Genome-wide analysis of DNA methylation and risk of cardiovascular disease in a Chinese population

The study samples were collected from the China PEACE Million Persons Project (MPP) cohort. The design and conduct of the project pilot have been previously described in detail [12]. In brief, the China-PEACE MPP is a national, population-based screening project funded by the Chinese government to identify individuals at high-risk of CVD. The project enrolled community population if they were 35–75 years old and could confirm their residence in a selected region including 141 primary health care sites (88 rural counties and 53 urban districts) from all 31 provinces in the mainland of China from September 15, 2014 to June 20, 2017. The participants were initially screened for high-risk of CVD by measurements of blood pressure, height, weight and blood lipid, and a questionnaire on cardiovascular-related health status. The subjects at high-risk of CVD were identified if they met at least one of four criteria: 1. Medical history of myocardial infarction (MI), percutaneous coronary intervention (PCI), coronary artery bypass grafting (CABG) treatment, or stroke; 2. High blood pressure defined as systolic blood pressure (SBP) ≥ 160 mmHg or diastolic blood pressure (DBP) ≥ 100 mmHg; 3. Dyslipidemia defined as low-density lipoprotein cholesterol (LDL-C) ≥ 160 mg/dL (4.14 mmol/L) or high-density lipoprotein cholesterol (HDL-C) < 30 mg/dL (0.78 mmol/L); 4. Risk of CVD in 10 years ≥ 20% based on WHO/ISH Cardiovascular Risk Prediction Charts for the Western Pacific Region B [13]. The high-risk participants received further health assessments, including electrocardiography, ultrasound scan, blood and urine analysis, and a questionnaire on lifestyle and medical history and were followed up every year to collect information on medication adherence, risk factor control, and any hospitalization. The data were collected from standardized in-person interviews by trained medical staff. The central ethics committee at the China National Center for Cardiovascular Disease (NCCD) approved this project. All participants had completed a written informed consent before participation in the project.

In the present study, we performed a nested case–control study based on the high-risk participants from the MPP cohort to examine and validate the association of genome-wide CpGs with the development of CVD. The inclusion criteria: 1. Blood samples were collected after January 1, 2016; 2. Two years follow-up data was available; 3. The questionnaire for the high-risk participants was available. Patients were excluded if they have self-reported any of medical history of cancer, chronic liver and kidney disease, infectious diseases, or any CVD at baseline.

Among the above eligible population, case (subjects who developed cardiovascular events) and control (subjects who did not develop cardiovascular events) were 1:1 matched for age (within 1 year), sex, BMI, the month of blood collection (within 1 month), medications treatment (antiplatelet medications and statins), ethnicity (han or non-han), current smoking status, and current alcohol intake at baseline. Then, 121 pairs were randomly sampled from the above eligible case–control pairs. Among them, 83 pairs case–control and 38 pairs case–control were randomly selected and regarded as the discovery set and the replicated set, respectively. Additional file shows this in more detail. [see Additional file 2: Figure S1].

Cardiovascular events were self-reported by participants during follow-up visits. In the present study, hypertension was defined as SBP ≥ 140 mmHg, or DBP ≥ 90 mmHg, or self-reported any antihypertensive drugs. Hyperlipidemia was defined as total cholesterol (TC) ≥ 6.2 mmol/L or LDL-C ≥ 4.1 mmol/L according to Chinese guideline for adult’s blood lipid, or self-reported lipid-lowering drug. Diabetes was defined as blood glucose greater than 7 mmol/L after at least 8 h after last meal or random blood glucose greater than 11.1 mmol/L, or any kind of hypoglycemic agents. Details of these variables are shown in Table 1.

In this study, TC, LDL-C, HDL-C, triglycerides (TG), and high-sensitivity C-reactive protein were measured by standardized enzymatic methods using Beckman Coulter AU680 analyzers (Beckman AU reagent). Hemoglobin A1c (HbA1c) was measured by high-performance liquid chromatography on the Arkray ADAMS-A1C HA-8180 analyzer. All tests were completed at the NCCD laboratory.

CVD events included CVD admission and CVD death. CVD admission was defined as any hospitalization for MI, ischemic stroke, CABG, PCI, and thrombolysis treatment for acute MI or stroke. CVD death was defined as death due to MI, angina, heart failure (HF), ischemic or hemorrhagic stroke reported by patient’s relatives or doctors. For the patients who have developed multiple cardiovascular events, only the earliest one was used in subsequent analyses.

Genomic DNA was isolated from human whole blood samples using Chemagic 360 and chemagic DNA Buffy Coat 200 Kit according to the manufacturer’s protocol. The quality and quantity of the extracted genomic DNA was analyzed with a DropSense 96 Spectrophotometer. The samples (500 ng genomic DNA) were treated for sodium bisulfite conversion using the EZ DNA methylation Gold Kit (Zymo Research, USA). And Genome-wide DNA methylation profiles were assessed using the Infinium Human Methylation EPIC BeadChip (850 K) (Illumina, USA) following manufacturers’ protocol, and DNA methylation level was quantified as a β value. β value represents the ratio of the fluorescent signal intensity measured by methylated and unmethylated probes and range from 0 (all copies of the CpG site in the sample are unmethylated) to 1 (all copies of the CpG site in the sample are methylated) [14].

Data quality control was performed using the chip analysis methylation pipeline (ChAMP) package [15] implemented in R both in discovery set and in replication set. Probes meeting one of following criteria were removed: (1) probes with detection p > 0.01; (2) probes with < 3 beads in at least 5% of samples; (3) all non-CpG probes contained in dataset; (4) all multi-hit probes; (5) probes containing single nucleotide polymorphisms [16]; and (6) all probes located on chromosome X and Y. After quality control, 769,031 CpGs were retained for analysis. The filtered data were normalized by Beta Mixture Quantile dilation (BMIQ) [17]. Additionally, batch effect was adjusted to reduce technical biases by COMBAT, which was evident in a singular value decomposition (SVD) analysis [18].

Differences in baseline characteristics between case and control groups were evaluated by Kruskal–Wallis test for continuous variables or Chi-Square test for categorical variables. Before genome-wide DNA methylation association analysis, we performed principal component analysis (PCA) using 769,031 CpGs in both discovery and replication set, and the first 5 principal components (PCs) with proportion of variance > 1% were included as covariates in model for minimizing the potential technical bias. To investigate the association of CVD outcomes and DNA methylation level at each CpG, we fitted linear regression models using CpGassoc package in R, with β as dependent variables and CVD outcomes as independent variables. The models were adjusted for covariates, including age, sex, BMI, current smoking status, hypertension, diabetes mellitus, hyperlipidemia, cell components and the first 5 PCs. Cell components were estimated by the proportions of whole blood (CD4 + T cells, CD8 + T cells, B cells, NK cells, monocytes, and neutrophile granulocyte) using the minfi [19] of R package based on Houseman approach. We used the FDR < 0.005 to identify significant CVD-associated CpGs in discovery set.

In replication set, the significant CpGs identified in discovery set were validated by linear regression models using CpGassoc package in R. The models were adjusted for the same factors as that in discovery set. The CpGs with p < 0.05 in replication set and the consistent direction of effect in both discovery and replication sets were considered as significant replication CpGs.

For significant CpGs, we used the annotation provided by Illumina and University of California Santa Cruz UCSC database (GRCh37/hg19). The genes annotated to CpGs with p < 10^–4 in discovery set were selected for the enrichment analysis. The enrichment analysis in GO terms were conducted with R package 'clusterProfiler', and FDR < 0.05 was used to identify significantly enriched categories.

Furthermore, ROC analysis was performed to assess the predictive value of significant CpGs in CVD outcomes. There were two prediction models included in ROC analysis: 1) model with baseline characteristics only; 2) model with both baseline characteristics and the significant CpGs. As a reference model, baseline characteristics included age, gender, BMI, current smoking status, current alcohol intake, hypertension, diabetes mellitus, hyperlipidemia, antiplatelet medication, stains treatment and family history of coronary heart disease. Statistical analysis were performed using the SAS version 9.4 and R version 3.5.

The baseline characteristics of the study subjects are shown in Table 1. In discovery set, there were 83 CVD cases (65 admissions for CVD, 18 died of CVD) and 83 controls with mean age of 62 years. The replication set included 38 CVD cases (33 admissions for CVD, 5 died of CVD) and 38 controls with mean age of 61 years. As the study subjects were from the high-risk participants of the MPP cohort, the proportion of hypertension subjects in our study cohort was approximately as high as 90%. There was no significant difference in most variables between the case and control group in two cohorts, except for hyperlipidemia (p = 0.038) and diastolic pressure (p = 0.019) in the discovery set.

Before performing a epigenome-wide association analysis, we did SVD analysis and PCA based on plate number, chip number and date of test for examining batch effect in the normalized data by BMIQ and the further processed data by COMBAT, respectively. SVA analysis showed that batch effect was almost completely controlled after adjustment in different chips and different sample locations by COMBAT [see Additional file 2: Figure S2, S3]. In the PCA, no significant aggregations were found between different plates, different chips and different date of test (see Additional file 2: Figure S4). After that, PCA based on study samples was conducted using 769,031 CpGs in both discovery and replication set, and the first 5 PCs with proportion of variance > 1% were included as covariates in the genome-wide association analysis by linear regression model [see Additional file 2: Table S1, S2]. The results showed that eight CpGs achieved a genome-wide significance level (FDR < 0.005) after adjustment for age, gender, BMI, current smoking status, hypertension, diabetes mellitus, hyperlipidemia, cell compositions and the first 5 PCs (Fig. 1 and Table 2). Six of 8 CVD-associated CpGs were located in SPON1 gene body (cg11651314), in PACS1 gene body (cg03914662), in TSS200 of the UACA gene (cg09306458), in CCDC50 gene body ( cg05946546), in 1stExon of the CYP8B1 gene ( cg07655795), and in HSD17B11 gene body ( cg02518222), respectively (Table 2). Of them, increased methylation level at CpG cg07655795 (CYP8B1) was associated with an increased risk of CVD, while decreased methylation level at the remaining 7 CpGs were associated with an increased risk of CVD.

For the 8 CVD-associated CpGs (FDR < 0.005) identified in discovery set, we performed replication analysis in an independent cohort (38 cases and 38 controls). Two CpGs of them (cg09306458, cg06901278) were replicated (p < 0.05) with consistent direction of effect in both sets (Table 2). In two replicated CpGs, cg09306458 was located in regulatory region of the UACA gene, and cg06901278 had no annotation gene.

Go term enrichment analysis were conducted to identify the possible functions affected by annotation genes. A total of 787 individual genes from 1036 CpGs with p < 10^–4 in discovery set were used for enrichment analysis. There were 17 significantly functional categories identified (FDR < 0.05), including 6 categories in biological processes (BP), and 11 categories in cellular component (CC) (Fig. 2) (see Additional file 1: Excel S1). The most enriched categories were significant enrichment for anatomical structure homeostasis as well as regulation of vesicle-mediated transport, which were located in BP group (Fig. 2).

ROC curves were drawn to assess the specificity and sensitivity of CVD prediction on the basis of 8 CVD-related CpGs identified in discovery set and baseline characteristics as a reference model. As shown in Fig. 3a, the model with both 8 CVD-related CpGs and baseline characteristics showed an excellent predictive effect for CVD occurrence and much better than the model with baseline characteristics only (AUC = 0.967, 95% CI (0.942 − 0.991); AUC = 0.621, 95% CI (0.536 − 0.706); p = 9.716E-15). In replication set, we also found the similar trend that the model combined 8 CVD-related CpGs and baseline characteristics showed significant predictive capacity compared with the reference model with baseline characteristics only (AUC = 0.786, 95% CI (0.676 − 0.896); AUC = 0.492, 95% CI (0.357 − 0.628); P = 1.178E-3) (Fig. 3b).