The impact of methylation quantitative trait loci (mQTLs) on active smoking-related DNA methylation changes

Abstract

Background

Methylation quantitative trait loci (mQTLs) are the genetic variants that may affect the DNA methylation patterns of CpG sites. However, their roles in influencing the disturbances of smoking-related epigenetic changes have not been well established. This study was conducted to address whether mQTLs exist in the vicinity of smoking-related CpG sites (± 50 kb) and to examine their associations with smoking exposure and all-cause mortality in older adults.

Results

We obtained DNA methylation profiles in whole blood samples by Illumina Infinium Human Methylation 450 BeadChip array of two independent subsamples of the ESTHER study (discovery set, n = 581; validation set, n = 368) and their corresponding genotyping data using the Illumina Infinium OncoArray BeadChip. After correction for multiple testing (FDR), we successfully identified that 70 out of 151 previously reported smoking-related CpG sites were significantly associated with 192 SNPs within the 50 kb search window of each locus. The 192 mQTLs significantly influenced the active smoking-related DNA methylation changes, with percentage changes ranging from 0.01 to 18.96%, especially for the weakly/moderately smoking-related CpG sites. However, these identified mQTLs were not directly associated with active smoking exposure or all-cause mortality.

Conclusions

Our findings clearly demonstrated that if not dealt with properly, the mQTLs might impair the power of epigenetic-based models of smoking exposure to a certain extent. In addition, such genetic variants could be the key factor to distinguish between the heritable and smoking-induced impact on epigenome disparities. These mQTLs are of special importance when DNA methylation markers measured by Illumina Infinium assay are used for any comparative population studies related to smoking-related cancers and chronic diseases.

Electronic supplementary material

The online version of this article (doi:10.1186/s13148-017-0387-6) contains supplementary material, which is available to authorized users.

Background

Active smoking has been recognized as a critical lifestyle factor for cardiovascular, respiratory, and neoplastic diseases and contributes to the leading causes of preventable morbidity and mortality [1, 2]. DNA methylation, one of the main forms of epigenetic modification, is involved in the pathways of smoking and smoking-induced diseases [3, 4]. Previous epigenome-wide association studies (EWASs) based on whole blood samples have successfully discovered an increasing number of tobacco smoking-related CpG sites in various genes, such as AHRR and F2RL3 [5–7]. These DNA methylation patterns have been shown to be useful as quantitative biomarkers to reflect both current and lifetime smoking exposure and to enhance the prediction of smoking-related risks [8–11].

DNA methylation of particular genomic loci might be influenced by neighboring genetic sequence variants [12]. The single nucleotide polymorphisms (SNPs) that are associated with methylation levels of CpG sites are known as methylation quantitative trait loci (mQTLs) [13]. This genetic effect has been determined across different tissues [13–16] and has been highlighted in several diseases, including neurological disorders, arthritis, and cancer [17–21]. Recently, the mQTLs have been further reported to play a modifying role in the associations between DNA methylation levels at specific CpG sites and environmental exposures. For instance, Zhang et al. identified 238 mQTLs that were associated with 65 alcohol dependence-related CpG sites in African Americans and 305 mQTLs for 44 unique CpG sites in European Americans [22]. In 2016, Gonseth et al. found out that three of the strongest maternal smoking-related CpG sites in newborns were significantly associated with SNPs located in the vicinity of each gene [23]. Thus, these hereditary traits provide a possible mechanism by which methylation patterns could be different under environmental exposures, if the distribution of risk alleles differs between the exposed and the unexposed. In addition, the linkages of epigenetic signatures to genotypes might also further provide more mechanistic evidence on the genetic and environmental risk factors for various forms of diseases [24].

However, such genetic influences have not been well addressed or even overlooked by previous EWASs of active smoking exposure; to our knowledge, no study has so far investigated their contributions to the methylation intensities of active smoking-related CpG sites and smoking-related health outcomes in the general population. Therefore, we conducted a comprehensive analysis in a large population-based study of older adults in Germany with the aim of exploring the hitherto unknown association between active smoking-related DNA methylation and individual genetic variations. In particular, we aimed to identify the mQTLs within ± 50 kb from each of 151 previously reported active smoking-related CpG sites in whole blood samples [25] and to assess their relationships with active smoking exposure and all-cause mortality.

Results

Participant characteristics

Characteristics of the study population in the discovery and the validation panel with respect to smoking behaviors and lifestyle factors are summarized in Table 1. The average age of the participants of both subsets at the baseline was about 61 years. About half of the participants in each subset were ever smokers (current or former smokers), and around 18% still smoked at the time of recruitment. Female participants included a larger proportion of never smokers than males (discovery set, 67.9 vs. 28.6%; validation set, 63.3 vs. 21.4%). Average cumulative smoking exposure (pack-years) of current smokers was considerably higher than that of former smokers in both subsets (discovery set, 34.6 vs. 22.0; validation set, 33.1 vs. 19.4). Average time after smoking cessation (years) of former smokers in both subsets was also similar, approximately 17 years. The majority of participants in both subsets of the study population were overweight or obese, reported no or only low physical activity, and no or low amounts of alcohol drinking. During a median follow-up time of about 12 years (discovery set, 12.6 years; validation set, 12.2 years), 94 participants died in the discovery set (CVD = 30, cancer = 46, other diseases = 18) and 49 died in the validation set (CVD = 17, cancer = 21, other diseases = 11).

Table 1.

Study population characteristics in discovery and validation panels (mean values (SD) for continuous variables and n (%) for categorical variables)

Characteristics	Discovery panel	Validation panel	p value
N	581	368
Age (years)	61.0 (6.3)	61.1 (6.4)	0.809
Sex (male)	241 (41.5%)	117 (31.8%)	< 0.001
Smoking status			0.864
Current smoker	108 (18.6%)	65 (17.7%)
Former smoker	173 (29.8%)	119 (32.3%)
Never smoker	300 (51.6%)	184 (50.0%)
Pack-years of smokinga
Current smokers	34.6 (18.2)	33.1 (18.2)	0.250
Former smokers	22.0 (17.5)	19.4 (15.5)	0.033
Smoking cessation time (years)b	16.5 (11.3)	17.2 (10.2)	0.742
Body mass indexc			0.248
Underweight or normal weight (< 25.0)	143 (24.7%)	116 (31.5%)
Overweight (25.0–< 30.0)	290 (50.2%)	151 (41.0%)
vObese (≥ 30.0)	145 (25.1%)	101 (27.5%)
Alcohol consumptiond			0.509
Abstainer	194 (36.3%)	128 (38.0%)
Low	301 (56.4%)	188 (55.8%)
Intermediate	30 (5.6%)	17 (5.0%)
High	9 (1.7%)	4 (1.2%)
Physical activitye			0.058
Inactive	109 (18.8%)	82 (22.3%)
Low	245 (42.2%)	176 (47.8%)
Medium or high	227 (39.0%)	110 (29.9%)
Prevalence of CVD at baselinef			0.621
Prevalent	86 (14.8%)	58 (15.8%)
Prevalence of diabetes at baselineg			0.617
Prevalent	86 (14.9%)	60 (16.6%)
Prevalence of cancer at baseline
Prevalent	33 (5.7%)	22 (6.0%)	0.744

^aFor subgroups of former and current smokers; data missing for 38 and 24 participants, respectively, in discovery and validation panels; a pack-year was defined as having smoked 20 cigarettes per day for 1 year

^bFormer smokers only, data missing for 5 and 2 participants, respectively, in discovery and validation panels; cessation time equals age at recruitment minus age at cessation

^cData missing for 3 participants in discovery panel

^dData missing for 47 and 31 participants, respectively, in discovery and validation panels. Categories defined as follows: abstainer, low [women, 0–< 20 g/d; men, 0–< 40 g/d], intermediate [20–< 40 g/d and 40–< 60 g/d, respectively], high [≥ 40 g/d and ≥ 60 g/d, respectively]

^eCategories defined as follows: inactive [< 1 h of physical activity/week], medium or high [≥ 2 h of vigorous or ≥ 2 h of light physical activity/week], low (other)

^fCVD cardiovascular disease. Data missing for 1 participant in discovery panel

^gData missing for 5 and 7 participants, respectively, in discovery and validation panels

Identification of mQTLs for smoking-related CpG sites

For the 1396 SNP-CpG pairs consisting of 150 smoking-related CpG sites and 909 corresponding SNPs (Fig. 1), 380 pairs were significant at a FDR < 0.05 in the discovery panel even after controlling for covariates (Additional file 1: Table S2; Additional file 2: Figure S1). These 380 pairs were then replicated in the validation panel by applying the fully adjusted mixed linear regression model. A subset of 246 pairs formed of 70 CpG sites and 192 SNPs reached the statistical significance level after FDR correction (FDR < 0.05; Table 2, Additional file 1: Table S3, Additional file 1: Table S4; Fig. 2). Eventually, 192 SNPs were designated as the mQTLs of 70 CpG sites. The pair cg23576855/rs75509302 showed the strongest inter-relationship (FDR-corrected p value = 8.86 e − 103). Among the 70 CpG sites, five were highly smoking-related loci (reported ≥ 6 times; Table 3), 14 were moderately smoking-related (4 or 5 times), and 51 were weakly smoking-related (2 or 3 times). These CpG sites with mQTLs were mainly located in the gene body (37/70), ten were located in transcription start sites (TSS1500) and 23 were in untranslated regions (UTR) or intergenetic regions (Additional file 1: Table S3). The largest number of mQTLs (n = 8) was found for locus cg06126421 within 6p21.33 (Table 3; Fig. 3). The coefficients of mQTLs ranged from −0.54 to 0.15.

Fig. 1.

Flowchart of selection of SNP-CpG pairs

Table 2.

List of 246 significant SNP-CpG pairs (chromosomal and CpG sites positions were based on GRCh37/hg19)

Chromosome	Gene	CpG site	Position	Number of SNP candidates	Number of mQTLs
1	AVPR1B	cg09069072	15,482,754	13	5
	GFI1	cg09662411	92,946,132	10	2
		cg09935388	92,947,588	10	2
		cg10399789	92,945,668	9	6
		cg12876356	92,946,825	10	2
		cg18146737	92,946,701	10	2
		cg18316974	92,947,035	10	2
	GNG12	cg25189904	68,299,493	5	2
	NOS1AP	cg11231349	162,050,657	4	2
	TMEM51	cg21913886	15,485,346	14	9
	unassigned	cg03547355	227,003,061	8	2
	unassigned	cg12547807	9,473,751	9	1
	unassigned	cg21393163	12,217,630	8	2
	unassigned	cg26764244	68,299,511	3	3
2	2q37.1	cg05951221	233,284,402	5	1
		cg03329539	233,283,329	5	2
	ALPP	cg23667432	233,244,439	5	2
	NFE2L2	cg26271591	178,125,956	8	6
	SNED1	cg26718213	241,976,081	9	4
	unassigned	cg27241845	233,250,371	5	2
3	GPX1	cg18642234	49,394,623	10	5
5	AHRR	cg03604011	400,201	15	5
		cg03991871	368,448	9	1
		cg11902777	368,843	9	4
		cg12806681	368,395	9	2
		cg14817490	392,920	15	4
		cg17287155	393,347	15	1
		cg23576855	373,300	9	7
		cg23916896	368,805	9	4
6	6p21.33	cg06126421	30,720,081	16	8
	CDKN1A	cg15474579	36,645,813	22	8
	IER3	cg15342087	30,720,210	16	2
		cg24859433	30,720,204	16	3
	TIAM2	cg00931843	155,442,993	6	1
	VARS	cg17619755	31,760,629	16	8
	unassigned	cg14753356	30,720,109	8	2
7	C7orf40	cg03440944	45,023,330	5	1
	CNTNAP2	cg11207515	146,904,206	14	7
		cg25949550	145,814,306	11	8
	GNA12	cg19717773	2,847,554	22	22
	HOXA7	cg08396193	27,193,709	7	1
	LRRN3	cg11556164	110,738,316	5	5
	MYO1G	cg12803068	45,002,919	10	1
		cg22132788	45,002,487	10	1
	TNRC18	cg09022230	5,457,226	24	3
8	ZC3H3	cg26361535	144,576,604	8	1
	unassigned	cg19589396	103,937,374	15	2
9	unassigned	cg01692968	108,005,349	2	2
10	ARID5B	cg25953130	63,753,550	6	3
11	KCNQ1	cg07123182	2,722,391	13	2
		cg26963277	2,722,408	13	3
		cg01744331	2,722,358	13	3
	KCNQ1OT1	cg16556677	2,722,402	13	2
	PRSS23	cg11660018	86,510,915	9	2
		cg23771366	86,510,999	9	2
	unassigned	cg16611234	58,870,075	10	10
14	C14orf43	cg01731783	74,211,789	6	1
		cg10919522	74,227,441	5	1
15	ANPEP	cg23161492	90,357,203	19	6
	SEMA7A	cg00310412	74,724,919	13	9
16	ITGAL	cg09099830	30,485,486	3	2
	XYLT1	cg16794579	17,562,419	3	1
17	LOC100130933	cg07251887	73,641,810	6	2
	STXBP4	cg07465627	53,167,407	8	4
19	CPAMD8	cg15159987	17,003,890	15	4
	CRTC1	cg23973524	18,873,223	12	1
	F2RL3	cg03636183	17,000,586	17	1
	PPP1R15A	cg03707168	49,379,127	10	4
21	ETS2	cg23110422	40,182,073	6	3
22	NCF4	cg02532700	37,257,404	8	2
Total				590	246

Fig. 2.

Manhattan plot of the results in validation panel. Red line, FDR-corrected p value = 0.05

Table 3.

Five frequently reported (≥ 6) CpG sites and corresponding mQTLs

CpG site	Frequencya	Gene	Chr	SNP	SNP positionb	Minor allele	Distance (bp)c	FDRd	MAFe
cg03636183	12	F2RL3	19	rs2227357	17,003,553	A	2967	0.048	0.125
cg05951221	8	2q37.1	2	rs790051	30,718,035	A	− 1866	6.2 e − 4	0.226
cg06126421	7	6p21.33	6	rs2535324	30,727,983	C	− 2046	1.8 e − 9	0.3
				rs3095339	30,728,290	G	7902	2.6 e − 4	0.252
				rs3131036	30,728,360	A	8209	2.6 e − 4	0.253
				rs3094122	30,737,552	G	8279	3.2 e − 3	0.206
				rs13217914	30,739,657	A	17,471	2.4 e − 21	0.157
				rs6911571	30,753,639	T	19,576	0.007	0.16
				rs4713361	30,756,066	A	33,558	1.3 e − 21	0.159
				rs13201769	30,718,035	A	35,985	6.9 e − 7	0.326
cg03329539	6	2q37.1	2	rs790051	233,282,536	A	− 793	0.031	0.226
				rs34547337	233,300,755	T	17,426	1.5 e − 5	0.314
cg14817490	6	AHRR	5	rs75509302	365,653	C	− 27,267	0.002	0.144
				rs11746079	410,980	C	18,060	1.5 e − 3	0.154
				rs72717419	431,996	T	39,076	0.021	0.207
				rs2672725	434,981	G	42,061	0.042	0.117

^aThe reported times of CpG in previous studies (based on systematic review [25])

^bPositions of CpG sites and SNPs were based on GRCh37/hg19

^cThe distance between SNP and CpG (SNP position–CpG position)

^dThe FDR-corrected p values of SNPs in fully adjusted mixed linear regression models, which controlled for age (years), sex, smoking status, random batch effect of methylation measurement, leukocyte distribution (Houseman algorithm), alcohol consumption (abstainer/low/intermediate/high), body mass index (BMI, underweight or normal weight/overweight/obese), physical activity (inactive/low/medium or high), prevalence of cardiovascular diseases (yes/no), prevalence of diabetes (yes/no), and prevalence of cancer (yes/no)

eMAF minor allele frequency

Fig. 3.

Locations of cg06126421 and eight mQTLs (carrier/non-carrier) (a) and distributions of methylation levels based on smoking status (b) in validation panel. Red line, FDR-corrected p value = 0.05; red dot, mQTLs; yellow triangle, cg06126421; blue bar, non-carriers of minor allele; red bar, carriers of minor allele

The 192 mQTLs were mainly mapped on chromosomes 1 (16%), 6 (10%), and 7 (25%). Three SNPs were the most frequently identified mQTLs, rs75509392 (MAF = 0.144) for eight CpG sites within AHRR, rs79050605 (MAF = 0.202), and rs34835481 (MAF = 0.210) for five and six CpG sites located in GFI1, respectively (Table 4). We assessed the effects of the mQTLs on the DNA methylation changes as the absolute values of the coefficient changes of smoking status between the models without and with adjusting for corresponding mQTLs (carrier/non-carrier) among the 246 SNP-CpG pairs (Additional file 1: Table S5). Part of the mQTLs had opposite effects on different CpG sites (Additional file 1: Table S5, Additional file 3: Figure S2). For example, the minor allele of the SNP rs75509302 attenuated the association of smoking exposure with the methylation of cg11902777 by 5.2% (Additional file 1: Table S5, Additional file 3: Figure S2). In contrast, this variant strengthened the demethylation of cg17287155 in response to different smoking behaviors by 2.44%.

Table 4.

Three most frequently identified mQTLs and corresponding CpG sites

SNP	Chr	SNP positiona	Minor allele	MAFb	CpG	Distance (bp)c	FDRd
rs75509302	5	365,653	C	0.144	cg23576855	− 7647	3.4 e − 100
					cg11902777	− 3190	1.4 e − 7
					cg17287155	− 27,694	7.8 e − 5
					cg03991871	− 2795	1.0 e − 4
					cg12806681	− 2742	1.2 e − 4
					cg23916896	− 3152	9.5 e − 4
					cg03604011	− 34,548	1.1 e − 3
					cg14817490	− 27,267	2.1 e − 3
rs34835481	1	92,991,624	T	0.210	cg10399789	45,956	2.2 e − 5
					cg12876356	44,799	1.3 e − 3
					cg09662411	45,492	1.9 e − 3
					cg18146737	44,923	2.0 e − 3
					cg18316974	44,589	3.0 e − 3
					cg09935388	44,036	0.016
rs79050605	1	92,925,962	G	0.202	cg12876356	− 20,863	3.7 e − 4
					cg18146737	− 20,739	1.1 e − 3
					cg18316974	− 21,073	1.7 e − 3
					cg09662411	− 20,170	1.9 e − 3
					cg09935388	− 21,626	2.2 e − 3

^aSNPs positions were based on GRCh37/hg19

^bMAF minor allele frequency

^cThe distance between SNP and CpG (SNP position–CpG position)

^dThe FDR-corrected p values of SNPs in fully adjusted mixed linear regression models, which controlled for age (years), sex, smoking status, random batch effect of methylation measurement, leukocyte distribution (Houseman algorithm), alcohol consumption (abstainer/low/intermediate/high), body mass index (BMI, underweight or normal weight/overweight/obese), physical activity (inactive/low/medium or high), prevalence of cardiovascular diseases (yes/no), prevalence of diabetes (yes/no), and prevalence of cancer (yes/no)

Genetic contributions of mQTLs to the DNA methylation changes

As shown in Additional file 1: Table S5, the associations between smoking exposure and DNA methylation were changed by between 0.01 and 18.96% by the mQTLs and were categorized by the distances between genetic variants and CpG sites and the reported frequencies of CpG sites. We observed that the closest SNPs (distance < 10 kb) had a slightly lower impact on DNA methylation levels than the mQTLs located ≥ 10 kb (Fig. 4a). Compared with the highly smoking-related CpG sites, the mQTLs affect the methylation levels of weakly smoking-related CpG sites the most, and the changes of moderately smoking-related CpG sites stayed in the intermediate position (Fig. 4b, F = 4.91, p value = 0.008). Additionally, potential gene-environment interactions of the 192 mQTLs with smoking behaviors (current/never smoking) were assessed. Only for rs75509302, a significant interaction with smoking was observed regarding the methylation levels of cg23576855 (Tables 5, Additional file 1: Table S6).

Fig. 4.

Percentage changes contributed by mQTLs to smoking-related DNA methylation changes based on SNP-CpG distance (a) and reported frequencies of CpG sites (b)

Table 5.

Impact of rs75509302-smoking interaction on the methylation level of cg23576855^a

Gene	CpG site	SNP	SNP-smoking interactionb				Smoking statusd
			Genotypec	Coefficient	SE	p value	Coefficient	SE	p value
AHRR	cg23576855	rs75509302	TT	Ref			−0.182	7.1 e − 3	3.8 e − 95
			CT	0.128	0.013	1.1 e − 20
			CC	0.268	0.033	2.5 e − 15

^aModel is fully adjusted for age, sex, BMI, smoking status (current and never smoking only), alcohol consumption, physical activity, prevalence of CVD, diabetes and cancer at baseline. The methylation levels of CpG sites were responses, the SNPs and SNP-smoking interactions were predictors;

^bThe never smoking * genotype groups and current smoking * TT group were used as references;

^cThe group of interaction between current smoking and listed genotype;

^dNever smoking was used as reference

Associations of mQTLs with active smoking exposure and all-cause mortality

Finally, we tested the associations of the 192 mQTLs (carrier/non-carrier) with different measurements of smoking exposure [ever smoking (current and former smoking) vs. never smoking, current smoking vs. never smoking, current smoking vs. former smoking, cumulative smoking (pack-years), durations of smoking (years), and the age of smoking initiation]. None of them was significantly associated with any smoking indicators after the correction of FDR (Additional file 1: Table S7). Similarly, the 192 SNPs were not significantly associated with all-cause mortality (death from CVD, cancer, and other chronic diseases) based on the results of the COX model (Additional file 1: Table S8).

Discussion

We conducted the first association study of 150 active smoking-related CpG sites and their corresponding SNPs located in the ± 50 kb region utilizing the genomic and epigenomic data of 949 participants from the ESTHER study. We found the DNA methylation levels of 70 CpG sites to be influenced by 192 proximal SNPs. These 192 mQTLs modified the DNA methylation changes in response to active smoking exposure, especially for the weakly/moderately active smoking-related CpG sites, but we did not observe any direct associations with active smoking exposure or all-cause mortality.

The mQTLs are presented in the vicinity of active smoking-related CpG sites. Locus cg06126421 (6p21.33) is one of the pronounced smoking-related CpG sites with eight mQTLs. Four of the SNPs could impair the hypomethylation of cg06126421, while others could accelerate this process. All of these mQTLs were located in a genomic region relating to inflammation and/or immune-related (malignant) diseases, such as allergies [26], multiple myeloma [27], diffuse large B cell lymphoma [28], and lung cancer [29]. We also identified a mQTL rs2227357 which could slightly promote the demethylation of cg03636183 (F2RL3), but no mQTLs were discovered near other well-established smoking-related loci, like cg05575921 (AHRR) or cg19859270 (GPR15). However, eight CpG sites within the AHRR region were found to be modified by mQTLs. Among them, cg23576855 manifested the strongest connection with rs75509302 and was also the only hit which could be affected by the interaction between smoking status and corresponding mQTL. This phenomenon might be highly contributed by the genetic feature of this locus, which is also a CG → CA SNP annotated as rs6869832 [5, 30], and shares a very high LD with rs75509302 in this study (R² = 0.94). Nevertheless, to our knowledge, the biological functions for most of the identified mQTLs in our study are not fully understood yet and need to be explored in further research.

The mQTLs may help to distinguish between the genetic and environmental effects on epigenome disparities. Researchers usually observed several outliers out of the predictive range of epigenetic signatures in EWASs [6, 7, 9, 31]. One of the most plausible explanations is measurement bias that may result from recall bias or intentional underreporting [32]. Our finding provides another possibility that the deviations of DNA methylation levels might be caused by neighboring genetic variants. For highly smoking-related CpG sites, active smoking is still the main driver of DNA methylation changes. For instance, the SNP rs2227357 only contributed to about 0.01% of the changes of the methylation level of cg03636183 (F2RL3), and the SNP rs790051 altered only 0.37% of the methylation level of cg05951221 (2q37.1). However, the mQTLs affected the methylation levels of less robustly smoking-related loci much more. For instance, the SNPs rs78131 and rs2741302 explained nearly 19% of the changes of cg26963277 (KCNQ1) and cg27241845, respectively. While additional external validation studies certainly are needed, we speculate that this strong diversity of results might be a result of undetermined biological interactions between the SNP and CpG sites.

Smoking-related CpG sites have been recognized as informative signatures of smoking exposure and smoking-related health outcomes [3, 4]. Part of these 70 CpG sites with mQTLs have been reported to be highly associated with long-term smoking exposure [cg03636183 (F2RL3) and cg06126421 (6p21.33)] [9, 10], aging-related health outcomes, such as telomere length (cg21393163) [33] and the development of frailty [cg14753356, cg19589396, cg23667432 (ALPP) and cg25189904 (GNG12)] [34], and were even employed to construct a comprehensive index to predict smoking impact in buccal cells [35]. Together with previous studies [8, 23], we suggest that future investigations utilizing smoking-related CpG sites might need to take the genotypes, especially the mQTLs of less robustly smoking-related loci, into consideration to account for their potential impact on DNA methylation levels.

Beyond the SNP-CpG associations, null associations of 192 mQTLs with active smoking and all-cause mortality additionally imply that these novel genetic variants might be independently associated with the DNA methylation changes and might not be involved in the pathophysiological development of smoking-related health outcomes. Therefore, these mQTLs might have the potential to be used in the causal inference tests between the CpG sites and smoking-related health outcomes as instrumental variables (Mendelian Randomization, MR) [36]. Recently, researchers have suggested a two-stage MR test to establish the causal role of epigenetic processes in pathways of diseases [37]. Larger population-based investigations with longitudinal design and repeated measurements of smoking exposure and epigenome data are warranted to evaluate these potential instrumental variables and obtain further insights into the plausibility of suggested causal effects of DNA methylation in the development of smoking-related diseases.

Major strengths of the present study include comprehensive information on a broad range of covariates in a population-based cohort and validation in an independent subgroup. Some limitations still have to be acknowledged in the interpretation of study results. First, smoking-related shifts in leukocyte distribution might affect the associations of DNA methylation in whole blood samples with active smoking [38]. Hence, we adjusted for leukocyte distribution by the Houseman algorithm to restrict potential confounding from differential blood counts to the greatest possible extent [39]. Further studies are also needed to evaluate to what extent our results can be generalized to middle-aged individuals or non-Caucasians, as the ESTHER study was conducted in the older (aged 50–75 years), almost exclusively Caucasian population in southern Germany during a routine screening program. In addition, our study had limited power for detecting direct associations of mQTLs with smoking exposure and all-cause mortality due to limited numbers of cases. Finally, we only measured a relatively small window of genetic regions (± 50 kb) in whole blood DNA due to the consideration of controlling for the pleiotropic effect or reverse causations from unknown genetic or epigenetic factors and the limited coverage of OncoArray [40], more mQTLs (cis- or trans-) for smoking-related CpG sites need to be established by expanding the search window. Since DNA methylation is highly tissue-specific, larger cohorts with various human tissues are also needed for more comprehensive evaluation of the whole landscape of genetic impact on the epigenome.

Conclusions

In conclusion, this study identified 192 mQTLs for 70 smoking-related CpG sites. These variants might theoretically reflect inherited differences in epigenetic states of people and their susceptibilities to smoking-related health outcomes. Incorporation of mQTLs might enhance the epigenetic-based assessments of smoking or smoking-related health outcomes by accounting for potential confounding from genetic background. Our results need to be further validated and confirmed in additional studies with larger number of participants and more detailed assessment of genomic and epigenomic data, including the CpG sites that have not previously been replicated. Along with previous investigations on the epigenetic changes related to other environmental exposures or lifestyle factors, our study adds evidence for the complex interplays among genetic traits, epigenetic signatures, and environmental factors.

Methods

Study design and population

Study subjects were selected from the ESTHER study, an ongoing statewide population-based cohort study conducted in Saarland, a state located in southwestern Germany. Details of the study design have been reported previously [41]. Briefly, 9949 older adults (aged 50–75 years) were enrolled by their general practitioners during a routine health check-up between July 2000 and December 2002 and followed up thereafter. The current cross-sectional analysis is based on data and biospecimen collected at baseline. Two independent subgroups were selected as the discovery and the validation panel for DNA methylation analyses as previously described [33]. Briefly, the discovery panel included 581 participants recruited consecutively at the start of the ESTHER study between July and October 2000. The validation panel included 368 participants randomly selected from the participants recruited between October 2000 and March 2001. The study was approved by the ethics committees of the University of Heidelberg and the state medical board of Saarland, Germany. Written informed consent was obtained from all participants.

Data collection

Information on socio-demographic characteristics, lifestyle factors, and health status at baseline was obtained by standardized self-administered questionnaires. In particular, detailed information on lifetime smoking history was obtained, including current smoking status and intensity, age at initiation, and smoking intensities at various ages, as well as the age of quitting smoking for former smokers [42]. Additional information on body mass index (BMI) was extracted from a standardized form filled by the general practitioners during the health check-ups. Prevalent cardiovascular disease (CVD) at baseline was defined by either physician-reported coronary heart disease or a self-reported history of a major cardiovascular event, such as myocardial infarction, stroke, pulmonary embolism, or revascularization of coronary arteries. Prevalent diabetes was defined by physician diagnosis or the use of glucose-lowering drugs. Prevalent cancer [ICD-10 C00-C99 except non-melanoma skin cancer (C44)] was determined by self-report or record linkage with data from the Saarland Cancer Registry (http://www.krebsregister.saarland.de/ziele/ziel1.html; in German). Deaths during the follow-up (between 2000 and end of 2014) were identified by record linkage with population registries in Saarland. Participants migrated out of Saarland were censored at the date last known to be alive. Information about the major cause of death was obtained from death certificates provided by the local public health offices and was coded with ICD-10 codes.

DNA methylation data

Blood samples were taken during the health check-up and stored at −80 °C until further processing. DNA from whole blood samples was collected using a salting out procedure [43]. DNA methylation profiles were extracted by the Illumina Human Methylation 450K BeadChip (Illumina, San Diego, CA, USA). As previously described [44], samples were analyzed following the manufacturer’s instruction at the Genomics and Proteomics Core Facility of the German Cancer Research Center, Heidelberg, Germany. Illumina’s GenomeStudio® (version 2011.1; Illumina, Inc.) was employed to extract DNA methylation signals from the scanned arrays (Module version 1.9.0; Illumina, Inc.). The methylation level of a specific CpG site was quantified as a β value ranging from 0 (no methylation) to 1 (full methylation). According to the manufacturer’s protocol, no background correction was done and data were normalized to internal controls provided by the manufacturer. All controls were checked for inconsistencies in each measured plate. Probes with a detection p value > 0.05 were excluded from analysis. We utilized the Illumina normalization and preprocessing method implemented in Illumina’s GenomeStudio®. We selected the profiles of 151 smoking-related loci which had been identified ≥ 2 times in previous smoking EWASs for the present analysis [25].

Genotyping data

Extracted DNA from blood cells was genotyped using the Illumina Infinium OncoArray BeadChip (Illumina, San Diego, CA, USA). General genotyping quality control assessment was as previously described [45]. Genotypes for common variants across the genome were imputed using data from 1000 Genomes Project (phase 3, Oct. 2014) with IMPUTE2 v2.3.2 after pre-phasing with SHAPEIT software v2.12. We set thresholds for imputation quality to retain both potential common and rare variants for validation. Specifically, poorly imputed SNPs defined by an information metric I < 0.70 were excluded. All genomic locations are given in NCBI Build 37/UCSC hg19 coordinates. All SNPs having a MAF < 1% were excluded. After imputation, the SNP set consisted of 9,198,808 genotyped and imputed SNPs. PLINK v1.90 was then used to extract SNPs for the required regions of interest [46]. As shown in Fig. 1, we first identified SNPs within 50 kb upstream and downstream from each of the 151 smoking-related CpG sites (470,582 SNP-CpG pairs), a window in which most SNPs with significant cis associations with CpG sites are located [13]. The locus cg01208318 was excluded without any corresponding SNPs in this restricted region. For each of the remaining 150 CpG sites, we excluded any SNPs with ≥ 1% missing values (n = 44,931), deviating from the Hardy-Weinberg equilibrium (HWE exact test’s p value < 0.0001, n = 1), with a minor allele frequency ≤ 0.1 (n = 414,671) or with high linkage disequilibrium (LD, R ² ≥ 0.5) (Additional file 1: Table S1). After the final quality control, 1396 SNP-CpG pairs with strongest SNPs remained for analysis, which were constituted of 150 CpG sites and 909 corresponding SNPs (Additional file 1: Table S2).

Statistical analyses

First, major socio-demographic characteristics, lifestyle factors, smoking behavior, and prevalence of major chronic diseases in both the discovery and the validation panel were summarized by descriptive statistics.

We then evaluated the associations between the methylation intensities of the 150 CpG sites and corresponding SNPs to identify mQTLs as follows. For all SNP-CpG pairs, we used a mixed linear regression model with methylation batch as a random effect in which the methylation level of CpG site was the outcome and each regional SNP was the predictor (categorical variable, coded into 0, 1, and 2 based on the numbers of the minor allele). The model was fully adjusted for the following covariates that have been shown to be associated with DNA methylation changes [47–54]: age (years), sex (male/female), smoking status (current/former/never smoking), alcohol consumption (abstainer, low [women, 0 to < 20 g/d; men, 0 to < 40 g/d], intermediate [20 to < 40 g/d and 40 to < 60 g/d, respectively], high [≥ 40 g/d and ≥ 60 g/d, respectively]), body mass index (BMI, kg/m², underweight [< 18.5, < 1% of the study population] or normal weight [18.5 to < 25], overweight [25 to < 30], obese [≥ 30]), physical activity (inactive [< 1 h of physical activity/week], medium or high [≥ 2 h of vigorous or ≥ 2 h of light physical activity/week], low [other]), the leukocyte distribution estimated by the Houseman algorithm [39], the prevalence of CVD (yes/no), diabetes (yes/no), and cancer (yes/no) at the baseline. After correction for multiple testing by false discovery rate (FDR, Benjamini-Hochberg method [55]), SNP-CpG pairs with a FDR < 0.05 were selected and then analyzed in the validation panel. SNPs of the pairs with a FDR < 0.05 in the validation panel were eventually identified as the mQTL for the corresponding CpG site.

Furthermore, we tested the contributions of the identified mQTLs to the DNA methylation levels of corresponding CpG sites. Due to the limited number of individuals in the subgroup of minor homozygotes, we recoded the SNPs in order to use the dominant model, in which the heterozygote and minor homozygotes were combined as the carrier of minor allele and the major homozygotes were considered as non-carrier of the minor allele. We compared the coefficients of active smoking exposure (current vs. never smoking) in the fully adjusted model without the mQTLs (β ₁) with the fully adjusted model including the mQTLs (β ₂). The changes of coefficients were calculated as 100% * (β ₁–β₂)/β ₁, and their absolute values were determined as the percentage change contributed by mQTLs. The percentage changes were categorized by the absolute distances (bp) between CpG sites and corresponding mQTLs and the reported frequencies of CpG sites. To explore whether the gene-environment interactions could modify the DNA methylation changes of smoking-related CpG sites, we also tested whether interactions between the identified mQTLs and active smoking exposure (current vs. never smoking) could affect the impact of smoking on the methylation levels of corresponding CpG sites. The mQTLs, smoking status, and their interaction (mQTLs*smoking status) were added in the model as predictors, and the methylation levels of CpG sites were outcomes. After controlling for all the potential covariates, the interactions with a FDR < 0.05 were considered as methylation-related interactions for corresponding CpG sites.

Finally, we examined whether the identified mQTLs (carrier/non-carrier) were associated with six active smoking indicators: ever smoking (current and former smoking) vs. never smoking, current smoking vs. never smoking, current smoking vs. former smoking, cumulative smoking (pack-years), durations of smoking (years), and the age of smoking initiation. The mixed linear models were fully adjusted for age (years), sex, smoking status, alcohol consumption, BMI, physical activity, the prevalence of CVD, diabetes, and cancer as described above. The mQTLs with a FDR < 0.05 were identified as smoking-related SNPs. We also assessed the associations of the significant mQTLs (carrier/non-carrier) with all-cause mortality in ESTHER study. Due to the limited number of deaths, we combined both subsets and performed the analysis using a multiple COX regression model. The model was adjusted for the above potential covariates and SNPs with a FDR < 0.05 were considered as all-cause mortality related variants.

Data cleaning and all aforementioned statistical analyses were performed by SAS version 9.4 (SAS Institute Inc., Cary, NC, USA). Manhattan plots for both panels were plotted by R package “ggplot2.”

Abbreviations

CpG sites

Cytosine-phosphate-guanine sites

CVD

Cardiovascular disease

EWASs

Epigenome-wide association studies

FDR

False discovery rate

mQTLs

Methylation quantitative trait loci

SNPs

Single nucleotide polymorphisms

Authors’ contributions

XG conceived the study, carried out the main data analyses, interpreted the data, and drafted the manuscript. HT conducted genotyping and imputation and contributed to the interpretation of the study. YZ and LPB contributed to the design of the study. HB conducted the ESTHER study and contributed to all aspects of this work. All authors contributed to revision of the manuscript and approved the final version for submission.

Ethics approval and consent to participate

The ESTHER study was approved by the ethics committees of the University of Heidelberg and the state medical board of Saarland, Germany. Written informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.