Validation and characterization of a DNA methylation alcohol biomarker across the life course

Recently, an alcohol predictor was developed using DNA methylation at 144 CpG sites (DNAm-Alc) as a biomarker for improved clinical or epidemiologic assessment of alcohol-related ill health. We validate the performance and characterize the drivers of this DNAm-Alc for the first time in independent populations. In N=1,049 parents from the Avon Longitudinal Study of Parents and Children (ALSPAC) Accessible Resource for Integrated Epigenomic Studies (ARIES) at midlife, we found DNAm-Alc explained 7.6% of the variation in alcohol intake, roughly half of what had been reported previously, and interestingly explained a larger 9.8% of AUDIT score, a scale of alcohol use disorder. Explanatory capacity in participants from the offspring generation of ARIES measured during adolescence was much lower. However, DNAm-Alc explained 14.3% of the variation in replication using the Head and Neck 5000 (HN5000) clinical cohort that had higher average alcohol consumption. To investigate whether this relationship was being driven by genetic and/or earlier environment confounding we examined how earlier vs. concurrent DNAm-Alc measures predicted AUDIT scores. In both ARIES parental and offspring generations, we observed associations between AUDIT and concurrent, but not earlier DNAm-Alc, suggesting independence from genetic and stable environmental contributions. The stronger relationship between DNAm-Alcs and AUDIT in parents at midlife compared to adolescents despite similar levels of consumption suggests that DNAm-Alc likely reflects long-term patterns of alcohol abuse. Such biomarkers may have potential applications for biomonitoring and risk prediction, especially in cases where reporting bias is a concern.

Introduction tissues can serve as sensitive and specific indicators of exposure. 10,11 Further, DNA methylation signals are stable over highly variable timescales, with some having been observed to persist well beyond the half-lives of existing targets for alcohol biomarkers. 12 For example, associations observed in relation to prenatal smoking exposure have been seen to endure into childhood 13 and some effect of the exposure appears detectable into adulthood. 14 Liu et al. developed four new blood DNA methylation-based alcohol models (DNAm-Alcs) as biomarkers of alcohol intake, using a large population sample (the CHARGE consortium 15 ). Of these four models, the one comprising the highest number of CpG sites (144) consistently showed the best performance. 6 The well powered analyses, good range of alcohol intake and promising proportion of the variance explained make this an interesting candidate biomarker. However, the reported performance of this model is likely inflated due to over-fitting. 16 Another limitation is the use of cross-sectional data makes it difficult to determine the direction of the association or to rule out genetic confounding, which could be an issue since many CpGs having been shown to be under genetic control 17 and alcohol use disorders are heritable. 18 Here and for the first time, we aim to assess the out-of-sample performance of the DNAm-Alcs by using data at multiple time points from an independent population, the subset of participants from the Avon Longitudinal Study of Parents and Children (ALSPAC) included in the Accessible Resource for Integrated Epigenomic Studies (ARIES), and further replication using the Head and Neck 5000 (HN5000) clinical cohort. Repeated measures of alcohol use and blood DNA methylation in two generations of the ARIES study and at different points in the life course, including DNA methylation profiles collected prior to own alcohol exposure, were used to address our aims. Our first aim was to test the performance of the DNAm-Alcs in ARIES and HN5000. Second, we aimed to further investigate the performance of DNAm-Alcs in ARIES at predicting measures of alcohol use disorders (AUDIT scores) in different groups of individuals (males and females at midlife and adolescence), which is designed to estimate the extent to which an individual engages in harmful patterns of alcohol consumption. 19 Lastly, we sought to address additional details regarding the nature of the association between the DNAm-Alcs and alcohol use, such as duration of alcohol use and the potential for genetic confounding, by using measurements of alcohol and DNA methylation at different times in the life course.

ARIES study
The Avon Longitudinal Study of Parents and Children (ALSPAC) is a population-based birth cohort drawn from the South West of England. Pregnant women resident in Avon, UK with expected dates of delivery 1st April 1991 to 31st December 1992 were invited to take part in the study. The initial number of pregnancies enrolled is 14,541 (for these at least one questionnaire has been returned or a "Children in Focus" clinic had been attended by 19/07/99). Of these initial pregnancies, there was a total of 14,676 foetuses, resulting in 14,062 live births and 13,988 children who were alive at 1 year of age. Detailed information on the mothers and their offspring has since been regularly collected. 20,21 The study website contains details of all the data that is available through a fully searchable data dictionary and variable search tool (http://www.bristol.ac.uk/alspac/researchers/our-data/). Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees (http://www.bristol.ac.uk/alspac/researchers/researchethics/). Informed consent for the use of data collected via questionnaires and clinics was obtained from participants following the recommendations of the ALSPAC Ethics and Law Committee at the time. Consent for biological samples has been collected in accordance with the Human Tissue Act (2004). Blood from 1 018 mother-child pairs (children at three time points and their mothers at two time points) were selected for analysis as part of the Accessible Resource for Integrative Epigenomic Studies (ARIES, http://www.ariesepigenomics.org.uk/). 22 ARIES consists of DNA methylation profiles obtained from 1 018 mother-child pairs measured at five time points (three time points for children: birth, childhood and adolescence; and two for mothers: during pregnancy and at midlife), and on approximately 600 fathers at midlife. In this analysis, we included participants from the ARIES parental generation at midlife (N = 1 049, mean age = 50.2 ± 5.4 SD) and the offspring generation at adolescence (N = 626, mean age = 17.4 ± 0.9 SD) with DNA methylation, alcohol, and covariate information available ( Figure 1). The subset of these participants with DNA methylation available at an earlier time point, during pregnancy in the parents (N = 518, mean age = 29.1 ± 4.2 SD) and in cord blood at birth (N = 438) and childhood (N = 601, mean age = 7.4 ± 0.1 SD) in the offspring, were included in some analyses.

HN5000 replication study
The Head and Neck 5000 (HN5000) clinical cohort study, a clinical cohort study of individuals with head and neck cancer, was used as an additional independent replication population. Full details of the study methods and overall population are described in detail elsewhere. 23,24 Briefly, between April 2011 and December 2014, 5511 individuals with HNC were recruited from 76 centres across the UK. Individuals were recruited before they started treatment unless their treatment was their diagnostic procedure. Full ethical approval was granted by the South West Frenchay Regional Ethics Committee (ref: 10/H0107/57).
At baseline, participants were asked to complete a series of questionnaires and information on diagnosis, treatment and co-morbidity was recorded on a short data capture form. Diagnoses were coded using the International Classification of Diseases (ICD) version 10 25 and clinical staging of the tumour was derived based on the American Head and Neck Society TNM staging. 26 Biological samples (blood, saliva and tissue) were collected from all consenting participants, and DNA methylation was measured for 448 individuals with an ICD-10 coding of oropharynx cancer and complete baseline questionnaire and data capture information. 23 Of those, participants of European ancestry with DNA methylation, alcohol, and all covariate information available were included in this analysis (N=281, mean age = 58.4 ± 9.9 SD)

Alcohol intake and disorder
Alcohol intake and AUDIT score 27 were available at multiple time points for the ALSPAC participants. To validate the performance of the DNAm-Alcs, we utilised measures of alcohol intake at midlife and adolescence taken as close as possible to the DNAm measurement, to replicate the cross-sectional design used in the derivation and original evaluation of DNAm-Alcs. For ARIES parents at midlife and the offspring generation at adolescence, alcohol drinking frequency was assessed by response to the question, "How often do you have a drink containing alcohol", with possible responses including "Never"; "Monthly or less"; "2 to 4 times a month"; "2 to 4 times a week"; "4 or more times a week". "Never" respondents were considered non-drinkers and were excluded from all analyses apart from examination of their influence in sensitivity analysis.
Participants were also asked to report the quantity of their typical alcohol consumption where "one drink referred to ½pint of beer/cider, a small (125ml) glass of wine or a single (25ml) measure of spirit", each of which is roughly equivalent to one UK alcohol unit (8 grams of ethanol). Alcohol intake was then calculated by multiplying participant drinking frequency by the typical quantity consumed and converting from UK units/week to g/day. In all statistical models, alcohol intake was log-transformed (log (g/day + 1)) to coerce its distribution to approximate normality. For categorical analyses, drinkers were grouped according to the following thresholds of alcohol intake: 'non-drinkers' reported no alcohol consumption (intake g/d = 0), 'light drinkers' consumed 0 < g/day ⩽14 for women and 0 < g/day ⩽28 in men, 'heavy drinkers' ⩾28 g/day in men and ⩾42 g/day in men. These categorical thresholds, which are the same as those used by Liu et al., were selected in order to maximize comparability between our results.
Alcohol use disorder identification test (AUDIT) scores were taken from the same time points as the measures of alcohol intake, cross-sectionally to DNAm at midlife in the ARIES parental generation and in adolescence in the offspring. AUDIT includes responses to 10 questions that evaluate participant alcohol misuse in the past year, producing scores that range from 0 to 40. Self-reported non-drinkers are included in this scale, receiving a score of 0.
In HN5000, detailed information on alcohol history was obtained at baseline via the self-reported questionnaire. Participants were asked about their current drinking status and their use of alcohol prior to receiving their HNC diagnosis. Respondents were asked to report their average weekly alcohol consumption of a range of beverage types (wine, spirits, and beer/larger/cider) before they were diagnosed with cancer. From these measures, average intake of alcohol consumption in units per week was derived, which we subsequently converted to log (g/day + 1).

DNA methylation measurements
In ARIES, cord blood and peripheral blood samples (whole blood, buffy coats or blood spots) were collected according to standard procedures. DNA  was performed to minimize non-biological variation in probes.
In HN5000, blood samples were processed and frozen upon receipt and stored at -80°C. Genome-wide DNA methylation data was generated in participants using Infinium MethylationEPIC BeadChip (EPIC array) (Illumina, USA) at the Bristol Bioresource Laboratories. Following extraction, DNA was bisulphite-converted using the Zymo EZ DNA MethylationTM kit (Zymo, Irvine, CA, USA) ang genome-wide methylation status of over 850,000 CpG sites was measured using the EPIC array according to the manufacturer protocol using an Illumina iScan. All data pre-processing and normalization was again performed using the meffil package, 29 which excluded 5 of the total samples for not meeting the QC criteria.

Covariates
For the ARIES parental generation at midlife, offspring generation in both childhood and adolescence, and HN5000 participants, DNA methylation levels were adjusted for age, sex, body mass index (BMI) and Houseman estimate white blood cell (WBC) counts by residualizing on these factors by ordinary least squares (OLS) regression prior to calculation of DNAm-Alcs. 31,32 Levels from pregnancy were not adjusted for sex, as only female participants had DNA methylation available, but were adjusted for age, WBC counts and prepregnancy BMI. The adjustment set for DNA methylation at birth included gestational age (weeks), child sex, birthweight (g), and WBC counts calculated using a cord blood reference set. 33 Indicator variables for self-reported current smoking at midlife for the ARIES parents and for having ever/never smoked a cigarette in adolescence for the offspring were utilised in sensitivity analyses.

DNA methylation alcohol score
We computed DNA methylation-derived alcohol scores (DNAm-Alcs) using coefficients made available from the lasso models estimated by Liu and colleagues, 6 predicting the relationship between DNA methylation and alcohol intake. Four DNAm-Alcs were examined using different sets of coefficients that varied according to the lasso tuning parameters DNA methylation levels were residualized from the linear effects of the covariates described above prior to computing DNAm-Alc. Software for producing all four DNAm-Alc estimates is freely available using the dnamalci R package (https://github.com/yousefi138/dnamalci).

Statistical analysis
Overall performance of DNAm-Alc scores in predicting continuous measures, including alcohol intake and AUDIT score, was assessed by R 2 or the percent of the variance in the outcome variable explained. For binary outcome measures, receiver operating characteristic (ROC) curves were generated and area under the curve (AUC) was calculated using the pROC package 34 to evaluate predictive performance. Linear relationships between DNAm-Alcs and AUDIT scores from different time points were evaluated using OLS regression. Models involving differing numbers of explanatory parameters were compared using the OLS-derived adjusted R 2 in order to account for the discrepancy in model degrees of freedom.
We further performed sensitivity analyses to check the robustness of our results to the impact of smoking and the inclusion of non-drinkers. To this end, we ran cross-sectional models using DNAm values that had been residualised using concurrent smoking as an additional covariate, and excluding participants identified as self-reported non-drinkers.
As an additional comparison, we re-fit linear models with the same sets of CpGs used in each DNAm-Alc, as well as age, sex, and BMI as explanatory variables. This provided an inflated assessment of prediction due to over-fitting, as has been described in detail by Hattab, Clark, and van den Oord, 16 but allowed direct comparison to the performance of the models originally implemented by Liu et al. In these re-estimated models, adjusted R 2 from the OLS model was again used to compare models with different numbers of covariates.
All statistical analysis was performed in R version 3.3.1. 35

Validation of DNAm-Alcs
The four DNAm-Alcs evaluated, using 5-144 CpGs, explained between 3.9 and 7.6% of the variation in alcohol intake in our sample of N=1 049 ARIES participants at midlife, using DNA methylation levels independent of age, sex, and BMI (   Figure   2), DNAm-Alcs explained a much lower percentage of the variation in adolescent alcohol intake, with a maximum of 0.8% for the 144 CpG DNAm-Alc. Again, this result did not change substantially when accounting for smoking or when excluding non-drinkers.
In each of these analyses, prediction performance was positively correlated with the number of CpG sites included in the model. That is, the 5 CpG DNAm-Alc consistently explained the least variation, the 144 CpG explained the most, with the 23 CpG and 78 CpG models ordered in between.
Replication in an independent UK-based population of adults with oropharyngeal cancer, N= 281 participants from HN5000, found each of the four DNAm-Alcs explained higher levels of variation in alcohol intake than observed in ARIES, with adjusted R 2 ranging from 8.8% to 14.3% (Supplementary Table 3). While this population was similar in age to the ARIES participants at midlife (mean age = 58.4 ± 9.9), their levels of alcohol consumption were considerably higher (mean intake = 28.0 ± 31.3g/d).
Similar trends but with consistently stronger correlations were observed in ARIES parents at midlife when the coefficients of the CpG sites included in the DNAm-Alcs were reestimated in ARIES, to allow direct comparison with performances based on re-estimated models reported by Liu et al. (Supplementary Table 4). By this approach, adjusted R 2 ranged from 8.8 to 17.3% for each of four CpG covariate sets, including the independent explanatory effects of age, sex, and BMI that jointly contributed an adjusted R 2 of 4.8%.
Assessment of the ability of DNAm-Alcs to distinguish heavy drinkers from non-

DNAm-Alc and AUDIT score
At midlife, DNAm-Alcs explained from 4.9 to 8.8% of the variation in participant AUDIT scores, a measure of alcohol use disorder (Table 1). In fact, every DNAm-Alc with a particular number of CpGs explained more of the variation in AUDIT than for alcohol intake. On average, DNAm-Alcs explained 1.3% more variation of AUDIT than alcohol intake in midlife. These DNAm-Alcs explained a comparable but slightly greater amount of variation in AUDIT when performing additional adjustment for smoking or following the exclusion of non-drinkers, from 5.0 to 9.0% and 5.5 to 9.8% respectively (Supplementary Tables 1 and 2).
In adolescents, DNAm-Alcs explained substantially less variation in AUDIT than at midlife, reaching only a maximum of 1.2% for the 144 CpG biomarker. Despite this decrease in total variance explained among adolescents, the DNAm-Alcs continued to correlate slightly better with AUDIT than with alcohol intake, explaining 0.05% more on average. Similar relationships with AUDIT were once again observed when adjusting for smoking and excluding non-drinkers.
As with alcohol intake, we observed a trend of consistent improvement in prediction performance for AUDIT with additional CpG sites: the 144 CpG DNAm-Alc score always performing the best and the 5 CpG DNAm-Alc performing worst.
The ability of the DNAm-Alcs to distinguish high (>15) versus low AUDIT score categories in ROC analysis produced AUCs that ranged from 0.67 with the 5 CpG DNAm-Alc to 0.80 for the 78 CpG score in ARIES participants at midlife (Figure 3). These AUC values were uniformly higher than those observed for all DNAm-Alcs at predicting heavy drinkers from light or non-drinkers using alcohol intake derived cut-offs. In ARIES adolescents, DNAm-Alc prediction of high versus low AUDIT individuals was much more comparable with performance observed using intake-based categories, with AUCs ranging from 0.51 to 0.56.

Earlier vs. cross-sectional DNAm-Alc and AUDIT
To determine whether the relationship observed between DNAm-Alc and AUDIT using cross-sectional data could be explained by genetic confounding, we compared the predictive performance of DNAm-Alc based on methylation of blood drawn earlier in life, at a period of expected low or no alcohol use, to the predictive performance of DNAm-Alc based on methylation of blood drawn later in the same participants, at the same time as completion of the AUDIT questionnaire. If we show that DNAm-Alc from blood drawn earlier in life is not associated with later AUDIT, then we've ruled out both genetic and earlier environment confounding of the cross-sectional DNAm-Alc -AUDIT relationship. A schematic flow-chart describing this analysis is provided in Figure 1.
In the ARIES mothers, we compared the effects of the 144 CpG DNAm-Alc measured during pregnancy to those of DNAm-Alc measured cross-sectionally to AUDIT, 21.1 years later on average during midlife (Table 2). For the N=518 participants with DNAm-Alc at both timepoints and AUDIT available, we found no evidence for a relationship between DNAm-Alc during pregnancy and AUDIT at midlife, both when considered as a single independent variable (β = 0.15, p = 0.56) and when adjusting for DNAm-Alc at midlife (β = 0.14, p = 0.57).
For ARIES children, we were able to compare the cross-sectional DNAm-Alc/AUDIT relationship in adolescence to DNAm-Alc measured at two earlier no/low drinking time points in the same individuals: in childhood (n=601) and at birth (n=438). AUDIT in adolescence appeared to be independent of DNAm-Alc in childhood (p = 0.88), including after adjusting for adolescent DNAm-Alc (p = 0.44). However, the cross-sectional relationship between DNAm-Alc and AUDIT in adolescence, while small in total magnitude (adjusted R 2 = 1.2), persisted with similar effect size when adjusting for childhood DNAm-Alc (β = 0.94 compared to β = 0.99). While DNAm-Alc at birth showed some suggestive evidence of a relationship with AUDIT in adolescence when considered alone (β =0.70, p = 0.06), this effect attenuated when adjusted for cross-sectional DNAm-Alc (β =0.56, p = 0.14). We considered whether the suggestive relationship between DNAm-Alc levels at birth and AUDIT in adolescence could have been driven by maternal drinking during in utero development, but this did not appear to be the case in follow up analysis that added examination of maternal DNAm-Alc during pregnancy and which failed to show effects independent of DNAm-Alc at birth (Supplementary Table 5).

Discussion
Using a population-based longitudinal study with repeated measures of alcohol use and blood DNA methylation in two generations and an additional independent replication population, we conducted in-depth validations of the DNAm-Alc biomarker scores developed by Liu et al., including their 144 CpG recommended DNAm-Alc score. 6 Our results confirmed the utility of this biomarker as a measure for alcohol consumption in the general population, albeit with slightly decreased explanatory power than initially reported in adults without substantial levels of consumption. Performance was closer to initial assessment in our replication analysis using a population with much greater average levels of intake. Further, we demonstrated the DNAm-Alc score shows stronger correlation for sustained alcohol use compared to shorter term alcohol use, as evidenced by the higher proportion of the variance explained at midlife than in adolescence, and for AUDIT score compared to cross-sectional weekly alcohol intake. Additionally, by observing that an earlier measure of DNAm-Alc repeated in the same participants appears independent of later drinking behaviour, we show that this DNAm-derived score captures information reflective of alcohol-use exposure and behaviour, beyond genetic determinants alone (i.e. ruling out genetic confounding as the primary explanation of the results).
We observed stronger relationships between the DNAm-Alcs and AUDIT score, that captured additional behaviours related to alcohol dependency, compared to simple intake While the DNAm-Alc showed limited explanatory capacity in ARIES, its performance was markedly improved in our HN5000 replication population, achieving 14% of variation explained (Supplementary Table 3). Participants in this study differed from our ARIES population in several important ways, including that they were all oropharyngeal cancer cases, slightly older, and had a little more than three times greater and more variable average alcohol intake. With the data available, we could not determine which factors were most responsible for the improved DNAm-Alc performance in HN5000, but it is at least possible that the biological impact of greater alcohol abuse in this group may be contributing.
We also observed reduced capacity of the DNAm-Alcs to predict categories of drinkers by alcohol intake in ARIES compared to the estimates by Liu et al. (Figure 3). In fact, the best performing DNAm-Alc in ARIES -AUC = 0.60 with the 144 CpG score for heavy versus nondrinkers -had a lower AUC than any DNAm-Alc for resolving heavy from light or non-drinkers at initial report. The lowest that had been observed there being AUC = 0.67 for heavy versus light using the 5 CpG score in the ARIC study. However, our ability to evaluate prediction performance was severely limited in power by the small number of participants falling into the higher categories of intake, with only a maximum of N=14 heavy drinkers for any of our comparisons. Interestingly, we did observe some ability of the DNAm-Alcs to resolve participants with different categories of AUDIT score, specifically high (>15) versus low AUDIT individuals, at midlife where we'd also seen the strongest relationship in continuous analysis.
The two best-performing DNAm-Alcs in this regard, the 78 and 144 CpG biomarkers, had AUCs of 0.80 and 0.79 which is in a range comparable to that reported by Liu et al. for intake categories.
Our assessment of the timing of the relationship between alcohol use and the DNAm-Alcs, the third aim of our study, observed that these scores captured an epigenetic response largely independent of confounding from genetic or environmental predisposition to alcohol dependency. This was most clearly observed in the ARIES mothers, who had a strong relationship between DNAm-Alc and AUDIT when both were measured in participants at the same time but for whom almost no relationship was observed for DNAm-Alcs calculated from DNAm measured roughly 20 years previously, at a time when alcohol consumption was much lower during the women's index pregnancy (Table 2). This was true both when the earlier DNAm-Alcs were considered on their own, or when mutually adjusted for the cross-sectional DNAm-Alc. If DNAm-Alc values were being driven largely by stable genotype-methylation relationships, as one might predict a priori given that DNAm levels are often highly correlated with nearby genetic polymorphisms, 17 we would expect to observe a consistent relationship between such biomarker scores and alcohol dependency in the same participants regardless of the timing of the DNAm collection. However, this was not the case: when DNAm-Alcs were drawn at the earlier time point, no independent relationship was observed. We saw a similar trend among ARIES offspring in childhood and adolescence.
Further, our analysis of repeated measures of DNAm-Alcs in the same participants across time, and the comparisons of these across populations with different drinking behaviours and a different drinking history (young and middle-aged), helped identify the duration of exposure potentially reflected by these biomarkers. The DNAm-Alcs appeared to reflect more than just concurrent intake over a short-term window, like that captured by a breathalyser. The correlation observed between DNAm-Alcs and intake/AUDIT in ARIES adolescents with comparable levels of alcohol use to adults over a shorter time frame was much smaller than that of ARIES parents at midlife with longer histories of alcohol use (Table   1). This suggests that DNAm-Alcs may capture longer-term patterns of alcohol abuse, perhaps cumulatively (the ALSPAC parents having been using alcohol for over two decades, while their children are very recent users). It also suggests that the alcohol dependence and problem drinking components of the AUDIT scores could be driving these associations more than the actual alcohol consumption levels per se (alcohol behaviour in midlife being more likely to score highly on these components in British populations). 36 However, this could not be evaluated specifically by our current study. While effects in adolescents were systematically smaller than for adults, they did appear to be more than just background noise: showing associations independent of the effects in those same participants at earlier time points with no likely alcohol exposure, at birth and 7 years of age ( Table 2).
The results of our current study should be interpreted in the context of the following limitations: 1) Tissue specificity-DNA methylation measured in peripheral blood WBCs was used both for the original development of the DNAm-Alcs and our current validation. Blood may not be the tissue most directly relevant for assessing the impacts of alcohol abuse (as opposed to e.g. liver or brain). However, alcohol is absorbed by peripheral-blood and is relevant to the direct assessment of alcohol exposure. Further, blood-based biomarkers may be more clinically useful, less invasive, and more readily accessible than liver or brain tissue.
2) Measurement bias and measurement error-Cross-sectional measurement of alcohol use by self-report may be too crude an approximation of the biologically relevant exposure, due to the variability of alcohol behaviour over time. This has the potential to bias our estimates towards the null, particularly for alcohol intake as it is more variable than the AUD risk assessed by AUDIT. The latter includes a measure of, but is not limited to, alcohol consumption, but also reflects symptoms of dependence and indicators of alcohol-related problems. Previous reports have shown that AUDIT scores are more stable (than alcohol use reports) and particularly so in midlife compared to young adulthood, and this is probably attributable to a more variable nature of alcohol use in younger age. 37  Our study also had notable strengths, including access to repeated DNAm measurements in participants drawn from a prospective, population-based, longitudinal birth cohort. These repeated observations, some of which were collected at time points prior to or during limited alcohol use, were unique in structure and allowed assessment of the role of genetic confounding that would not have been possible with many other study designs.
In conclusion, while we find that the DNAm-Alcs do not always predict alcohol intake as well as initially hoped, our revised evaluation in a large population-based sample demonstrates that they do still capture non-trivial amounts of information about the biological impacts related to alcohol use. Importantly, our results using AUDIT score rather than intake alone suggest that these biomarkers may explain additional variation in alcohol use disorder beyond alcohol consumption.    . 'Heavy drinkers' were participants who consumed ≥ 42 g per day in men and ≥ 28 g per day in women; 'non-drinkers' consumed 0 g per day; 'light drinkers' consume 0 < g per day ≤ 28 in men and 0< g per day ≤ 14 in women; AUDIT was dichotomized into participants with a score >15 and those with a score ≤ 15.