Study population

Data from nine population-based cohort studies were used in the analysis. In addition to the FHS [20], our investigation included the Agricultural Lung Health Study (ALHS) [21], the Cooperative Health Research in the Region Augsburg (KORA) [22], the Genetic Epidemiology Network on Arteriopathy (GENOA) Study [23], the Health and Retirement Study (HRS) [24], the Multi-Ethnic Study of Atherosclerosis (MESA) Study [25], the Rhineland Study [26], the Rotterdam Study [27], and the Study of Health in Pomerania (SHIP) [28]. Institutional review committees of all cohorts approved this study, and all study participants provided written informed consent.

In each cohort, participants with prevalent CVD, heart failure, and atrial fibrillation were excluded. Prevalent CVD includes the following conditions: angina pectoris, coronary insufficiency, cerebrovascular accident, atherothrombotic infarction of the brain, transient ischemic attack, cerebral embolism, intracerebral hemorrhage, subarachnoid hemorrhage, or intermittent claudication. After excluding participants without DNA methylation data, 3,898 participants in FHS and 11,544 participants in eight independent external cohorts were included in cross-sectional association analyses, while 3260 participants in FHS and 3910 participants in five external cohorts were included in longitudinal association analyses (Fig. 1).

Clinical and behavioral data collection

Overall, clinical data for traits such as age, BMI, SBP, DBP, and the use of antihypertensive medication were collected at in-person examinations. Stage 2 HTN was defined as SBP ≥ 140 mm Hg, DBP ≥ 90 mm Hg, or use of antihypertensive medication for treating HTN at the examination. We added 15 mm Hg and 10 mm Hg to a measured SBP and DBP values, respectively, for participants currently using antihypertensive medication.

Cigarette smoking status was determined based on self-reported smoking behavior. Current smokers were participants who smoked on average at least one cigarette per day in the past year. Self-reported alcohol intake was captured via questionnaires wherein the participants reported the frequency with which they consumed various alcoholic beverages (i.e., beer, liquor, or wine). A standard drink is 12 oz of beer, 4–5 oz of wine, or 1.5 oz of liquor, which is equivalent to appropriately 14 g of ethanol [16, 29]. We summed the total alcohol consumption across all alcoholic beverages and utilized “drink” (i.e., one drink = 14 g of ethanol) as the unit for the alcohol consumption. This study included nine population-based cohorts, and therefore, we focused on habitual alcohol consumption in general populations rather than examining specifically for alcohol disorder. Study-specific methods for clinical data collection are presented in Supplemental Text and Supplemental Table 1.

DNA methylation data collection and processing

DNA methylation was measured using blood samples collected at the same time when alcohol consumption data were assessed in all cohorts. Whole blood samples were assayed for DNA methylation via the Infinium Human Methylation 450 BeadChip platform or Infinium MethylationEPIC platform (San Diego, CA) (Supplemental Text). The methylated probe intensity and total probe intensities were extracted using the Illumina Genome Studio (version 2011.1) with the methylation module (version 1.9.0). Preprocessing of the methylated (M) signal and unmethylated signal (U) was conducted; methylation beta-value (βM) was defined as \(\beta =\frac\). Further information regarding DNA extraction and processing has been outlined [16] and described in Supplemental Text.

Derivation of epigenetic risk score (ERS) for alcohol consumption

We implemented an ERS score based on 144 alcohol-associated CpGs previously reported in a meta-analysis of 6926 participants of European ancestry [15]. The previous study generated the regression coefficients (βi, i = 1–144) for these 144 CpGs using the Least Absolute Shrinkage and Selection Operator (LASSO) method. This method performs both variable selection and regularization to enhance the prediction accuracy and interpretability of the resulting statistical model by shrinking the coefficients of some variables to exactly zero [30]. An ERS score was calculated for each participant by summing the regression coefficient-weighted DNA methylation levels of the 144 CpGs: ERS score = \(\sum_^_\times _\). The ERS score represents personal DNA methylation levels in response to alcohol consumption. Across five cohorts, one drink of alcohol consumption was associated with 0.09 higher unit of ERS (Supplemental Table 2). Methods for calculating the ERS for cohorts missing certain CpGs can be found in Supplemental Text.

Discovery association analysis of ERS with BP traits in FHS

We performed both cross-sectional and longitudinal regression analyses in FHS to examine the association between the ERS (independent variable) and BP traits: SBP (continuous), DBP (continuous), and HTN (dichotomous) (dependent variables). Linear mixed regression models were used to evaluate the association of the ERS with the two continuous BP traits. Generalized estimating equations (GEE) were used to evaluate the association of the ERS with dichotomous HTN. A total of 3,898 participants were included in the cross-sectional analysis from the FHS Offspring cohort (n = 2393; examination 8) and FHS Third Generation cohort (n = 1505; examination 2) participants. All models were adjusted for age, age squared, sex, BMI, and current smoking status. Covariates were selected based on their significant correlation with BP traits and/or DNA methylation based on previous studies [31,32,33,34,35] and our own data (Supplemental Fig. 1–3). Age, sex, and BMI are important risk factors for BP traits [31,32,33]. The age-squared term was included as a covariate due to the quadratic relationship between age and BP traits (Supplemental Fig. 1–2). Current smoking (versus current none smokers) status was included for its association with BP [34] and DNA methylation [35]. The familial correlation (for family data) was further adjusted for the random effect in models [36]. The familial correlation, or genetic correlation, was calculated based on the self-reported pedigree file to quantify the proportion of shared genetic material or the degree of trait similarity due to genetic factors [37]. Correlation is on average 0.5 for a parent–child relationship, 0.25 for a grandparent–grandchild relationship, and 0.125 for the first-degree siblings.

Longitudinal analyses of all BP traits included FHS Offspring cohort participants (n = 1932) who attended both examinations 8 and 9 and third-generation participants (n = 1328) who attended both examinations 2 and 3. Our linear mixed regression models evaluated the association of change in BP over time (i.e., ΔSBP and ΔDBP) with the baseline ERS after adjusting for baseline age, baseline age-squared, sex, baseline BMI, baseline smoking status, baseline SBP/DBP (i.e., if the model’s outcome was ΔDBP, we adjusted for baseline DBP), and time between baseline and the follow-up examination. Our GEE models evaluated the association of incident HTN with the baseline ERS after adjusting for baseline age, baseline age-squared, baseline BMI, baseline smoking status, and time between baseline and follow-up examination; in addition, these GEE models excluded all participants with HTN at baseline examination. In the sensitivity analysis, we defined participants with stage 1 HTN using the 2017 guideline (i.e., ≥ 130/80 SBP/DBP mm Hg or with antihypertension treatment) [38]. We performed cross-sectional and longitudinal GEE models to investigate the associations of ERS with prevalent and incident HTN using the new definition.

Replication association analysis of ERS with BP traits in external cohorts

For replication, independent external participants from eight cohorts (n = 11,544) were included in cross-sectional association analyses, while participants from five cohorts (n = 3910) were included for longitudinal association analyses, using the same methods described for the discovery stage in the FHS. We summarized the results from the association analyses using an inverse-variance weighted, fixed-effects meta-analysis, assuming a single true effect between the ERS and a BP trait.

Analysis of epigenetic risk score with BP traits in participants without antihypertension medication

To minimize the possible effects of antihypertension medication on DNA methylation, we conducted a sensitivity analysis among participants without antihypertension medication in five cohorts (i.e., FHS, GENOA, HRS, Rhineland Study, and SHIP). Similar to the primary analysis, we conducted the cross-sectional analysis using linear mixed effects model with ERS as the independent variable and BP traits as the dependent variables in each cohort. The priori power analysis was conducted with a range of assumed effect sizes (i.e., 0.001, 0.0025, 0.005, 0.01, 0.1, 0.2, 0.5), default type I error rate (\(\alpha =0.05\)), and actual sample sizes to check if the type-II error was consistent between primary and sensitivity analyses.

Association analysis of ERS with alcohol consumption

We used a linear mixed regression model to test the cross-sectional association between the ERS (outcome) and self-reported alcohol intake (exposure) in each of the five cohorts (i.e., FHS, GENOA, HRS, Rhineland Study, and SHIP). The change in the ERS associated with one drink of alcohol consumption per day was calculated with adjustment for age, age-squared, sex, BMI, current smoking status, and familial correlation.

Association analysis of alcohol consumption with BP traits

To compare the association of BP traits with ERS and questionnaire-based alcohol consumption, we performed cross-sectional (i.e., FHS, GENOA, HRS, Rhineland Study, and SHIP) and longitudinal (i.e., FHS, GENOA, and SHIP) analyses between BP traits and alcohol consumption. We used linear mixed effects or GEE models to quantify the associations between SBP/DBP/HTN (outcome variables) and alcohol consumption (predictor). All models were adjusted for age, age-squared, sex, BMI, current smoking status, and familial correlation.

Association analysis of ERS with biochemical biomarkers of alcohol intake

We tested the association of the ERS with two established biomarkers of chronic alcohol consumption: aspartate aminotransferase (AST) and alanine aminotransferase (ALT) concentrations. Separate linear mixed regression models were used with each enzyme as the dependent variable. Serum AST and ALT were measured on fasting morning samples using the kinetic method (Beckman Liquid-State Reagent Kit) [39]. Model 1 (i.e., the reduced model) quantified the association between the self-reported alcohol intake and liver enzyme concentrations after adjusting for age, sex, BMI, and current smoking status. Model 2 (i.e., the full model) further adjusted for the ERS. To compare the two models, we also performed a likelihood ratio test (LRT) to gauge whether the addition of the ERS significantly improved model fit.

Association analysis of individual alcohol-associated cpgs with BP traits in FHS

We examined the cross-sectional association of 144 CpGs that were used to calculate ERS with BP traits in the FHS. We applied the linear mixed effect model to account for the pedigree with each CpG probe as the predictor variable and SBP/DBP as the outcome variable. Covariates included age, age-squared, sex, BMI, and current smoking status.

All statistical analyses were conducted using the R (version 4.0.3) software package [40]. Meta-analyses were conducted with the metafor package (version 3.0.2) [41]. The priori power analysis was performed using ‘pwr.f2.test’ function in the pwr R package (version 1.3–0) [42]. LRT was performed using the ‘lrtest’ function in the lmtest R package (version 0.9.39) [43]. Statistical significance was defined as two-sided p < 0.05.