valueGENE	Top-ranked SNP	Major/minor	MAF Cases	MAF Controls	OR (95% CI)	P-valueSNP
SUV420H1	Chromatin remodelling	11q13.2	0.006	rs10896300	G/A	0.11	0.13	0.86 (0.75–0.99)	0.036
ASH1L	Chromatin remodelling	1q22	0.010	rs12724079	T/C	0.17	0.20	0.84 (0.75–0.94)	0.003
KAT5	Chromatin remodelling	11q13	0.011	rs1151500	C/T	0.13	0.11	1.23 (1.07–1.42)	0.004
JARID2	Chromatin remodelling	6p24-p23	0.023	rs7769291	C/T	0.26	0.24	1.14 (1.02–1.26)	0.016
PRMT1	Chromatin remodelling	19q13	0.029	rs3745469	C/T	0.06	0.08	0.79 (0.66–0.95)	0.011
MBD3L1	DNA methylation	19p13.2	0.029	rs10412487	T/C	0.31	0.34	0.87 (0.79–0.96)	0.005
SRRT	microRNA biosynthesis	7q21	0.030	rs13245899	A/G	0.42	0.39	1.14 (1.04–1.25)	0.006
SIRT3	Chromatin remodelling	11p15.5	0.033	rs1045454	C/T	0.18	0.16	1.17 (1.03–1.32)	0.013
KDM5A	Chromatin remodelling	12p13.33	0.033	rs527118	T/C	0.31	0.34	0.87 (0.79–0.96)	0.004
DROSHA	microRNA biosynthesis	5q11.2	0.035	rs7735863	G/A	0.41	0.43	0.89 (0.81–0.97)	0.011
KAT6A	Chromatin remodelling	8p11	0.040	rs7008906	G/A	0.42	0.39	1.14 (1.04–1.24)	0.006

Gastric cardia adenocarcinoma (GCA)
Gene (n = 14)	Gene function	Genomic location	P-valueGENE	Top-ranked SNP	Major/minor	MAF Cases	MAF Controls	OR (95% CI)	P-valueSNP
GSG2	Chromatin remodelling	17p13	0.003	rs7220026	G/A	0.28	0.25	1.19 (1.05–1.33)	0.005
MBD3L1	DNA methylation	19p13.2	0.004	rs10412487	T/C	0.30	0.34	0.82 (0.74–0.92)	0.001
SIRT6	Chromatin remodelling	19p13.3	0.009	rs352493	T/C	0.27	0.24	1.21 (1.08–1.37)	0.002
KDM4A	Chromatin remodelling	1p34.1	0.013	rs304303	C/A	0.18	0.22	0.81 (0.71–0.93)	0.002
PRMT7	Chromatin remodelling	16q22.1	0.016	rs2863973	T/G	0.07	0.09	0.73 (0.60–0.89)	0.002
SUV420H1	Chromatin remodelling	11q13.2	0.016	rs10896300	G/A	0.11	0.13	0.87 (0.74–1.03)	0.105
ASH1L	Chromatin remodelling	1q22	0.018	rs8179271	A/G	0.23	0.26	0.85 (0.75–0.96)	0.007
SRRT	microRNA biosynthesis	7q21	0.021	rs13245899	A/G	0.42	0.39	1.16 (1.05–1.29)	0.005
SMARCD1	Chromatin remodelling	12q13-q14	0.026	rs836178	G/T	0.14	0.16	0.83 (0.71–0.96)	0.011
DROSHA	microRNA biosynthesis	5q11.2	0.034	rs7735863	G/A	0.40	0.43	0.86 (0.77–0.95)	0.004
EZH1	Chromatin remodelling	17q21.1-q21.3	0.034	rs4792953	T/C	0.41	0.44	0.87 (0.78–0.97)	0.011
KAT2A	Chromatin remodelling	17q12-q21	0.034	rs1122326	A/C	0.08	0.10	0.80 (0.66–0.96)	0.015
DICER1	microRNA biosynthesis	14q32.13	0.036	rs7157322	A/C	0.42	0.38	1.17 (1.05–1.30)	0.004
CARM1	Chromatin remodelling	19p13.2	0.048	rs1541596	G/A	0.20	0.23	0.86 (0.75–0.97)	0.019

Gastric non-cardia adenocarcinoma (GNCA)
Gene (n = 12)	Gene function	Genomic location	P-valueGENE	Top-ranked SNP	Major/minor	MAF Cases	MAF Controls	OR (95% CI)	P-valueSNP
SMARCC1	Chromatin remodelling	3p21.31	0.012	rs13094264	G/A	0.32	0.28	1.18 (1.02–1.35)	0.022
PPP1CA	Chromatin remodelling	11q13	0.016	rs1790733	T/C	0.32	0.29	1.16 (1.01–1.33)	0.038
RPS6KA5	Chromatin remodelling	14q31-q32.1	0.017	rs3783834	C/T	0.42	0.37	1.25 (1.10–1.43)	0.001
GEMIN4	microRNA biosynthesis	17p13.3	0.018	rs910924	C/T	0.10	0.13	0.73 (0.59–0.89)	0.002
KAT8	Chromatin remodelling	16p11.1	0.018	rs2855475	C/T	0.10	0.07	1.34 (1.07–1.68)	0.010
ASH1L	Chromatin remodelling	1q22	0.019	rs12724079	T/C	0.16	0.20	0.79 (0.67–0.94)	0.007
KAT5	Chromatin remodelling	11q13	0.029	rs1151500	C/T	0.14	0.11	1.29 (1.07–1.56)	0.009
PRMT1	Chromatin remodelling	19q13	0.033	rs3745469	C/T	0.06	0.08	0.74 (0.57–0.96)	0.023
KDM8	Chromatin remodelling	16p12.1	0.037	rs12051243	T/G	0.17	0.20	0.80 (0.68–0.95)	0.011
SETD7	Chromatin remodelling	4q31.1	0.042	rs720257	G/A	0.36	0.32	1.22 (1.07–1.40)	0.003
SETD1A	Chromatin remodelling	16p11.2	0.046	rs897986	G/A	0.08	0.06	1.33 (1.04–1.70)	0.021
SMAD3	microRNA biosynthesis	15q21-q22	0.047	rs1465842	G/A	0.21	0.17	1.27 (1.08–1.50)	0.003

Abbreviations: CI, confidence intervals; MAF, minor allele frequency; OR, odds ratio; UGI, upper gastrointestinal.

Note: Per-allele ORs and P-value^SNP were estimated for minor allele among controls in combined population adjusted for age, sex, study, and eigenvectors. The table was sorted by P-value^GENE.

SNP-based analyses

Overall, SNP associations with risk (P-value^SNP less than 0.05) were seen for 132 SNPs in 49 genes with ESCC and for 123 SNPs in 57 genes with GC (Supplementary Tables 3 and 4, available as Supplementary data at IJE online). However, none of the individual SNP P-values for associations with any of the four cancers examined here reached the Bonferroni correction threshold for multiple comparisons.

Table 2 shows the top-ranked SNP in each of genes with P-value^GENE< 0.05 for associations with risk of ESCC, GC, GCA and GNCA. Our top-ranked SNPs were: rs6983924 in AGO2 (OR = 1.16; 95% CI = 1.06–1.27; P-value^SNP= 9.7E-04) for ESCC; rs12724079 in ASH1L (OR = 0.84; 95% CI = 0.75–0.94; P-value^SNP= 0.003) for GC; rs10412487 in MBD3L1 (OR = 0.82; 95% CI = 0.74–0.92; P-value^SNP= 7.2E-04) in GCA; and rs3783834 in RPS6KA5 (OR = 1.25; 95% CI = 1.10–1.43; P-value^SNP= 6.3E-04) for GNCA. The top-ranked SNPs in ASH1L differed by GC subtype: ASH1L rs12724079 was the top-ranked SNP with GC (OR = 0.84; 95% CI = 0.75–0.94; P-value^SNP  = 0.003) and GNCA (OR =  0.79; 95% CI = 0.67–0.94; P-value^SNP = 0.007) risk associations, whereas rs8179271 was the top-ranked SNP with GCA risk (OR = 0.85; 95% CI = 0.75–0.96; P-value^SNP= 0.007).

In silico functional annotation and eQTL analyses of SNPs in genes (P-value^GENE < 0.05)

Top-ranked SNPs (38 SNPs) in 37 genes (P-value^GENE< 0.05) were mapped for their putative functional/regulatory sites (Supplementary Table 5, available as Supplementary data at IJE online). There were 28 SNPs in intronic regions, 5 SNPs in 3’ UTRs and 4 SNPs in 5’ UTRs in these genes of interest. Rs352493 was the only non-synonymous SNP (pro456Arg) in SIRT6. Several SNPs, including rs4889898 (CBX4), rs1045454 (SIRT3), rs304303 (KDM4A), rs8179271 (ASH1L), rs910924 (GEMIN4) and rs2855475 (KAT8) were located proximal to non-coding RNA genes. A number of SNPs mapped to strong enhancer and transcription-like enhancer regions and DNaseI/open chromatin sites, including rs6983924 (AGO2) in normal oesophageal tissue and rs1541596 (CARM1), rs1465842 (SMAD3) and rs1151500 (KAT5) in stomach tissues. Many SNPs had the potential to alter the DNA-binding motifs of a number of proteins and transcription factors.

Table 3 shows results for eQTLs with P-values less than 0.05. Rs10898459 (variant T) in EED was associated with lower mRNA levels (rho = –0.22; P = 0.029) in normal oesophagus, and the same trend for T variant was seen in GTEx data (P = 0.060). In normal gastric tissue, rs12724079 (variant C) in ASH1L was associated with higher mRNA levels (rho = 0.35; P = 6.3E–04), a result which remained even after adjustment for multiple comparisons. In contrast, rs8179271 (variant G, pairwise LD r²= 0.70 with rs12724079) was associated with lower mRNA levels of ASH1L (rho = –0.35; P = 0.045) in normal gastric cardia tissues. Figure 1A shows eQTL results from the dominant model for ASH1L SNPs, and indicated that correlation coefficients for normal gastric (rho = 0.37, P = 0.0004) and normal non-cardia gastric (rho = 0.20, P = 0.1413) were both positive, whereas the coefficient for normal cardia was negative (rho = –0.35, P = 0.045). We also found that rs7157322 (variant C) in DICER1 and rs1790733 (variant C) in PPP1CA were correlated with reduced mRNA levels of DICER1 in normal cardia (rho = –0.42; P = 0.014) and of PPP1CA in normal non-cardia gastric tissues (rho = –0.41; P = 0.002), respectively (Table 3). However, these associations were not replicated in GTEx data.

Table 3.

Selected eQTL analysis results (P < 0.05) of genes (P-value^GENE< 0.05) in normal oesophageal and stomach tissues

			Shanxi normal esophagus (n = 100)				GTEx normal esophagus mucosa (n = 95)b
Gene	Top-ranked SNP	OR (95% CI)	U133A mRNA probe	Major/minor	rhoa	P-value	Ref/Alt	Direction	P-value
EED	rs10898459	0.88 (0.81–0.96)	209572_s_at	C/T	−0.22	0.029	T/C	+	0.060
			Shanxi normal stomach (n = 90)				GTEx normal stomach (n = 76)b
			U133A mRNA probe	Major/minor	rhoa	P-value	Ref/Alt	Direction	P-value
ASH1L	rs12724079	0.84 (0.75–0.94)	218554_s_at	T/C	0.35	0.0006	T/C	+	0.800
			Shanxi normal cardia stomach (n = 34)				GTEx normal stomach (n = 76)b
			U133A mRNA probe	Major/minor	rhoa	P-value	Ref/Alt	Direction	P-value
DICER1	rs7157322	1.17 (1.05–1.30)	206061_s_at	A/C	−0.42	0.014	A/C	±	0.200
ASH1L	rs8179271	0.85 (0.75–0.96)	218554_s_at	A/G	−0.35	0.045	No SNP	No SNP	No SNP
			Shanxi normal non-cardia stomach (n = 56)				GTEx normal stomach (n = 76)b
			U133A mRNA probe	Major/minor	rhoa	P-value	Ref/Alt	Direction	P-value
PPP1CA	rs1790733	1.16 (1.01–1.33)	200846_s_at	T/C	−0.41	0.002	A/G	±	0.400

Abbreviations: eQTL, expression quantitative trait loci; Ref/Alt, reference/alternative allele.

^a Statistical significance for correlation (rho) was determined using nonparametric Spearman's rank test; Direction is based on minor allele.

^b GTEx Portal [http://www.gtexportal.org/home/]; Direction is based on alternative allele.; Symbols: +, positive influence; ±, no change.

Figure 1.

Expression quantitative trait loci (eQTL) analysis of ASH1L mRNA expression in normal gastric tissues, and genomic locations for ASH1L and SNPs. A: eQTL analyses were conducted using dominant model for rs12724079 in 90 gastric tissues, rs8179271 in 34 cardia tissues and rs12724079 in 56 non-cardia tissues. ASH1L mRNA levels were assessed using the Affymetrix_U133A probe 218 554_s_at. The major/minor allele for each SNP in our population is shown. B: Genome Browser [http://genome.ucsc.edu/] image of ASH1L gene region on human assembly hg19 based on NIH Epigenomics Roadmap data [http://www.genboree.org/epigenomeatlas/] and ENCODE data [http://genome.ucsc.edu/ENCODE/]. The position of the Affymetrix_U133A probe 218 554_s_at, which detects variant mRNAs 1, 2 and 3 of ASH1L and CpG islands in this region are shown. Regulatory domains [chromatin state segmentation using a hidden Markov Model (ChromHMM)] and core histone marks for normal gastric tissue (Gastric), stomach mucosa (SM) and stomach smooth muscle (SSM) tissue are shown: Red, active transcriptional start site (TSS); Dark Green, transcription elongation/transition; Orange, active-to-weak enhancer. H3K4Me3 and H3K27Ac activation marks in this region and mRNA levels from a large number of ENCODE cells lines is also shown. SNPs rs12724079 and rs8179271 are highlighted by a light blue-coloured box. For clarity, not all SNPs contained in this region (dbSNP) are shown. ASH1L-AS1 and other non-coding RNA genes within ASH1L are indicated and text is coloured grey.

Discussion

In this study we systematically examined the cumulative effects of variation in multiple epigenetic machinery genes acting in functional mechanistic pathways as well as individual gene effects, to understand genetic susceptibility to ESCC and GC in a high-risk Chinese population.

We found pathway-based associations for the overall epigenetic pathway and the chromatin remodelling subpathway with GCA risk. Given the relatively large proportion of chromatin remodelling genes contributing to the overall pathway, the chromatin remodelling subpathway (including genes encoding enzymes involved in histone modifications) is likely driving the association of GCA risk with the overall epigenetic pathway. The absence of pathway-based associations for GC and GNCA may be due to reduced power from the smaller number of GNCA cases genotyped here. Alternatively, this result may reflect aetiological differences in the gastric carcinogenesis pathway^11,45–47 and in epigenetic mechanisms important for the development of GC subtypes in this high-risk population. We found no evidence for epigenetic pathway-based associations with ESCC.

The functions of the 37 genes we observed as associated with UGI cancer risk are summarized in Supplementary Table 7, available as Supplementary data at IJE online. In relation to chromatin remodelling genes (histone and nucleosome modifiers), we identified six ESCC-associated genes encoding products involved in the methylation (SETD1B and EED), sumoylation (CBX4), ubiquitination (RING1), biotinylation (BTD) and acetylation (KAT5) of histone tails, and a further two associated genes that encode proteins involved in nucleosome (chromatin) remodelling/dynamics (POLE3 and SMARCD1).^48,49 KAT5 was also associated with GC and GNCA risk, and SMARCD1 was associated with GNCA risk. Twelve different chromatin remodelling genes were associated with GC, CGA and GNCA risk. These genes encode unique histone modifying demethylases (KDM5A, KDM4A and KDM8), acetyltransferases (KAT2A, KAT6A and KAT8), and methyltransferases or protein units (EZH1 and SUV420H1) and a deacetylase enzyme (SIRT6), as well as arginine methyltransferases (PRMT1, PRMT7 and CARM1).^48–50 Post-translational modification of histones and nucleosome remodelling are important epigenetic mechanisms regulating gene expression,⁴⁹ and alterations of these genes have been associated with the development of many cancers.⁵⁰

Interestingly, 14 of the ESCC- and GC-associated chromatin remodelling genes encode proteins whose function has been linked to the regulation of HOX genes (Supplementary Table 7, available as Supplementary data at IJE online), suggesting that genetic variation in these genes may result in altered HOX expression in oesophageal and gastric tissues. HOX genes encode important transcription factors involved in developmental processes and cell differentiation, and dysregulation of specific HOX genes has previously been reported in ESCC and GC.^51,52

EED encodes a member of the polycomb repressive complex 2 (PRC2) that interacts with EZH2, a methyltransferase which negatively regulates HOX gene expression.^49,50 In our study, EED was associated with risk of ESCC. The T variant of its top-ranked SNP (rs10898459) was associated with a protective effect for ESCC and correlated with reduced EED mRNA levels in normal oesophageal tissues. This finding is consistent with previous data that EED mRNA and protein are expressed at low to medium levels in normal oesophageal tissues, but are overexpressed in a number of cancers [http://www.proteinatlas.org/ENSG00000074266-EED/gene], including ESCC.⁵³ In 133 tumour and normal paired oesophageal tissues from our study population, we also observed upregulation of EED mRNA in ESCC tumour compared with normal oesophageal tissues (fold-change = 1.34, P = 3.3E-17). Collectively, these data are consistent with the possibility that rs10898459 and low basal expression of EED may be protective against ESCC development.

ASH1L encodes a TrxG methyltransferase that is required for maximal expression and H3K4 methylation of HOX genes.^54,55 In this study, ASH1L was the only gene associated with risk of all three gastric cancer sites examined. We also identified two potential eQTLs (rs12724079 and rs8179281) influencing ASH1L mRNA levels in normal gastric tissues. However, the putative protective genotypes of both SNPs had opposite effects on ASH1L mRNA in cardia and non-cardia normal gastric tissues (Figure 1A), suggesting the possibility that the effect of this SNP on ASH1L expression may differ by anatomical subtype. Interestingly, rs8179271 maps 3’ to ASH1L antisense RNA 1 (ASH1L-AS1), which encodes a long non-coding RNA (lncRNA) of unknown function (Figure 1B). Thus, confirming the influence of rs12724099 and rs8179271 on ASH1L (and ASH1L-AS1) expression in gastric tissue remains an interesting question for future research.

PPP1CA encodes a catalytic subunit of protein phosphatase 1 (PP1), whose activity is important for chromatin structure and inflammation. The GNCA risk-associated rs1790733 variant in PPP1CA was significantly associated with reduced levels of PPP1CA mRNA in normal noncardia gastric tissue. Downregulation of PPP1CA has been shown to promote proliferation, migration and invasion of gastric cancer cells in vitro,⁵⁶ suggesting that dysregulation of PPP1CA may contribute to GNCA susceptibility.

Just one DNA methylation machinery gene (MBD3L1) was related to UGI cancer in our study, and it was associated with risks for ESCC, GC and GCA but not GNCA. MBD3L1 encodes a transcriptional repressor whose expression appears to be restricted to germ cells [http://www.proteinatlas.org/ENSG00000170948-MBD3L1/tissue], thus the significance of MBD3L1 in relation to ESCC and GC is unclear.

Lastly, six genes (AGO2, DICER1, DROSHA, GEMIN4, SMAD3 and SRRT) involved in microRNA biosynthesis were found to be associated with one or more UGI cancers. It has been suggested that the global downregulation of microRNAs in tumours compared with normal tissues⁵⁷ may be due in part to reduced expression of DROSHA and DICER.⁵⁸ In our study, DICER1 was associated with GCA risk and the risk allele of its top-ranked SNP (rs7157322) was correlated with decreased mRNA levels in normal cardia tissues, an observation consistent with a potential cis-regulatory effect of rs7157322 on DICER1 expression.

This study had several strengths and limitations. The literature on genetic variants in genes encoding epigenetic regulators and their association with UGI cancer is very limited^25,59,60 (Supplementary Table 7, available as Supplementary data at IJE online). Our study is the first to perform a comprehensive examination of genetic variants in epigenetic machinery genes at multiple levels of association (i.e. pathways, genes and SNPs) using new state-of-the-art biostatistical methodologies. Additional in silico and eQTL analyses provided complementary functional characterization to support observed risk associations. The relatively large number of cases studied allowed us to assess associations with good power. However, examination of large numbers of genes and SNPs in multiple types of cancers raises concerns about multiple comparisons. Further, our sample size was only modest for eQTL analyses, which limited our power to see true associations. Lack of replication in GTEx data could be also partly due to limited statistical power as well as the lack of information on anatomical location (cardia versus non-cardia). Many associations were found with P-values less than the nominal 0.05 level, but none of our pathway, gene or SNP association P-values passed Bonferroni adjustment; thus, we are unable to exclude the potential role of chance as an explanation for our results. Furthermore, the lack of H. pylori data is a potential concern. However, the genetic loci that have previously been related to H. pylori status were not among the epigenetic pathway genes assessed here,⁶¹ thus confounding by H. pylori status seems unlikely. Finally, the generalizability of our findings to other ethnic populations remains to be determined.

In conclusion, our study provides suggestive but limited evidence for the involvement of epigenetic pathways and genes in UGI cancer susceptibility, particularly for the chromatin remodelling pathway and risk of gastric cardia cancer. Future studies are warranted to replicate our findings and to further address functional roles for potentially important epigenetic machinery genes in UGI cancers.

Supplementary Data

Supplementary data are available at IJE online.

Funding

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

Conflict of interest: None declared.