Genome wide and Digital PCR Epigenetic Assessments of Alcohol Consumption

Abstract

The lack of readily employable biomarkers of alcohol consumption is a problem for clinicians and researchers. In 2014, we published a preliminary DNA methylation signature of heavy alcohol consumption that remits as a function of abstinence. Herein, we present new genome wide methylation findings from a cohort of additional subjects and a meta-analysis of the data. Using DNA from 47 consecutive heavy drinkers admitted for alcohol detoxification in the context of alcohol treatment and 47 abstinent controls, we replicate the 2014 results and show that 21,221 CpG residues are differentially methylated in active heavy drinkers. Meta-analysis of all data from the 448,058 probes common to the two methylation platforms show a similarly profound signature with confirmation of findings from other groups. Principal components analyses show that genome wide methylation changes in response to alcohol consumption load on two major factors with one component accounting at least 50% of the total variance in both smokers and non-smoking alcoholics. Using data from the arrays, we derive a panel of 5 methylation probes that classifies use status with a Receiver Operator Characteristic Area Under the curve (AUC) of 0.97. Finally, using droplet digital PCR, we convert these array based findings to two marker assay with an AUC of 0.95 and a four marker set AUC of 0.98. We conclude that DNA methylation assessments are capable of quantifying alcohol use status and suggest that that readily employable digital PCR approaches for substance consumption may find widespread use in alcohol related research and patient care.

Introduction

Alcohol consumption is common and not always harmful. A large number of epidemiological studies have demonstrated that moderate alcohol use is associated relatively healthy outcomes as compared to those who either abstain or drink heavily (Ronksley, Brien, Turner, Mukamal, & Ghali, 2011; Sun et al., 2011).

However, chronic heavy alcohol consumption is invariably associated with increased risk for undesirable outcomes (Centers for Disease Control, 2014). Heavy drinking increases risk for hypertension, stroke, liver disease, as well as various forms of cancers (International Agency for Research on Cancer, 2012; Rehm et al., 2010; World Health Organization, 2014). Additionally, drinking while pregnant may cause Fetal Alcohol Syndrome. Finally, from the personal and societal perspective, 1.1 million Americans are arrested annually for drunk driving, (Federal Bureau of Investigation, 2014) with the overall impact of excessive alcohol intake on the incidence of crime, divorce, accidental death or injury being perhaps beyond simple mathematical imputation. When all things are taken into consideration, excessive alcohol consumption is consistently one of the largest preventable causes of misery in the world (Centers for Disease Control, 2014).

This need not be the case. As conceptualized in the ICD 10 diagnostic category of Alcohol Use Disorder (AUD) or more generally, alcoholism, excessive alcohol consumption is treatable and screening for it in routine clinical encounters is routinely recommended (Jonas, Amick, Feltner, & et al., 2014). Unfortunately, its detection in the clinical setting is problematic with 84% of U.S. adults reporting that they have never discussed alcohol consumption with a health professional (McKnight-Eily et al., 2014). However, even if clinicians ask about alcohol intake, the resulting patient self-report data with respect to alcohol consumption is notoriously unreliable (Magura & Kang, 1996).

In order to overcome the inherent difficulties in self-report and provide a more objective assessment of alcohol intake that can supplement or replace self-report, a number of tools have been developed. Acute alcohol consumption is readily detectable by sampling breath, urine or blood (Vearrier, Curtis, & Greenberg, 2010). However, most heavy drinkers do not willingly present acutely intoxicated to settings amenable to the use of these tools and these methods cannot be used to assess alcohol intake in the days, weeks, or months prior to the sampling. Instead, the clinical laboratory assessment of chronic alcohol consumption currently relies on a set of algorithms that incorporate levels of liver proteins (e.g., alanine aminotransferase (ALT), γ-glutamyl transferase (GGT), aspartate aminotransferase (AST) or carbohydrate-deficient transferrin (CDT)) or alcohol metabolites such as ethyl glucuronide (EtG) (Kharbouche et al., 2012; Tavakoli, Hull, & Okasinski, 2011). In general, serum enzyme screening is regarded as insensitive and non-specific (Helander, 2003; Strid & Litten, 2003) with CDT levels being regarded by many as the most accurate currently available biomarker (Strid & Litten, 2003; Tavakoli et al., 2011). Even so, CDT levels are affected by a variety of common medications, smoking status and medical conditions (Bortolotti, De Paoli, & Tagliaro, 2006). Therefore, although early studies showed considerable promise, more recent studies have concluded that the overall sensitivity of CDT assessments for moderate AUD is only 60% with markedly lower sensitivity and specificity for milder forms of AUD (Tavakoli et al., 2011). Finally, EtG and fatty acid esters (FAEE) content of hair have found some utility. Unfortunately, high cost, concerns about non-beverage sources of alcohol and the dependency on the availability of straight hair, have limited the use of this approach (Cabarcos, Álvarez, Tabernero, & Bermejo, 2015; Vearrier et al., 2010). Hence, many clinicians and scientists feel that additional mechanisms for assessing alcohol use are needed.

Epigenetic methods may offer a more scalable, sensitive and specific approach to assessing alcohol consumption. In 2014, we published the first genome wide study on the relationship of alcohol consumption to heavy alcohol consumption (Philibert et al., 2014). Since that time, these findings have been partially reproduced at the single and genome wide level (Brückmann, Di Santo, Karle, Batra, & Nieratschker, 2016; Liu et al., 2016). In this communication, we confirm and extend those prior findings and reduce the methylation assessment process to a set of digital PCR assays.

Methods

The current study uses clinical and methylation (Illumina Humanmethylation450 BeadChip; a.k.a. 450K array) data from our prior analyses (Philibert et al., 2014) and from newly collected subjects. The procedures used in the recruitment and characterization of the new subjects were approved by the Western Institutional Review Board (www.wirb.com).

Subject Ascertainment and Characterization

The heavy drinking subjects were collected from one of three Iowa substance use treatment organizations; Center for Alcohol and Drug Services (CADS, Davenport, IA), Prelude Behavioral Services (campuses in Iowa City and Des Moines, IA) and Alcohol and Drug Dependency Services of Southeast Iowa (ADDS, Burlington, IA). In brief, between 1 and 7 days after their last intake of alcohol, subjects over the age of 18 who were admitted for acute intoxication in the context of severe alcohol use disorder were referred for participation in the study by staff or via printed posters concerning the study. Subjects were not considered for the study if still intoxicated or under the influence of other substances. After full informed written consent was received, each subject was interviewed with a series of instruments including a modified form of Version II of the Semi Structured Assessment for Genetic Studies (SSAGA-II), the Michigan Alcohol Screening Test, and our Substance Use Questionnaire, an inventory that assesses substance consumption over recent time periods (Bucholz et al., 1994; Philibert et al., 2014; Selzer, 1971). After interview, the subjects were phlebotomized to provide biomaterial for the study. Serum was isolated then frozen at −80°C until used. Whole blood DNA was prepared using cold protein precipitation (Lahiri & Nurnberger, 1991).

Control subjects were collected from the University of Iowa site using our previously described protocol (Philibert et al., 2014). In brief, control subjects who denied consumption of alcohol in the past year and a lifetime history of any substance use disorder, with the exception of tobacco use disorder were solicited for the study. After consent was received, the subjects were interviewed with both the SSAGA and the Substance Use Questionnaire, and then phlebotomized to provide sera and DNA for the study.

Genomewide Assessments

Genome wide methylation status for whole blood DNA from the 94 newly collected subjects was determined by the University of Minnesota Genomics Center using the Infinium MethylationEpic Beadchip (a.k.a. Epic array) using reagents and protocols from the manufacturer (Illumina, San Diego, USA).(Pidsley et al., 2016) Probe filtering, background correction and adjustment for probe types on the Epic arrays were performed on the methylation intensity data (IDAT) files using the MethyLumi, WateRmelon and FDb.InfiniumMethylation.hg19 R packages (Davis S, Du P, Bilke S, Triche, & M, 2017; Pidsley et al., 2013; Triche, 2014). Quality control was performed on the sample and probe levels. For samples, those with >1% CpG sites with a detection p-value >0.05 were removed while CpG sites with a bead count <3 and/or >1% samples with a detection p-value >0.05 were removed. All samples passed quality control whereas 860,056 CpG sites remained after quality control (vs. 865,918 before quality control). Subsequently, data for each of the 448,058 probes common to Epic and 450K array was extracted and combined into a single file for data analysis. This Epic data set (Accession # pending acceptance of manuscript) can be obtained through the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) website. The data set from our 2014 subjects (450K) is already publically available via GEO GSE57853)

The association between methylation and alcohol use status was analyzed using MethLAB while controlling for batch, chip and gender. Correction for genome wide comparison was performed using the False Discovery Rate method (Benjamini & Hochberg, 1995). As per the previous recommendations for biomarker searches, analyses were not corrected for cellular heterogeneity (Bauer et al., 2016; Philibert & Glatt, 2017). Power for the meta-analysis was calculated using the method of Valentine and colleagues (Valentine, Pigott, & Rothstein, 2010).

Methylation Sensitive ddPCR Assessments

The methylation status of individual CpG residues of interest was determined using a droplet digital PCR paradigm as previously described (Andersen, Philibert, Gibbons, Simons, & Long, 2017). In brief, 1 μg of DNA from each subject was bisulfite converted using the Fast 96 Bisulfite Conversion kit (Qiagen, Germany). An aliquot of each sample was pre-amped, diluted 1:3000, and then PCR amplified using primer probe sets for each locus from Behavioral Diagnostics (Coralville, IA, USA) and Universal Digital PCR reagents and protocols from Bio-Rad (Carlsberg, CA, USA). The number of droplets containing amplicons that have at least one “C” allele (corresponding to the methylated cytosine residue), at least one “T” allele, at least one “C” and “T” allele, or no amplifiable alleles was determined using a QX-200 droplet counter and QuantiSoft software (Bio-Rad, USA), and expressed as a percent methylation.

Quantitative non-genome wide analyses of both array and ddPCR derived methylation data were conducted using JMP Version 10 software (SAS, Cary, NC USA) and the specific routines outlined in the text. Pathway analysis was conducted using GoMiner™ package and the default settings (Zeeberg et al., 2003). All comparisons between ddPCR assessments and Illumina array values were conducted using biomaterial and clinical data from the 2017 cohort.

Results

The clinical and demographic variables of the subjects from our 2014 analyses and the current study whose DNA was examined with methylation arrays are described in Table 1. The recruitment criteria for the two cohorts are identical with the exception that the 2014 subjects could not have any significant medical illness. Consistent with this difference, the subjects from the 2014 study tend report less alcohol (13 ± 11 vs 19 ± 10) drinks per day and cigarette consumption (13 ± 11 vs 19 ± 15 cigs per day) over the month prior to admission than the current cohort despite being similar in age. Notably, 6 of the 33 subjects from the earlier study were biochemically confirmed non-smokers while only 2 of the 47 case subjects from the second cohort were non-smokers.

Table 1.

Clinical Characteristics of Subjects

	2014 Study		New Sample
	Case	Control	Case	Control
n =	33	33	47	47
Age	45.5 ± 7.8	46.7 ± 7.8	45.7 ± 14.5	45.9 ± 14.2
Ethnicity
White	31	32	42	46
African American	2	1	5	1
Gender
Male	25	25	35	30
Female	8	8	12	17
Days since last drink	4.0 ± 1.8		4.2 ± 1.8
Drinks per Day
Prior Week	13 ± 13	0	17 ± 9	0
Prior Month	13 ± 11	0	19 ± 10	0
Smoker
Yes	27	1	45	2
No	6	32	2	45

Genome Wide Results

A total of 448,058 probes are common to the Epic and the 450K platform. Using a standard genome wide approach, we analyzed these data with respect to alcohol consumption status for each of these cohorts separately, and together (n=160). Predicted power for the analysis was excellent. Assuming low sample heterogeneity, the predicted power to detect an effect size of 2.5, 1 and 0.5 at a given locus in the meta-analysis was 99%, 99% and 82%, respectively.

After False Discovery Rate correction (FDR), a total of 2460, 5028 and 21,221 probes were significantly associated with heavy alcohol consumption in the 2014, the 2017 and the combined cohorts, respectively, at the p<0.05 level. Table 2 lists the most significant findings from the meta-analysis with reference to the findings in the individual cohorts. Please see Supplemental Table 1 for a complete listing of the significant markers for each analysis. In addition, in order to better understand whether the changes at a given locus were related to alcohol or tobacco consumption, the ranking of a given probe with respect to smoking status in the Framingham Heart Study (n=1599) is also provided (Dogan, Beach, & Philibert, 2017).

Table 2.

Results of Initial Meta-Analyses

Merged Dataset Analyses				Association with Smoking in FHS*	2017 Samples (n=94)		2014 Data (n=66)
Rank	Probe	FDR pvalue	β Coefficient		FDR pvalue	β Coefficient	FDR pvalue	β Coefficient
1	cg05575921	2.11E-21	−1.15	1	3.69E-19	−1.29	0.006895	−0.95
2	cg01940273	9.68E-18	−0.40	4	2.45E-13	−0.46	0.00766	−0.32
3	cg21566642	8.68E-17	−0.52	2	2.45E-13	−0.61	0.01258	−0.40
4	cg02583484	2.18E-16	−0.33	57	6.19E-09	−0.34	0.000984	−0.32
5	cg03636183	1.11E-15	−0.44	3	8.44E-13	−0.48	0.012232	−0.40
6	cg09935388	1.36E-14	−0.66	11	1.53E-07	−0.63	0.000984	−0.72
7	cg21161138	9.83E-13	−0.29	5	1.01E-11	−0.36	0.069614	−0.18
8	cg06644428	9.83E-13	−0.63	33	1.24E-07	−0.69	0.009737	−0.55
9	cg07986378	3.09E-12	−0.38	60	1.37E-07	−0.37	0.00766	−0.41
10	cg04987734	5.89E-11	0.29	#N/A	1.22E-06	0.30	0.013966	0.27
11	cg20732076	7.37E-11	−0.31	#N/A	1.90E-06	−0.28	0.00816	−0.35
12	cg18917643	1.15E-10	−0.29	#N/A	2.38E-07	−0.35	0.034976	−0.20
13	cg25998745	2.15E-10	−0.30	#N/A	0.000119	−0.30	0.00327	−0.29
14	cg23779890	8.71E-10	−0.23	#N/A	0.000494	−0.20	0.000984	−0.28
15	cg11660018	1.54E-09	−0.27	26	1.43E-06	−0.29	0.031264	−0.25
16	cg23193759	3.28E-09	−0.24	#N/A	0.001907	−0.20	0.000463	−0.31
17	cg24437104	3.32E-09	−0.12	#N/A	4.69E-05	−0.11	0.012226	−0.14
18	cg27241845	3.32E-09	−0.27	37	1.72E-07	−0.35	0.146194	−0.15
19	cg14753356	3.42E-09	−0.35	50	6.78E-06	−0.35	0.023246	−0.34
20	cg10691866	4.14E-09	−0.30	332	0.000224	−0.32	0.005631	−0.28
21	cg23090529	5.32E-09	−0.34	240	0.000194	−0.33	0.009705	−0.36
22	cg02708705	6.78E-09	−0.29	297	0.000209	−0.28	0.011256	−0.29
23	cg14622549	7.00E-09	0.43	#N/A	0.000189	0.43	0.00759	0.46
24	cg12490835	7.17E-09	−0.17	#N/A	3.75E-05	−0.16	0.015507	−0.20
25	cg03329539	7.17E-09	−0.23	22	2.11E-06	−0.27	0.05659	−0.18
26	cg14580211	7.17E-09	−0.28	34	0.000239	−0.25	0.006895	−0.33
27	cg24859433	7.17E-09	−0.22	10	2.05E-06	−0.26	0.075908	−0.15
28	cg02003183	7.17E-09	0.44	#N/A	7.67E-07	0.52	0.098987	0.31
29	cg03450635	9.78E-09	−0.22	#N/A	3.83E-06	−0.25	0.045636	−0.19
30	cg19459094	1.02E-08	−0.33	#N/A	0.000256	−0.28	0.00766	−0.41

^*rank, where given, is taken from Dogan and associates (2017)

Given the high smoking intensity of the alcohol consuming subjects of the latest cohort, it is not surprising to see that cg05575921, a well-established methylation biomarker of smoking status, is the top ranked probe in the meta-analysis (Gao, Jia, Zhang, Breitling, & Brenner, 2015). In fact, each of the top nine most significantly associated probes was also significantly associated with smoking in the FHS cohort with the strength of the associations being largely driven by the signal from the 2017 cohort. In contrast, in the “healthier” 2014 cohort, cg05575921 was the 28^th most significant probe.

By annotating the current findings with information from the FHS study with respect to smoking, an approximate understanding of the relative influence of smoking can be derived. However, in order to get a better understanding of the effects of only alcohol on DNA methylation, we compared the methylation status of the 8 biochemically confirmed non-smokers to those of the biochemically confirmed non-smoking and alcohol abstinent controls. After FDR correction, 507 probes were significant associated with heavy alcohol consumption. Consistent with the non-smoking status of the cohort, cg05575921 methylation was not associated with case status (i.e. 242,587^th ranked probe). Interestingly, as opposed to combined dataset where the direction of change is towards demethylation (i.e. –β coefficient; 27 of 30 probes), the directionality of the significantly associated changes in the non-smoking heavy alcohol consumers is more mixed with 16 of the 30 top probes being hypermethylated in association with alcohol consumption (Fisher’s Exact; p<0.001).

The difference in the directionality of methylation changes suggests that the characteristics of the epigenomic response to heavy drinking in the presence and absence of smoking may differ (Beach et al., 2015). To further explore this, we conducted principal components analysis (PCA) for the signal from the top 100 probes for all subjects combined or just for those who were biochemically confirmed non-smokers. PCA analysis of the combined analyses showed one major component that accounted for 65% of the variance with six other components with eigenvalues between 7 and 1 accounting for another 19.4% of the variance (supplemental Figure 1). In contrast, the PCA of data from the non-smokers showed a major component that accounted for 52% of the variance with a second component accounting for another 21% of the variance and six components with eigenvalues greater than one accounting for another 9.6% of the variance.

Additional exploration of potential differences in the characteristics of the epigenomic response was obtained via pathway analysis of the annotation data from the 500 most significantly associated probes for all subjects or just the non-smoking subjects. Sixteen Gene Ontology pathways from the Gene Ontology were significantly (FDR p<0.05) enriched using the results from all subjects (See Table 4). In contrast, only the top pathway from Table 4, the Protein Binding pathway, was significantly enriched in the 500 probes most differentially methylated in the contrasts using the data only from those who did not smoke (results not shown).

Table 4.

The Ten Most Differentially Regulated Gene Ontology Pathways

GO Category	Category Name	Genes		Log10 P-Value	FDR
		Total	Changed
GO:0005515	protein binding	6815	173	−9.55835	0.00
GO:0005488	binding	11509	242	−6.81424	0.00
GO:0050794	regulation of cellular process	6319	146	−4.68958	0.03
GO:0007165	signal transduction	2324	66	−4.37442	0.03
GO:0050789	regulation of biological process	6728	152	−4.37266	0.02
GO:0023034	intracellular signaling pathway	1708	52	−4.24371	0.03
GO:0006996	organelle organization	1582	49	−4.19934	0.03
GO:0071842	cellular component organization	2175	62	−4.16194	0.02
GO:0035556	intracellular signal transduction	1454	45	−3.87411	0.04
GO:0009987	cellular process	11702	234	−3.82927	0.04

FDR=false discovery rate.

Single and Multiple Marker Analyses

To identify CpG loci suitable for translation, based on Akaike’s Information Criterion (AIC) (Akaike, 2011) of each probe, we filtered the results described in Table III to identify probes highly associated with alcohol consumption in the presence and absence of smoking, then used stepwise forward logistic regression to identify a sets of CpG probes predictive of alcohol status in all subjects. A number of three marker sets had Receiver Operation Characteristic (ROC) Area Under the Curve (AUC) of 0.93 or better. For example, a three marker set consisting of cg02583484, cg23779890 and cg20732076 had an AUC of 0.95 while a five marker set consisting of cg02583484, cg23779890, cg13126206, cg06690548 and cg16743273 had an AUC of 0.97. The sequence surrounding the CpG sites targeted by each of these five probes and two additional markers with highly ranked information content, cg04987734 and cg07465627, were then reviewed for suitability for digital PCR assay design. Although a number of parameters were considered in this process, particularly focus was given to the average methylation difference between cases and controls (e.g. Δβ) and the level of sequence complexity present in the region surrounding the candidate CpG site.

Table 3.

Analysis of Only Non-Smoking Subjects

Rank	Probe ID	FDR pvalue	Beta Coefficient	Rank in all subjects
1	cg06690548	0.001	−0.46	306
2	cg07768103	0.001	0.87	125
3	cg00827581	0.001	0.69	191
4	cg02583484	0.001	−0.36	4
5	cg07721872	0.001	0.59	799
6	cg06579345	0.001	−1.21	22100
7	cg04895225	0.002	−0.42	1778
8	cg07841371	0.003	0.50	1276
9	cg04753163	0.003	−0.39	2679
10	cg01445100	0.003	0.37	40
11	cg19459094	0.003	−0.57	30
12	cg21881034	0.003	−0.96	36620
13	cg15133201	0.003	0.34	11685
14	cg10072464	0.003	0.67	5856
15	cg21373806	0.003	0.49	2217
16	cg00464927	0.003	0.63	126
17	cg21248060	0.003	0.67	90
18	cg26292521	0.003	−0.82	93805
19	cg02212339	0.004	0.61	174
20	cg10278046	0.004	0.42	479
21	cg13103051	0.004	0.33	517
22	cg13076829	0.004	0.36	319
23	cg06375580	0.004	−0.76	34412
24	cg16743273	0.004	0.47	1840
25	cg02734358	0.005	−0.58	16252
26	ch.3.55501R	0.005	−1.20	72720
27	cg07664370	0.005	−0.90	25352
28	cg25115537	0.005	0.32	4173
29	cg01199327	0.005	0.68	402
30	cg00872580	0.005	0.37	98

Droplet digital PCR (ddPCR) assays were successfully designed and tested for cg02583484, cg13126206, cg04987734 and cg07465627. Figure 1 illustrates the correlation between methylation status at each of these four loci for 92 of the 94 subjects profiled by ddPCR and the Epic array (n.b. 4 wells of each ddPCR plate are consumed by controls). Despite the small dynamic range, the correlation between the Epic array and ddPCR cg02583484 and cg04987734 were particularly high (0.89 and 0.91, respectively) with the results for cg13126206 and cg07465627 (0.54 and 0.36, respectively) being more modest. What is more, when the values for methylation as determined by Epic Array and ddPCR are regressed against alcohol status, in each case the ddPCR values account for a greater portion of the variance than that for the array methylation values suggesting that the ddPCR assays more accurately assess methylation at these loci than the Illumina probes (cg02583484 R² 0.44 vs 0.40; cg13126206 R² 0.13 vs 0.08; cg04987734 R² 0.47 vs 0.38; and cg07465627 R² 0.33 vs 0.12).

Figure 1.

The relationship between methylation values as determined by the Epic array or ddPCR. Epic values are expressed as fractional methylation whereas ddPCR values are expressed as percent methylation. The Pearson correlation co-efficient (R²) for cg02583484, cg04987734, cg13126206 and cg07465627 are 0.89 and 0.91, 0.54 and 0.36, respectively. N=92 for all four assessments.

Based on their AIC, the four ddPCR markers incorporated into models for the prediction of alcohol use status in all 160 subjects. By itself, Dcg04987734 (i.e. ddPCR assessment of methylation of the CpG locus denoted by cg04987734) had a ROC AUC of 0.91. A two marker set composed of Dcg04987734 and Dcg02583484 had an AUC of 0.95. A three marker set consisting of Dcg04987734, Dcg02583484 and Dcg07465627 had an AUC of 0.97. However, addition of Dcg13126206 the three marker set did not increase the AUC further.

Discussion

Although these findings need to be replicated and extended by others before they can be generally accepted, we believe that it is likely that the present results will be affirmed. The results of our genome wide analyses are internally consistent with 29 of 30 of most highly associated markers in the combined dataset being associated with an FDR <0.1 in both of the individual datasets. Furthermore, all of the loci successfully targeted in our ddPCR assays were significantly associated with alcohol use in all analyses of our datasets. Our results are also consistent with those of others. Of the 28 most highly CpG sites reported by Liu and colleagues (see Table 3 of their 2016 paper) that are available for examination in our combined dataset, 12 are also significantly associated with alcohol status in our analyses (Liu et al., 2016). Methylation at the second most powerfully predictive site used in our three marker assay set, cg02583484, was the 2^nd ranked probe in the 2016 Liu meta-analyses. Finally, methylation at cg23779890, one of the loci for which we were unable to construct a ddPCR assay, has been independently shown to be associated with alcohol use status by Nieratschker and colleagues, as well as Liu and colleagues work (Brückmann et al., 2016; Liu et al., 2016). Hence, there is reason for optimism.

The results also speak to the power of well-designed, biologically informed case and control analyses in the development of diagnostic tests. Case and control approaches have long been the favorite of geneticists (Risch, 2000). However, for stigmatized behaviors such as heavy alcohol consumption, their skillful employment may particularly important because misclassification of subjects can have profound effects on the power of both genetic and epigenetic case and control analyses (Philibert & Glatt, 2017; Silverberg et al., 2001). The extent to which this takes place with respect to studies of alcohol consumption is not clearly established. However, for smoking, the potential for high rates of unreliable self-report have been documented in single investigator and national cohorts have been clearly demonstrated (Andersen et al., 2017; Caraballo, Sharapova, & Asman, 2016). Hence, our decision on recruiting case subjects directly from inpatient alcohol treatment centers and biochemically verified substance free controls from low risk settings which generated marker sets that have AUC above 0.95 and account for 67% of the variance seems to have worked well. In contrast, using data from 13317 subjects collected using a variety of other study designs almost all of which do not have serological confirmation of substance use status, Liu and colleagues reported a set of 144 CpG residues that together with other variables resulted in an AUC of 0.90 to 0.99 for heavy drinkers. However, their null model, which included gender and age, alone had an AUC of 0.8, with the top 144 CpG residues accounting for only ~15% of the variance for heavy drinkers (Liu et al., 2016). Since our goal to develop unbiased clinical tests applicable to both genders and all ages and ethnicities, use of demographic predictor variables as outlined by Lui and colleagues made necessary by the lack of signal to noise ratio in these studies is not acceptable. Therefore, for these and other reasons, we recommend that those contemplating similar searches to develop epigenetic tests for other environmental or substance consumption variables, such as cannabis use, employ the more powerful case and control approach with independent validation of substance use status for their initial marker development. In addition, because ethnic specific confounding can have profound effects on the reproducibility of results, we also recommend that sufficient numbers of each ethnicity be present to independently test each locus in each ethnicity (Dogan et al., 2015).

Although pathway analyses are frequently difficult to interpret, the finding that the Gene Ontology pathway of Protein binding, which is defined as interacting selectively and non-covalently with any protein or protein complex, is plausible. First and foremost, ethanol is a solvent and at the concentrations found acutely in heavy drinkers (0.3-0.4%), it is not surprising to see the alcohol induced changes in the dielectric constant of the cytoplasm (Jones, 1989) or production of free radicals (Manzo-Avalos & Saavedra-Molina, 2010) results in the alteration of the epigenetic signature of genes of large macromolecular complexes that depend upon selective, non-covalent interaction with other proteins for assembly and function. The finding by Liu and colleagues that the epigenetic changes that they observed in association with alcohol were reflected in changes in gene transcription of peripheral blood suggests that many of the changes we observed are also coupled to changes in gene expression (Liu et al., 2016). However, what is truly needed is an understanding of the epigenetic changes that take place in the brain.

Although smoking and drinking are well known to co-segregate, (Bierut, Dinwiddie, Begleiter, & et al., 1998) the extent of the co-morbidity surprised us with nearly 90% of the first 100 subjects recruited through our current project (R44AA022041) being daily smokers. Given the profound effects of smoking on genome wide methylation, (Dogan et al., 2017; M. V. Dogan et al., 2014; Zeilinger et al., 2013) this comorbidity has significant implications on strategies for identifying loci capable of measuring alcohol intake in all potential subjects. In that regard, we note that none of our candidate markers associated with smoking status in any of 5 genome wide methylation datasets in our possession (Dogan et al., 2017; M.V. Dogan et al., 2014; Monick et al., 2012; Philibert, Beach, Li, & Brody, 2013; Philibert, Beach, & Brody, 2012). However, that does not mean that other factors such as nutritional status on alcohol associated methylation changes can be ignored. Over the past two years, we have become increasingly aware of a modest but yet significant relationship between methylation of the methylene tetrahydrofolate reductase (MTHFR) promoter and smoking associated methylation changes (Beach et al., 2017). Since 1) MTHFR is a key regulator of methyl group donor availability, 2) substance users often do not consume diets rich in folate and 3) alcoholics are often deficient in folate, we remain wary of possible dietary influence on alcohol related methylation changes.

The strength of many of the associations noted in Table 2 appear to be stronger in the 2017 cohort than in the 2014 cohort. The most likely explanation, other than larger sample size of the 2017 cohort, is that the subjects in the latter cohort drank nearly 6 drinks per day in the month prior to admission than those in the 2014 cohort. However, it also may be that the response of methylation to alcohol consumption is non-linear for some or all of the key loci. Further large scale studies to determine the shape of the dose response curve are needed.

The next few immediate steps for the testing and development of these markers are relatively clearly cut. First, the ethnic representation of study subjects needs to be broadened. Although very representative of the treatment population at the centers included in our study, the ethnic diversity of the subjects included in this study is low. Only those of European and African ancestry were represented and the number of those with African ancestry was low (n=8). Given the strong impact of genetic variation, some of which is ethnic specific, on smoking induced methylation signal, it will be prudent to examine the performance of these and other alcohol biomarker candidates large numbers of subjects in a wide variety of ancestries to minimize possible confounding by cis and trans genetic variation. Second, the dose response and time dependency characteristics of the methylomic response need to be examined. Careful attention should be paid to co-factors that may alter the linearity of response at a given locus. Third, and most importantly, the reversion curve for methylation status as a function of abstinence needs to be established. Reversion of substance induced changes has already been documented for both cigarette and alcohol associated methylation changes (Bauer et al., 2015; Brückmann et al., 2016; Philibert et al., 2016; Philibert et al., 2014; Wilson et al., 2017). However, this preliminary understanding of the reversion process is woefully short of the knowledge base needed to monitor the effectiveness and guide the improvement of existing therapies.

In summary, we report the use of genome wide meta-analysis to develop a set of digital PCR assays highly predictive of recent heavy alcohol consumption status. We suggest that these digital assay approaches will find utility in clinical research and that further efforts may lead to the introduction of generally usable clinical biomarkers for the diagnosis and treatment of excessive alcohol consumption. In order to achieve this goal, testing and extension of the current results by independent research groups will be necessary.