DNA methylation levels and long-term trihalomethane exposure in drinking water: an epigenome-wide association study

Abstract

Trihalomethanes (THM) are undesired disinfection byproducts (DBPs) formed during water treatment. Mice exposed to DBPs showed global DNA hypomethylation and c-myc and c-jun gene-specific hypomethylation, while evidence of epigenetic effects in humans is scarce. We explored the association between lifetime THM exposure and DNA methylation through an epigenome-wide association study. We selected 138 population-based controls from a case-control study of colorectal cancer conducted in Barcelona, Spain, exposed to average lifetime THM levels ≤85 μg/L vs. >85 μg/L (N = 68 and N = 70, respectively). Mean age of participants was 70 years, and 54% were male. Average lifetime THM level in the exposure groups was 64 and 130 µg/L, respectively. DNA was extracted from whole blood and was bisulphite converted to measure DNA methylation levels using the Illumina HumanMethylation450 BeadChip. Data preprocessing was performed using RnBeads. Methylation was compared between exposure groups using empirical Bayes moderated linear regression for CpG sites and Gaussian kernel for CpG regions. ConsensusPathDB was used for gene set enrichment. Statistically significant differences in methylation between exposure groups was found in 140 CpG sites and 30 gene-related regions, after false discovery rate <0.05 and adjustment for age, sex, methylation first principal component, and blood cell proportion. The annotated genes were localized to several cancer pathways. Among them, 29 CpGs had methylation levels associated with THM levels (|Δβ|≥0.05) located in 11 genes associated with cancer in other studies. Our results suggest that THM exposure may affect DNA methylation in genes related to tumors, including colorectal and bladder cancers. Future confirmation studies are required.

Introduction

Disinfection byproducts (DBPs) appear during water treatment when organic matter reacts with the disinfectant. Trihalometanes (THMs) are one of the most abundant classes in chlorinated waters; in the past, Spain has shown high levels of THMs compared to Northern Europe.¹ Lifetime exposure to DBPs has been related to bladder cancer^2,3 and, possibly, colorectal cancer⁴ in epidemiological studies. However, the mechanisms involved in these deleterious effects are not clear.

Some mechanisms of action have been proposed for DBPs from experimental models. Cytotoxicity, defined as an alteration of the cell integrity with or without direct DNA damage, has been shown by some compounds (i.e., chloroform and trichloroacetic acid), probably dependent on the release of oxidative free radicals during liver reductive metabolism through cytochromes.^5,6 For some brominated compounds, it has been proposed that in vivo bioactivation through glutathione S transferases (GSTT1) generates carbonyl radicals, which could be genotoxic (as they attach as adducts to DNA).⁷ However, these 2 mechanisms do not completely explain the effects observed when doses are low and exposure is chronic. In chronic low doses, toxicants may generate non-mutational changes as a consequence of repetitive cytotoxicity/regeneration cycles that may produce somatic changes in the methylation of DNA in the DNA daughter strands, leading to DNA instability and induction of apoptotic pathways.⁸ In mice, the exposure to several trihalometanes and haloacetic acids (the second most abundant DBP in chlorinated waters) reduced global DNA methylation levels. Specifically, hypomethylation of some protooncogenes (c-myc and c-jun) associated with increased mRNA expression has been documented. The exposed mice had higher than expected rates of kidney and liver tumors.^8-11

The mechanistic evidence in humans is limited. Epidemiological studies have shown an interaction with polymorphisms of GSTT1 and CYP2E1 for bladder cancer in subjects chronically exposed to THM in drinking water.¹² Furthermore, levels of DNA methylation at some transposons (LINE-1) in granulocytes increased with lifetime THM exposure in controls.¹³ To our knowledge, there are no studies exploring THM chronic exposure and changes in genome-wide DNA methylation. In this context, we aim to explore the association between lifetime exposure to THM and DNA methylation levels through an epigenome-wide association study.

Results

Mean age and sex ratio of study participants was similar between exposure groups for the 138 analyzed subjects (Table 1). Only 6 subjects in the low-exposure group and 10 in the high-exposure group were current smokers. Most of the subjects (>70%) had elementary studies or less. The estimated blood cell proportions of CD8 + T-, CD4 + T-, NK-, B-lymphocytes and monocytes were not significantly different between exposure groups, while the proportion of granulocytes was significantly higher in the low-exposure group. In the high-exposure group, most of THM exposure was due to brominated THM, while chloroform predominated in the low-exposure group.

Table 1.

Characteristics of the study subjects by exposure groups

	THM* ≤ 85 µg/L	THM* > 85µg/L
Variables	n = 68	n = 70	P-value
Sex, n(%)
Male	34 (50%)	41 (59%)	0.3
Female	34 (50%)	29 (41%)
Age years, mean (SD)	70.5 (5.9)	70.1 (5.6)	0.7
Trihalomethane levels (µg/L), median (IQR)
Chloroform	21.7 (18.7, 24.1)	17.3 (16.4, 17.6)	<0.001
Bromodichloromethane	21.0 (19.1, 23.2)	35.1 (32.2, 35.6)	<0.001
Dibromochloromethane	11.7 (10.6, 13.1)	28.1 (25.5, 29.3)	<0.001
Bromoform	12.0 (9.8, 14.1)	49.2 (45.6, 51.9)	<0.001
Total THM	64.0 (58.8, 72.7)	129.9 (118.4, 134.1)	<0.001
Smoking status, n(%)
Never smoker	37 (54%)	35 (50%)	0.6
Former smoker	25 (37%)	25 (36%)
Current smoker	6 (9%)	10 (14%)
Municipality, n(%)
A	10 (15%)	0 (0%)	<0.001
B	57 (84%)	3 (4%)
C	1 (1%)	17 (24%)
D	0 (0%)	50 (71%)
Education level, n(%)
Elementary or less	54 (79%)	52 (74%)	0.6
High school	11 (16%)	12 (17%)
Universitary or more	3 (4%)	6 (9%)
Estimated proportions of blood cell types, median (IQR)
CD8 T-lymphocytes	3.27 (1.32, 5.67)	4.80 (1.89, 6.87)	0.07
CD4 T-lymphocytes	13.36 (10.73, 16.79)	14.46 (9.60, 17.31)	0.6
NK lymphocytes	8.62 (5.98, 11.39)	10.33 (6.51, 13.35)	0.06
B lymphocytes	3.17 (2.25, 4.14)	3.61 (2.30, 5.23)	0.08
Monocytes	8.23 (7.00, 9.75)	8.26 (6.56, 9.57)	0.8
Granulocytes	61.63 (57.35, 69.07)	59.74 (54.56, 64.41)	0.01

^*average lifetime trihalomethanes level

The methylation intensities in both Infinium I and II assays showed the expected bimodal distribution when all probes were plotted together, but approximately normal distribution for most of the CpGs when considered individually. From the original 485,577 probes, we excluded potentially unreliable probes leaving 252,156 for further analyses (Fig. 1). Among these, crude mean methylation levels were significantly different between exposure groups in 23,302 CpG sites after multiple comparison correction (FDR < 0.05, λ  = 1.207). Using the selectModel command, the variables that best adjusted all the models were age, sex, the first principal component (surrogate variable), and the estimated blood cell proportions. After adjustment for these covariables, 1840 CpG sites showed different mean methylation levels between exposure groups after FDR correction (FDR < 0.05, λ = 1.048). Using the observed mean methylation values, 140 CpG sites had an absolute observed Δβ > 0.05 between exposure groups (Table S1). Among them, in 29 CpG sites, the covariate adjusted β coefficient showed differences > 0.05 due to THM exposure (absolute expected Δβ between 0.051 and 0.150 Table 2). A cluster of high-exposure subjects was associated with differential methylation of these 29 CpG sites (Fig. 2).

Figure 1.

Flowchart of 450K Infinium Methylation BeadChip sample analyses. Note: *Differential methylation was defined as an |Δβ|>0.05 and false discovery rate-FDR<0.05.

Table 2.

Top CpG sites (n = 29) with methylation levels associated with THM exposure (|Δβ|>0.05 and FDR<0.05), after adjusting for covariables (age, sex, first principal component and blood cells proportion)

IllumID	Gene symbola	Gene name	Gene features	EntrezID	Mean methylation if THM <85	Observed Δβ (95%CI)b	Expected Δβ (95%CI)c	FDR	Associated diseases-functiond
cg12949141	PCBD2	pterin-4 α-carbinolamine dehydratase	Body_none	84105	0.556	0.095 (0.07,0.12)	0.113 (0.09,0.136)	3.72E-11	Cluster of genes of Colorectal cancer
cg07613278	API5	apoptosis inhibitor 5	TSS200_shore	8539	0.116	0.067 (0.051,0.082)	0.054 (0.039,0.07)	2.34E-06	Cell cycle activation of protooncogenes, Hepatocellular carcinoma
cg27143246	MYNN	myoneurin	TSS1500_shore	55892	0.145	0.096 (0.073,0.119)	0.071 (0.051,0.091)	4.55E-06	Colorectal cancer, Bladder cancer, ovarian cancer
cg08538416			IGR_shelf		0.881	−0.119 (−0.15,−0.089)	−0.083 (−0.108,−0.059)	7.35E-06
cg24866706			IGR_none		0.908	−0.08 (−0.099,−0.06)	−0.061 (−0.079,−0.043)	7.88E-06
cg14162627	ZNF322B	zinc finger protein 322 pseudogene 1	TSS1500_shore	387328	0.489	0.065 (0.047,0.083)	0.054 (0.038,0.07)	9.61E-06	Pseudogenes, duplications, MHC
cg05757007	ASCC3	activating signal cointegrator 1 complex subunit 3	Body_none	10973	0.787	−0.127 (−0.161,−0.093)	−0.085 (−0.11,−0.059)	9.74E-06	Repair of N-Alkylated nucleotides, DNA demethylation
cg17558214	PHF14	PHD finger protein 14	Body_none	9678	0.884	−0.105 (−0.133,−0.077)	−0.075 (−0.098,−0.052)	1.17E-05	Biliary tract cancer, colorectal cancer
cg26435481	SNORD114–9	small nucleolar RNA, C/D box 114−9	TSS1500_none	767585	0.866	−0.083 (−0.105,−0.061)	−0.055 (−0.072,−0.038)	1.22E-05	Acute promyelocytic leukemia
cg22838357			IGR_shelf		0.838	−0.117 (−0.148,−0.085)	−0.076 (−0.1,-0.052)	2.17E-05
cg15739904			IGR_island		0.562	−0.055 (−0.07,−0.039)	−0.051 (−0.067,−0.034)	6.23E-05
cg27447696			IGR_none		0.831	−0.069 (−0.087,−0.051)	−0.053 (−0.071,−0.035)	9.73E-05
cg00338852	ARL13B	ADP-ribosylation factor-like 13B	3′UTR_none	200894	0.893	−0.083 (−0.107,−0.059)	−0.054 (−0.072,−0.036)	1.01E-04	Ciliopathies, Joubert syndrome
cg13325919	SOX2	SRY (sex determining region Y)-box 2	3′UTR_shore	6657	0.778	−0.071 (−0.092,−0.051)	−0.06 (−0.081,−0.04)	1.09E-04	Gastric cancer, cervical squamous cell carcinoma, non-small lung cell carcinomas, colorectal cancer, non-invasive bladder cancer, head and neck squamous cell, ovarian cancer, invasive gliomas, glioblastomas, esophageal carcinomas, testis and germ cell neoplasia, hepatocellular carcinoma, anaplastic thyroid carcinoma, osteosarcomas, upper urinary tract transitional carcinomas
cg02279953	SNORD114–11	small nucleolar RNA, C/D box 114-11	TSS1500_none	767589	0.915	−0.084 (−0.109,−0.059)	−0.055 (−0.075,−0.036)	2.53E-04	No information available
cg16896134	ZNF714	zinc finger protein 714	3′UTR_none	148206	0.851	−0.085 (−0.111,−0.06)	−0.052 (−0.07,−0.033)	4.20E-04	non-syndromic cleft lip with or without cleft palate
cg17256697			IGR_shore		0.743	−0.084 (−0.113,−0.055)	−0.069 (−0.095,−0.044)	7.34E-04
cg21653586			IGR_none		0.330	0.103 (0.068,0.138)	0.099 (0.062,0.136)	8.75E-04
cg04141141	PPAP2A	phosphatidic acid phosphatase type 2A	Body_none	8611	0.866	−0.092 (−0.12,−0.064)	−0.055 (−0.076,−0.034)	8.90E-04	Prostate cancer
cg22840583	EMR4P	egf-like module containing, mucin-like, hormone receptor-like 4 pseudogene	TSS1500_none	326342	0.850	−0.096 (−0.125,−0.067)	−0.063 (−0.088,−0.037)	2.80E-03	Pseudogene, leukocyte migration in non human species
cg27103937			IGR_shelf		0.722	0.068 (0.04,0.096)	0.075 (0.044,0.106)	3.09E-03
cg19637330			IGR_island		0.577	−0.11 (−0.173,−0.046)	−0.15 (−0.216,−0.084)	7.61E-03
cg02698886			IGR_shore		0.591	0.072 (0.049,0.095)	0.052 (0.028,0.075)	1.08E-02
cg23048036	CEP97	centrosomal protein 97kDa	TSS1500_shore	79598	0.788	−0.078 (−0.105,−0.052)	−0.052 (−0.076,−0.028)	1.21E-02	Ciliogenesis, centromere migration
cg00389785			IGR_shelf		0.646	−0.071 (−0.099,−0.044)	−0.056 (−0.083,−0.029)	1.63E-02
cg21004684	IGHMBP2	immunoglobulin mu binding protein 2	Body_none	3508	0.673	−0.027 (−0.053,−0.0004)	−0.051 (−0.076,−0.026)	2.12E-02	Breast cancer
cg22244039			IGR_shore		0.780	−0.068 (−0.109,−0.026)	−0.088 (−0.131,−0.045)	2.23E-02
cg25592910	PCDH15	protocadherin-related 15	TSS200_none	65217	0.512	0.056 (0.03,0.082)	0.054 (0.027,0.08)	2.37E-02	Usher syndrome I, NK/T cell lymphomas
cg07707039	COX11	cytochrome c oxidase assembly homolog 11 (yeast)	Body_shelf	1353	0.612	0.057 (0.033,0.082)	0.051 (0.026,0.076)	2.40E-02	Breast cancer, type 2 diabetes mellitus

^aGenes were sorted by false discovery rate-FDR from the regression models adjusted for age, sex, the first principal component and the estimated white blood cell proportion.

^bThe observed deltabeta corresponds to the mean difference of high exposed vs. low exposed group to THM, confidence interval is based on a empirical Bayes moderated t-test.

^cThe expected deltabeta corresponds to the β coefficient of high exposed vs. low exposed group to THM adjusted for age, sex, the first principal component and the estimated white blood cell proportion, confidence interval is based on a empirical Bayes moderated linear regression.

^dFrom PubMed search (gene expression, protein levels or genetic polymorphisms associated with specific molecular mechanisms or diseases in humans).

Figure 2.

Heatmap of CpG sites with methylation levels associated with THM exposure (|Δβ|>0.05 and FDR<0.05). Note: DNA methylation heatmap of methylated genes passing FDR<0.05, in white blood cells DNA of persons exposed during lifetime to trihalomethanes. Each row represents a CpG site with columns representing each sample. The top dendrogram shows the results of an unsupervised hierarchical clustering of 138 samples based on 29 CpG sites, which separates those subjects flagged as highly exposed in average to lifetime THM levels >85 μg/L (marked as black in the bottom box), from those exposed to lower levels (the remainder columns). In the right box each site is marked to its corresponding region. A scatterplot of the actual lifetime THM levels is shown at the bottom box.

For pathway analyses, we included all the genes annotated by Illumina to specific CpG sites (intergenomic regions were not included, as they cannot be properly annotated to a contiguous gene). Using GSEA, 96 pathways were enriched using KeGG (P-value of the family wise error rate (FWER) < 0.001), and 128 using Reactome (Tables S2 and S3). Induced networks of the 140 CpG sites showed multiple interrelationships with SOX2 and RB1 pathways (Fig. 3) which are respectively an oncogene and a tumor suppressor that are associated with different classes of cancer, including bladder and colorectal cancer (Table 2).

Figure 3.

Network analysis of differentially methylated genes after adjustement by covariables. Expected Δβ is the methylation change expected after adjusting for age, sex, the first principal component and the estimated white blood cell proportion.

A specific evaluation of the 29 CpG sites with absolute β coefficients above 0.05 in the adjusted model was also explored. As the pathways were not enriched using this shortlist, a manual PubMed search of the related genes was performed. Eleven genes (PCBD2,¹⁴ API5,¹⁵ MYNN,¹⁶ ASCC3,¹⁷ PHF14,¹⁸ SNORD114–9,¹⁹ SOX2,^20–24 PPAP2A,²⁵ IGHMBP2,²⁶ PCDH15,²⁷ and COX11²⁸) were related to different classes of cancer (colorectal and bladder cancer, among others) and 2 to ciliopathies (ARL13B^29,30 and CEP97³¹) (see Table 2). Other diseases (type 2 diabetes mellitus, Usher syndrome, and non-syndromic cleft lip), and some cell specific pathological and physiological mechanisms (duplications, repair of N-alkylated nucleotides, leukocyte migration) were also found for some of these genes. Differentially methylated regions (Table 3) overlap several genes (CEP97, API5, SNORD114–9, SNORD114–11, MYNN, and PCDH15).

Table 3.

Differentially methylated regions (n = 30) identified using a Gaussian kernel (DMRcate) with methylation levels associated with trihalomethane exposure ((|Δβ|>0.05 and FDR<0.05) after adjusting for covariables (age, sex, first principal component and blood cells proportion)

Associated gene symbol(s)	Genomic area(s)	Coordinates (hg19)	no. probes	Min P-value	Mean P-value	Max Δβ
		chr8:125313573–125313940	3	3.06E-31	2.60E-28	−0.05
CEP97	TSS1500,Body	chr3:101442954–101443851	4	5.44E-22	8.32E-13	−0.05
API5	TSS1500,TSS200,1stExon,Body	chr11:43333145–43333782	7	1.13E-21	2.28E-19	0.05
SNORD114–9,SNORD114–10,SNORD114–11	TSS1500,TSS200	chr14:101432120–101433601	4	3.67E-15	4.04E-12	−0.06
MYNN	TSS1500,TSS200,1stExon,5′UTR	chr3:169489583–169491183	8	1.23E-13	5.29E-08	0.07
		chr6:24360304–24360603	3	9.85E-06	1.77E-05	0.05
		chr1:19110734–19110922	2	1.28E-05	1.40E-05	−0.15
EXOC3L2	5′UTR,TSS200,TSS1500	chr19:45737011–45738115	9	1.05E-04	5.98E-04	0.06
PCDH15	TSS200	chr10:56561096–56561124	2	7.46E-04	7.48E-04	0.05
DNHD1	Body	chr11:6592066–6592745	3	1.22E-03	1.66E-03	0.06
PCDHGA5,PCDHGA4,PCDHGA2,PCDHGB2,PCDHGA1,PCDHGB1,PCDHGA3	TSS1500,Body,TSS200,1stExon	chr5:140743575–140744556	6	1.45E-03	7.12E-03	0.06
C22orf27	TSS1500,TSS200,Body	chr22:31317764–31318546	10	2.91E-03	4.95E-03	−0.06
GPX6	1stExon,5′UTR,TSS200	chr6:28483537–28483691	2	3.66E-03	5.22E-03	−0.07
NXPH2	TSS1500	chr2:139538356–139539001	4	3.73E-03	1.13E-02	0.09
SMAD3	TSS1500	chr15:67356310–67356942	3	5.66E-03	7.02E-03	0.09
PRSS21	TSS1500,Body	chr16:2866901–2868001	5	6.08E-03	1.30E-02	−0.05
		chr8:599963–600488	4	6.68E-03	2.46E-02	−0.07
SLFN12	1stExon,5′UTR,TSS1500	chr17:33759484–33759986	4	9.44E-03	1.97E-02	−0.06
LMTK3	Body	chr19:49000743–49002477	6	1.18E-02	1.83E-02	−0.06
KCNMA1	Body	chr10:79110632–79111034	2	1.19E-02	1.92E-02	0.06
		chr4:25090198–25090665	5	1.36E-02	1.93E-02	−0.06
		chr2:47799165–47799268	2	1.44E-02	1.49E-02	−0.05
		chr3:133502540–133503437	6	1.49E-02	2.07E-02	0.06
KIF25	Body	chr6:168435636–168435923	3	2.48E-02	2.54E-02	−0.09
DUSP22	TSS200,1stExon,5′UTR,Body	chr6:291903–292596	5	2.89E-02	3.62E-02	−0.08
PPP4R2	TSS1500	chr3:73045556–73045686	2	2.94E-02	2.98E-02	0.07
		chr4:1512820–1513089	3	3.31E-02	3.78E-02	−0.06
C21orf56	Body	chr21:47581558–47582049	2	3.33E-02	3.81E-02	0.07
STK32C	Body	chr10:134045578–134046066	3	4.17E-02	4.66E-02	0.06
		chr8:125313573–125313940	3	3.06E-31	2.60E-28	−0.05

All sensitivity analyses provided larger top lists with slightly different variations on the order of top hits. However, the differentially methylated top hits were preserved independently of the strategy used (normalization, robust modeling, or less stringent filtering of probes). We opted to keep the shortest most conservative list.

Discussion

This is the first DNA methylation epigenome-wide study linking chronic THM exposure to changes in human DNA methylation. We found that subjects with an average lifetime exposure to THM higher than 85 µg/L showed differential methylation of 29 CpG sites (FDR < 0.05 and absolute expected Δβ > 0.05) compared to those with less exposure. Some of the genes annotated to these sites are associated with different cancers, including bladder and colorectal.

Epidemiological studies have suggested that chronic exposure to DBPs increases bladder cancer risk.^2,3 Experimental evidence has suggested that the mechanisms behind DBP carcinogenicity may partially be explained by epigenetic alterations that result from chronic cytotoxicity mediated by a mixture of oxidative metabolites release, reductive free radicals, and cell integrity disruption.^5,32 For non-genotoxic compounds, cycles of mitotic regeneration in response to cytotoxicity may produce cumulative epigenetic changes.⁸ These cumulative epigenetic changes and, particularly, global DNA hypomethylation, may lead to genomic instability and cell apoptosis.³³ In particular, global DNA hypomethylation and hypomethylation of several protooncogenes after exposure to several THM and HAA has been described in rodents.^8–11 This specific derepression of protooncogenes may lead to overcome the apoptotic pathways and generate survival of transformed neoplastic cells.³⁴ In our data, the induced networks and several individual CpG sites, annotated to specific genes, suggest a mechanism related to tumor suppressor release (RB1), or oncogene activation (SOX2), which may explain part of the THM mechanism in humans. Moreover, some of the differentially methylated sites and regions are located in genes related to both or, specifically, to either bladder cancer^16,20 or colorectal cancer^21,35 (MYNN, SOX2, RB1, and SMAD3), which supports this hypothesis. In addition, the cancer-related KeGG pathway was ranked 2nd in the GSEA analysis (colorectal cancer was ranked 22nd and bladder cancer 58th, both significant). Using Reactome, EGFR signaling in cancer was ranked 45th. Other cancer-related pathways that were enriched include p53, MAPK, and PPAR signaling. Other metabolic, immunological, and neurodevelopment pathways were also non-specifically enriched using both curated databases.

We used total THM as a DBP exposure proxy and classified subjects into 2 exposure categories. We followed a simplistic approach in order to optimize the statistical power that ignored potential differences attributable to THM composition and quantitative estimates of exposure. We found that THM exposure did not perfectly classify the subjects DNA methylation levels (Fig. S1). Limited sample size, exposure misclassification, and individual unmeasured characteristics may explain these classification differences. Given that exposure was assigned based on residence, unaccounted socioeconomic factors (other than educational level) might explain some of the observed differences.³⁶ DBPs and, particularly, THM as proxy of the whole mixture may behave differentially according to the specific route of exposure. Given the different molecular weights and polarity of the DBPs, potential exposure routes include not only ingestion but also inhalation and/or dermal absorption during showering and water related activities (cleaning, swimming, flushing the toilet, etc.).^37–41 These routes, plus the subject water handling and consumption behaviors, may produce differential internal doses and partially explain why subjects apparently exposed to similar THM levels had a wide range of changes in DNA methylation levels in our sample. Exposure misclassification due to unaccounted non-residential exposure (e.g., workplace) is expected to be low, only affecting the ingested THM fraction. On the other hand, individual DNA methyltransferases activity or specific protective genetic polymorphisms may enhance or reduce the epigenetic changes and reduce the potential for new DNA methylation changes. Unfortunately, we do not have data about these specific variables.^42,43 Finally, as these processes may be perpetuated through inflammatory pathways, the role of chronic inflammation and its modification through anti-inflammatory medications (either intended or pleiotropic) is another potential target of future research in the area.⁴⁴

We used microarrays as a cost-effective approach to interrogate epigenome-wide DNA methylation changes associated with THM exposure. However, there are several analytical gaps in the Infinium HumanMethylation450 BeadChip analyses, as there is no consensus for preprocessing and downstream analysis of the data.⁴⁵ We decided to be as conservative as possible and compared different approaches to observe internal reproducibility of results. The results were concordant using different approaches and we are confident that we tried to maximize our results while reducing potential analytical bias. This is the main strength of our study. In addition, the design was intended to maximize the power even under the small sample size constrain. In this first exploratory study with limited sample size and absence of experimental validation, we generate some hypotheses that require replication in new studies with larger sample size.

The use of whole blood samples instead of the target organ is a limitation in our study. As DNA methylation is time and organ specific we cannot assure that the differences found can be replicated on the target organs. Second, the cross sectional blood sampling does not allow us to observe the evolution of changes in time, as the samples were collected from a case-control study. However, there is a limited amount of longitudinal studies of cancer and, to our knowledge, none has evaluated THM exposure, ours being the first study to provide some potential hypotheses about these mechanisms in humans. Third, the changes found in methylation levels are relatively small (utmost 15%), which increases the probability of positive results by chance, even after adjusting for multiple comparison. Fourth, differentially methylated regions methods and pathway analyses for DNA methylation data are active and rapidly evolving research areas and results may differ by method applied. Our differentially methylated region results seem to be driven by the most differentially methylated probes, even if other probe differences were modest (absolute Δβ < 0.01) or non-significant (FDR > 0.05) in the differential methylation site analysis. Pathway analyses try to summarize common pathways assuming large gene-specific interactions included on curated databases but do not account for unknown intergenomic regions without annotated genes. In addition, without gene expression data, DNA methylation pathway analyses only may suggest but cannot provide exact information about the gene network underlying the observed methylation changes. Finally, we have not performed a laboratory validation using a different technique. This and the replication of the results are future steps that should be performed for confirmation. Thus, we call for a cautious interpretation of the results, given that some of them may be the product of chance.

In summary, our results suggest that long-term THM exposure affect DNA methylation in genes related to tumors, including bladder and colorectal cancer. These findings, if confirmed or validated in other populations, may contribute to understanding the molecular mechanisms or THM pathogenesis.

Materials and Methods

Study population

This study is part of a population-based case-control study of colorectal cancer, which in turn is part of a larger multicase-control study (MCC-Sp) conducted in Spain between 2007 and 2012.⁴⁶ A subset of population-based controls who provided a blood sample (81% among the recruited) were selected using a stratified random selection based on residence in Barcelona metropolitan area (4 municipalities), age at recruitment (>60 years), and availability of lifetime estimates of THM levels (see below). Study subjects were frequency matched by age group (60–70 vs. 70–80 y) and THM levels (<85 µg/L vs. ≥85 µg/L, the median). A total of 70 subjects (males and females) were selected in each exposure group (50% 60–70 y old, 50% 70–80 y old). The MCC-Sp project and the present study have been approved by the Investigation Review Boards of the different participant hospitals in Barcelona. All the participants had signed the informed consent of the main MCC study and agreed to the molecular analyses.

Questionnaires

Trained interviewers administered a comprehensive computer-assisted questionnaire to the study subjects in the primary care centers. The interview took around 90 min and included personal, socio-demographic, lifestyle, family history, and medical history variables. A validated self-administered food frequency questionnaire was also filled in. The questionnaire is available at the study website: www.mccspain.org.

Lifetime trihalomethane exposure

Exposure estimates were based on levels in the residence of the subject, as a proxy of exposure through all the routes (ingestion, inhalation, dermal contact) in different situations (drinking water and water-based fluids, showering, bathing, dish-washing, etc.). Residential history, including complete address of all residences where study subjects had lived at least during one year since age 18 was requested. Trihalomethane levels in the study municipalities were estimated back to 1940 using available measurements. Historical data on water source and the available THM measurements were used to estimate annual average THM levels for years when measurements were absent. Available THM measurements were averaged and imputed to the past when water source and treatment were unchanged. Proportion of surface water was used as a weight to this average in the event of changes in water source. Before chlorination started, THM levels were assumed to be zero. Residential histories and estimated THM levels were merged by zip code and year to estimate average THM levels from age 18 to the time of interview in study subjects, according to previously published methodology.⁴⁷ Subjects included in this analysis had THM exposure data covering at least 70% of years from age 18 until the time of interview, with an average (range) of available data of 95.2% (74.1–100) of the years.

DNA extraction and genome-wide DNA methylation array

DNA was extracted from EDTA-treated blood using the Chemagic DNA Blood5k Kit (Perkin Elmer) following manufacturer's protocol at the Spanish National Genotyping Center (CEGEN-Barcelona) and using a manual DNA extraction kit (PROMEGA) at the Bellvitge Biomedical Research Institute (IDIBELL). None of the samples presented visual signs of DNA degradation (smear bands or bands below 10,000 bp) as observed after running 100 ng of DNA on a 1.3% agarose gel. The isolated genomic DNA was stored at −80°C until use. For every sample, 1 µg of DNA was bisulfite-converted using the Zymo EZ DNA methylation kit (ZYMO Research Corporation, Irvine, CA) according to manufacturer instructions. Converted DNA was eluted in 22 µL elution buffer. DNA methylation level was assessed using the Infinium HumanMethylation450 BeadChip (Illumina Inc., San Diego, CA)⁴⁸ at the CEGEN facility at the Barcelona Biomedical Research Park (PRBB) according to manufacturer's instructions.⁴⁹ Briefly, 4 µL of bisulfite-converted DNA was isothermally amplified overnight (20–24 h) and fragmented enzymatically. DNA was precipitated using isopropanol and collected after centrifugation at 4°C. Precipitated DNA was resuspended in hybridization buffer and dispensed onto the Infinium HumanMethylation450 BeadChips using a robot (Tecan Group Ltd., Männedorf, Switzerland). Samples were distributed in random blocks according to inclusion criteria to reduce potential chip associated batch effects. Two samples were run by duplicate and one in triplicate, plus a Jurkat DNA control. In total, 12 chips were run (12 samples/chip). Every Infinium slide has 8 chips, thus 2 slides were run simultaneously in a single laboratory batch. Hybridization was performed at 48°C overnight (16–20 h) using an Illumina hybridization oven. Amplified and fragmented DNA samples annealed to locus-specific 50mers (covalently linked to some of the bead types) during hybridization. After hybridization, free DNA was washed away and the BeadChips were processed through a single nucleotide extension following immunohistochemistry staining (ddNTP) in capillary flow through chambers (Tecan GenePaint automated slide processor) using the Freedom Evo robot. Fluorescence signal was captured as an image on an Illumina iScan system using Illumina iScan software. After background subtraction, 2 raw intensity data (idat) files were produced (one per channel) using Illumina GenomeStudio software.

Illumina 450K data preprocessing considerations

The 450K array combines 2 assays in one platform: the Infinium I (type I probes) and Infinium II (type II probes).⁴⁸ The probes used to assess methylation levels are technically different. In consequence, preprocessing of 450K data should also control for potential differences between assays. In addition, 2 other technical biases are possible: probes that cross-hybridize, and probes mapping in polymorphic residues.^50,51 Probes that cross-hybridize account for about 8.6% of the 450K array. Methylation levels measured in these probes likely reflect a combination of levels of methylation at the various locations to which these probes hybridize. On the other hand, polymorphic target probes (4.3% of 450K probes) are probes with polymorphisms at the target C or G at the extension point. Since the 450K platform quantitatively genotypes level of C/T SNPs after bisulfite conversion, these probes may assess a difference in genotype rather than a true difference in methylation levels.

Bioinformatics and statistical analysis

The array data were preprocessed and then processed using R version 3.1.2⁵² and different Bioconductor packages.⁵³ The pipeline used (adapted from RnBeads)⁵⁴ included the following steps: (1) loading raw intensity data (idat);⁵⁵ (2) prefiltering (removal of SNP-enriched probes, greedycut algorithm removal of unreliable measurements, and removal of predefined blacklisted probes—all crossreactive probes and polymorphisms as reported by Chen et al.⁵¹); (3) normalization (methylumi-noob for background correction⁵⁶ and dasen normalization⁵⁷); (4) quality control (detection and exclusion of technical failures during bisulfite conversion, hybridization, extension and staining, detection of potential sample mixups using the default 65 SNPs included in the array); (5) postfiltering (removal of non CpG probes, removal of sex chromosomes); (6) negative control batch effect correction⁵⁸ and surrogate variable analysis-sva multidimensional reduction adjusting for residual confounding; (7) visualization of the general unadjusted methylation profile. Once data were clean, the downstream data analysis was performed using β values. The β value index was calculated using both intensities of DNA methylation fraction at a specific CpG site: β = M/(M + U + α), where M represents methylated and U unmethylated signal intensities and α is an arbitrary offset (100) to stabilize β values whether the intensities are low. β values are bounded between 0 and 1.⁵⁹

Epigenome wide DNA methylation association analyses

In total, 144 samples were processed. From the original 140 study subjects, one was processed in duplicate and one in triplicate, and a Jurkat sample was added as a technical control. The duplicates correlation and triplicates correlation was between 0.995 and 0.998. One of the samples was selected at random to keep for further analyses. The other duplicates, the Jurkat sample, and 2 mislabeled (cancer cases) were discarded, leaving 138 samples for analyses. We used an empirical Bayes moderated t-test and/or empirical Bayes moderated linear regression models using limma⁵⁸ to test the associations between THM levels and DNA methylation (measured as β values). We selected some potential adjustment covariables: age, sex, municipality, highest education reached (elementary or less, high school, and university or more), and tobacco smoking (classified as never, former, and current smokers). Using a surrogate variable approach, we detected a potential residual uncorrected batch effect and we included a surrogate variable in the adjusted models.⁵⁸ Finally, we used the Houseman algorithm (included in minfi) to estimate the proportions of white blood cells in our samples. In brief, this algorithm uses ∼473 most informative CpG probes to estimate the proportions of T- (CD8, CD4), NK-, and B-lymphocytes, monocytes, and total granulocytes. Adjustment by cell mixtures control population cell stratification in the model due to specific cells populations DNA methylation epigenetic landmarks.^55,60,61 We used the selectModel option from limma to reduce multiple comparisons. This command compares nested models using the Akaike Information Criteria-AIC. Genomic inflation factor (λ)⁶² was calculated for adjusted and unadjusted models. Statistical significance after multiple testing comparison was established using the Benjamini-Hochberg false discovery rate (FDR).⁶³ Statistically significant differential methylation between groups was defined as an absolute Δβ ≥ 0.05. To define cut-offs, we defined the absolute observed Δβ as the difference between the average methylation of the comparison groups and the absolute expected Δβ as the β coefficient of the adjusted models. Finally, to capture differentially methylated regions we used DMRcate,⁶⁴ which uses the limma empirical Bayes t-moderated statistic calculated per probe and a Gaussian kernel smooth using a 1000 window. The estimate is smoothed based on the varying density of CpG sites given the irregular spacing of the data on the genome interrogated by the Infinium HumanMethylation450 BeadChip.

Pathway analyses

Pathways associated with differentially methylated sites were interrogated using ConsensusPathDB and GSEA^65,66 Three different approaches were used: (1) Overrepresentation analyses (number of overrepresented selected genes per set); (2) Gene set enrichment analyses (GSEA) using the KeGG and Reactome curated databases (full list of genes pre-ranked using the t-statistic of the adjusted model); and (3) Induced networks (induced relationships among genes-proteins filling the gaps of unmeasured products not represented/included on the differentially methylated list).^66,67 For overrepresentation analyses, the background gene pool was limited to those genes annotated by Illumina in the array. A manual PubMed search of the top CpG sites was performed to annotate the genes associated with specific diseases/conditions.

Sensitivity analyses

We tested our preprocessing approach and modeling compared to other potential approaches to test reproducibility of the top hits. The following analyses were performed: β values normalized using BMIQ,⁶⁸ a longer probe list with a less conservative filtering excluding only those probes with MAF > 5% in European population, M-values (logit₂ transformation of β values)⁵⁹ used as the outcome, and robust regression instead of linear regression.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Acknowledgments

Authors thank our fieldworkers Lourdes Arjona and Estela Carrasco. We also thank the CEGEN and Magda Monfort for processing the blood samples and performing the microarray laboratory analyses. Thanks to Nadia Vilahur and Eva Morales for their initial feedback in some issues related to the microarray analyses. We thank the Parc de Salut Mar Biobank (MARBiobanc) and Núria Somoza for their support to the project.

Funding

This study was funded by the Spanish Health Ministry (Fondo de Investigaciones Sanitarias–FIS, Instituto de Salud Carlos III, Spain number FIS PI11–226). Lucas A Salas received a Colciencias PhD Scholarship, Colombia (Grant: 529/2011). Parc de Salut Mar Biobank (MARBiobanc) provided biobanking facilities for to the project. MARBiobanc is supported by Instituto de Salud Carlos III FEDER (RD09/0076/00036).

Supplemental Material

Supplemental data for this article can be accessed on the publisher's website.

Ethics Committee Approval

The study was approved by the Ethics Committee of the Research Center (IMIM-IMAS).