Transancestral GWAS of alcohol dependence reveals common genetic underpinnings with psychiatric disorders.

Liability to alcohol dependence (AD) is heritable, but little is known about its complex polygenic architecture or its genetic relationship with other disorders. To discover loci associated with AD and characterize the relationship between AD and other psychiatric and behavioral outcomes, we carried out the largest genome-wide association study to date of DSM-IV-diagnosed AD. Genome-wide data on 14,904 individuals with AD and 37,944 controls from 28 case-control and family-based studies were meta-analyzed, stratified by genetic ancestry (European, n = 46,568; African, n = 6,280). Independent, genome-wide significant effects of different ADH1B variants were identified in European (rs1229984; P = 9.8 × 10-13) and African ancestries (rs2066702; P = 2.2 × 10-9). Significant genetic correlations were observed with 17 phenotypes, including schizophrenia, attention deficit-hyperactivity disorder, depression, and use of cigarettes and cannabis. The genetic underpinnings of AD only partially overlap with those for alcohol consumption, underscoring the genetic distinction between pathological and nonpathological drinking behaviors.

INTRODUCTION

Excessive alcohol use is a leading contributor to morbidity and mortality. One in 20 deaths worldwide is attributable to alcohol consumption, as is 5.1% of the global burden of disease[1]. Alcohol dependence (AD), as defined by the Fourth Edition of the American Psychiatric Association Diagnostic and Statistical Manual of Mental Disorders (DSM-IV)[2], is a serious psychiatric disorder characterized by tolerance, withdrawal, loss of control over drinking and excessive alcohol consumption despite negative health and social consequences. Among alcohol drinkers, 12% meet criteria for DSM-IV AD during their lifetimes[3]. In the United States, only 25% of those with AD ever receive treatment[4].

AD is moderately heritable (49% by a recent meta-analysis)[5] and numerous genome-wide association studies (GWAS) have aimed to identify loci contributing to this genetic variance (see[6] for a review). According to one study, common SNPs are responsible for as much as 30% of the variance in AD[7], but few have been identified to date. Variants in the genes responsible for alcohol metabolism, especially ADH1B and ALDH2, have been strongly implicated[8-13]. The association between AD (and related drinking phenotypes) and rs1229984, a missense SNP (Arg48His) in ADH1B that affects the conversion of alcohol to acetaldehyde, represents one of the largest common-variant effect sizes observed in psychiatry, with the His48 allele accelerating ethanol metabolism and affording approximately 3-fold reduction in likelihood of AD across numerous studies[8,10]. Another functional polymorphism, rs671 in ALDH2 (Glu504Lys), strongly affects alcohol metabolism by blocking conversion of acetaldehyde to acetate and has an even stronger effect on risk for AD, but is rare except in some Asian populations[8,12,13] ADH1B and ALDH2 polymorphisms, however, only explain a small proportion of the heritable variation in AD in populations of European or African ancestry.

In this study, the Substance Use Disorders working group of the Psychiatric Genomics Consortium (PGC-SUD[14]) compiled the largest numbers of carefully diagnosed alcohol dependent individuals and alcohol-exposed controls to date, from both case-control and family studies. These included substantial numbers of both European ancestry (EU, N = 46,568, including 38,686 unrelated individuals) and admixed African-American ancestry (AA, N = 6,280, including 5,799 unrelated individuals) subjects. AD diagnoses were derived from clinician ratings or semi-structured interviews following DSM-IV[2] criteria. Each study was subjected to stringent quality control (QC) before conducting GWAS within each population of each study, followed by a genome-wide meta-analysis. We estimated the SNP-heritability (h) of AD and examine the extent to which aggregate genetic variation in AD is related to traits from 45 other GWAS, including continuous measures of alcohol consumption. We also examined whether polygenic risk scores (PRS) derived from these analyses predicted alcohol dependence and related measures of problem drinking in three independent samples.

RESULTS

GWAS meta-analyses:

The trans-ancestral discovery meta-analysis of GWAS of AD in 28 cohorts (Table 1; Supplementary Table S1) identified a genome-wide significant (GWS; p < 5E-8) association in the ADH gene cluster on chromosome 4 (Figure 1; Table 2). Examining this locus in each population (Figure 2), rs1229984 in ADH1B was the strongest associated variant from the analysis in EU (z = −7.13, p = 9.8E-13), while rs2066702, also in ADH1B, was the most significant variant in AA (z = −5.98, p = 2.2E-9). Trans-ancestral modelling reinforced the robust effects of rs1229984 and other ADH1B SNPs on liability to AD across inverse-variance weighted, random effects, and Bayesian models (Supplementary Figure S1, Supplementary Table S2).

Table 1:

Descriptive statistics for cohorts in the meta-analysis of AD.

				European (EU)				African - American (AA)
		Male (%)	Ages (years)	N Total		N Unrelated		N Total		N Unrelated
Dataset	PMID			Case	Control	Case	Control	Case	Control	Case	Control
Case-control: Logistic Regression
Comorbidity and Trauma Study (CATS)	23303482	56%	18-67	572	817	572	817	--	--	--	--
Christchurch Health and Development Study (CHDS)	23255320	48%	16-30	112	500	112	500	--	--	--	--
Collaborative Study of the Genetics of Alcoholism - case-control cohort (COGA-cc)	20201924	54%	18-79	583	363	583	363	--	--	--	--
Family Study of Cocaine Dependence (FSCD)	18243582	51%	18-60	266	174	266	174	255	241	255	241
German Study of the Genetics of Alcoholism (GESGA)	19581569	65%	18-84	1314	2142	1314	2142	--	--	--	--
Gene-Environment Development Initiative - Great Smoky Mountains Study (GEDI-GSMS)	8956679	57%	9-26	42	565	42	565	--	--	--	--
Center on Antisocial Drug Dependence (CADD)	25637581	70%	13-20	400	577	400	577	51	51	51	51
Phenomics and Genomics Sample (PAGES)	28371232	57%	18-74	37	523	37	523	--	--	--	--
Collaborative Study on the Genetics of Nicotine Dependence (COGEND Nico)	17158188	34%	25-82	135	272	135	272	46	232	46	232
COGEND - Study of Addiction: Genetics and Environment (COGEND SAGE)	20202923	37%	18-77	311	225	311	225	104	103	104	103
Spit For Science	24639683	36%	>18	252	1863	252	1863	74	841	74	841
National Institute on Alcohol Abuse and Alcoholism Intramural (NIAAA)	n/a	67%	>18	442	206	442	206	404	110	404	110
Mayo Clinic Center for the Individual Treatment of Alcohol Dependence (CITA)	25290263	55%	≥18	378	646	378	646	--	--	--	--
Alcohol Dependence in African Americans (ADAA)	n/a	57%	18-69	--	--	--	--	794	297	794	297
Family-based, twins and sibs: Generalized Estimating Equations (GEE)
Brisbane Longitudinal Twin Study (BLTS)	23187020	43%	18-30	60	938	51	546	--	--	--	--
GEDI - Virginia Twin Study on Adolescent Behavioral Development (GEDI-VTSABD)	9294370	38%	8-32	209	503	188	318	--	--	--	--
Minnesota Center for Twin and Family Research (MCTFR)	23942779	41%	16-21	609	2100	553	1274	--	--	--	--
Center for Education and Drug Abuse Research (CEDAR)	21514569	63%	16-34	59	200	54	152	--	--	--	--
Swedish Twin Registry (STR)	23137839	47%	40-83	76	8311	76	6112	--	--	--	--
Yale-Penn	24166409	58%	16-79	1094	301	1004	252	--	--	--	--
Family-based, large/complex pedigrees: Logistic Mixed Model
Collaborative Study of the Genetics of Alcoholism - family cohort (COGA-fam)	23089632	45%	12-88	605	682	168	138	--	--	--	--
Australian Alcohol and Nicotine Studies (OZ-ALC-NAG)	21529783	45%	18-82	1571	3069	1111	805	--	--	--	--
Irish Affected Sib Pair Study of Alcohol Dependence (IASPSAD)	15770118	50%	17-84	721	1814	436	1802	--	--	--	--
Yale-Penn	24166409	51%	16-79	--	--	--	--	1607	1070	1263	933
Summary statistics
Netherlands Study of Depression and Anxiety / Netherlands Twin Registry (NESDA/NTR)	18197199	31%	>18	390	1633	390	1633	--	--	--	--
Finnish Nicotine Addiction Genetics Project (NAG-Fin)	17436240	52%	30-92	439	1137	439	1137	--	--	--	--
FinnTwin12 (FT12)	17254406	47%	20-27	88	874	88	874	--	--	--	--
National Longitudinal Study of Adolescent to Adult Health (Add Health)	25378290	47%	24-34	768	2981	768	2981	--	--	--	--
Helsinki Birth Cohort Study (HBCS)	16251536	43%	56-70	36	1583	36	1583	--	--	--	--
Total				11569	34999	10206	28480	3335	2945	2991	2808

Overview of numbers of alcohol dependent cases and controls from each cohort in the current analysis, including the number of genetically unrelated individuals. Cohorts are listed by study design and analysis method. Sample sizes are listed after QC exclusions and stratified by ancestry group. PubMed identifiers (PMID) are listed for previous publications describing each cohort, along with the percentage of male samples and the age range in the cohort.

Figure 1:

Manhattan plot of discovery trans-ancestral meta-analysis showing strong evidence for rs1229984 in ADH1B.

Results from the discovery meta-analysis of all cohorts (Ncase=14,904, Ncontrol=37,944) for association of genome-wide SNPs with AD under a fixed effects meta-analysis weighted by effective sample size. Dashed red reference line indicates genome-wide significance after correction for multiple testing (p < 5E-8).

Table 2:

Top 10 loci from the meta-analyses of alcohol dependence by ancestry

						A1 Allele Freq.		INFO score		Effect size (OR)		Discovery meta-analysis p-value
SNP	CHR	BP	A1	A2	Gene	EU	AA	EU	AA	EU	AA	EU	AA	Trans
Top clumped variants in trans-ancestral meta-analysis (14,904 cases, 37,944 controls)
rs7644567*	3	29201672	A	G	RBMS3	0.964	0.705	0.96	1.00	--	1.229	3.94E-04	6.64E-06	1.36E-08
rs2066702	4	100229017	A	G	ADH1B	--	0.215	--	0.99	--	0.731	--	2.21E-09	2.21E-09
rs1229984	4	100239319	T	C	ADH1B	0.040	0.014	0.90	0.91	0.486	0.912	9.79E-13	3.48E-01	2.18E-11
rs1789912	4	100263942	T	C	ADH1C	0.418	0.132	1.00	1.02	1.106	1.211	1.98E-07	1.32E-03	1.47E-09
rs6827898	4	100295863	A	G	(ADH region)	0.123	0.112	0.96	0.94	1.145	1.270	5.21E-07	9.31E-04	2.97E-09
rs894368	4	100309313	A	C	(ADH region)	0.309	0.386	0.99	0.96	0.887	0.981	1.93E-08	9.73E-01	3.30E-07
rs2461618	7	68667233	A	G	--	--	0.088	--	0.98	--	0.669	--	6.30E-07	6.30E-07
rs116338421	8	145761256	C	G	ARHGAP39	--	0.172	--	0.97	--	0.755	--	4.86E-07	4.86E-07
rs79171978	12	17798824	C	G	--	0.099	0.027	0.99	0.99	1.201	1.016	5.47E-08	8.18E-01	5.98E-07
rs8017647	14	32456358	T	C	--	0.792	0.565	1.00	0.99	0.901	0.923	8.05E-06	4.71E-02	1.03E-06
Top clumped variants in African ancestry meta-analysis (3,335 cases, 2,945 controls)
rs5781337	1	223883425	CA	C	--	0.263	0.212	0.98	0.93	1.007	0.664	8.85E-01	1.62E-07	6.59E-02
rs143258048	3	75982870	A	AC	ROBO2	--	0.028	--	0.88	--	0.490	--	1.86E-06	--
rs3857224	4	100129685	T	C	ADH6	0.315	0.585	0.99	1.00	0.970	0.814	2.40E-01	5.86E-07	2.36E-03
rs2066702	4	100229017	A	G	ADH1B	--	0.215	--	0.99	--	0.731	--	2.21E-09	2.21E-09
rs2461618	7	68667233	A	G	--	--	0.088	--	0.98	--	0.669	--	6.30E-07	6.30E-07
rs116338421	8	145761256	C	G	ARHGAP39	--	0.172	--	0.97	--	0.755	--	4.86E-07	4.86E-07
rs79016499	11	93010988	T	C	--	--	0.066	--	0.93	--	1.729	--	1.36E-06	--
rs10784244	12	62035165	G	A	--	0.153	0.484	1.00	1.00	1.041	1.226	6.26E-02	1.04E-06	2.49E-04
rs17199739	16	25444288	G	A	--	0.176	0.096	0.99	0.96	0.994	0.693	4.25E-01	1.11E-06	8.66E-03
rs740793	17	3846353	G	A	ATP2A3	0.453	0.350	0.97	0.97	0.996	1.370	4.66E-01	1.48E-06	3.44E-01
Top clumped variants in European ancestry meta-analysis (11,569 cases, 34,999 controls)
rs1229984	4	100239319	T	C	ADH1B	0.040	0.014	0.90	0.91	0.486	0.912	9.79E-13	3.48E-01	2.18E-11
rs3811802	4	100244221	G	A	ADH1B	0.454	0.529	0.96	0.96	1.162	0.914	2.40E-08	2.19E-02	1.22E-04
rs113659074	4	100252308	T	G	ADH1B	0.068	0.093	0.98	0.95	0.800	1.166	1.54E-06	6.63E-02	2.99E-04
rs1229863	4	100252386	A	T	ADH1B	0.174	0.038	0.99	0.99	1.145	1.254	7.80E-07	4.26E-02	9.28E-08
rs1154445	4	100288521	G	T	(ADH region)	0.425	0.134	0.97	0.99	1.137	1.211	1.80E-07	2.63E-02	1.48E-08
rs6827898	4	100295863	A	G	(ADH region)	0.123	0.112	0.96	0.94	1.145	1.270	5.21E-07	9.31E-04	2.97E-09
rs894368	4	100309313	A	C	(ADH region)	0.309	0.386	0.99	0.96	0.887	0.981	1.93E-08	9.73E-01	3.30E-07
rs79171978	12	17798824	C	G	--	0.099	0.027	0.99	0.99	1.201	1.016	5.47E-08	8.18E-01	5.98E-07
rs4388946	12	17935154	C	A	--	0.240	0.297	0.99	0.98	1.137	0.950	7.14E-07	1.87E-01	7.05E-05
rs34929220	15	69769635	T	C	DRAIC	0.690	0.937	0.90	0.94	0.893	1.028	1.02E-06	8.38E-01	7.38E-06

Top 10 nominally independent variants from the discovery trans-ancestral (Trans.; Ncase=14,904, Ncontrol=37,944) meta-analysis and the discovery meta-analyses in African (AA; Ncase=3,335, Ncontrol=2,945) and European (EU; Ncase=11,569, Ncontrol=34,999) ancestry cohorts, respectively. Independent variants are identified based on clumping for LD (pairwise r2 < 0.1) in 1000 Genomes Project Phase 3 data[21]. EU results are clumped using European (EUR) ancestry reference samples, AA results are clumped using African ancestry reference samples from the American Southwest (ASW), and trans-ancestral results are clumped using merged EUR and African ancestry (AFR) reference samples. P-values and allele frequencies (Freq.) are reported from two-tailed tests of association with AD in fixed effects meta-analyses weighted by effective sample size. Bold p-values indicate genome-wide significance after correction for multiple testing within the analysis (p < 5E-8). Odds ratios (OR) and INFO scores are reported from the meta-analyses of the subset of unrelated individuals within each ancestry. Variants are sorted by chromosome (CHR) and base pair (BP) position for genome build hg19, with genes annotated by Ensembl VEP[49]. Allele frequency and OR are given with respect to allele 1 (A1).

SNPs included in the trans-ancestral meta-analysis were not conditioned on being analyzed in both the EU and AA analyses. For instance, a SNP of strong effect in one group may not be sufficiently common or well-imputed for analysis in the other ancestral group (e.g., rs2066702 is not found in non-African populations but is among the top 10 in the trans-ancestral analysis due to strong effects in the AA group). For rs7644567 (denoted with *), the SNP does not passed QC in a sufficient number of cohorts to meet the minimum sample size requirement for inclusion in the EU-only analyses – it is only represented among EU cohorts by summary statistics from two Finnish cohorts – but allele frequency, INFO score, and meta-analyzed p-values from the Finnish summary statistics are reported since they contribute to the trans-ancestral meta-analysis.

Figure 2:

Regional plots for the ADH1B locus (rs1229984) in the trans-ancestral discovery, African-American (AA), and European (EU), meta-analyses.

Results of fixed effects meta-analysis with effective sample size weights for the ADH1B locus in (A) all cohorts (Ncase=14,904, Ncontrol=37,944); (B) AA cohorts (Ncase=3,335, Ncontrol=2,945); and (C) EU cohorts (Ncase=11,569, Ncontrol=34,999). Red reference line indicates the genome-wide significance threshold after correction for multiple testing within each analysis (p < 5E-8). Within ancestry, colored points reflect the degree of LD (pairwise r2) to the index variant (indicated by a purple diamond) in 1000 Genomes Project reference data[21] for individuals of (B) African or (C) European ancestry, respectively. LD structures in the two ancestries differ, so for the trans-ancestral sample (A) LD is not given, indicted by gray points. Two-tailed tests used for all analyses.

Clumping the ADH locus for linkage disequilibrium (LD; r2 < 0.1 within 500 kb) suggested multiple independent signals in both populations, with the differing leading alleles reflecting different LD structures and allele frequencies in each population (Table 2, Supplementary Figure S2). Conditional analyses controlling for rs1229984 and rs2066702 had limited power, but results showed limited attenuation of effect sizes between marginal and conditional analyses, consistent with the existence of additional independent effects in the region (Supplementary Table S3; Supplementary Figure S3). Suggestive independent signals in the genotyped cohorts included trialleleic variant rs894368 (marginal z = −4.57, p = 4.9E-6; conditional z = −4.53, p = 5.8E-6) and insertion rs112346244 (marginal odds ratio = 0.912, SE = .024, z = −3.81, p = 1.4E-4; conditional odds ratio = 0.883, SE = .025, z = −5.05, p = 4.5E-7; Supplementary Table S3). Several additional variants that were prioritized in the conditional analysis, while not significant, were in moderate to strong LD with rs698 (marginal odds ratio = 1.115, SE = .021, z = 5.19, p = 2.1E-7; conditional odds ratio = 1.084, SE = .021, z = 3.78, p = 1.6E-4), a functional ADH1C variant with a role in AD[8,11].

A single novel SNP on chromosome 3, rs7644567, also reached GWS in the meta-analysis (z = 5.68, p = 1.36E-8; Supplementary Figure S4). Potential biological associations with rs7644567, including chromatin contacts (Supplementary Figure S5) and cerebellar expression of RBMS3, are summarized in Supplementary Information A9. However, rs7644567 did not replicate in two independent AA samples (Yale-Penn2 and COGA AAfGWAS) or the independent FINRISK cohort; all three replication cohorts estimating effects of the minor allele in the opposite direction of the discovery meta-analysis (Supplementary Table S4). The SNP is also rare in most EU samples (minor allele frequency [MAF] < 0.01), with the current GWAS results primarily attributable to AA cohorts, along with FinnTwin and NAG-Fin. The EU cohorts in the discovery meta-analysis show no evidence of association of AD with the SNPs in strongest LD with rs7644567 in African (rs13098461; z = 0.27, p = 0.79) or Finnish (rs9854300; z = 0.10, p = 0.92) reference samples (Supplementary Information A9). Based on the clear lack of replication there is insufficient evidence to conclude rs7644567 is associated with AD based on the current results.

There was limited genome-wide evidence for heterogeneity across all cohorts, within ancestry, between ancestries, or between study designs within ancestry (Supplementary Information A8; Supplementary Figures S6–S8). Evidence for inflation from population stratification or other confounding was also limited in the discovery meta-analysis (lambda = 0.962; Supplementary Figure S9) and within EU (lambda = 1.053, LD score regression [LDSR] intercept = 1.018) and AA (lambda = 1.007, LDSR intercept = 0.991-0.997; Supplementary Information A11). Gene-level association testing with MAGMA[15] did not identify any additional significant genes in EU or AA (Supplementary Table S5, Supplementary Information A12), likely due to lack of power.

Heritability and genetic correlations:

Liability-scale SNP-heritability of AD was estimated at h = 0.090 (SE = 0.019, z = 4.80, p = 8.02E-7) in the meta-analysis of unrelated EU samples. Exclusion of the ADH1B locus did not substantially modify this estimate (h = 0.089, SE = 0.0185). Nominally significant polygenic signal for the meta-analysis of unrelated AA individuals was observed based on LDSR with scores computed from 1000 Genomes African populations (z = 2.12, p = 0.017), but the quantitative estimate of h was unstable depending on the choice of reference panel, reflecting the challenge of correctly specifying LDSR and robustly modelling LD for the AA population (Supplementary Information A11).

Significant genetic correlation with AD in EU was observed for 17 traits after correction for multiple testing (p < 1.11E-3 for 45 tested traits; Figure 3; Supplementary Table S6). The largest positive correlations were with ever smoking tobacco (rg = 0.708, SE = 0.134, p = 1.3E-7) and lifetime cannabis use (rg = 0.793, SE = 0.217, p = 2.5E-4), and with other psychiatric disorders, especially schizophrenia (rg = 0.357, SE = 0.054, p = 3.2E-11), ADHD (rg = 0.444, SE = 0.097, p = 4.2E-6), and depression (rg = 0.561, SE = 0.085, p = 3.5E-11). Educational attainment (rg = −0.468, SE = 0.066, p = 9.7E-13) and age at first birth (higher values indicate that participants were older when they had their first child; rg = −0.626, SE = 0.104, p = 2.0E-9) showed significant inverse genetic correlation with AD suggesting that liability to AD risk was genetically related to lower educational attainment and lower age at which the participant had his or her first child.

Figure 3:

Genetic correlations between 45 traits and alcohol dependence in Europeans.

Genetic correlation results from LD score regression (LDSR) with the meta-analysis of AD in unrelated EU individuals (Ncase=10,206, Ncontrol=28,480). After Bonferroni correction, significant correlations are observed with 17 traits and disorders (p < 1.1E-3; bold), with nominally significant results for 8 additional traits and disorders (p < .05; italics) based on two-tailed tests of the estimated genetic correlation with block jackknife standard errors. Error bars indicate 95% confidence intervals, with arrows indicating intervals extending above 1 or below −1. Vertical gray reference line corresponds to the null hypothesis of no genetic correlation with AD. Phenotypes are organized by research domain.

Unexpected patterns of genetic correlation were observed when comparisons were made to other alcohol-related measures, indicating that those measures reflect aspects of alcohol use that are genetically distinguishable. AD was genetically correlated with alcohol consumption in a meta-analysis of the Alcohol Genome-wide Association (AlcGen) and Cohorts for Aging and Research in Genomic Epidemiology Plus (CHARGE+) consortia[16] (rg = 0.695, SE = 0.155, p = 6.9E-6) but only modestly with alcohol consumption from the recent large UK Biobank analysis[17] (rg = 0.371, SE = 0.092, p = 5.2E-5). No significant genetic correlation was observed between AD and a recent GWAS of the Alcohol Use Disorders Identification Test (AUDIT) in a 23andMe cohort[18] (rg = 0.076, SE = 0.171, p = 0.65), perhaps due to the low levels of drinking and drinking-related problems in that population[18]. AD is, however, nominally genetically correlated with GWAS of delay discounting in the 23andMe sample[19] (rg = 0.487, SE = 0.178, p = 6.0E-3).

Association with ADH1B expression:

Based on the strong observed association with rs1229984 and rs2066702 we examined whether other variants affecting ADH1B expression (eQTLs) were also associated with AD using GTEx V7 results (https://www.gtexportal.org/)[20]. Three variants, rs11939328 (EU p = 0.78, AA p = 0.98, Trans p = 0.78), rs10516440 (EU p = 3.97E-6, AA p = 1.97E-3, Trans p = 4.72E-8), and rs7664780 (EU p = 0.87, AA p = 0.083, Trans p = 0.405), were selected after LD-informed clumping and the exclusion of variants in LD (r2>0.1) with the GWS coding alleles rs1229984 and rs2066702. Of these, only rs10516440 (AD conditional analyses: EU p = 1.34E-3, AA p = 0.013, Trans p = 7.44E-5) was a significant multi-tissue eQTL in random effects analysis for ADH1B (SFE = 319.4, SHet = 27.6, p = 1.4E-76), ADH1A (SFE = 139.4, SHet = 6.6, p = 6.72E-33), and ADH1C (SFE = 167.3, SHet = 8.9, p = 1.9E-39). Rs10516440 is a LD proxy (r2 > 0.9) of rs6827898 (Table 2) in populations of European and African descent. These variants are both located in an intergenic region in the ADH gene cluster between ADH1C and ADH7. In line with the fact that the protective coding alleles are associated with increased activity of the enzyme encoded by ADH1B, the major allele rs10516440*A was associated with increased ADH1B expression and reduced AD risk.

Associations with other GWS loci:

We examined results for the eight independent variants associated at GWS levels with alcohol consumption in the UK Biobank[17] (Supplementary Table S7). Among the UK Biobank findings, three of the four reported variants in the ADH region of chromosome 4 (rs145452708 – a proxy for rs1229984 with D’=1, rs29001570 and rs35081954) were associated in the present study with AD (p ranging from 3.5E-5 – 2.3E-10) with sign concordant effects; the remaining variant was excluded from our analysis due to MAF < 0.01. The UK Biobank lead variant in KLB, rs11940694, was nominally associated with AD (p = 0.0097), though this does not surpass multiple testing correction for the eight GWS alcohol consumption loci. We see little evidence (p > 0.2) for association of AD with the reported loci at GCKR and CADM2, which may be due to differences in power for the given effect size or because these genes exert an influence on liability to consume alcohol but not later problems. The locus on chromosome 18 showed limited regional association with AD, but the index variant was not present in our analysis because it no longer appears in the 1000 Genomes Phase 3 reference panel[21].

Polygenic Risk Score (PRS) analyses:

PRS based on our meta-analysis of AD were significantly predictive of AD outcomes in all three tested external cohorts. PRS derived from the unrelated EU GWAS predicted up to 0.51% of the variance in past month alcohol use disorder in ALSPAC (p = 0.0195; Supplementary Figure S10A) and up to 0.3% of problem drinking as indexed by the CAGE screener in GS (p = 7.9E-6; Supplementary Figure S10B). PRS derived from the unrelated AA GWAS predicted up to 1.7% of the variance in alcohol dependence in the COGA AAfGWAS cohort (p = 1.92E-7; Supplementary Figure 10C).

Importantly, PRS derived from the unrelated EU GWAS showed much weaker prediction (maximum R2 = 0.37%, p = 0.01; Supplementary Figure S10D) in the COGA AAfGWAS than the ancestrally-matched AA GWAS-based PRS despite the much smaller discovery sample for AA. In addition, the AD PRS also still yielded significant variance explained after controlling for other genetic factors. Prediction of CAGE scores in GS remained significant and showed minimal attenuation (R2 = 0.29%, p = 1.0E-5) after conditioning on PRS for alcohol consumption derived from UK Biobank results[17]. In COGA AAfGWAS, the AA PRS derived from our study continued to predict 1.6% of the variance in alcohol dependence after inclusion of rs2066702 genotype as a covariate, indicating independent polygenic effects beyond the lead ADH1B variant (Supplementary Information A14).

Power analysis:

Power analyses indicated that the current meta-analysis is expected to have at least 41% power to detect very common variants (MAF ≥ 0.25) with odds ratios ≥ 1.10 at p < 5E-8 and 63% power for p < 1E-6 (Supplementary Figure S11). Power at p < 1E-6 is relevant because only 5 loci reach that threshold in the current meta-analysis. Power is lower for less common variants (MAF ≤ 0.05) even with odds ratios ≥ 1.20 at p < 1E-6 (60% power) and p < 5E-8 (38% power).

For perspective, power computations using the observed distribution of top effects for other large GWAS of polygenic traits suggest that we observe significantly fewer genome-wide significant loci for AD than would be expected if the loci had true effect sizes and allele frequencies similar to schizophrenia (expected: 25.4 loci, 95% CI 21-30) or obesity (expected: 8.9 loci, 95% CI 6-12), but not fewer than would be expected for effect sizes similar to major depression (Supplementary Information A10, Supplementary Table S8).

DISCUSSION

To our knowledge, this is the largest GWAS of rigorously-defined AD, comprising 14,904 AD individuals and 37,944 controls. We identified known loci in ADH1B that differed between EU and AA, as well as novel genetic correlations between AD and psychiatric disorders (e.g., schizophrenia), tobacco and cannabis use, and social (e.g., socio-economic deprivation) and behavioral (e.g., educational attainment) outcomes. Analyses also revealed a genetic distinction between GWAS results for alcohol consumption and AD. Although larger sample sizes can be amassed by focusing on quantitative measures of consumption, only the upper tail is relevant to AD (as a medical diagnosis) and even that does not capture other aspects of disordered drinking (e.g., loss of control, withdrawal) directly. Conversely, cases derived from electronic medical records (e.g., ICD codes) result in a high rate of false negatives, while self-screening instruments (e.g. AUDIT scores) are best suited to analyses of disordered drinking when a sufficiently high threshold or score cut-off is applied to focus on severity. Our study has the advantage of greater diagnostic precision via use of semi-structured interviews to diagnose AD systematically in a majority of the constituent studies, and therefore greater interpretability in the context of clinically-important AD.

The genome-wide significant SNPs reaffirm the importance of functional variants affecting alcohol metabolism to the risk of AD. The top association in ADH1B, rs1229984, is a missense variant that is amongst the most widely studied in relation to alcohol use, misuse and dependence[8-10]. The resulting amino acid substitution (Arg48His) increases the rate at which alcohol dehydrogenase 1B oxidizes ethanol to acetaldehyde[8]. Studies on Asian populations in which the derived allele is common demonstrated strong protection against the development of AD[8,9,13]. In EU and AA, the protective allele is present at much lower frequencies (EU MAF = 0-4%, AA MAF < 1%), nevertheless, recent large-scale studies have shown an association between this locus and alcohol consumption and problems at GWS levels in EU with similar effect size[8-10]. The lead variant in AA cohorts, rs2066702 (Arg370Cys), is another functional missense variant in ADH1B, and it also encodes an enzyme with an increased rate of ethanol oxidation[8]. The allele encoding Cys370 is common in AA, but rare in other populations[8]. Our results clearly show that these two different functional SNPs in ADH1B both affect risk for alcoholism, with their relative importance dependent upon allele frequency in the population studied. There is a suggestion of additional independent effects in the chromosome 4 region, but larger studies will be needed to evaluate this.

The only other locus to reach significance was rs7644567 on chromosome 3, primarily driven by AA cohorts. The locus failed to replicate in two small, independent AA samples, and in the only European cohort with even a modest allele frequency (FINRISK) the effect was in the opposite direction. There have also been discussions about whether the standard GWAS significance threshold should be applied to the more genetically diverse African-ancestry cohorts[22,23] and the possibility of confounding from non-linear relationships between the phenotype and ancestry-informative markers like rs7644567 in admixed samples[24], all of which increase our skepticism regarding this finding. There is, therefore, insufficient evidence at this time to conclude that rs7644567 is associated with alcohol dependence. Analyses of much larger samples of African ancestry will be needed to resolve this.

Despite limited SNP-level findings, there is significant evidence for polygenic effects of common variants in both EU and AA cohorts. The estimated h2g = 0.09 for AD in EU is only modestly lower than those recently reported for alcohol consumption (h2g = 0.13)[17] and AUDIT scores (h2g = 0.12)[18], and comparable to estimates derived for cigarettes-per-day[25]. Our h2g estimate is lower than a prior report[7], likely reflecting a combination of differences in estimation method (GREML vs. LDSR) and greater heterogeneity in ascertainment strategy across samples in the current study (see[26-28]). The latter is especially relevant in comparing h from that prior single cohort to our meta-analysis that included cohorts with a wide range of ages at ascertainment, cultural environments, and ascertainment strategies, including enrichment for other substance use disorders. Similar to other psychiatric disorders (e.g. schizophrenia), a much larger sample size will potentially aid in overcoming across-sample heterogeneity and capture a greater proportion of genetic variance.

Comparing our GWAS to recent GWAS of alcohol consumption measures suggests that the liability underlying normative patterns of alcohol intake and AD are only partially overlapping. Genome-wide, genetic correlations were significantly < 1 with log-scaled alcohol consumption by participants in AlcGen and CHARGE+ Consortia cohorts[16] (rg = 0.695) and in the UK Biobank[17] (rg = 0.371). We also observe only partial replication of the 8 loci significantly associated with consumption in the UK Biobank, with strongest results from SNPs in the ADH region, including a proxy for rs1229984. In addition there was no significant correlation with GWAS of log-scaled AUDIT scores in 23andMe participants[18] (rg = 0.076). Subsequent analyses suggest these estimates are sensitive to sample characteristics, with somewhat higher genetic correlations reported in analysis of alcohol consumption in the full UK Biobank[29] (rg = 0.75) and of AUDIT in combined data from 23andMe participants and UK Biobank[30] (rg = 0.39). Importantly, initial UK Biobank data inclusion of a subset of participants recruited for a study of smoking and lung function in the first analysis[17], which may have resulted in collider bias[31] and contributed to the initial lower genetic correlation.

One key factor in interpreting the differences between these traits and AD is that the distribution of consumption levels and AUDIT scores can be highly skewed in population samples, with most individuals at the low (non-pathological) end of the spectrum. This effect may be especially pronounced among the older, healthy volunteers of the UK Biobank cohort[32] and in the 23andMe cohort, which is more educated and has higher socioeconomic status than the general US population[18]. We hypothesize that the variants that affect consumption at lower levels may differ substantively from those that affect very high levels of consumption in alcohol dependent individuals, who are also characterized by loss of control over intake[33]. This appears to be the case in studies that used specific cut-offs to harmonize AUDIT scores with AD data[30,34]. The larger of these studies[30] reports that the genetic correlation between AD and AUDIT scores is maximized at an AUDIT cutoff ≥ 20 (with controls defined as those scoring ≤ 4; rg = 0.90). Interestingly, that study also found that a score reflecting items related to problem drinking (AUDIT-P) resulted in a stronger genetic correlation (rg = 0.64) than a score related to alcohol consumption alone (rg = 0.33). The strong genetic correlation of AD with lower educational attainment and lower socio-economic status (i.e. higher Townsend deprivation), in contrast to positive genetic correlations of education with consumption[17] and AUDIT scores related to consumption[30], further underscore this distinction between normative/habitual levels of alcohol intake and diagnosed AD, at least in the respective populations studied.

The current analysis identified robust genetic correlation of AD with a broad variety of psychiatric outcomes. This correlation is strongest for aspects of negative mood, including neuroticism and major depression, as also seen in twin studies[35,36] and through recent specific molecular evidence for pleiotropy[37,38]. Taken together with evidence from other recent genomic studies[37], and null correlations for other GWAS of alcohol consumption, but not for measures of problem drinking (e.g., AUDIT-P), these findings suggest that major depression may primarily share genetic liability with alcohol use at pathological levels.

AD was also strongly genetically correlated with poor educational and socioeconomic outcomes, and marginally correlated with measures of risk-taking. Nominally significant genetic correlations with delay discounting (i.e. favoring immediate rewards), risk-taking, and the strong genetic correlation of AD with ADHD, cigarette smoking and cannabis use may similarly reflect a shared genetic factor for risk-taking and reduced impulse control. Common genetic liability to early, risky behaviors is characteristic of both AD[39] and age of first birth[40]. The observed negative genetic correlation with age of first birth is consistent both with risk-taking and with the significant genetic correlations of AD with lower socioeconomic status, as indexed by higher neighborhood Townsend deprivation score, and lower educational attainment. Lower socioeconomic status is correlated with both AD[41] and age at first birth[42] and the current study suggests that shared genetic liabilities may be one potential mechanism for their observed relationship. However, the question of whether these genetic correlations represent causal processes, horizontal pleiotropy, or the impact of unmeasured confounders should be explored in the future[43].

Lower genetic correlations were observed for most biomedical and anthropometric outcomes. Liver enzymes GGT and ALT, once proposed as possible biomarkers for alcohol abuse[44], showed only nominal evidence for genetic correlation with AD and neither survived multiple testing correction. Notably, we did not find any association between AD and body-mass index (BMI). Negative genetic correlations with BMI were previously reported for both alcohol consumption[17] and AUDIT scores[18], but there is prior evidence that BMI has differing underlying genetic architecture in the context of AD and outside of that context[45]. The negative genetic correlations observed in those studies are consistent with studies of light to moderate drinking, which is also associated with healthier lifestyle behaviors, while heavy and problematic drinking is typically associated with weight gain[46].

This study benefits from precision in diagnostic assessment of AD, known alcohol exposure in a majority of the controls, and careful quality control that excluded overlap of individuals between studies. Despite these strengths, our sample size was insufficient to identify additional GWS loci robustly. Power analyses indicate that additional SNPs associated with AD are likely to have small effect sizes, smaller than schizophrenia[47] and more consistent with more common psychiatric disorders (e.g. major depression[48]). This supports the pressing need for collection of large numbers of well characterized cases and controls. The differences between our results and the study of AUDIT scores[18] highlight that ascertainment and trait definition are critically important and must be taken into account. Careful study of how screening tools, such as the AUDIT, correlate to genetic liability to AD (as defined by DSM-IV or similar) could substantially boost sample sizes for future AD GWAS. There is also a continued need to characterize the genetic architecture of AD in non-EU populations.

We show a novel genetic distinction between drinking in the pathological range (AD) and habitual drinking that does not cross the threshold into pathology or dependence nor captures behavioral aspects of disordered drinking. Larger future samples will allow us to uncover additional pleiotropy between pathological and non-pathological alcohol use as well as between AD and other neuropsychiatric disorders.

METHODS

Samples:

The Substance Use Disorders working group of the Psychiatric Genomics Consortium (PGC-SUD[14]) collected individual genotypic data from 14 case/control studies and 9 family-based studies and summary statistics from GWAS of AD from 5 additional cohorts (Table 1). AD was defined as meeting criteria for a DSM-IV[2] (or, for one cohort, DSM-IIIR[50]; a very similar construct; Supplementary Note B1) diagnosis of AD. Diagnoses were derived either from clinician ratings or semi-structured interviews. Excepting three cohorts with population-based controls (N=7,015), all controls were screened for AD. Individuals with no history of drinking alcohol and those meeting criteria for DSM-IV alcohol abuse were excluded as controls where possible (Supplementary Information A1; Life Sciences Reporting Summary). This study was approved by the institutional review board (IRB) of Washington University in St. Louis and was conducted in accordance with all relevant ethical regulations. Each contributing cohort obtained informed consent from their participants and received ethics approvals of their study protocols from their respective review boards in accordance with applicable regulations.

Quality Control and Imputation:

Data for the cohorts that shared raw genotypes were deposited to a secure server for uniform quality control (QC). QC and imputation of the 14 case/control studies was performed using the ricopili pipeline (https://github.com/Nealelab/ricopili). For the 9 family-based cohorts, an equivalent pipeline, picopili (https://github.com/Nealelab/picopili), was developed for QC, imputation, and analysis appropriate for diverse family structures, including twins, sibships and extended pedigrees (Supplementary Information A2).

After initial sample and variant QC, principal components analysis (PCA) was used to identify population outliers for exclusion and to stratify samples in each study by continental ancestry. Identified EU and AA ancestry populations were confirmed by PCA using the 1000 Genomes reference panel[21] (Supplementary Figure S12). Ancestry within these 2 groups was accounted for with principal components. Final sample and variant QC, including filters for call rate, heterozygosity, and departure from Hardy-Weinberg equilibrium (HWE), was then performed within each ancestry group in each cohort. Samples were also filtered for cryptic relatedness and for departures from reported pedigree structures (Supplementary Information A3; Life Sciences Reporting Summary).

Each cohort was imputed using SHAPEIT[51] and IMPUTE2[52], using the cosmopolitan (all ancestries) 1000 Genomes reference panel consistent with prior recommendations[53] (see also[47,54,55]). Concordance of minor allele frequencies (MAF) with the reference panel was verified prior to imputation, with SNPs in EU cohorts compared to MAF in European population samples and AA cohorts compared to MAF in African population samples (Supplementary Information A4). Cryptic relatedness between cohorts was excluded after imputation (Supplementary Information A5). Imputed SNPs were then filtered for INFO score > 0.8 and allele frequency > 0.01 prior to analysis.

Association Analysis:

A GWAS of AD status was performed within each ancestry stratum of each sample using an association model appropriate for the study design (Table 1, Supplementary Table S1). For case/control studies, GWAS was performed using logistic regression with imputed dosages. For family-based studies of small, simple pedigrees (e.g., sibships), association with imputed genotypes was tested using generalized estimating equations (GEE). For more complex pedigrees, imputed genotypes were tested using logistic mixed models. Sex was included as a covariate, along with principal components to control for population structure (Supplementary Information A6, Supplementary Note B2, Supplementary Figures S13–S14).

In addition to this primary analysis, subsets of genetically unrelated individuals were selected from each family-based cohort (i.e. the most severe case in each family, by symptom count, was selected, followed by selection of unrelated/married-in controls) and used to perform a conventional case/control GWAS using logistic regression. This was used in place of the family-based GWAS for estimation of effect sizes and LD score regression analyses (Supplementary Table S2).

Genome-wide Meta-Analysis:

The primary discovery meta-analysis of all ancestry-stratified GWAS (Ncase = 14,904; Ncontrol = 37,944) was conducted in METAL[56]. As the different study designs (family vs. case-control) produced effect sizes that were not comparable, results were combined using weighting by effective sample size (Supplementary Information A7, Supplementary Note B3). Separate ancestry-specific discovery meta-analyses of EU (N = 46,568) and AA (N = 6,280) cohorts were also performed. Heterogeneity was evaluated across all cohorts and between study designs (Supplementary Information A8).

In addition to the discovery meta-analyses, we conducted meta-analyses for two design subsets. First, we performed sample size weighted meta-analysis of the subset of genetically unrelated individuals in EU (N = 38,686) and AA (N = 5,799) cohorts for use in LD score regression (LDSR) analysis. Second, we performed inverse-variance weighted meta-analysis of genetically unrelated individuals in genotyped cohorts to estimate within-ancestry effect sizes for EU (N = 28,757) and AA (N = 5,799). These effect sizes were then used to compare trans-ancestral fine mapping results using inverse-variance weighted fixed effects, random effects[57], and Bayesian[58] models (Supplementary Information A7). Supplementary Table S2 summarizes all of the meta-analytic models considered in the current analysis.

Replication:

As described below, a novel locus on chromosome 3 was genome-wide significant (GWS) in the trans-ancestral discovery meta-analysis. To seek replication, we examined the association between this locus and DSM-IV AD in two independent AA samples: Yale-Penn 2 (911 cases, 599 controls; tested using GEE) and COGA AAfGWAS (880 cases, 1,814 controls; tested using GWAF[59]). Association with AD status, broadly defined using hospital and death records, was also examined in the FINRISK cohort (1,232 cases, 22,614 controls) using Firth logistic regression[60] (Supplementary Information A1.4; Life Sciences Reporting Summary).

Power Analysis:

Post-hoc power analysis was performed for odds ratios ranging from 1.05 to 1.30 and across allele frequencies using CaTS[61] with the estimated effective sample size. Power analysis identifies the range of SNP effect sizes the current study was likely to detected at genome-wide significance if such effects exist. Additionally, we made specific comparisons to the distribution of effects for schizophrenia[47], obesity[62] and major depression[48] as meaningful benchmarks to understand the magnitude of effect sizes plausible for AD (Supplementary Information A10; Life Sciences Reporting Summary).

Heritability and Genetic Correlation Analysis:

LDSR analysis[63] was performed to estimate the heritability explained by common SNPs in meta-analyses of unrelated EU and AA samples, respectively. LDSR was performed using HapMap3 SNPs and LD scores computed from 1000 Genomes reference samples corresponding to each population (Supplementary Information A11). Conversion of h estimates from observed to liability scale[64] was performed assuming population prevalences of 0.159 and 0.111 for AD in alcohol-exposed EU and AA individuals, respectively[3]. Gene-level enrichments were also tested with MAGMA[15] (Supplementary Information A12).

Genetic correlation between AD and 45 traits from LD Hub[25] and other published studies[16-19,65-71] was examined using LDSR with the same unrelated EU meta-analysis (10,206 cases and 28,480 controls) and precomputed European LD scores. LDSR compares GWAS results for pairs of traits to estimate the correlation in the genetic liabilities explained by all common SNPs in the LD reference panel. To avoid increasing the multiple testing burden, redundant or highly-correlated phenotypes were reduced by manually selecting the version of the phenotype with the greatest predicted relevance to AD, largest sample size, or highest heritability (Supplementary Information A13).

Polygenic Risk Scores:

To test the generalizability of the current GWAS results, polygenic risk scores (PRS) were computed in three external cohorts (Supplementary Information A1.5; Life Sciences Reporting Summary). PRS computed from EU ancestry results were used to predict alcohol dependence in ALSPAC[72,73] and COGA AAfGWAS, and CAGE screener scores in Generation Scotland (GS)[74]. PRS based upon the AA results were used to predict alcohol dependence in COGA AAfGWAS (Supplementary Information A14).

Data availability:

Summary statistics from the genome-wide meta-analyses are available on the Psychiatric Genomics Consortium’s downloads page (http://www.med.unc.edu/pgc/results-and-downloads), including the source data for Figures 1 and 2. Individual-level data from the genotyped cohorts and cohort-level summary statistics will be made available to researchers following an approved analysis proposal through the PGC Substance Use Disorder group with agreement of the cohort PIs; contact the corresponding authors for details. Cohort data are also available from dbGaP except where prohibited by IRB or European Union data restrictions. Expression data used to evaluate variants in ADH1B is available from GTEx (https://gtexportal.org/home/). Hi-C data used to evaluate the chromosome 3 variant can be queried with HUGIn (https://yunliweb.its.unc.edu/hugin/). Publicly available genome-wide summary statistics used for testing genetic correlations are accessible through LD Hub (http://ldsc.broadinstitute.org/), or from the Psychiatric Genomics Consortium (http://www.med.unc.edu/pgc/results-and-downloads), the Social Science Genetic Association Consortium (SSGAC; https://www.thessgac.org/data), Enhancing Neuro Imaging Genetics through Meta Analysis (ENIGMA; http://enigma.ini.usc.edu/research/download-enigma-gwas-results/), and the Neale Lab (http://www.nealelab.is/uk-biobank); for availability of summary statistics from other studies contact the respective authors. The source data for Figure 3 is included in Supplementary Table S6.

Code availability:

Code for GWAS of case/control cohorts with ricopili is available at https://github.com/Nealelab/ricopili. Code for GWAS of family-based cohorts with picopili is available at https://github.com/Nealelab/picopili. Code and reference data for LD score regression analyses are available at https://github.com/bulik/ldsc. Effective sample size calculations were implemented using output from PLINK (https://www.cog-genomics.org/plink2), and GMMAT (https://content.sph.harvard.edu/xlin/software.html#gmmat) and geepack (https://cran.r-project.org/web/packages/geepack/index.html) in R (https://cran.r-project.org/); stand-alone software for this purpose hasn’t been written but example code is available from the first author by request.