Epigenetics of alcohol use disorder—A review of recent advances in DNA methylation profiling

Abstract Alcohol use disorder (AUD) is a major contributor to morbidity and mortality worldwide. Although there is a heritable component, the etiology of AUD is complex and can involve environmental exposures like trauma and can be associated with many different patterns of alcohol consumption. Epigenetic modifications, which can mediate the influence of genetic variants and environmental variables on gene expression, have emerged as an important area of AUD research. Over the past decade, the number of studies investigating AUD and DNA methylation, a form of epigenetic modification, has grown rapidly. Yet we are still far from understanding how DNA methylation contributes to or reflects aspects of AUD. In this paper, we reviewed studies of DNA methylation and AUD and discussed how the field has evolved. We found that global DNA and candidate DNA methylation studies did not produce replicable results. To assess whether findings of epigenome‐wide association studies (EWAS) were replicated, we aggregated significant findings across studies and identified 184 genes and 15 gene ontological pathways that were differentially methylated in at least two studies and four genes and three gene ontological pathways that were differentially methylated in three studies. These genes and pathways repeatedly found enrichment of immune processes, which is in line with recent developments suggesting that the immune system may be altered in AUD. Finally, we assess the current limitations of studies of DNA methylation and AUD and make recommendations on how to design future studies to resolve outstanding questions.

ate the influence of genetic variants and environmental variables on gene expression, have emerged as an important area of AUD research. Over the past decade, the number of studies investigating AUD and DNA methylation, a form of epigenetic modification, has grown rapidly. Yet we are still far from understanding how DNA methylation contributes to or reflects aspects of AUD. In this paper, we reviewed studies of DNA methylation and AUD and discussed how the field has evolved. We found that global DNA and candidate DNA methylation studies did not produce replicable results. To assess whether findings of epigenome-wide association studies (EWAS) were replicated, we aggregated significant findings across studies and identified 184 genes and 15 gene ontological pathways that were differentially methylated in at least two studies and four genes and three gene ontological pathways that were differentially methylated in three studies. These genes and pathways repeatedly found enrichment of immune processes, which is in line with recent developments suggesting that the immune system may be altered in AUD. Finally, we assess the current limitations of studies of DNA methylation and AUD and make recommendations on how to design future studies to resolve outstanding questions.

K E Y W O R D S
addiction, alcohol abuse, alcohol dependence, alcohol use disorder, DNA methylation, epigenome-wide association study

| INTRODUCTION
Alcohol use disorder (AUD) is a disease characterized by an inability to stop or control alcohol use despite negative social, occupational, or health consequences. AUD affects over 100 million individuals worldwide and is associated with significant morbidity and mortality. 1 Nearly 100 million disability-adjusted life years, which are estimated years lost due to illness, disability, or early death, were attributed to alcohol misuse in 2016. 1 Despite the widespread prevalence of the disease, there are few effective interventions for prevention and treatment, and little is known about its underlying etiology. This may be due in part to the wide range of alcohol-related phenotypes that AUD can represent.
Most studies in this review classified AUD according to the Diagnostic and Statistical Manuel of Mental Disorders, Fourth Edition (DSM-IV), which divides AUD into alcohol dependence based upon satisfaction of three or more criteria relating to physical or psychological dependence and alcohol abuse, which requires satisfaction of at least one criterion involving alcohol use despite physical or mental damage to the individual. 2 The most recent edition of the manual, the DSM-5, groups the same criteria (with the exception of one criterion) into AUD, which is classified as mild, moderate, or severe. Both across and within these diagnostic categories, AUD can represent many combinations of symptoms, symptom severity, and alcohol consumption patterns, leading to a wide range of alcohol-related phenotypes.
Despite the phenotypic variability in AUD, studies have consistently found genetic factors to influence an individual's risk of developing the disorder. For example, family and twin studies have shown AUD to be 50%-70% heritable. [3][4][5][6] Although genome-wide association studies to date have succeeded in identifying some variants associated with AUD, they are not able to fully explain the complex etiology of AUD. 7,8 Epigenetic modifications, which are changes in gene expression that are not attributable to changes in DNA sequence, can contribute to gene expression and thus may provide insight into physiological causes and consequences of AUD. Epigenetic modifications can also be influenced by genetic variants and environmental exposures and have been shown to play a role in determining when and to what extent gene transcription occurs. 9 A growing recognition of the importance of epigenetics and advances in assays for DNA methylation, a well-known form of epigenetic modification, have led to a rapid rise in the number of studies investigating AUD and DNA methylation in the past decade ( Figure 1).
In this review, we focus on DNA methylation, as it is the best studied form of epigenetic regulation in human studies of AUD. DNA methylation refers to methylation at the fifth carbon in cytosine bases and most often occurs in CpG dinucleotides, which consist of a cytosine followed by a guanine. DNA methylation is typically carried out by proteins in the DNA methyltransferase family. 10 On the other hand, demethylation can occur either through deamination or removal of methylated bases via activation-induced deaminases or ten-eleven translocations, respectively. The removed or deaminated bases are then replaced through the base excision repair pathway. 11 Though the biological roles of DNA methylation are not fully understood, it has several frequently discussed roles in regulating gene expression. DNA methylation in the promoter region has been shown to promote or inhibit transcription factor binding, thereby activating or silencing transcription, respectively. 10 DNA methylation in the gene body has been positively correlated with gene expression and intergenic DNA methylation may also affect transcription. 10,12 DNA methylation can also recruit methyl CpG binding proteins that in turn recruit histone deacetylase complexes that repress gene expression. 11 Demethylation, or the replacement of a 5-methylcytosine with a cytosine, likely has the opposite effect as DNA methylation and has been associated with some diseases.
Many standard DNA methylation assays do not distinguish between methylation and hydroxymethylation, which means that values that are being recorded as differential methylation may be due to varying degrees of differential hydroxymethylation rather than methylation. Structurally, the two are similar-hydroxymethylation refers to the covalent addition of a hydroxymethyl group, rather than a methyl group, to the C5-position of cytosine in CpG dinucleotides. Functionally, however, they may be different. Hydroxymethylation is not fully understood, but it is common in the brain and may play a role in DNA repair or transcription factor binding or may be a transition state of DNA methylation. [12][13][14] Our review identified only one epigenome-wide association study (EWAS) that used array technology and also differentiated between methylation and hydroxymethylation. In this study, some of the top hits had higher proportions of hydroxymethylated sites than methylated CpG sites, which highlights the importance of distinguishing between the two types of methylation as the signal may at times be a hydroxymethylation rather than methylation signal.
Specific to AUD, alcohol has been shown to alter DNA methylation patterns in rodents, coinciding with increases in voluntary alcohol administration that mimic a transition from light "social" drinking to excessive alcohol consumption in humans. [15][16][17][18][19] Thus, DNA methylation could play a role in the biological consequences of alcohol consumption F I G U R E 1 Timeline of studies on AUD and DNA methylation and may also contribute to the neurobiological architecture of AUD pathology. In this review, we summarize how our understanding of the associations between DNA methylation and the disease-specific phenotype AUD has evolved. We performed a search in PubMed using the terms "DNA methylation" and "alcohol abuse," "alcohol dependence," or "alcohol use disorder" from January 2000 to April 2020. We reviewed all papers that resulted from this search and included in this review original articles on DNA methylation in human specimens (n = 27). We excluded studies for which the DNA methylation and AUD data were not in the same individuals, for example, studies of maternal AUD and offspring DNA methylation patterns. Though included in our discussion, we also excluded studies on alcohol consumption from our primary analysis. Based on our findings, we discuss the types of DNA methylation studies, including global DNA methylation, candidate gene methylation, and EWAS, and what we have learned from each form of inquiry. We also aggregate available EWAS data to identify molecular targets and pathways that had significant differential methylation in two or more studies. We used this as a measure of replication in light of generally small sample sizes in currently available EWAS data. Finally, we discuss challenges involved in interpreting DNA methylation data in AUD and how these may be resolved in the future.  (Table 1). Although most studies used DSM-IV diagnostic criteria, some classified AUD using the Alcohol Use Disorder Identification Test (AUDIT), which is a 10-item questionnaire that categorizes alcohol use as harmful or hazardous based on alcohol-related problems and behaviors. 20 We also included studies that use the Semi-Structured Assessment for the Genetics of Alcoholism (SSAGA-II), which is based upon DSM-III and IV criteria, and the International Classification of Diseases, tenth revision (ICD-10), which is influenced by but not identical to the DSM-IV (Table 1 and supporting information Table S1). Five studies in our search examined global DNA methylation or the average methylation of cytosine bases across the genome in cases compared to controls.
However, because they provided little insight into the disorder and all studies had small sample sizes, varying assays and differing results, we decided not to include them in the primary analysis.

| Candidate gene methylation studies
Candidate gene methylation studies are based upon a priori hypotheses that a disease is associated with variation of DNA methylation in the promoter region or around transcription start sites of a specific gene (supporting information Table S1). These studies primarily used pyrosequencing or bisulfite sequencing. Though bisulfite sequencing has been shown to be more sensitive than pyrosequencing, the two methods generally yield similar results. 21 Although reviews investigating genetic studies of candidate genes have shown them to not be replicable, it has not been clear if that is also true for candidate gene methylation studies. 22 In our review of candidate gene methylation studies of AUD, we also found conflicting results. Certain findings were replicated in subsequent EWAS. Aldehyde Dehydrogenase 2 Family Member (ALDH2) 23 and Opioid Receptor Mu 1 (OPRM1) 24 were both found to be hypermethylated in candidate gene methylation studies and also significantly hypermethylated in EWAS of AUD. 25,26 Orexin 27 was neither shown to be differentially methylated in a candidate gene methylation study nor differentially methylated in any EWAS we reviewed. The majority of candidate gene methylation studies, however, were not replicated, with many contradicted by subsequent findings. For example, Glutamate Ionotropic Receptor NMDA Type Subunit 2B (GRIN2B), 25,28 Ganglioside Induced Differentiation Associated Protein 1 (GDAP1), 29,30 Somatostatin Receptor 4 (SSTR4), 31,32 and Solute Carrier Family 6 Member 3 (SLC6A3) 26,[33][34][35][36] were all found to have conflicting results across studies.

| EWAS
EWAS refer to studies that survey specific CpG sites across the genome without a priori hypotheses. Given the failures to replicate candidate gene methylation studies and the limitations they have in explaining AUD, there has been a recent emphasis on EWAS in better understanding DNA methylation in AUD. This has been made possible by the development of array-based platforms that offer relatively easy, high through-put methods for assessing CpG sites in nearly 99% of Reference Sequence database genes, with valid and standardized methods for processing, quality control, and analysis. [37][38][39] The chips range in their coverage, from the first-generation Illumina 27K introduced in 2008 to the Illumina 450K in 2011 and the EPIC Chip in 2016, which cover roughly 27,000, 450,000 and 850,000 CpG sites, respectively. Unlike Sanger sequencing of bisulfite-converted DNA, which allows for assessment of DNA methylation across the entire genome, it is important to note that the Illumina arrays used in the studies we reviewed only survey a fraction of the genome with a bias toward CpG sites in promoter regions. 40 EWAS can generate thousands of significant hits, so it is much more difficult to determine if they are replicable. Still, it is equally important to determine how reliable this mode of inquiry is. To assess whether there was overlap between significant CpG sites in EWAS, we aggregated all available CpG sites and identified which genes had significantly differential methylation in two or more studies. We found 180 CpG sites that were hits in two studies and four that were hits in three studies (p value < 0.05, Tables 1, 2 and S2). Although these replicable hits provide targets for molecular follow-up experiments, it is important to note that this represents a small amount of overlap between significant CpGs, because several studies had over 1000  significant hits. Discrepancies in significant CpGs may be in part due to differences in tissue types as methylation signatures vary by cell type. 41 Of the EWAS investigating AUD, seven were in blood, five were in brain, one was in liver, and one was in buccal cells with some using more specific components of blood or brain (Table 1). Other factors that may explain these discrepancies include differences in AUD diagnostic tools, differences in demographic characteristics of patient populations and small sample sizes.

| RECENT MOLECULAR DISCOVERIES
Despite the low replicability across EWAS, several potential targets for further study emerged from our analysis. In particular, there were four genes that were shown to be differentially methylated in three different studies (Table 2). We searched major studies of alcohol consumption for all of the genes replicated in three AUD studies. 26 49 and increases in NF-κB have been shown to be associated with neuroinflammation, chronic alcohol consumption and alcohol-induced hepatic inflammation. [50][51][52] Epigenetic modifications in HNRNPA1 may help mediate this relationship between NF-κB and AUD, as well as explain some of the transcriptional differences associated with alcohol consumption. 53 Lipase maturation factor 1 (LMF1) was also differentially methylated in three studies of AUD in blood, brain and buccal cells 44,54,55 along with one study of alcohol consumption in blood. 46 LMF1 is a chaperone that activates vascular lipases necessary for lipid clearance, and thus, mutations have been associated with hypercholesteremia. 56,57 Little is known about the other two most common differentially methylated genes. Leucine rich repeat containing 20 (LRRC20) was differentially methylated in both studies of AUD 44,54,55 and a study of alcohol consumption. 46 It has been associated with autoantibodyindependent changes in interferon-alpha levels, and leucine-rich repeats have been associated with immune responses, but the function of this protein is far from understood. 58  were 15 pathways that were significant in at least two studies and three that were significant in three studies (Table 3). Aside from pathways related to immune regulation, there are several related to basic biological processes, such as signaling, that are involved in immune responses as well as many other cellular functions.
The finding that immune system processes emerge across molecular and pathway analyses provides an interesting human parallel to research into the role of the immune system in rodent models of AUD. As most of the studies in this review with meaningfully large sample sizes were in blood, however, these results should be interpreted with caution. DNA in blood is extracted from leukocytes, so there is a bias in our results toward sampling immune cells for methylation analyses in blood. Still, the importance of immune pathways was relevant across tissue types. Like epigenetics, the immune system represents an interface between the genome and the environment and thus may also integrate genetic variants with environmental signals to affect neurobiological processes. 62 Though far from painting a full picture of AUD pathology, animal studies have found alcohol to cause neuroinflammation, which has in turn been associated with AUD-related psychiatric symptoms. In rodents, chronic, heavy alcohol exposure has been shown to lead to alcohol dependence and voluntary self-administration, which suggests that understanding the physiological effects of ethanol may lead to a better understanding of AUD psychopathology. In parallel, these models have shown neuroinflammation to develop as a result of ethanol exposure, and neuroinflammation has been independently associated with symptoms of psychiatric disorders often comorbid with AUD such as depression. 63,64 Evidence suggesting that the inflamma- which was also repeatedly found to be differentially methylated in pathway analyses of EWAS data. However, it should be noted that these genes are just the first surfacing from initial EWAS studies, and likely there will be multiple others, as the epigenetic architecture of AUD is complex.

| LIMITATIONS AND FUTURE DIRECTIONS
Moving forward, there are several limitations to the studies we reviewed that may have straightforward solutions. For example, most of the studies suffer from small sample sizes, which may explain the lack of replicability across studies. There are also differences in tissue type, diagnostic criteria, preprocessing, and array technology, which precluded meta-analysis in this instance. In addition, statistical methods varied widely across studies, with many not controlling for smoking, which is often comorbid with AUD and which also has strong epigenetic effects. 67 Studies also varied on their significance thresholds, as well as whether they used standard corrections for cell type proportion, multiple testing, batch effects and false discovery rates. If standardized procedures are developed for quality control, normalization, exclusion of low-quality samples and methylation outliers, and statistical controls for confounding variables, the limits of small sample sizes may be overcome with valid and comprehensive meta-analyses.
Other barriers will require more profound advances to overcome. More broadly, given that alcohol consumption is a core component of AUD, it will be important to distinguish epigenetic signatures that may predispose individuals to AUD from those that may result from alcohol consumption. As Witt et al. demonstrated, many of the AUD-associated DNA methylation sites in peripheral tissues were reversed after a short period of abstinence, which suggests many sites identified in studies of AUD are likely associated with alcohol consumption rather than the etiology of AUD. This is important because genetic studies of AUD have found the genetic variants associated with AUD and alcohol consumption to be distinct. 69 More specifically, genetic variants associated with AUD appear to be more closely related to variants associated with psychiatric disorders, while those associated with alcohol consumption appear to be more closely associated with metabolism. 69 Because data on alcohol consumption are much easier to collect, studies of DNA methylation in alcohol consumption have, on the whole, been much larger than those in AUD.
For example, Liu et al. 42 looked at alcohol consumption using a sample size (n = 13,317) much larger than all of the AUD studies combined.
Dugue et al., 46 Wilson et al., 43   Going forward, it will also be important to understand the functional relevance of the molecular targets identified in these studies.
This will require integration of genetic, epigenetic and gene expression data. One strategy to better investigate the relationship between genetic variants and epigenetic changes is to incorporate methylation quantitative trait loci (mQTLs), which are genetic variants that may affect DNA methylation patterns. EWAS could integrate genetic data to determine if mQTLs, or genetic variants associated with DNA methylation in specific regions, are related to DNA methylation patterns in AUD. Metastable epialleles, which are alleles that are particularly vulnerable to environmental exposures and show differential expression in genetically identical individuals, could also be explored.
It will also be important to better understand how DNA methylation relates to gene expression. DNA methylation has traditionally been thought to repress transcription. However, increasing evidence shows that its role may be more complicated and it will be important to not only identify DNA methylation but also begin to understand how this methylation affects transcription. This will also require distinguishing between methylation and hydroxymethylation, because hydroxymethylation may be associated with demethylation or have a different effect on gene expression than methylation at any given site.
Although existing bisulfite sequencing methods do not distinguish between the two, oxidative bisulfite sequencing can be used to remove hydroxymethylated sites in order to assay methylation alone.
The methylation results can then be compared to traditional bisulfite sequencing to discern levels of hydroxymethylation. Though more labor intensive, use of both bisulfite and oxidative bisulfite sequencing in AUD may help to distinguish between these two types of methylation.
In addition to understanding how DNA methylation affects gene transcription, we must understand the temporal order of methylation and functional significance of affected proteins to understand its biological effects. Animal models of AUD or AUD-related symptoms such as compulsive alcohol administration can provide a system in which to where AUC represents a measure of biomarker accuracy on a scale from 0 to 1, with 1 being the most accurate. Liang et al. 76 also found methylation to have potential as a biomarker. They identified 143 CpG sites that predicted phosphatidylethanol (PeTH), a phospholipid metabolite shown to persist in blood for up to 3 weeks after alcohol consumption, with an AUC of 0.9 in a training dataset and an AUC of 0.8 in a replication dataset. 76 The studies had few overlapping predictors and both require further validation. It also remains unclear whether epigenetic biomarkers would be stable over longer periods of time or have higher resolution than PeTH. Still, biomarkers of alcohol consumption are a promising area of research.
Recently, methylation risk scores, the DNA methylation equivalent to polygenic risk scores, have emerged as a specific type of biomarker. With regard to alcohol consumption, methylation risk scores have been shown to be much better at predicting consumption than polygenic risk scores, explaining roughly 12.5% of phenotypic variance compared to polygenic risk scores, which explain roughly 0.7% of variance. 77 Larger EWAS may enable the development of AUD biomarkers and methylation risk scores. Biomarkers of AUD may help clinicians identify AUD in patients so they can better implement appropriate interventions. In research, biomarkers could lead to more objective classifications of AUD, which would in turn improve the validity of studies. In addition, biomarkers of AUD and alcohol consumption could be used in tandem to better address patients' needs and to better parse alcohol consumption and AUD psychopathology in research studies.

| CONCLUSIONS
The assessment of DNA methylation in AUD is a rapidly evolving and promising new area of research. In our review, we identified 184 genes that were consistently differentially methylated in at least two EWAS, with four of them, HNRNPA1, LMF1, LRRC20 and PLEKHG4B, replicated in three EWAS. In addition, we compared gene ontological analyses and found 15 pathways overlapping between studies, the most common relating to immune regulation and basic cellular processes.
Though differences in diagnostic criteria, methylation assays and preprocessing precluded a meta-analysis in this instance, our analysis identified promising targets for future epigenetic and genetic studies of AUD and provided an update on where the field has come and important questions that must be resolved in the future. The biggest challenge going forward with EWAS will be identifying the functional significance of DNA methylation patterns and parsing the epigenetic signatures that may predispose individuals to AUD from those that result from confounding variables such as smoking or from alcohol consumption itself. The most immediate application of DNA methylation data appears to be in studying the function and therapeutic relevance of the genes that are differentially methylated across studies and in the development of biomarkers of AUD that may yield more accurate tools for clinical and research-related diagnostic purposes.