Genome-wide DNA methylation differences according to oestrogen receptor beta status in colorectal cancer

ABSTRACT

Involvement of sex hormones in colorectal cancer (CRC) development has been linked to oestrogen receptor β (ERβ). Expression of ERβ is found reduced in tumour tissue and inversely related to mortality. However, mechanisms are not well understood. Our study aimed to detect differentially methylated genes associated with ERβ expression, which could point to mechanisms by which ERβ could influence risk and prognosis of CRC. Epigenome-wide DNA methylation profiling was performed using Illumina HumanMethylation450k BeadChip arrays in two independent tumour sample sets of CRC patients recruited in 2003–2010 by the German DACHS study (discovery cohort n = 917, replication cohort n = 907). ERβ expression was measured using immunohistochemistry and scored as negative, moderate and high. Differentially methylated CpG sites and genomic regions were determined using limma in the R-package RnBeads. For the comparison of tumours with moderate/high ERβ versus negative expression, differentially methylated CpG sites were identified but not confirmed by replication. Comparing tumours of high with tumours of negative ERβ expression revealed 2,904 differentially methylated CpG sites of which 403 were replicated (FDR adjusted p-value<0.05). Replicated CpGs were annotated to genes such as CD36, HK1 or LRP5. A survival analysis indicates that 30 of the replicated CpGs are also associated with overall survival (FDR-adjusted p-value<0.05). The regional analysis identified 60 differentially methylated promotor regions. The epigenome-wide analysis identified both novel genes as well as genes already implicated in CRC. Follow-up mechanistic studies to better understand the regulatory role of ERβ could inform potential targets for improving treatment or prevention of CRC.

Introduction

Incidence rates for colorectal cancer (CRC) are higher in men than in women [1]. Observational and experimental evidence suggest that this gender-specific difference could be attributed to the differential exposure to sex-hormones, especially to oestrogen [2]. This is supported by findings of epidemiologic studies as well as randomized controlled trials, which show that menopausal hormone therapy (MHT) reduces CRC risk [3]. Oestrogen exerts its effects on colon cells predominantly through the nuclear receptor oestrogen-receptor beta (ERβ), the predominant oestrogen-receptor expressed in the colon mucosa, which is mainly involved in anti-proliferative and apoptotic effects [4]. The expression of ERβ decreases during carcinogenesis [5] and patients with ERβ positive tumours seem to have better survival rates compared to patients with ERβ negative tumours [6]. There is also indication that CRC risk associated with exogenous hormone use varies by expression status of ERβ [7,8].

Mechanisms underlying the relationship between ERβ expression and prognosis of CRC are not well understood. Oestrogen receptors are nuclear receptors, so they bind to the DNA once they are activated [9]. As a transcription factor, ERβ can have an influence on DNA methylation upon DNA binding [10].

The human genome is predominantly methylated at the cytosine of dinucleotide sequences CpG (cytosine and guanine), which often cluster as CpG islands in regulatory regions of genes [11]. DNA hypermethylation in the promotor region of a gene is usually associated with silencing of the gene [12]. Aberrant DNA methylation is a characteristic of many cancers including CRC [13]. Absence of ERβ has been found to lead to hypermethylation in the promotor region of glucose transporter 4 [10]. ERβ can regulate DNA methylation at further specific genomic loci by recruiting thymine DNA glycosylase, a factor which is involved in demethylation [14]. We hypothesize that differences in DNA methylation patterns exist between ERβ positive tumours and ERβ negative tumours of CRC patients, which could modulate gene expression patterns and thereby influence prognosis. Here, a genome-wide approach was used to identify differentially methylated sites and regions according to ERβ expression status, which could bring light into mechanisms behind ERβ actions.

Results

Study population – descriptive characteristics

For the discovery analysis, data from 917 patients was available after filtering (exclusion of 15 samples) and exclusion of patients with missing values for ERβ expression status (n = 199) or any of the other covariates (CIMP status: two patients, smoking: two patients, stage: two patients). Of these patients, 401 had ERβ negative tumours, 380 had ERβ moderate tumours, and 136 had ERβ high tumours. The final replication set consisted of 907 CRC patients (excluding 140 patients with missing values for ERβ expression status, 14 patients with missing values for CIMP status and three patients with missing values for smoking status). Of these 907 patients of the replication set, 433 had ERβ negative tumours, 320 had ERβ moderate tumours, and 154 had ERβ high tumours. Patient characteristics were similar between discovery and replication sets except for CIMP status and smoking, which differed significantly. Both factors were adjusted for in the statistical analyses (Table 1).

Table 1.

Characteristics of the discovery cohort and the replication cohort investigating DNA methylation according to ERβ expression.

	Discovery				Replication
	Total	ERβ negative	ERβ positive	P-valueb	Total	ERβ negative	ERβ positive	P-valueb	P-valuec
N (%)	917	401 (43.7)	516 (56.3)		907	433 (47.7)	474 (52.3)
Age (Median (range))	70.0 (33–96)	71 (33–91)	70 (34–96)	0.12	69 (33–94)	70 (33–94)	69 (37–91)	0.23	0.47
Sex, n (%)
Male	554 (60.4)	243 (60.6)	311 (60.3)	0.92	508 (56.0)	239 (55.2)	269 (56.8)	0.64	0.06
Female	363 (39.6)	158 (39.4)	205 (39.7)		399 (44.0)	194 (44.8)	205 (43.2)
Smoking, n (%)
Never	401 (43.7)	171 (42.6)	230 (44.6)	0.82	452 (48.8)	220 (50.8)	232 (48.9)	0.73	0.01
Ever	384 (41.9)	170 (42.4)	214 (41.5)		322 (35.5)	148 (34.2)	174 (36.7)
Current	132 (14.4)	60 (15.0)	72 (13.9)		133 (14.7)	65 (15.0)	68 (14.4)
Tumor location, n (%)
Colon	587 (64.0)	240 (59.9)	347 (67.3)	0.02	603 (66.5)	280 (64.7)	323 (68.1)	0.26	0.27
Rectum	330 (36.0)	161 (40.1)	169 (32.7)		304 (33.5)	153 (35.3)	151 (31.9)
Tumor stage, n (%)
I-III	792 (86.4)	351 (87.5)	441 (85.5)	0.37	773 (85.2)	371 (85.7)	402 (84.8)	0.71	0.48
IV	125 (13.6)	50 (12.5)	75 (14.5)		134 (14.8)	62 (14.3)	72 (15.2)
Tumor purity (Median (range))a	0.74 (0.30–0.94)	0.73 (0.32–0.94)	0.74 (0.30–0.92)	0.79	0.74 (0.37–0.94)	0.73 (0.37–0.94)	0.75 (0.48–0.93)	0.003	0.97
CIMP status, n (%)
negative	714 (77.9)	311 (77.6)	403 (78.1)	0.84	787 (86.8)	368 (85.0)	419 (88.4)	0.13	<0.0001
high	203 (22.1)	90 (22.4)	113 (21.9)		120 (13.2)	65 (15.0)	55 (11.6)

Abbreviations: ERβ, estrogen receptor beta; SD, standard deviation; CIMP, CpG island methylator phenotype.

Wilcoxon rank sum tests were used for calculation of p-values for age and tumor purity; Chi-square tests were used for calculation of p-values for sex, smoking, tumor location, tumor stage and CIMP status.

Variables with missing values: discovery set: smoking (N = 2), stage (N = 2), CIMP (N = 2); replication set: smoking (N = 3), CIMP (N = 14).

^atumor purity was estimated by the method published by Aran et al. (57) using 44 CpGs consistently unmethylated in immune cells but consistently methylated in tumor cells for estimation.

^bdifference between ERβ negative and ERβ positive samples.

^cdifference between discovery and replication set.

Clustering of tumours, comparison of datasets

To compare the datasets, unsupervised clustering of the tumours was performed. In the discovery dataset six clusters of samples were found. Clusters 1 and 2 were enriched for CIMP positive colorectal tumours and BRAF mutation (Supplementary Figure 1). In the replication dataset five clusters were found; here as well, two clusters were enriched for CIMP positive colorectal tumours and BRAF mutation (Supplementary Figure 2). ERβ status did not seem to yield any specific pattern. The clustering detected clinically relevant cancer subtypes and thereby provided assurance for the data quality and the comparability of the discovery and replication datasets.

Differences in DNA methylation according to ERβ expression

The primary analysis focused on exploration of differences in DNA methylation between ERβ negative tumours and ERβ positive tumours with moderate or strong ERβ expression. In this analysis, 1,220 differentially methylated sites annotated to 738 different genes associated with ERβ expression were discovered (FDR-adjusted p-value<0.05) (see Supplementary Table 1 and Figure 1). Nearly half of the significant CpG sites were hypermethylated (49%) in ERβ positive (moderate/strong) tumours compared to ERβ negative tumours. The absolute difference in mean DNA methylation beta values of significant CpGs ranged between 0.3% and 6.4%. The significantly differentially methylated CpG sites between ERβ positive and negative tumours were mainly located in OpenSea (54.0%), further CpGs are located in CpG-shores (21.2%), CpG-islands (13.2%) and in CpG-shelfs (11.6%) (Figure 2). Whereas the proportion of hypo-/hypermethylation is around 50% in CpG-shores, CpG-shelves and OpenSea (Figure 2), CpG-islands are more often hypomethylated (76.4%) in ERβ negative samples. Gene expression data for the DACHS samples were not available, so gene expression profiles from CRC were retrieved from the BioXpress database. These expression profiles indicate that many of the genes associated with differentially methylated sites are expressed in CRC and could therefore be of functional relevance (Supplementary Table 1).

Figure 1.

Manhattan plot of p values for the association of DNAmethylation with moderate/high ERβ expression status (primary analysis) in the discovery cohort.

Figure 2.

Relation of the 1220 statistically significant differentially methylated probes between ERβ moderate/high and ERβ negative samples to CpG islands, shelves, shores and OpenSea. Numbers of significant probes in CpG islands, shelves (north and south), shores (north and south) and open sea. 13.2% of significant CpGs are located in CpG islands (76.4% of these CpGs are hypomethylated in ERβ negative samples), 54.0% of the significant CpGs are located in OpenSea (42.6% of these are hypomethylated), 11.6% of significant CpGs are located in CpG-shelves (48.6% of these are hypomethylated) and 21.2% are located in CpG-shores (54.7% of these CpGs are hypomethylated).

None of the 1,220 significant CpG sites of the discovery cohort were found significant in the replication cohort after adjustment for multiple testing (Supplementary Table 1). Applying the nominal p-value of 0.05, 176 CpG sites were replicated. Among them, CpG sites annotated to the genes HK1, CD36 or LRP5 can be found.

In the secondary analysis, comparing tumours with high ERβ expression and those without, 2,904 significant CpG sites (FDR p-value <0.05) annotated to 1,538 different genes were identified. In the replication analysis, 403 of these 2,904 CpG sites were significant after adjustment for multiple testing. Of the replicated sites, 291 (72.2%) are hypomethylated and 112 (27.8%) hypermethylated in ERβ positive tumours. These 403 CpGs are annotated to 252 different genes. The absolute mean difference in DNA methylation of these replicated CpGs based on the beta value ranged from 1.3% to 9.6%. The top 20 replicated CpG sites are shown in Table 2 and all replicated CpG sites are shown in Supplementary Table 2. The top differentially methylated and replicated CpGs are located in genes like SNHG3-RCC1, CD36, GSDMB or HK1. It was investigated whether the 252 replicated genes contain a transcription factor-binding site for ERβ in their promotor region. The TF2DNA database (http://fiserlab.org/tf2dna_db//search_genes.html) showed that nine of 252 genes harbour transcription factor binding sites for ERβ, namely HK1, MLPH, PRDM2, SAMD14, SLC44A4, TRIM2, UNC5D, XRCC3 and ZG16 (Supplementary Table 2).

Table 2.

Top 20 of the 403 replicated CpG sites according to ERβ expression (ERβ high, score 2, vs. ERβ negative, score 0).

	Discovery						Replication set
CpG site	Chr	Gene	Position in gene	Mean difference of beta values (ERβ pos- ERβ neg)	p-valuea	FDR- adjusted p-value	Mean difference of beta values (ERβ pos- ERβ neg)	p-valuea	FDR-adjusted p-value	Gene expression from BioXpress databaseb
cg00452755	chr1	SNHG3-RCC1	Body	−0.04	1.97E-10	1.12E-05	−0.02	2.09E-03	0.03	741
cg26138637	chr7	CD36	5ʹUTR	0.09	5.55E-10	2.75E-05	0.05	4.15E-03	0.04	172
cg05842113	chr17	GSDMB	TSS1500	−0.06	1.00E-09	3.32E-05	−0.03	3.79E-03	0.04	1783
cg06748146	chr10	HK1	Body	−0.08	1.01E-09	3.32E-05	−0.04	5.84E-03	0.05	5144
cg15942979	chr8	BIN3	Body	−0.09	1.18E-09	3.59E-05	−0.06	3.56E-04	0.02	1065
cg01398428	chr11	LRP5	Body	−0.05	2.04E-09	4.73E-05	−0.04	9.83E-05	0.02	7473
cg26405128	chr3	STT3B	Body	−0.05	1.04E-08	1.70E-04	−0.03	1.53E-03	0.03	5598
cg02606728	chr11	LRP5	Body	−0.04	1.26E-08	1.79E-04	−0.02	3.94E-03	0.04	7473
cg05593336	chr6			0.08	1.60E-08	1.92E-04	0.05	3.28E-03	0.04
cg18173263	chr17	PEMT	Body	−0.07	2.02E-08	2.11E-04	−0.05	2.66E-04	0.02	636
cg14214797	chr14	CCDC88C	Body	−0.05	2.54E-08	2.36E-04	−0.03	5.23E-03	0.04	2208
cg01566460	chr1	NOL9	Body	−0.07	2.63E-08	2.36E-04	−0.05	3.55E-04	0.02	343
cg26308909	chr7	MAD1L1	Body	−0.05	3.31E-08	2.73E-04	−0.04	1.87E-04	0.02	1002
cg08186892	chr16	CMIP	Body	−0.06	3.52E-08	2.80E-04	−0.03	3.56E-03	0.04	4759
cg06574229	chr9	LRRC26	TSS1500	−0.05	4.48E-08	3.23E-04	−0.03	6.54E-03	0.05	42
cg02405662	chr19	DAZAP1	Body	−0.07	5.51E-08	3.48E-04	−0.04	8.61E-04	0.03	5256
cg16372648	chr3	SELT	Body	−0.06	5.53E-08	3.48E-04	−0.03	1.80E-03	0.03	3527
cg19776833	chr3	FOXP1	Body	−0.05	7.70E-08	4.18E-04	−0.03	2.82E-05	0.01	2617
cg17240454	chr6	SPDEF	TSS200	−0.05	8.15E-08	4.25E-04	−0.03	4.96E-03	0.04	453
cg00542261	chr10	HK1	Body	−0.08	8.97E-08	4.54E-04	−0.05	1.76E-03	0.03	5144

Abbreviations: Chr, chromosome; FDR, false-discovery rate; TSS, transcription start site; TSS1500, CpG is 1500 basepairs away from the TSS and located within the 5`region;

TSS200, CpG within 200 basepaires from TSS; 5`UTR, untranslated region.

Tumor purity was estimated by the method published by Aran et al. (57) using 44 CpGs consistently unmethylated in immune cells but consistently methylated in tumor cells for estimation.

All analyses adjusted by age, sex, smoking, cancer stage, tumor location, CIMP status, tumor purity.

^aderived using M-values.

^bRNAseq gene expression values (median log2 (read counts)) for colon cancer were retrieved from the BioXpress database (57&59).

As differential methylation seems more pronounced in the secondary analysis comparing tumours with high ERβ expression to those without ERβ expression, these samples were used for the subsequent analyses.

DNA methylation of CpGs located in the ERβ gene (ESR2)

Nineteen CpG sites located in the ERβ gene were measured by the 450k array. One of them (cg01015652) was found nominally differentially methylated between tumours with high ERβ expression and ERβ negative tumours (Supplementary Table 3). This CpG site was however not found differentially methylated in the replication set. So no confirmed differentially methylated site in the ERβ gene was found.

Differentially methylated regions

As DNA methylation CpG sites can be highly correlated in specific regions, an additional analysis with focus on promotor regions was performed. This analysis considered both statistical significance (p-value) and biological importance (by considering mean difference in methylation) and ranks the regions accordingly. Using the cutoff proposed by the R program RnBeads, 60 promotor regions containing 10 or more CpGs were identified (Table 3). Table 3 shows the regions sorted by combinedRank, a metric given to the sites out of mean difference in means across all sites in a region, mean of ratios in mean methylation and the combined p-value of the region. The top promotor regions are annotated to the genes HOXB5, GPC5 or SHOX2, which are expressed in CRC as indicated by expression levels from the BioXpress database (RNAseq gene expression values (median log2 (read counts)) for colon cancer were retrieved from the BioXpress database). None of the 60 differentially methylated promotor regions were confirmed in the replication set after adjustment for multiple testing. The genes EPM2AIP1, MUC2, MLH1, CBLN1, NHLRC1 were however replicated at nominal significance (p-value<0.05). Two of the 60 genes (MUC2, KDR) were also found among the genes annotated to the 403 significantly replicated differentially methylated CpGs associated with high ERβ positive expression.

Table 3.

Differentially methylated promotor regions according to ERβ expression sorted by combined rank (out of mean difference, mean quot log2 and combined p-value).

	Discovery							Replication
Chr	Gene	Mean difference of beta values (ERβ pos- ERβ neg)	Mean quot log2	Combined p-valuea	FDR adjusted p-value	Number of sites in region	Combined Rankb	Mean difference of beta values (ERβ pos- ERβ neg)	Mean quot log2	Combined p-valuea	FDR adjusted p-value	Gene expression from BioXpress databasec
chr17	HOXB5	−0.061	0.23	2.77E-04	0.12	13	98	−0.0046	0.010	0.35	0.57	769
chr12		0.054	−0.26	8.88E-04	0.20	25	124	0.0269	−0.129	0.14	0.50
chr19		0.047	−0.26	1.56E-03	0.24	10	182	−0.0141	0.093	0.53	0.68
chr13	GPC5	0.042	−0.33	1.62E-03	0.25	11	185	0.0089	−0.081	0.09	0.50	1
chr3	SHOX2	0.034	−0.26	1.58E-03	0.24	10	198	0.0043	−0.039	0.32	0.57	10
chr19	ZNF470	0.045	−0.25	2.41E-03	0.28	10	237	−0.0177	0.104	0.61	0.71	244
chr17	TEKT3	0.031	−0.34	1.25E-03	0.22	10	285	−0.0025	0.026	0.71	0.73	3
chr6	HLA-L	0.041	−0.24	4.97E-03	0.39	17	352	0.0135	−0.090	0.09	0.50	62
chr11	RDX	0.032	−0.33	4.97E-03	0.39	10	354	0.0071	−0.073	0.14	0.50	723
chr22	KCNJ4	0.029	−0.18	2.88E-03	0.31	10	432	0.0075	−0.050	0.07	0.50	1
chr4		0.038	−0.35	7.03E-03	0.44	11	450	0.0029	−0.030	0.41	0.57
chr4	ZNF141	0.038	−0.35	7.03E-03	0.44	11	450	0.0029	−0.030	0.41	0.57	122
chr3	EPM2AIP1	0.034	−0.35	7.17E-03	0.44	27	456	0.0081	−0.087	0.03	0.41	874
chr19	ZNF569	0.038	−0.22	8.09E-03	0.47	10	486	−0.0004	0.007	0.69	0.72	61
chr5	MAST4	0.029	−0.25	8.29E-03	0.47	10	493	0.0032	−0.036	0.16	0.50	443
chr17	TMEM106A	−0.054	0.23	1.04E-02	0.49	10	599	0.0105	−0.045	0.79	0.79	145
chr1	DMRTA2	0.030	−0.23	1.14E-02	0.50	10	634	0.0057	−0.039	0.38	0.57	5
chr11	MUC2	−0.042	0.13	3.68E-06	0.01	10	697	−0.0228	0.080	0.02	0.41	10,143
chr6	ELOVL5	0.030	−0.24	1.44E-02	0.53	11	763	−0.0047	0.051	0.67	0.72	2903
chr6		0.030	−0.24	1.44E-02	0.53	11	763	−0.0047	0.051	0.67	0.72
chr11	RAB39A	0.029	−0.22	1.48E-02	0.53	12	779	0.0051	−0.036	0.16	0.50
chr6		0.025	−0.22	5.53E-03	0.41	15	783	0.0008	0.001	0.30	0.57
chr20	COL9A3	0.025	−0.15	1.59E-02	0.54	34	816	0.0029	−0.009	0.33	0.57	373
chr3	MLH1	0.031	−0.28	1.62E-02	0.55	36	832	0.0189	−0.117	0.02	0.41	1170
chr13	POU4F1	0.039	−0.24	1.64E-02	0.55	13	837	0.0119	−0.078	0.09	0.50	11
chr2	CACNB4	0.030	−0.26	1.67E-02	0.55	11	851	0.0099	−0.089	0.07	0.50	19
chr17	HAP1	0.025	−0.19	1.73E-02	0.56	10	874	0.0015	0.000	0.24	0.57	11
chr19	ZFP30	0.037	−0.15	1.76E-02	0.56	12	883	−0.0096	0.035	0.42	0.57	328
chr19	ZNF781	0.037	−0.15	1.76E-02	0.56	12	883	−0.0096	0.035	0.42	0.57	12
chr1	CHD5	0.026	−0.20	1.92E-02	0.58	11	931	0.0025	−0.009	0.29	0.57	9
chr5	RNF180	0.037	−0.16	1.96E-02	0.58	10	943	−0.0131	0.059	0.62	0.71	35
chr6	SPATS1	0.028	−0.11	1.64E-02	0.55	10	950	0.0058	−0.019	0.43	0.57	0
chr16	CBLN1	0.028	−0.19	2.05E-02	0.60	10	966	0.0108	−0.077	0.04	0.45	22
chr14	VASH1	0.029	−0.21	2.12E-02	0.60	10	987	−0.0018	0.011	0.67	0.72	721
chr12	WNT1	0.023	−0.22	1.29E-02	0.52	13	990	0.0072	−0.047	0.09	0.50	2
chr17	SPATA32	−0.036	0.11	9.11E-03	0.47	14	991	−0.0075	0.021	0.14	0.50
chr13	HS6ST3	0.025	−0.20	2.15E-02	0.60	10	1005	0.0012	−0.008	0.32	0.57	1
chr18	MAPK4	0.024	−0.16	2.18E-02	0.60	11	1016	0.0086	−0.035	0.17	0.50	4
chr17		0.027	−0.11	1.56E-02	0.54	10	1018	0.0155	−0.055	0.32	0.57
chr11	SLC35F2	0.028	−0.20	2.25E-02	0.60	14	1044	0.0021	−0.023	0.18	0.51	2085
chr18	DSC3	0.027	−0.18	2.25E-02	0.60	15	1045	−0.0071	0.047	0.70	0.72	88
chr18	GNAL	0.024	−0.25	2.26E-02	0.60	11	1048	0.0008	−0.005	0.23	0.57	67
chr7	TMEM176A	0.025	−0.21	2.27E-02	0.60	10	1049	0.0042	−0.048	0.26	0.57	6021
chr19	ZNF570	0.041	−0.21	2.34E-02	0.61	11	1078	0.0057	−0.025	0.44	0.58	52
chr17		0.026	−0.16	2.34E-02	0.61	10	1080	−0.0014	0.033	0.63	0.72
chr17	SEZ6	0.026	−0.18	2.47E-02	0.61	12	1123	0.0031	−0.029	0.20	0.53	1
chr1	SUSD4	0.028	−0.17	2.47E-02	0.61	11	1125	−0.0060	0.035	0.59	0.71	10
chr20	ISM1	0.031	−0.22	2.49E-02	0.61	30	1136	−0.0021	0.029	0.41	0.57	48
chr17	HOXB6	−0.026	0.10	4.01E-03	0.35	10	1139	−0.0140	0.065	0.39	0.57	1641
chr14	KCNK13	0.038	−0.22	2.51E-02	0.61	15	1143	0.0083	−0.048	0.22	0.57	24
chr15	ISL2	0.024	−0.17	2.53E-02	0.61	12	1147	0.0131	−0.087	0.12	0.50	17
chr2	MFSD2B	0.026	−0.18	2.79E-02	0.63	13	1239	0.0029	−0.026	0.32	0.57	217
chr6	NHLRC1	0.032	−0.17	2.81E-02	0.63	13	1248	0.0212	−0.131	0.02	0.41	145
chr20	SOX18	0.025	−0.15	2.86E-02	0.63	14	1264	0.0096	−0.047	0.16	0.50	185
chr14		0.027	−0.19	2.86E-02	0.63	11	1265	−0.0014	0.009	0.59	0.71
chr22	ADORA2A-AS1	0.031	−0.10	2.92E-02	0.64	16	1287	−0.0153	0.056	0.30	0.57
chr8	EBF2	0.031	−0.18	2.99E-02	0.64	10	1307	−0.0111	0.068	0.41	0.57	3
chr4	KDR	0.028	−0.09	1.34E-02	0.52	12	1317	0.0290	−0.084	0.12	0.50	719
chr1	HMCN1	0.021	−0.20	2.95E-02	0.64	12	1329	−0.0030	0.025	0.59	0.71	212
chr11	MIR1908	0.029	−0.23	3.17E-02	0.65	16	1362	0.0039	−0.029	0.38	0.57

Abbreviations: Chr, chromosome; FDR, false-discovery rate.

Tumour purity was estimated by the method published by Aran et al. (57) using 44 CpGs consistently unmethylated in immune cells but consistently methylated in tumour cells for estimation.

All analyses adjusted by age, sex, smoking, cancer stage, tumour location, CIMP status, tumour purity.

^acalculated using M-values .

^bregions sorted by combinedRank, a metirc given to the sites out of mean difference in means across all sites in a region, mean of ratios in mean methylation and the combined p-value of the region.

^cRNAseq gene expression values (median log2 (read counts)) for colon cancer were retrieved from the BioXpress database (59).

Pathway enrichment analysis

Pathway enrichment analysis was conducted to determine whether the genes associated with the 2904 significant CpG sites identified in the secondary analysis (comparing tumours with high ERβ expression to tumours of negative expression) were enriched for particular pathways. Results of the pathway enrichment analysis are shown in Supplementary Table 4 for all pathways with enrichment p-value <0.05. Several enriched pathways are known to be involved in carcinogenesis, for example Wnt signalling pathway, Rap1 and Ras signalling pathway.

An overview of all analyses and results is given in Figure 3.

Figure 3.

Overview of analyses and results.

Survival analysis

We tested whether the 403 replicated CpG sites are associated with overall survival in the discovery set (Supplementary Table 5). Of the 915 included patients, 287 died during follow-up. Median follow-up time was 56.2 months. Cox-proportional hazard models were used for estimation of hazard ratios (HR) and 95% confidence intervals, models were adjusted by age, sex, cancer site, smoking, CIMP status, cancer stage and tumour purity.

The survival analysis yielded 62 CpG sites (40 genes) being nominally associated with overall survival (p-value<0.05). Table 4 shows 30 CpGs (22 genes) which remained significantly associated with overall survival after adjustment for multiple testing. These CpGs are annotated to genes such as for example FOXK1 or TMEFF2.

Table 4.

Significant associations of 30 ERβ associated CpG sites with overall survival after CRC diagnosis.

CpG	Chr	Gene	N	Events	beta	SE	p-value	FDR adjusted p-value	HR (CI)
cg26177213	chr7	FOXK1	915	287	−2.22	0.59	1.82E-04	0.0007	0.11 (0.03–0.35)
cg21547763	chr1	C1orf226	915	287	2.03	0.58	4.33E-04	0.0017	7.61 (2.46–23.55)
cg00397714	chr2	TMEFF2	915	287	1.77	0.53	8.35E-04	0.0033	5.88 (2.08–16.63)
cg20637199	chr19	STK11	915	287	−1.77	0.53	8.95E-04	0.0036	0.17 (0.06–0.48)
cg22467470	chr19	STK11	915	287	−1.48	0.45	1.13E-03	0.0045	0.23 (0.09–0.55)
cg12492273	chr7	MAD1L1	915	287	−1.19	0.37	1.47E-03	0.01	0.31 (0.15–0.63)
cg07536072	chr19	STK11	915	287	−1.68	0.53	1.59E-03	0.01	0.19 (0.07–0.53)
cg25570453	chr7	DGKB	915	287	1.17	0.38	2.07E-03	0.01	3.22 (1.53–6.78)
cg09418321	chr12	DYRK4	915	287	−1.41	0.46	2.23E-03	0.01	0.25 (0.10–0.60)
cg04001880	chr15	POLG	915	287	−1.71	0.57	2.64E-03	0.01	0.18 (0.06–0.55)
cg05818515	chr1	SDF4	915	287	−2.15	0.73	3.10E-03	0.01	0.12 (0.03–0.48)
cg24437104	chr1	SNAP47	915	287	−2.05	0.70	3.15E-03	0.01	0.13 (0.03–0.50)
cg18173263	chr17	PEMT	915	287	−1.24	0.42	3.26E-03	0.01	0.29 (0.13–0.66)
cg13436968	chr7	VSTM2A	915	287	1.02	0.35	3.56E-03	0.01	2.76 (1.39–5.47)
cg05694563	chr17	PEMT	915	287	−2.11	0.73	3.73E-03	0.01	0.12 (0.03–0.50)
cg02625024	chr17	PEMT	915	287	−1.42	0.49	3.82E-03	0.01	0.24 (0.09–0.63)
cg27498434	chr8	GSR	915	287	−1.46	0.51	4.10E-03	0.02	0.23 (0.09–0.63)
cg02400474	chr19	ZSCAN22	915	287	−1.22	0.42	4.19E-03	0.02	0.30 (0.13–0.68)
cg19373760	chr1		915	287	1.59	0.56	4.53E-03	0.02	4.90 (1.64–14.69)
cg05211192	chr7	MAD1L1	915	287	−1.37	0.49	4.82E-03	0.02	0.25 (0.10–0.66)
cg01267270	chr5	PHAX	915	287	1.61	0.57	4.97E-03	0.02	5.01 (1.63–15.41)
cg05235525	chr14	WDR20	915	287	−1.41	0.51	5.09E-03	0.02	0.24 (0.09–0.65)
cg00337703	chr10	TDRD1	915	287	1.90	0.68	5.30E-03	0.02	6.71 (1.76–25.59)
cg26309566	chr11	LRP5	915	287	−1.83	0.66	5.97E-03	0.02	0.16 (0.04–0.59)
cg13712852	chr2	VAMP5	915	287	−1.66	0.61	6.37E-03	0.02	0.19 (0.06–0.63)
cg26680995	chr2		915	287	1.16	0.43	7.30E-03	0.03	3.18 (1.37–7.42)
cg22009751	chr2	HDAC4	915	287	1.33	0.52	1.01E-02	0.04	3.79 (1.37–10.44)
cg22542139	chr17	PEMT	915	287	−1.16	0.45	1.06E-02	0.04	0.31 (0.13–0.76)
cg21010701	chr4	DTHD1	915	287	−1.30	0.52	1.24E-02	0.05	0.27 (0.10–0.75)
cg26138637	chr7	CD36	915	287	0.95	0.38	1.27E-02	0.05	2.60 (1.23–5.50)

Abbreviations: Chr, chromosome; CI, confidence interval; FDR, false-discovery rate; HR, hazard ratio; SE, standard error;

Analyses adjusted by age, sex, smoking, cancer stage, tumour location, CIMP status, tumour purity.

Discussion

Oestrogen and its receptor ERβ are considered to play an important role in CRC development and loss of expression of the latter has been associated with poorer survival among CRC patients [6]. To gain insight into possible mechanisms underlying the relationship between ERβ expression and survival after CRC, genome-wide DNA methylation differences between ERβ positive and ERβ negative tumours of CRC patients were investigated.

In the primary analysis, comparing samples with moderate/high expression of ERβ and without ERβ expression, 1,220 CpG sites differentially methylated were identified in the discovery dataset. While none of those sites were replicated after multiple testing adjustments using an independent dataset, 176 of these CpG sites showed nominal p-value<0.05. These CpGs were annotated to genes such as CD36, HK1 or LRP5, which were also found to be associated with high ERβ expression in the secondary analysis.

In the secondary analysis comparing tumours of high ERβ expression with tumours of negative expression, 403 differentially methylated CpGs were found replicated. Most of these replicated CpGs are hypomethylated in ERβ positive tumours, which is in line with the study by Liu et al. [14] which reported recruitment of thymine DNA glycosylase by ERβ leading to demethylation. The CpG sites cg06748146 and cg00542261 located in the gene body of HK1 gene (hexokinase 1) were less methylated in ERβ positive tumours compared to ERβ negative tumours in the current analysis. Gene body methylation is usually positively correlated with expression [15]. Former studies reported that increased expression of HK1 was associated with TNM stage and poorer overall survival [16]. So, lower gene body methylation of HK1 could be associated with lower expression and therefore increased overall survival of ERβ positive tumours. In addition to this evidence, the promotor region of HK1 contains a transcription factor-binding site for ERβ, which makes it even more likely that ERβ could be involved.

Some genes contained several of these 403 CpGs. Among these genes was CD36 (CpGs: cg26138637, cg18508525, cg27625491) which showed higher methylation levels in ERβ positive tumours compared to ERβ negative tumours. CD36 is a transmembrane glycoprotein receptor that has recently been found as a common marker of metastatic cells and plays a role in tumour angiogenesis and in stromal transformation [17]. Low levels of CD36 have been reported to be correlated strongly with improved prognosis of CRC [18]. Since higher methylation levels are expected to result in lower levels of CD36, this gene could be involved in mechanisms leading to better prognosis in ERβ positive tumours.

Also five CpGs (cg01398428, cg02606728, cg26309566, cg27426151, cg07056967) in the body of LRP5 (Low-density lipoprotein receptor-related protein 5), a receptor involved in the Wnt signalling pathway [19], were less methylated in ERβ tumours. Aberrant activation of the Wnt signalling pathway is related to CRC development [20]. Thus lower methylation of gene body of LRP5 might lead to lower expression and thereby regulate the Wnt signalling pathway and lower the risk of ERβ positive tumours. In addition, differentially methylated CpG sites were found associated to genes, which have not been previously reported to be associated with colorectal carcinogenesis such as STT3B (Catalytic Subunit Of The Oligosaccharyltransferase Complex) or CMIP (C-Maf Inducing Protein), which warrant further investigation.

Nine replicated genes were found to have transcription factor binding sites for ERβ. Among these genes are the putative tumour suppressor genes PRDM2 [21], SAMD14 [22] and UNC5D [23]. In addition, XRCC3 [24] has previously been reported to be associated with time to metastasis in CRC and loss of ZG16 [25] found related with progression of CRC. So further understanding of the interaction of these genes with ERβ in disease progression is needed.

The 60 promotor regions identified to be differentially methylated between ERβ positive and negative tumours based on the combined rank calculated in RnBeads were not confirmed in the replication set after adjustment for multiple testing. The gene expression profiles of these genes, retrieved from the BioXpress database, indicate that many of these genes are expressed in colon and rectal cancer. Five promotor regions were replicated at nominal significance (p-value<0.05), namely MUC2, EPM2AIP1, MLH1, CBLN1 and NHLRC1. MUC2 (Mucin 2) silencing has been associated with carcinogenesis and poor survival in CRC [26]. MUC2 showed lower methylation in ERβ positive tumours and therefore might be upregulated. EPM2AIP1 is higher methylated in positive tumours. This is a gene regulated by the same promotor region as MLH1 [27], for which higher methylation levels were also found in our dataset. The function of EPM2AIP1 is so far unknown. The MLH1 gene encodes a mismatch repair protein which is downregulated by promotor methylation in tumours emerging through the microsatellite instability pathway (MSI) (15–20% of all CRCs). These tumours have a slightly better prognosis compared to microsatellite stable tumours [28]. Furthermore, MLH1 is downregulated by promotor hypermethylation in CIMP [29] The higher promotor methylation of MLH1 associated with ERβ positive tumours found in this analysis might play a role for differences in ERβ expression associated survival. Furthermore, there is a positive correlation between oestrogen levels and MLH1 expression [30]. Oestrogen has been found to induce MLH1 expression through ERβ, and thereby enhances mismatch repair activity. As a consequence, ERβ may exert anti-tumorigenic effects on colon cells [31].

Two further nominally replicated promotor regions, CBLN1 (cerebellin precursor protein), a secreted cerebellum-specific protein [32] and NHLRC1 (NHL Repeat Containing E3 Ubiquitin Protein Ligase 1 or malin gene), associated with a rare neurodegenerative disease [33], have unknown function in CRC tissue. Two promotor regions (of genes MUC2 and KDR) included replicated CpGs detected in the secondary analysis. The gene KDR (kinase insert domain receptor) was found stronger methylated in ERβ positive tumours. KDR, which is also known as vascular endothelial growth factor receptor 2 (VEGFR-2), functions as the main mediator of VEGF-induced endothelial proliferation, survival and migration [34]. Overexpression is associated with metastasis and poor prognosis [34]. So higher methylation in ERβ positive tumours and thereby downregulation of KDR could be involved in improved prognosis.

We did not find an association between DNA methylation in the ESR2 promotor region and expression of ERβ measured by immunohistochemistry. Expression of ERβ could have been influenced by other mechanisms than methylation of the promotor region. Expression of ERβ could have been regulated by non-coding RNAs like microRNAs or long non-coding RNAs [35] or other mechanisms influencing translation [36].

The pathway analysis points to an involvement of the Wnt signalling pathway, which regulates cell growth and thereby plays a critical role for gastrointestinal tract homeostasis [20]. Further pathways, involved in carcinogenesis are the Rap1 (Ras-related protein 1) signalling pathway and the Ras (Rat sarcoma) signalling pathway. Impairment of the Rap1 signalling pathway or downregulation of Rap1 has be reported to be associated with a poor prognosis [37]. Deregulated Ras signalling is associated with increased proliferation, angiogenesis and decreased apoptosis [38]. About 45% of all CRC tumours display an activating K-Ras mutation [39]. Thus a better understanding of the role of ERβ in regulation of these pathways could identify potential targets to improve the prognosis of CRC patients with ERβ negative tumours.

As differences in survival between patients with ERβ positive and ERβ negative tumours have been reported [6], we were interested to know whether the 403 CpG sites associated with ERβ expression from this analysis are also associated with survival in CRC patients. We found that 30 CpGs were associated to overall survival after FDR-adjustment. The most significant CpG is annotated to the FOXK1, a gene with reported association with proliferation, invasion and metastasis in colorectal cancer [40–42]. Also the genes TMEFF2 [43] and STK11 [44] have been found associated with survival in CRC patients.

The clinical characteristics of the two datasets were comparable and differences found in smoking and CIMP status were adjusted for in the statistical analysis. Using RPMM clustering based on methylation beta values, clusters of tumours in both data sets were identified, which correspond to the CIMP molecular subtypes as previously described [45], thus demonstrating the validity of discovery and replication datasets.

Limitations of our study include the fact that methylation patterns are strongly cell type specific and contaminations of tumour tissue sections with non-tumour cell types such as immune cells, fibroblast or others can have an influence on the measured methylation values, which might not have been entirely accounted for by adjusting for tumour purity in the data analysis. The considerable intra-tumour heterogeneity of the ERβ staining [6] (moderate versus strong) observed by immunohistochemistry might also affect methylation patterns as the tumour tissue used in the methylation analysis was not specifically selected for homogeneous ERβ staining. This might have affected the results of the comparison of the methylation pattern, especially for the primary analysis where all tumours showing moderate to high ERβ expression were combined. However, the misclassification of ERβ expression, which is likely to have been random, would have led to attenuation of methylation differences associated with ERβ expression. To minimize this problem, the secondary analysis was conducted, where only tumours with high expression were compared to tumours with negative expression of ERβ. Thus, these results based on this analysis are likely to be more robust against misclassification of ERβ expression.

Some patients had to be excluded because of missing information of ERβ expression (n = 199 of the discovery set and n = 140 of the replication set). The information was missing because ERβ measurement was unsuccessful (staining unsuccessful, loss of cores etc.) for 13.9% of available patient samples recruited between 2007 and 2010 and 12.8% of patient samples recruited between 2003 and 2006. These proportions look high but are not uncommon for immunohistochemistry [46]. By using the available samples, selection bias could have occurred but is more likely to have been random so that replication results should not have been substantially affected. This is the first study exploring epigenome-wide differential methylation according to ERβ expression. The strengths of the study include the use of large and comparable discovery and replication sets derived from a population based cohort of CRC cases, ERβ expression data that were scored independently by two pathologists, and the availability of information on several important clinical and molecular factors for adjustment in the statistical analysis to minimize confounding and enhance validity of results.

In conclusion, we found differentially methylated genes according to ERβ expression status and were able to replicate 403 CpGs associated with strong ERβ expression. Besides novel genes related to these sites, such as NOL9 or HEATR2, some of the genes, like CD36, HK1 or LRP5, are known to be associated with CRC carcinogenesis. Thus these genes should be of interest for follow-up mechanistic studies to better understand the role of ERβ in regulation of specific genes and their influence on development and progression of CRC.

Material and methods

Study design and study population

For this analysis, data of an ongoing population based case-control study, DACHS-study (Darmkrebs: Chancen der Verhütung durch Screening), was used, details have been previously described [47–49]. Briefly, patients were recruited as cases in all 22 hospitals in the study area (Rhine-Neckar region, Southern Germany, approximately 2 million inhabitants) providing initial treatment for CRC. Patients were eligible for the study if they were at least 30 years old and diagnosed with histologically confirmed CRC, and they needed to be physically and mentally able to participate in a personal interview. Extensive epidemiologic information has been collected through face-to-face interviews at recruitment and clinical information was obtained from medical records. All participants gave written informed consent. The study was approved by the ethics committee of the University of Heidelberg and the medical boards of Baden-Wuerttemberg and Rhineland-Palatinate. Eligible for the present analysis was a series of 1,137 CRC patients diagnosed between January 2007 and December 2010 with available DNA methylation data, used as discovery set. Additional 1,064 patients diagnosed between January 2003 and December 2006 were eligible for replication analyses.

Immunohistochemistry

Formalin-fixed paraffin-embedded tumour samples were obtained from the pathology institutes of the cooperating clinics and transferred to the tissue bank of the National Center for Tumor Diseases (Heidelberg, Germany), where tissue microarrays were constructed. ERβ-staining was carried out as described in Rudolph et al. [6]. In short, the anti-ERβ antibody (primary mouse monoclonal, 14C8, Abcam) was applied at a dilution of 1/50 at room temperature for 30 min. After incubation with the appropriate biotinylated secondary antibody (Dako antimouse, dilution 1/200, Dako), an incubation with the streptavidin avidin-biotin complex kit (Dako) was performed. Following endogenous peroxidase blocking, the antibody reactions were revealed using the Dako EnVision + System-HRP. Staining was performed with a Dako autostainer (DakoCytomation) based on the avidin-biotin-complex method. Two pathologists assessed the expression of ERβ independently (C.T. and W.R assessed samples resected between 2003 and 2006 and C.T. and H.Bl. assessed samples resected between 2007 and 2010) using a three level categorical scoring system based on Konstantinopoulos et al. [5]: 0 = negative expression (< 10% of cell nuclei positive), 1 = moderate expression (>50% of cell nuclei moderate positive or 10–50% of cell nuclei with strong positive staining), 2 = high expression (>50% of cell nuclei with strong positive staining). For samples resected between 2003 and 2006 discrepant scores were resolved in an additional joint review. For samples resected between 2007 and 2010 discrepant scores of 0/1 or scores 1/2 were resolved in favour of the score of the pathologist (C.T.) who also evaluated the first set. Discrepant scores of 0/2 were resolved in an additional review.

DNA methylation profiling

DNA was isolated from formalin-fixed, paraffin-embedded tumour samples using the DNeasy Blood & Tissue Kit (Qiagen, Hilden, Germany), for further details see Gündert et al. [50]. Infinium HumanMethylation 450k Bead Chip arrays (Illumina, San Diego, CA, USA) were used to measure methylation levels in CRC tissue DNA following the manufacturer’s protocol [51]. The array measures methylation at 485,577 cytosine-guanine dinucleotide (CpG) sites, using bisulfite-converted DNA. It has coverage of 99% of RefSeq genes (mean of 17 CpG sites per gene). The measured CpG sites are concentrated around promotor regions and genes [12]. Profiling of the tumour samples included in the discovery set and in the replication set were both carried out at the DKFZ (German Cancer Research Center) Genomics and Proteomics Core Facility. Methylation beta values (ratio of methylated probe intensity versus methylated + unmethylated intensities) were used for describing the level of methylation at a locus and for graphical representations. Beta values have a range between 0 (unmethylated) and 1 (methylated). Due to distributional properties, M-values (defined as log2 ratio of methylated and unmethylated probe intensities) were used for differential methylation analyses [52] as done by the R-package RnBeads by default.

DNA methylation of samples in the replication set was measured in two batches more than one year apart, therefore these datasets were combined with correction for batch. All samples of the discovery set were measured in one batch.

Preprocessing and normalization of methylation data

For quality control, preprocessing and for conducting the analyses the R package RnBeads [53] (R version 3.2.3/Bioconductor packages) was used. The discovery methylation dataset initially contained 1,137 samples with 485,577 probes each. In the first filtering step, 24,837 probes and 15 samples were removed by the greedycut algorithm identifying them as unreliable measurements (based on detection p-value >0.05). Greedycut implemented in RnBeads is an iterative algorithm which filters out probes and samples with a high fraction of unreliable measurements [53]. Additionally 10,119 CpG sites within three basepairs of common single nucleotide polymorphisms (SNPs) were excluded because they can affect probe hybridization [54]. Furthermore, 29,928 cross-reactive probes mapping to multiple sites at the genome were filtered out [55]. Raw intensities were normalized using SWAN (Subset-quantile within array normalization) [56] to correct for the technical differences between type 1 and type 2 array designs in the 450k array. After normalization a second filtering step was performed where 9,108 probes on sex chromosomes were removed, 1,014 probes without acceptable context, and 14,523 probes with missing values were removed. At the end 396 048 probes remained for analysis. The replication set was normalized in the same way (using SWAN), but omitting filtering. The same CpGs as in the discovery set were retained manually to allow for an independent replication.

Genomic annotation of CpG sites

The following categories were used for describing the location of CpG sites relative to a gene: TSS1500 (1,500 bp upstream from transcription start site – TSS), TSS200 (200bp upstream from TSS), 1^st exon, 5`UTR (5`untranslated region), body (gene body) and 3`UTR (3`untranslated region). For CpG locations relative to CpG islands the following categories were used: CpG island, S-Shore, N-Shore (up to 2kb up- and downstream of CpG-islands), S-Shelf and N-Shelf (2–4 kb up- and downstream of CpG-islands), OpenSea (all other locations).

Statistical analysis

For the primary analysis, samples with moderate or high ERβ expression (scores 1 and 2) were considered as ERβ positive and compared to samples with ERβ negative expression (score 0). To detect differentially methylated CpG sites between ERβ positive and ERβ negative samples, we used linear models (limma, implemented in RnBeads). Differential methylation analyses were adjusted by age (continuous), sex, cancer site (colon/rectum), smoking (ever/never/current), CpG island methylator phenotype (CIMP) status, cancer stage (stages I-III/stage IV) and tumour purity. CIMP status was defined by the number of hypermethylated loci based on the genes MLH1, MINT1, MINT2, MINT31 and MGMT, whereby hypermethylation at three or more loci was classified as CIMP-high and methylation at less than three loci was classified as CIMP-negative (as described in Jia et al. [29]). Tumour tissue purity was estimated using the method published by Aran et al. [57], a computational method which uses 44 CpGs consistently unmethylated in immune cells but consistently methylated in tumour cells for estimation. The method yields continuous values for tumour purity between 0 and 1. P-values were corrected for multiple testing using Benjamini-Hochberg adjustment and significance was assessed at a false-discovery rate (FDR) threshold of 0.05. For a secondary analysis, samples with high ERβ expression (score 2) were compared to ERβ negative (score 0) samples, excluding the samples with moderate ERβ expression (score 1). Adjustment and further procedures were identical as for the primary analysis.

As methylation of CpG sites can be highly correlated by function and genomic location, differences in promotor regions according to ERβ expression status were additionally investigated using RnBeads` internal method for identifying differentially methylated genomic regions (DMRs) (predefined regional clusters of neighbouring CpGs). For this analysis, promotor regions are defined as the regions 1.5 kb up-stream and 0.5 kb downstream of transcription start sites. To consider both statistical significance and biological relevance, RnBeads calculates a priority ranked list of DMRs using three different criteria. To measure statistical significance a combined p-value per region is calculated out of the uncorrected p-values of CpGs within a region using a generalization of Fisher’s method [58]. These combined p-values are corrected for multiple testing using FDR-adjustment. Biological relevance is measured using mean absolute difference in methylation and relative ratio of mean methylation. Ranks are assigned to the DMRs based on each of the three metrics. The combined rank is the maximum (i.e. worst) of the three individual ranks. A rank cutoff is provided by RnBeads. A lower combined rank indicates a higher chance of biologically relevant differential methylation.

The replication dataset was used for independent replication of the results, performing the same analysis for differential methylation analyses and DMRs on the significant CpG sites and regions identified in the discovery dataset. Multiple testing adjustment was carried out using the Benjamini-Hochberg method employing a threshold of 0.05.

Gene expression profiling from BioXpress database

To investigate whether differentially methylated genes found in the discovery analysis are expressed in colon cancer, gene expression (median log2 (read counts)) of these genes was retrieved from the BioXpress database [59] (https://hive.biochemistry.gwu.edu/tools/bioxpress/). BioXpress is a database of expression data from the Cancer Genome Atlas (TCGA: http://cancergenome.nih.gov/), the International Cancer Genome Consortium (ICGC: https://icgc.org/) and manually curated data from publications.

Unsupervised clustering

For quality control and to compare the discovery and replication sets, we performed unsupervised clustering of the two datasets separately. Clustering was performed as described in Hinoue et al. [45] using recursively partitioned mixture model (RPMM) as implemented in the R-package RPMM [60]. After preprocessing of the datasets (see above) and excluding probes with missing values, unsupervised clustering was performed based on the top 1,500 probes with the most variable methylation beta values (based on standard deviation). Graphical representation was made using the heatmap.3 function in R. The samples were ordered within the RPMM classes using the function Seriate of the seriation package in R with default parameters.

Pathway enrichment analysis

Pathway enrichment analysis was performed with the genes associated with the 2,904 differentially methylated CpG sites (FDR adjusted p-value <0.05) from the secondary analysis using the Database for Annotation, Visualization and Integrated Discovery (DAVID) web tool [61,62].

Survival analysis

Recruited patients were followed up with respect to survival at 3 and 5 years after CRC diagnosis and vital status was determined using population registries. To investigate whether the replicated CpG sites are associated with overall survival, Cox proportional hazards models were employed to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for all replicated CpGs in 915 patients of the discovery set with respect to overall survival. Models were adjusted by the same adjustment variables (age, sex, cancer site, smoking, CIMP status, cancer stage and tumour purity) as used for the differential methylation analysis. Survival time was calculated from date of diagnosis until date of death from any cause or date of last contact. Resulting p-values were adjusted for multiple testing using FDR-adjustment.

Funding Statement

This work was supported by the German Federal Ministry of Education and Research under Grants 01KH0404, 01ER0814, 01ER0815, 01ER1505A, 01ER1505B; German Research Council under Grants BR 1704/6-1, BR 1704/6-3, BR 1704/6-4, CH 117/1-1, HO 5117/2-1, HE 5998/2-1, KL 2354/3-1, RO 2270/8-1, BR 1704/17-1; and the Interdisciplinary Research Program of the National Center for Tumor Diseases (NCT), Germany.

Acknowledgments

We are very grateful to the study participants and interviewers who collected the data. We would also like to thank the following hospitals and cooperating institutions that recruited patients for this study: Chirurgische Universitätsklinik Heidelberg, Klinik am Gesundbrunnen Heilbronn, St Vincentiuskrankenhaus Speyer, St Josefskrankenhaus Heidelberg, Chirurgische Universitätsklinik Mannheim, Diakonissenkrankenhaus Speyer, Krankenhaus Salem Heidelberg, Kreiskrankenhaus Schwetzingen, St Marienkrankenhaus Ludwigshafen, Klinikum Ludwigshafen, Stadtklinik Frankenthal, Diakoniekrankenhaus Mannheim, Kreiskrankenhaus Sinsheim, Klinikum am Plattenwald Bad Friedrichshall, Kreiskrankenhaus Weinheim, Kreiskrankenhaus Eberbach, Kreiskrankenhaus Buchen, Kreiskrankenhaus Mosbach, Enddarmzentrum Mannheim, Kreiskrankenhaus Brackenheim, Cancer Registry of Rhineland-Palatinate, Mainz. We are also very grateful for the support of the pathologies in the provision of tumor samples: Institut für Pathologie, Universitätsklinik Heidelberg; Institut für Pathologie, Klinikum Heilbronn; Institut für Angewandte Pathologie, Speyer; Pathologisches Institut, Universitätsklinikum Mannheim; Institut für Pathologie, Klinikum Ludwigshafen; Institut für Pathologie, Klinikum Stuttgart; Institut für Pathologie, Klinikum Ludwigsburg. In addition, we like to thank the Genomics and Proteomics Core Facility for provision of the methylation analysis services, especially Matthias Schick, Dr. Melanie Bewerunge-Hudler and Prof. Dr. Stefan Wiemann.

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary material

Supplementary data for this article can be accessed here.