Alcohol and aging: Next‐generation epigenetic clocks predict biological age acceleration in individuals with alcohol use disorder

DNA methylation levels from whole blood samples were assessed using an Infinium MethylationEPIC BeadChip microarray (manufactured by Illumina Inc., San Diego, CA, USA). The wateRmelon package in R was used to process the raw data. Cross‐reactive probes and probes that failed quality assessment were removed, and then a scale‐based correction was applied to Illumina type I probes relative to type II probes. A quantile normalization approach was executed for standardizing methylated and unmethylated allele intensities, and then β‐values were calculated using the ratio of intensities. Estimates of blood cell proportions of six cell types (monocytes, granulocytes, natural killer cells, B cells, CD4+ T cells, and CD8+ T cells) were generated using the Houseman estimation method (Houseman et al., 2016). GrimAge V1 and GrimAge V2 were calculated using Horvath's epigenetic age calculator software (https://dnamage.clockfoundation.org/↗).

The causal epigenetic clock models (CausAge, AdaptAge, and DamAge), developed with an epigenome‐wide Mendelian randomization (EWMR) approach and reliance on age‐associated differential methylation, offer distinguishments of helpful changes from harmful ones (Ying et al., 2024). CausAge was estimated through 586 CpGs selected by a causality‐enriched elastic net model. DamAge contains only damaging CpG sites (n = 1090) and estimates higher mortality risk in individuals expected to obtain more harmful changes during aging. Meanwhile, AdaptAge consists of only adaptive CpGs (n = 1000) and predicts lower mortality risk in individuals expected to accumulate more protective changes during aging (Ying et al., 2024). In other words, a higher AdaptAge value represents the aggregation of more beneficial changes and presumably enhanced longevity, while a greater DamAge value represents the accrual of more deleterious changes and presumably degraded longevity (Ying et al., 2024). In our sample, we utilized 396 CausAge CpGs, 1089 DamAge CpGs, and 998 AdaptAge CpGs. There were some discrepancies in the CpGs available to implement due to the clocks being trained on the Illumina 450K chip platform. Nevertheless, by integrating the weight of each available CpG, we computed causality‐enriched epigenetic ages as represented by CausAge, DamAge, and AdaptAge.

To investigate differences in age‐adjusted GrimAge V1, GrimAgeV2, CausAge, DamAge, and AdaptAge between AUD and HC cohorts, a linear model was utilized with adjustment for the covariates sex, race, body mass index (BMI), smoking status, and estimated proportions of five blood cell types (monocytes, natural killer cells, B cells, CD4+ T cells, and CD8+ T cells). Granulocytes were not included in the model following the execution of variance inflation factor analysis (Jung et al., 2023). The age‐adjusted epigenetic clocks (defined by epigenetic age acceleration, EAA) were calculated by taking the residual resulting from a linear regression of the DNAm age on chronological age. Furthermore, we additionally controlled for AUD diagnosis to test associations with alcohol consumption behaviors, LFE biomarkers (i.e., GGT, AST, ALT), and alternate clinical biomarkers (i.e., A1C, CRP). We first conducted our analyses with the total sample (N = 615) and then divided the total sample into a young cohort (<40 years old) and an old cohort (≥40 years old) to investigate differences across generations. We tested to see whether there was an interaction effect between AUD diagnosis (AUD vs. HC) and age group (young vs. old) in the linear model; subsequent analyses within each age subgroup were carried out once a significant interaction effect between AUD diagnosis and age group was observed. All alcohol‐related clinical variables were standardized (mean = 0, SD = 1) to generate interpretable results across a wide spectrum of scales. There was no evidence to suggest a non‐Gaussian distribution of the residuals from the models (Shapiro–Wilk test, p > 0.05).

There may be potential systematic differences between the AUD and HC groups in terms of several confounding covariates (e.g., smoking status). To control for bias in the comparison of these two groups, propensity scores for retrospective matching analysis were calculated (Chen et al., 2022). With the total sample (372 AUD and 243 HC participants), a propensity score was generated for each participant using a logistic regression model with the same covariates (age, sex, race, BMI, smoking status, five blood cell compositions). Based on these propensity scores, the optimal approach with a greedy matching algorithm matched 243 AUD patients and 243 HC individuals (1:1 matching), which left 129 AUD patients unmatched. The balance in the matched group was assessed before and after matching. Using the matched sample, the differences in epigenetic age acceleration (EAA) between the two groups as well as the associations with the alcohol consumption‐based measures were tested after adjusting for propensity scores.

Finally, a formal statistical test was carried out to investigate whether there is a significant difference in the effect size of age‐adjusted GrimAge V1 and V2, in which the models were tested with the difference between age‐adjusted DNAm age as the dependent variable and the alcohol use measure of interest as the independent variable with adjustment for the covariates mentioned above (the theoretical formula is available in Appendix). All statistical significance was defined by a Bonferroni‐corrected significance level (0.01 = 0.05/5, testing five epigenetic clocks). Statistical analyses were performed in R version 4.2. S1

We performed analyses modeling the relationship between alcohol consumption‐related metrics and EAA calculated from both versions of GrimAge, CausAge, DamAge, and AdaptAge. Comparing the two GrimAge clocks, we first found AUD status to be significantly associated with GrimAge V1 (β = 2.01, p = 3.70E‐08) and GrimAge V2 (β = 2.43, p = 1.50E‐09) (Table 2, Figure 1). In other words, GrimAge V1 was 2.01 years advanced, and GrimAge V2 was 2.43 years advanced in the AUD group as compared to the HC group. However, the causality‐enriched clocks (CausAge, DamAge, and AdaptAge) did not show any statistically significant differences in EAA between AUD patients and HC individuals (Table 2, Figure 1). Further statistical testing suggested a lack of statistically significant difference in the effect sizes of AUD between GrimAge V1 and V2 (Table S1).

We also found the following alcohol consumption behavioral variables to be significantly associated with both versions of GrimAge: total number of drinks in the past 90 days, average number of drinks per day, and number of heavy drinking days (ps < 0.0001) (Table 2). In particular, the total number of drinks in the past 90 days and average drinks per day were more significantly associated with GrimAge V2 than GrimAge V1 given their effect sizes and p‐values (Table 2); however, the association with the number of heavy drinking days was slightly more significant for GrimAge V1 (β = 1.07, p = 1.99E‐06) than GrimAge V2 (β = 1.17, p = 2.30E‐06) (Table 2). None of the causality‐enriched epigenetic clocks met the significance threshold for the alcohol behavioral variables (ps > 0.01) (Table 2). Liver function enzyme levels, as characterized by GGT, AST, and ALT, were identified to be significantly associated with GrimAge V2 (ps < 0.001), while GrimAge V1 had significant associations with GGT and AST (ps < 0.001) (Table 2). Overall, the effect sizes with their p‐values indicate that all three of these variables (GGT, AST, and ALT) were more significantly associated with GrimAge V2 than GrimAge V1 (Table 2). Meanwhile, DamAge had significant associations with the liver enzyme variables GGT and AST (ps < 0.001), while CausAge had a significant association with only GGT (p < 0.01) (Table 2).

Looking at the alternative clinical variables, both GrimAge clocks had significant associations with CRP, but GrimAge V2 (β = 0.77, p = 2.91E‐05) had a slightly stronger association than GrimAge V1 (β = 0.64, p = 1.18E‐04) (Table 2). For A1C, neither GrimAge V1 nor V2 met the significance threshold. The subsequent analyses of comparison in effect sizes across various alcohol consumption measures between GrimAge V1 and V2 showed no statistically significant differences (Table S1). Nonetheless, GrimAge V2 tends to slightly outperform GrimAge V1 in modeling EAA in the context of AUD. CausAge, DamAge, nor AdaptAge met the significance threshold for the alternative clinical variables (ps ≥ 0.01) (Table 2).

We observed an interaction effect on age‐adjusted GrimAge clocks between AUD diagnosis (AUD vs. HC) and age group (young vs. old) and thus conducted association analyses within subgroups of age. In the young (<40 years old) subgroup, GrimAge V2 supported a slightly more significant association with AUD (β = 1.99, p = 2.67E‐04) than GrimAge V1 (β = 1.59, p = 2.06E‐03) (Table 3) even though the statistical significance in an effect size difference between them was not supported (Table S1). Regarding alcohol consumption variables, GrimAge V2 maintained a more significant association with total drinks, but GrimAge V1 retained more significant associations with average drinks per day and number of heavy drinking days (ps < 0.01) (Table 3). Amongst the causality‐enriched clocks, DamAge established a significant association with total number of drinks (β = 1.11, p = 5.41E‐03) (Table 3). With respect to the liver function enzymes, only GrimAge V2 had a significant association with GGT (p < 0.01) (Table 3). In addition, DamAge exhibited significant associations with GGT and AST (ps < 0.01). GrimAge V2 also had a stronger association with CRP (β = 0.77, p < 0.01) in comparison with GrimAge V1 (β = 0.66, p = 0.01) (Table 3).

In the old (≥40 years old) subgroup, GrimAge V2 also maintained a bit more of a significant association with AUD (β = 3.13, p = 4.34E‐08) compared with GrimAge V1 (β = 2.73, p = 1.13E‐07) (Table 4). Associations with total drinks, average number of drinks per day, and number of heavy drinking days were more significant for GrimAge V2 (ps < 0.01) (Table 4). Significant associations with GGT, AST, ALT, and CRP reflect this pattern as well (Table 4). Altogether, GrimAge V2 was typically slightly more associated with alcohol consumption behaviors, LFEs, and CRP compared with GrimAge V1 (Table 4). For the causality‐enriched clocks, none of the causality‐enriched epigenetic clocks showed any significant associations with LFEs or the alternative clinical variables except AdaptAge with A1C (β = 1.44, p < 0.01) (Table 4).

In expanded analyses evaluating sex‐and race‐specific subgroups, we observed similar patterns of associations between alcohol consumption‐related metrics and EAA (Tables), with the exception that African American participants typically showed weaker associations across alcohol consumption‐related phenotypes. S2.1–S3.2

The propensity score analysis with the optimal matching approach selected 243 AUD patients and 243 HC individuals (1:1 matching), which left 129 AUD patients unmatched (Chen et al., 2022). The distribution of propensity scores in the two groups before and after matching suggested that the matched group was well‐balanced. Compared with the results of the total sample (Table 2), the results of the matched sample showed similar patterns of association with AUD status (with smaller effect sizes) and the alternate clinical biomarkers (i.e., CRP with smaller effect sizes). However, the results revealed dissimilar patterns of association with alcohol consumption behaviors, differing in magnitude (i.e., stronger associations of GrimAge V1 with total number of drinks, number of drinking days, average number of drinks per day) and LFEs (significant association of GrimAge V2 with GGT only) (Table S4). In a similar manner as the total sample analysis, the causality‐enriched clocks generally did not show any statistically significant associations with alcohol‐related metrics, with the only exception being DamAge and CausAge with only GGT (Table S4).

In this study, we formulated more comprehensive comparisons between newly reported epigenetic clocks and well‐characterized alcohol‐related phenotypes. Previously, our studies compared associations of EAA with AUD and its associated endophenotypes, focusing on the first and second‐generation epigenetic clocks (Jung et al., 2023). Interestingly, only EAA from PhenoAge (defined as a methylation‐based prediction of composite phenotypic age) and GrimAge V1 were associated with severe alcohol‐related phenotypes, even though PhenoAge and GrimAge V1 lack overlap in the types of biomarkers used (Levine et al., 2018; Lu et al., 2019). The two clocks were shown to better differentiate morbidity and mortality risk among individuals of the same chronological age than the first‐generation epigenetic clocks. Furthermore, in contrast with GrimAge V1, both PhenoAge and GrimAge V2 integrate CRP levels; however, PhenoAge employs serum glucose levels while GrimAge V2 applies hemoglobin A1C levels (Levine et al., 2018; Lu et al., 2019, 2022).

In terms of predictive power, GrimAge V1 outperforms PhenoAge in regard to its extent of associations with both mortality and morbidity parameters, including time‐to‐death, time‐to‐cancer, time‐to‐CHD (coronary heart disease), age at menopause, and various comorbidities (Lu et al., 2019). Moreover, GrimAge V2 surpasses GrimAge V1 in predicting time‐to‐death resulting from all‐cause mortality, time‐to‐CHD, time‐to‐CHF (congestive heart failure), type 2 diabetes, hypertension, disease‐free status, physical functioning level, and again various comorbidities (amongst others) (Lu et al., 2022). Although a statistical test did not confirm a significant difference in effect sizes between GrimAge V1 and V2, we observed that GrimAge V2 generally outperformed GrimAge V1 across most alcohol behaviors as well as LFT measures. This was evident in both the magnitude of the effect sizes and their p‐values. Our findings align with previous research indicating that GrimAge V2 shows stronger associations with age‐related phenotypes compared with GrimAge V1. This improvement may be attributed to the incorporation of DNAm logCRP and DNAm logA1C, which are linked to numerous age‐related conditions (Lu et al., 2022) as well as alcohol‐associated behaviors.

Out of the three fourth‐generation clocks, DamAge features the higher degree of significance in positive association with mortality, with comparable efficacy to PhenoAge (Ying et al., 2024). Meanwhile, AdaptAge maintains a significant negative association with mortality (Ying et al., 2024). Looking at some contexts of clinical validation, with pluripotent stem cell reprogramming, DamAge and AdaptAge sustain clinically substantial associations, with DamAge showcasing the most substantial association out of the previously ratified clock models (Ying et al., 2024). Despite both DamAge and AdaptAge demonstrating significant associations (DamAge—strong positive associations, AdaptAge—strong negative associations) with several age‐related diseases such as hypertension, cancer, and Hutchinson–Gilford progeria (HGP), our findings showed that these clocks were not associated with alcohol‐induced phenotypes and AUD.

By additionally incorporating DNAm‐based estimators of high sensitivity‐CRP, a widely used biomarker of inflammation, and hemoglobin A1C, a common biomarker for assessing blood glucose levels, GrimAge V2 showed a more comprehensive and sensitive prediction of human mortality and morbidity than GrimAge V1. To start, both plasma proteins are involved in alcohol metabolism pathways (Bell et al., 2000; Lindtner et al., 2013; Wang et al., 2010). For example, AUD corresponds to short‐and long‐term alcohol intake that disrupts the management of glucose and consequently A1C levels (Wiss, 2019). However, the dynamics between alcohol and blood sugar concentration appear to fluctuate depending on drinking intensity and duration (Athyros et al., 2007; Facchini et al., 1994). Similar to A1C, CRP indicates a relationship between chronic alcohol use and inflammatory pathways and also shows associations with age‐related health conditions (Rosen et al., 2018; State, 2021). Our studies may suggest incorporation of the A1C and CRP biomarkers in future epigenetic clock models to potentially provide a better understanding of the effects of alcohol on advanced age acceleration (Bell et al., 2000; Caruana et al., 2015). However, further studies are necessary to see whether A1C and CRP directly contribute to the mechanisms of alcohol‐related metabolism.

More broadly, the GrimAge models outperform the fourth‐generation clocks to a considerable degree. Given the differences in construction and CpG sites selected, the GrimAge foundation appears to incorporate CpGs with more predictive power. One potential candidate would be cg05575921, which is found on the aryl‐hydrocarbon receptor (AHRR) gene and has been linked to smoking (Bravo‐Gutierrez et al., 2021). Smoking pack‐years was included as a clinical biomarker in both GrimAge models, to no surprise given the well‐established associations of smoking with mortality (Bravo‐Gutierrez et al., 2021). However, smoking has also been found to interact with alcohol consumption (Verplaetse & McKee, 2017). Studies have shown that alcohol can elevate the desire to smoke at both smaller and larger quantities; meanwhile, nicotine in tobacco can dilute the soporific effects of alcohol, thus suggesting a complex dynamic between the two substances that may be better encapsulated by the GrimAge model (Verplaetse & McKee, 2017). Furthermore, the GrimAge clocks were designed to predict time to death (due to all‐cause mortality) using DNAm‐based age‐related plasma protein biomarkers and DNAm‐smoking packs per year, while the causality‐enriched epigenetic clocks were developed to predict chronological age using CpGs identified by an epigenome‐wide Mendelian randomization (EWMR) approach and reliance on age‐associated differential methylation. Many studies have shown that the first‐generation clocks predicting chronological age tend to underperform in comparison to the second‐generation clocks in regard to associations with various health conditions (Levine et al., 2018; Robinson et al., 2020).

The fourth‐generation clocks (CausAge, DamAge, and AdaptAge) were founded upon CpG sites deemed causal to aging as represented by lifespan, extreme longevity, health span, frailty index, self‐rated health, aging‐GIP1, socioeconomic traits‐adjusted aging‐GIP1, and healthy aging (Ying et al., 2024). To clarify, GIP1 (genetically independent phenotype, principal component 1) is derived from health span, mother's lifespan, father's lifespan, remarkable longevity, index of frailty, and self‐assessed health (Timmers et al., 2022). Despite the secondary performance of these models in comparison with the GrimAge clocks, they may be useful in whittling down the list of significant candidates regarding health outcome‐associated epigenetic influences on biological age.

Our study has several strengths, including a well‐characterized sample and enrichment of alcohol‐related clinical phenotypes. Our sample consists of individuals from European and African American ancestries. In the AUD group, 47.58% identified as African Americans and 51.34% as European Americans (Table 1). The over‐representation of African Americans in our sample extends our findings' generalizability to ancestral groups that are not often represented in research studies. In addition, the enrichment of alcohol‐related clinical phenotypes provides opportunities for conducting various endophenotypic analyses, including disease severity analyses.

While our investigation had adequate power to detect the association of EAA with alcohol consumption‐related behaviors, we cannot confirm causal relationships between these variables. Specifically, we cannot definitively demonstrate whether AUD is a risk factor for accelerated epigenetic aging or a consequence of chronic alcohol use using the epigenetic clock models. Our study involves a cross‐sectional analysis, where several variables are considered at a singular point in time. As such, the study is unable to analyze two separate time points for the same individual, so there is a lack of longitudinal character that ultimately prevents any solidified conclusions in regard to causality prediction by the epigenetic clock models (Caruana et al., 2015). To address this limitation of our cross‐sectional study, future studies could incorporate the collection of methylation data at multiple time points to investigate the effect of long‐term alcohol use on aging processes. In addition, there may be potential bias due to the self‐reporting nature of the questionnaires filled out by participants in the studies covered by the data set (Jung et al., 2023). Furthermore, our study was limited in collecting data on other environmental/lifestyle factors (e.g., exercise and diet) that may influence DNA methylation levels, and therefore, these factors were not included in our analyses. Nevertheless, we did account for age, sex, race, BMI, and smoking status, which all have been shown to be significantly associated with epigenetic aging (Hannum et al., 2013; Horvath, 2013; Levine et al., 2018; Lu et al., 2019, 2022). Although our study examined methylation in whole blood only and controlled for blood cell type composition, expansion to tissues and cell methylation samples may be necessary to validate alcohol‐related aging. Furthermore, the differences in chip compatibilities with the causality‐enriched clocks, where CausAge was missing a considerable number of probes, make comparison of performance a challenge. As such, the CausAge associations may have been much more skewed (Ying et al., 2024).

The data that support the findings of this study are available from the corresponding author upon reasonable request.