Alcohol consumption and epigenetic age acceleration across human adulthood

Sample characteristics

This study included 3823 participants with both DNAm measurements and alcohol consumption (Table 1). The three age groups contained similar proportions of women and men. Older participants (> 64 years) tended to have a lower education level (p < 0.001) compared to young (22–44 years) and middle-aged (45–64 years) groups (Table 1). For example, about 31.8% of participants in the older group, 47.5% of participants in the middle-aged group, and 63.5% of participants in the young group were 4-year college graduates. Additionally, the older group (80.5%) had a higher prevalence of hypertension than that in the young (29.1%) and middle-aged (56.1%) groups (p < 0.001) (Table 1).

A higher proportion of heavy drinkers (> 2 drinks per day for women and > 3 drinks per day for men) was found in the older group (5.6%) compared to middle-aged (3.5%) and younger (0.7%) groups based on the average total alcohol consumption over long-term (p < 0.001). Greater long-term average consumption of wine and liquor was observed in middle-aged and older participants compared to young participants, while the same trend was not observed with long-term average beer consumption.

Association of long-term average alcohol consumption with GAA and PAA

Before conducting association analyses in each age group, we first investigated the interaction between EAA and three age groups. A significant interaction was observed between long-term average alcohol consumption and age groups for PAA (p = 9.4e-4) but not for GAA (Supplementary Table 1). To be consistent, we explored the associations of alcohol consumption with PAA and GAA within the age groups. In the primary analyses, we quantified an association of long-term average alcohol consumption (the main predictor variable) with each EAA metric (the outcome variable) in linear regression, adjusting for sex, physical activity index (PAI), education level, body mass index (BMI), smoking (pack-year), chronological age, and lab index (for batch effects) as covariates. We observed significant associations of the long-term average alcohol consumption with GAA and PAA in middle-aged and older age groups, while not in the young age group (Figure 1). For example, one additional drink of total alcohol consumption per day was associated with a 0.43-year increase in GAA in middle-aged participants (p = 5.4e-6) and a 0.37-year increase in GAA in older participants (p = 0.001). Similarly, one additional drink of total alcohol consumption per day was associated with a 0.71-year increase in PAA in middle-aged participants (p = 2.1e-6) and a 0.60-year increase in PAA in older participants (p = 7.5e-4). No significant association was observed between long-term average total alcohol consumption and the other three EAA metrics (IEAA, EEAA, and EAASkinBlood) in any of the three age groups (Supplementary Table 2).

Figure 1. Association analyses between long-term average alcohol consumption and EAAs in each age group and in pooled samples in the Framingham Heart Study. Age groups: young (24–44 years), middle-aged (45–64 years), older (65–94 years). The x-axis represents the effect size of alcohol consumption on GAA or PAA. Results are adjusted for sex, physical activity score, education level, BMI, smoke pack-year, chronological age, and lab. The long-term average drinking was calculated as the average of consumption (total alcohol or each type of alcoholic beverages) across up to 26 years. For continuous consumption variables, the effect size was in response to a standard drink per day. The drinking category was grouped based on the long-term average alcohol consumption, with light drinkers as the reference. Non-drinkers were defined as participants with average alcohol consumption equal to zero; light drinkers were defined as less than 1 drink per day for women and less than 2 drinks per day for men; at risk drinkers were defined as 1–2 drinks per day for women and 2–3 drinks per day for men; heavy drinkers were defined as more than 2 drinks per day for women and more than 3 drinks per day for men. Abbreviations: GAA: GrimAge acceleration; PAA: PhenoAge acceleration. Effect sizes and p-values can be found in Supplementary Tables 2, 3.

To investigate if EAA displayed a linear relationship with alcohol consumption in the age groups, we compared EAA metrics in non-drinkers, at-risk drinkers, and heavy drinkers with light drinkers defined using long-term average total alcohol consumption (Figure 1). Categorical alcohol consumption variables were defined based on the following rules: non-drinkers included participants with average total alcohol consumption equal to zero; light drinkers included women consuming < 1 drink per day and men consuming < 2 drinks per day; at-risk drinkers included women consuming 1–2 drinks per day and men consuming 2–3 drinks per day; heavy drinkers contained women consuming > 2 drinks per day and men consuming > 3 drinks per day. We observed that heavy drinkers displayed a stronger association of GAA and PAA than at-risk drinkers in middle-aged and older participants, while non-drinkers possessed a similar GAA and PAA as light drinkers. As shown in Figure 1, the relationship between total alcohol consumption and age acceleration is linear for both GAA and PAA. For example, compared to light drinkers, we observed an increase in GAA in at-risk drinkers (β = 0.70, p = 0.007) and in heavy drinkers (β = 2.17, p = 8.0e-8) in the middle-age group. Similarly, compared to light drinkers, we observed a 1.52-year increase in PAA in at-risk drinkers (p = 1.6e-4) and a 2.54-year increase in PAA in heavy drinkers (p = 5.4e-5) in the middle-age group (Supplementary Table 3).

Association of epigenetic age acceleration with different alcoholic beverages

For each type of alcoholic beverage (i.e., beer, wine, and liquor), we explored its association with EAAs using the same statistical models with the same set of covariates adjusted for the analyses of long-term average total alcohol consumption. We observed that three types of alcoholic beverages displayed different effect sizes in association analyses with GAA and PAA in middle-aged and older participants (Figure 1). Both wine (β = 0.91, p = 0.008) and liquor (β = 1.45, p = 7.4e-5) consumptions were associated with PAA, while no significant association was observed between beer consumption and PAA in middle-aged participants (Supplementary Table 2). In addition, liquor consumption displayed the greatest accelerative change compared to the other two alcoholic beverages (Figure 1). For example, we observed a 0.82-year increase in GAA with one additional drink of liquor consumption per day in middle-aged participants (p = 4.8e-4) and a smaller increase in GAA with one additional drink of beer consumption per day (β = 0.45, p = 5.2e-4) (Supplementary Table 2). The similar trend could also be observed in older participants. Compared to a 0.91-year increase in GAA with one additional drink of liquor consumption per day (p = 2.5e-5) in older participants, smaller increases were observed in GAA with one addition drink of beer or wine (beer: β = 0.27, p = 0.10; wine: β = −0.03, p = 0.90) (Supplementary Table 2).

Secondary analyses of alcohol consumption variables with epigenetic age acceleration

Interaction between sex/smoking status and long-term average alcohol consumption

To investigate whether sex and smoking status modified the association between alcohol consumption and EAAs, we explored the interaction between sex and smoking status with long-term average total alcohol consumption in each age group and the pooled sample. No significant interaction was observed between sex and long-term average total alcohol consumption (Supplementary Table 13). Smoking status did not modify the associations of long-term average total alcohol consumption with most EAAs. The only exception was smoking status modified the association of IEAA with total alcohol consumption in the middle-aged participants and the pooled sample (Supplementary Table 14). One additional drink of total alcohol consumption in former smokers increased 0.73-year in IEAA compared to the increase in non-smokers in the middle-aged group (p = 0.004).

Mediation analysis of epigenetic age acceleration

To investigate whether EAA mediates the association of long-term average alcohol consumption with hypertension, we conducted mediation analyses in each of three age groups and the pooled sample. We found long-term average total alcohol consumption as well as long-term average wine and liquor consumption were associated with higher odds of hypertension in the middle-aged group (Supplementary Table 15). For example, one additional drink of long-term average total alcohol consumption per day was associated with 1.27 times of odds of having hypertension (95% CI: 1.09 to 1.48; p = 0.002) in middle-aged participants, adjusting for sex, PAI, education, BMI, smoke pack-year, chronological age, and lab index. Similarly, one additional drink of beer (OR = 1.19, p = 0.03), wine (OR = 1.28, p = 0.04), and total alcohol consumption (OR = 1.19, p = 0.002) was associated with higher odds of hypertension in the pooled sample. Both GAA and PAA were also significantly associated with hypertension adjusting for the same set of covariates in the middle-aged group and the pooled sample (Supplementary Table 15). For example, a one-year increase in GAA and PAA was associated with 1.09 times and 1.05 times of odds of having hypertension in the middle-aged participants, respectively.

With mediation analyses, we found that GAA and PAA displayed mediating effects in most of the associations between long-term average alcohol consumption variables and hypertension in the middle-aged group and the pooled sample (p < 0.01) (Supplementary Table 16). For example, in the middle-aged group, 12.62% and 19.30% of associations of hypertension with long-term average wine and liquor consumption were mediated by PAA, respectively. GAA mediated 16.43% of the association between total alcohol consumption and hypertension in middle-aged participants. Similarly, 13.30% and 13.33% of the association of total alcohol consumption with hypertension was mediated through GAA and PAA, separately, in the pooled sample. The largest proportion of mediation in the pooled sample was observed by GAA in the association of liquor consumption with hypertension, that is, 27.55% of the association between liquor consumption and hypertension was mediated by GAA.

Study participants

The Framingham Heart Study (FHS) is a longitudinal cohort study initially established to identify cardiovascular risk factors. From 1948 to 1953, the FHS recruited 5209 residents of Framingham, Massachusetts into the Original cohort [44]. Children of the Original cohort and the spouses of the children (n = 5124) were enrolled in the Offspring cohort since 1971 [45]. The Third Generation cohort was enrolled since 2002 (n = 4095) [46]. Starting from enrollment, Offspring and Third Generation participants underwent regular health examinations every 4–8 years to collect demographic information and cardiovascular risk factors [47]. The Offspring cohort has undergone nine examinations and the Third Generation cohort has undergone three examinations. Peripheral blood samples have been collected from study participants at each examination. This study utilized participants with DNAm measurements from the Offspring participants at exam 8 and the Third Generation participants at exam 2. Because this study used longitudinal alcohol measures, we removed Offspring participants without alcohol consumption measurements at exam 1 and exam 8, then retained those who had at least one alcohol consumption measurement between exam 2 and exam 7. All Third Generation participants with DNAm measurement had alcohol consumption measurements at both exam 1 and exam 2. In addition, we excluded participants with missing values in any of the covariates. Finally, 2321 participants in the Offspring cohort and 1502 participants in the Third Generation cohort remained for subsequent analyses (Figure 2). Among the entire study participants (women 53.8%), ages ranged from 24 to 94 years with an average age of 58.1 years when DNAm was measured.

Figure 2. The flow chart of study design. We categorized the Framingham Heart Study Offspring and Third Generation participants into three age groups: young (24–44 years, n = 690), middle-aged (45–64 years, n = 1866), older (65–94 years, n = 1267). We conducted association analyses of long-term average and cross-sectional alcohol consumption with EAA metrics in each age group. We used the pooled sample to estimate to what extent GAA and PAA mediates the association between alcohol consumption and hypertension.

DNA methylation measurement in whole blood and quality control

DNAm profiling and quality control procedures have been described previously [10]. Briefly, 2846 participants from the Offspring cohort attending exam 8 and 1549 participants from the Third Generation cohort attending exam 2 received DNAm measurement using the Infinium HumanMethylation 450K BeadChip array (Illumina, Inc., San Diego, CA, USA). Genomic DNA was extracted from whole blood samples collected at routine exams. DNAm profiling was conducted by bisulfite conversion, followed by whole genome amplification, fragmentation, array hybridization, and single-base pair extension (the manufacturer’s protocols) [48]. The DNAm levels in Offspring participants were measured in two separate laboratories (n = 576 and 2270) and the DNAm levels in Third Generation participants were measured by the third laboratory in Illumina. Several quality control procedures were applied separately to each batch of DNAm data. Briefly, we removed cross-reactive probes that mapped to multiple locations at the CpG level [49]. We also removed low-quality probes with a high missing rate (> 20%), with detection p > 0.01, and with single nucleotide polymorphisms (SNPs) at CpG sites or ≤ 10 bp of single base extension [49, 50]. At the participant level, we excluded individuals if they had a missing rate > 1% or were outliers according to multi-dimensional scaling (MDS) analysis [51] and with a poor match to the 65 SNP genotypes on the Infinium HumanMethylation 450K BeadChip array.

Epigenetic age acceleration

The primary outcome variables were age accelerations based on PhenoAge [14] and GrimAge [15], which are the second generation of epigenetic age variables estimating biological age based on aging-related phenotypic biomarkers [52]. Biological age predicts chronological age with clinical biomarkers and reveals differences in biological aging levels by their significant associations with disease-specific morbidity and mortality [53, 54]. PhenoAge was calculated from 513 CpGs, reflecting morbidity and mortality of diseases [14]. GrimAge specially considered smoking pack-year-related CpGs and was calculated from 1030 CpGs [15]. As secondary analyses, three additional EAA variables based on Horvath’s age [11], Hannum’s age [12], and skin and blood clock [13] were used as outcome variables. Horvath’s age utilized 353 CpGs based on DNAm measurements from multi-tissues, and the corresponding age acceleration reflects IEAA in cells [11, 55]. Hannum’s age utilized 71 CpGs from blood samples, and its age acceleration, EEAA, tracked aging-related changes in blood cell components compared to IEAA [12, 55]. Skin and blood clock and its age acceleration (EAASkinBlood) focused on skin and blood samples and were estimated from 391 CpGs [13]. We used Horvath’s online calculators to compute lab-specific age accelerations for two primary and three secondary epigenetic age variables [11]. EAA was estimated as the residual from a linear model of chronological age regressed on each of the epigenetic age metrics. Of note, the epigenetic age variables were highly correlated with chronological age while EAA metrics were not correlated with chronological age (Supplementary Figures 4, 7).

Alcohol consumption measurement

Hypertension

To explore whether EAA metrics mediated the association between alcohol consumption and prevalent hypertension, we defined hypertension as participants with systolic blood pressure (SBP) ≥ 130 mmHg or diastolic blood pressure (DBP) ≥ 80 mmHg or any medication to lower blood pressure in Offspring participants at exam 8 and Third Generation participants at exam 2 [57].

Covariates

In order to control for confounding, the following covariates at exam 8 (Offspring) and at exam 2 (Third Generation) were considered: sex, PAI, education, BMI, chronological age, smoking pack-year, lab, and surrogate variables (SVs). The PAI score combined weighted hours of slight, moderate, and heavy activity [58] to capture the total energy expenditure from body movements. The hours to conduct the slight, moderate, and heavy activities were recorded by the questions “the number of hours with slight/moderate/heavy activities for a typical day.” We modified the metabolic equivalent score (MET-score) formula used in the International Physical Activity Questionnaire (IPAQ) to calculate PAI in this study [59]:

$$\text{PAI}=3.3\times \text{slight}\text{hours}+4\times \text{moderate}\text{hours}+8\times \text{heavy}\text{hours}$$

We categorized education into three levels: without a college degree, some college degree, and a four-year college degree or above. Smoking status was categorized into three levels: non-smokers, former smokers, and current smokers. Pack-year of smoking variable measured both smoking quantity and smoking duration by multiplying the number of cigarette packs per day by years of smoking [60]. The number of cigarettes consumed per day was assessed with the following question “How many cigarettes do you smoke per day now” for each participant in the Offspring and the Third Generation cohorts and transformed into the number of packs per day by assuming 20 cigarettes in each pack. Age at initiation and end of smoking were additionally collected to capture the total years of smoking.

Statistical analysis

For primary analyses, the participants (2321 Offspring participants and 1502 Third Generation participants) were divided into three age groups: young (24–44 years, n_young = 690), middle-aged (45–64 years, n_middle-aged = 1866), and older age (> 64 years, n_older = 1267) groups. We formally assessed if the age group modified the association between long-term average alcohol consumption and EAAs in participants of three age groups. For all analyses, lab-specific SVs were imputed to account for unobserved confounding effects by the package “sva” [61]. Lab-specific EAA residuals were obtained by regressing EAA metrics on SVs and were used as outcome variables. In primary analyses, we used linear regression models to quantify the association of long-term average alcohol consumption (i.e., total alcohol consumption and each type of alcoholic beverage consumption) with EAA residuals in each of the three age groups. All regression models adjusted for covariates including sex, PAI, education level, BMI, smoking pack-year, chronological age, and lab index variable. To investigate whether the association between alcohol consumption and EAA residuals was linear, we compared the mean differences in EAA residual variable between drinking categories, i.e., non-drinkers, at-risk drinkers, and heavy drinkers, versus light drinkers as the reference group (the drinking group with the largest number of participants), adjusting for the same set of covariates.

We conducted several secondary analyses. First, we investigated the associations of cross-sectional alcohol consumption and recent binge drinking with EAA residuals (outcome variables) in each of the three age groups. We conducted association analyses of long-term average and cross-sectional alcohol consumption with EAA residuals in the pooled sample (i.e., pooling participants across the three age groups), adjusting for the same set of covariates. We conducted secondary association analyses, with additional adjustments for imputed white blood cell compositions (including CD8+ T cells, CD4+ T cells, natural killer cells, B cells, monocytes, and granulocytes) based on the Houseman’s method [10, 62]. In addition, we conducted secondary association analyses of long-term average alcohol consumption with EAAs adjusting for long-term average values of covariates (i.e., average BMI, PAI, and smoke pack-year). We also evaluated whether sex and smoking status modified the association between long-term average total alcohol consumption and EAA residuals in each age group and the pooled sample, by including an interaction term (i.e., sex × alcohol or smoking × alcohol) in the regression model. To account for multiple testing, we used p < 0.01 (= 0.05/5) for significance; here we considered accounting for two primary and three secondary EAA measures.

Previous studies have shown that alcohol consumption increases the odds of hypertension [21, 56, 63]. We, therefore, examined the association of long-term average alcohol consumption and residuals of GAA and PAA with hypertension utilizing the mediation analysis “PROC CAUSALMED” in SAS. All statistical analyses were performed with R software (version 4.1.1) and SAS software (version 9.3).