Genomewide Association Study of Maximum Habitual Alcohol Intake in >140,000 US European- and African-American Veterans Yields Novel Risk Loci

Abstract

BACKGROUND

Habitual alcohol use can be an indicator of alcohol dependence, associated with a wide range of serious health problems.

METHODS

We completed a genomewide association study in 126,936 European-American (EUR) and 17,029 African-American (AFR) subjects in the Veterans Affairs Million Veteran Program (MVP) for a quantitative phenotype based on maximum habitual alcohol consumption (“MaxAlc”).

RESULTS

ADH1B, on chromosome 4, was the lead locus for both populations: for EUR, rs1229984 (p=4.9 × 10⁻⁴⁷); for AFR, rs2066702 (p=2.3 × 10⁻¹²). In the EUR, we identified three additional genomewide-significant (GWS) MaxAlc loci: on chromosome 17, rs77804065 (p=1.5 × 10⁻¹²), at CRHR1 (corticotropin-releasing hormone receptor 1); the protein product of this gene is involved in stress and immune responses; and on chromosomes 8 and 10. EUR and AFR samples were then meta-analyzed; the associated region at CRHR1 increased in significance to 1.02 × 10⁻¹³, and we identified two additional genomewide-significant loci, FGF14 (p= 9.86 × 10⁻⁹) (chromosome 13), and a locus on chromosome 11. Besides ADH1B, none of the five loci have prior GWS support. Post-GWAS analysis identified genetic correlation to other alcohol-related traits, smoking-related traits, and many others. Replications were observed in UKBiobank data. Genetic correlation between MaxAlc and alcohol dependence was 0.87 (p=4.78 × 10⁻⁹). Enrichment for cell types included dopaminergic and GABAergic neurons in midbrain, and pancreatic delta cells.

CONCLUSIONS

The present study supports five novel alcohol use risk loci, with particularly strong statistical support for CRHR1. Additionally, we provide novel insight regarding the biology of harmful alcohol use.

Introduction

The Million Veteran Program (MVP) is a US Department of Veterans Affairs (VA) initiative with a goal of recruiting at least one million VA healthcare beneficiaries, creating a database of genomic and phenotypic information useful for increasing understanding of health and disease (1). The sample is linked both to the VA’s extensive electronic health record (EHR) and to self-report survey information specific to the MVP study. The MVP is particularly valuable for elucidating health problems that are highly prevalent in military veterans, including alcohol use disorder (AUD) and harmful alcohol use.

DSM-IV alcohol dependence (AD), which in DSM-5 is the more severe type of AUD, is moderately heritable; genomewide association studies (GWAS) of AD and habitual alcohol use have been conducted in European (2–8), African (2, 5, 6), and East Asian (6, 9–13) ancestry populations. Most studies of AD diagnosis have been in small samples, but one reported on ~16,000 subjects (2), and the Psychiatric Genomics Consortium has completed a mega-analysis for AD (14). This AD mega-analysis included 14,904 AD cases and 37,944 controls from 28 case-control and family-based studies. Although this study consistently detected AD polygenic architecture, ADH1B risk alleles were the only loci identified, perhaps due to the heterogeneity across the cohorts included (14). Alcohol consumption is the major risk factor for AD and has medical importance per se. For example, alcohol consumption, even in the normal range (“social” drinking) bears a direct relationship to decline in several cognitive measures (15). Studies of large database samples, including the UK Biobank (4), have focused on alcohol consumption and their findings have implications for AD risk. Associations with variants mapped to genes that encode alcohol metabolizing enzymes – generally ADH1B variants in European- and African-ancestry (EUR and AFR) subjects (16, 17) as well as ALDH2*rs671 (18) in Asians – have been observed consistently. Some studies have reported associations at other loci with various alcohol-related traits (2, 4, 8); these reports are comparatively few. One meta-analysis of alcohol drinking (>105,000 EUR individuals) identified associations of daily alcohol intake with KLB, GCKR, and CDH13 (8). The GWAS of alcohol consumption in the UK Biobank sample (4) of >112,000 is the largest to date; this study considered only EUR subjects, and the phenotype was based on average weekly alcohol consumption. Genome-wide significant (GWS) associations were identified at several alcohol dehydrogenase (ADH) loci, in addition to other loci including GCKR, CADM2, and FAM69C.

In the present investigation, we studied the genetic architecture of an alcohol consumption phenotype – maximum habitual (“in a typical month…”) alcohol use, or MaxAlc -- in the MVP sample (19). We used two strategies to increase power for risk variant identification: a large sample size and substantial informativity of the phenotype. We included 143,965 MVP participants, and we used MaxAlc defined as a quantitative phenotype. A different phenotype, “maximum number of drinks consumed in any 24-hour lifetime period” often called MAXDRINKS, has previously been studied (5, 20). The trait definitions differ in that MaxAlc reflects typical habitual (daily) maximum usage, as opposed to the maximum use ever, which might be on a single occasion. Heaviness of habitual alcohol use may be more correlated with risk of AD than MAXDRINKS (21). Accordingly, we expected that our analysis would be informative regarding the mental and physical consequences of excessive alcohol consumption and alcohol dependence.

Methods and Materials

Subject recruitment

Participants were enrollees in the MVP (1) (Table 1). Users of the VHA healthcare system received invitational mailings, encounters with MVP staff while receiving clinical care, or both. Inclusion criteria were ability and willingness to provide informed consent. Research involving MVP in general is approved by the VA Central IRB; the current project was approved by local IRBs in Boston, San Diego, and West Haven.

Table 1.

Demographic characteristics of Million Veteran Program (MVP) European American (EUR) (N=126,936) and African American (AFR) (N=17,029) enrollees with completed Baseline and Lifestyle surveys

Age (years):	EA (N=126,936)	AA (N=17,029)
18–29	937 (0.7)	97 (0.6)
30 – 39	2,667 (2.1)	524 (3.1)
40 – 49	6,250 (4.9)	1,702 (10.0)
50 – 59	16,407 (12.9)	5,207 (30.6)
60 – 69	56,805 (44.8)	6,708 (39.4)
70 – 79	28,237 (22.2)	2,139 (12.6)
80+	15,519 (12.2)	646 (3.8)
missing	114 (0.1)	6 (0.0)
mean (SD)	66.2 (11.4)	60.3 (10.6)
median	66	61
Sex:
male	118,752 (93.6)	14,981 (88.0)
female	8,070 (6.4)	2,041 (12.0)
missing	114 (0.1)	7 (0.0)
Ethnicity (self-identified):
Hispanic	1,325 (1.0)	204 (1.2)
non-Hispanic	124,603 (98.2)	16,603 (97.5)
unknown	894 (0.7)	216 (1.3)
missing	114 (0.1)	6 (0.0)
Marital status:
Married	72873 (57.4)	6278 (36.9)
Divorced	21294 (16.8)	3603 (21.2)
Civil commitment	521 (0.4)	79 (0.5)
Never married	7814 (6.2)	1812 (10.6)
Widowed	8269 (6.5)	848 (5.0)
Separated	1713 (1.3)	941 (5.5)
Cohabitating	3006 (2.4)	257 (1.5)
missing	11446 (9.0)	3211 (18.9)

Two optional surveys were designed to augment data contained in the electronic health record. The MVP Baseline Survey elicits information regarding demographic factors, family pedigree, health status, lifestyle habits, military experience, medical history, family history of specific illnesses, and physical features. The MVP Lifestyle Survey contains questions from validated instruments, in domains selected to provide information on environmental exposures, dietary and other habits, sleep and exercise habits, and sense of well-being. This latter instrument includes the following item: “In a typical month, what is/was the largest number of drinks of alcohol (beer, wine, and/or liquor) you may have had in one day?” The response to this item was used to define the phenotype in the present study, referred to here as MaxAlc. All EUR and AFR subjects who responded to the questionnaire were included. Differences between respondents and non-respondents among MVP participants are shown in Supplementary Table 1.

Phenotype distribution is shown in Supplementary Figure 1.

Genotyping and Microarray

Genotyping was accomplished via a 723,305-SNP Affymetrix Axiom biobank array, customized for the MVP (1, 22). Additional information is provided in Supplementary Methods.

GWAS Analyses

We performed single variant tests using RVTEST(23) software, including the first 10 principal components, age, and sex as covariates in the linear regression association analyses, separately for EUR and AFR. The significance threshold was p=5×10⁻⁸.

Post-GWAS analyses

To investigate shared genetic and molecular mechanisms, we tested genetic overlap (i.e., shared risk alleles) of MaxAlc with a wide range of phenotypes. Genetic correlations were calculated using the LD score regression method (https://github.com/bulik/ldsc) (24). LDSC results regarding 232 traits were extracted from the data available at LD Hub v1.4.0 (http://ldsc.broadinstitute.org/ldhub/)(25). Genetic correlations for an additional 1,547 traits were calculated using the GWAS summary association results available at https://sites.google.com/broadinstitute.org/ukbbgwasresults; these GWAS used data from ~337,000 unrelated British individuals from the UK Biobank (26).

To explore further the functional role of the GWS variants identified, we conducted an expression quantitative trait locus (eQTL) analysis using GTEx V7 data (30). FDR correction, MAGMA (28), FUMA (31), and eQTL analyses are described further in Supplementary Methods.

Results

We identified an unusual instance of Hardy-Weinberg disequilibrium (HWD). ADH1B rs1229984, the most consistently associated alcohol risk variant in European populations (16), was initially excluded from analysis because it deviated from Hardy-Weinberg equilibrium expectations (HWEE) (p=1.46e-43). This variant is functional (32), presents very strong allele frequency differences among human populations (33), and has undergone selection in Asian and European populations (17, 34), although there is an open debate about the presence of convergent evolution in Europeans (35). Since ADH1B rs1229984 is the most relevant locus associated with alcohol drinking behaviors that has a very well established causative mechanism (17), we investigated the cause for HWD further to avoid unnecessary exclusion of this variant, which would have highlighted the association of other variants in the same region due to the LD without reflecting the real causal mechanism. This is described in Supplementary Results.

Primary GWAS Analysis

We observed 7.8% SNP-based heritability (p = 1.01×10⁻⁴⁰) calculated on the basis of the summary association data in “G1” EUR via LD score regression (LDSC). As with other large-scale GWAS(25), an inflated lambda_gc value was observed in the summary association data (λ_gc=1.16; Supplementary Figure 2). The LDSC intercept was 1.011 (SE = 0.0091), however, demonstrating that this inflation was due to polygenicity and not to population stratification, phenotype distribution, or other confounders. (25). In the smaller AFR sample (n=17,029), no effect of polygenicity was observed in the summary association data (λ_gc=1.01; Supplementary Figure 3). Four independent GWS regions were identified in “G1” EUR (Figure 1). The lead region was on chromosome 4, lead SNP rs1229984 (p=4.9×10⁻⁴⁷; Figure 2a); gene ADH1B, (beta subunit, class I alcohol dehydrogenase). GWS SNPs mapped to numerous loci in the region, so we performed conditional analysis for these loci using GCTA with EUR summary statistics and 1000G data as reference LD. This analysis confirmed that there are only four independent signals, i.e. no associated region reflected more than one independent signal. The other three associated regions map to chromosome 17, lead SNP rs77804065 (p=1.5×10⁻¹²; Supplementary Figure 4a), at CRHR1, corticotropin-releasing hormone receptor 1, with the protein product of this gene involved in stress and immune responses (numerous additional GWS SNPs were found in the chromosome 17 region, including variants that map to KANSL1, KAT8 Regulatory NSL Complex Subunit 1); chromosome 8, lead SNP rs7821592 (p=3.6×10⁻⁰⁸; Supplementary Figure 4b), closest gene XPO7, exportin 7, the protein product of which mediates nuclear export of proteins; and chromosome 10, lead SNP rs1577857 (p=4.2×10⁻⁰⁸; Supplementary Figure 4c), at LOC105378478, which has unknown function (closest gene, RNU6–53P).

Figure 1.

Manhattan plot

Figure 2. Regional Manhattan Plots:

A. Regional Manhattan plot, chromosome 4 ADH genes, EUR

B. Regional Manhattan plot, chromosome 4 ADH genes, AFR

C. Regional Manhattan plot, meta-analysis of EUR and AFR, chromosome 17 (CRHR1) region

The MVP includes mostly male subjects (93.6%). Although sex was included as a covariate, males and females differ in their prevalence of and genetic liability to AUDs (36, 37), so we evaluated whether inclusion of females affected the results substantively by repeating the analysis excluding females. No major differences were observed between GWAS of both-sexes and male-only samples (Supplementary Table 2).

In the AFRs, one GWS region was identified, lead SNP ADH1B*rs2066702 (2.29×10⁻¹²; Figure 2b). Conditional analysis, with AFR LD reference, confirmed that this reflects a single peak.

When EUR and AFR results were meta-analyzed (n=143,965 subjects total), we identified two additional GWS loci, uncharacterized LOC105376602 (p= 4.63×10⁻⁰⁸) on chromosome 11, and FGF14 (p= 9.86×10⁻⁰⁹) on chromosome 13 (Supplementary Figure 5b and 5e). In addition, the associated region at CRHR1 increased in statistical significance to p=1.02×10⁻¹³. Comparing EUR results (Supplementary Figure 4a) with EUR-AFR meta-analysis (Figure 2c), we observed different lead variants on chromosome 17, but they both indicated CRHR1 as a credible gene responsible for the association observed. Results are summarized in Table 2 and more extensively in Supplementary Table 3.

Table 2.

Most significant SNPs

EA
uniqID	rsID	chr	pos	P-value	start	end	nSNPs	nGWASSNPs	LeadSNPs	Gene
4:100239319:C:T	rs1229984	4	100239319	4.91E-47	100019089	100638613	111	107	rs1229984	ADH1B
8:21811530:C:G	rs7821592	8	21811530	3.63E-08	21777476	21869727	43	30	rs7821592	XPO7
10:110572259:G:T	rs1577857	10	110572259	4.15E-08	110462973	110635222	116	90	rs1577857	RNU6–53P
17:43810896:C:T	rs77804065	17	43810896	1.54E-12	43463493	44865603	3414	2979	rs77804065;rs199447	CRHR1
AA
uniqID	rsID	chr	pos	P-value	start	end	nSNPs	nGWASSNPs	LeadSNPs	Gene
4:100229017:A:G	rs2066702	4	100229017	2.29E-12	99986965	100263535	54	51	rs2066702	ADH1B
META
uniqID	rsID	chr	pos	P-value	start	end	nSNPs	nGWASSNPs	LeadSNPs	Gene
4:100239319:C:T	rs1229984	4	100239319	1.07E-49	100019089	100638613	110	106	rs1229984	ADH1B
8:21827162:C:T	rs2291317	8	21827162	2.48E-08	21777476	21869727	43	30	rs2291317	XPO7
10:110572259:G:T	rs1577857	10	110572259	4.15E-08	110462973	110635222	116	90	rs1577857	RNU6–53P
11:28648185:C:T	rs7931459	11	28648185	4.63E-08	28591168	28704399	172	84	rs7931459	LOC105376602
13:102868108:C:T	rs1360983	13	102868108	9.86E-09	102860527	102911712	51	44	rs1360983	FGF14
17:43785349:C:T	rs61667602	17	43785349	1.02E-13	43463493	44865603	3423	2976	rs61667602;rs1378358	CRHR1

To verify our results in an independent sample, we used summary association data from the AD GWAS conducted by the Psychiatric Genomics Consortium (PGC) (38). Although to date this is the largest AD GWAS, its effective sample size (39) is much smaller than the one used in our analysis (PGC = 31,819; MVP = 143,965) so there is low statistical power to replicate our findings. Nevertheless, considering our six GWS results in trans-ancestry meta-analysis, we observed genome-wide significant replication of the chromosome 4 ADH1B*rs1229984 association (p=2.18×10⁻¹¹), a nominal replication of chromosome 10 rs1577857 (p=2.44×10⁻³) and direction replication (i.e., the loci showed the same effect direction in both MaxAlc and AD) for all loci (Supplementary Table 4). We estimate that the probability to observe a direction replication of all six MVP-identified loci in PGC AD GWAS by chance is 1.7% (Supplementary Figure 6). Leveraging the polygenic architecture of the complex traits investigated, in the EUR sample, MaxAlc in MVP showed r_g=0.87 with AD in the PGC cohort (p=4.78×10⁻⁹) by LDSC. For additional replication, we investigated UK Biobank data regarding nine traits related to alcohol use (Supplementary Table 5). To identify the phenotypes most closely related to MaxAlc, we performed a genetic-correlation analysis and observed the strongest correlation with “Amount of alcohol drunk on a typical drinking day” (r_g=0.81, p=5.83×10⁻⁴⁰). Significant correlations were also observed with the other traits, including “Frequency of consuming six or more units of alcohol” (r_g=0.70, p=2.72×10⁻³⁰), “Ever been injured or injured someone else through drinking alcohol” (r_g=0.84, p=8.56×10⁻⁵), and “Ever had known person concerned about, or recommend reduction of, alcohol consumption” (rg=0.64, p=3.79×10⁻¹⁴). Considering the most strongly genetically correlated alcohol-use trait (i.e., “Amount of alcohol drunk on a typical drinking day”), we observed replications (Supplementary Table 4) for chromosome 4 rs1229984 (p=3.77×10⁻³²), chromosome 10 rs1577857 (p=0.027), and chromosome 17 rs77804065 (p=2.67×10⁻⁶) and rs61667602 (p=1.25×0⁻⁶).

We evaluated possible association of genes identified as associated in previous investigations of alcohol consumption phenotypes: GCKR, CADM2, FAM69C, KLB, and CDH13. No GWS results were observed, but suggestive results were observed in EUR at two of these loci, GCKR (min p=5.78×10⁻⁶) and KLB (min p=5.54×10⁻⁶), and nominally significant signals were observed in the remaining genomic regions (Regional Manhattan Plots for all five of these are in Supplementary Figure 7). This could be attributable to the polygenic architecture of complex traits, where loci have very small effect sizes, and a much larger sample size will be needed to replicate these loci at a genome-wide significance level; or to the difference between MaxAlc and AD, which has a high correlation with PGC AD (see above) and the consumption phenotypes wherein these other markers were identified.

Phenome-wide Genetic Correlations

LDSC revealed significant genetic correlations (FDR q<0.05) with 238 of nearly 1800 traits (Figure 3; Supplementary Table 6). The most significant observed correlations (Supplementary Table 3 shows all results at FDR q <0.05) were with respect to smoking and alcohol-drinking traits, where the top correlations were with current smoking status (positive correlation, r_g= 0.55, p=1.30×10⁻³⁹), the degree of past smoking (past tobacco smoking; negative correlation, r_g= −.46, p=5.49×10⁻³⁶) and “healthy” alcohol-drinking behaviors (e.g., alcohol usually taken with meals; negative correlation, r_g= −.50, p= 5.44×10⁻³⁴). Among the other highly significant correlations, several were related to level of education (e.g. years of schooling, r_g= −.37 p=1.53×10⁻²⁵) and socio-economic status (Townsend deprivation index, r_g= 0.53, p=3.69×10⁻²⁷). Numerous correlations were also found with measures of physical activity (e.g., no physical activity in the last four weeks, r_g=0.41, p=3.06×10⁻¹⁷). Other noteworthy correlations included mood swings (r_g=0.20, p=1.05×10⁻⁵) and risk taking (r_g=.20, p=2.74×10⁻⁵). Considering psychiatric traits, we observed significant genetic correlations with depressive symptoms (r_g = 0.22, p=4×10⁻⁴), schizophrenia (r_g = 0.13, p=0×10⁻⁴), and attention deficit hyperactivity disorder (r_g = 0.32, p = 0.023).

Figure 3.

Phenome-wide genetic-correlation analysis. Blue shades corresponds to significance strength, from white, non-significant (p > 0.05), to very light blue (p < 0.05), light blue (FDR q < 0.05), to blue (Bonferroni correction p < 2.81 × 10⁻⁴⁰), and dark blue (top-20 results). Phenotype labels are included for the top 20 results.

Gene-based association and Tissue and Cell-type Enrichment analysis

Gene-based association analysis and tissue and cell type enrichment results are shown in Figures 1 and 5 and described in Supplementary Methods.

Figure 5.

A. Statistical significance of the enrichments for tissue-specific gene expression. Detailed results are reported in Supplementary Table 4.

B. Statistical significances for cell types in human cortex from adult samples. “Hybrid” refers to a mixture of oligodendrocyte progenitor cells (OPC), oligodendrocytes, and neurons. Detailed results are reported in Supplementary Table 5.

C. Statistical significances for cell types in human midbrain. Detailed results and acronym legends are reported in Supplementary Table 6.

D. Statistical significances for cell types in human pancreas. Detailed results are reported in Supplementary Table 7.

E. Statistical significances for cell types in conventional dendritic cells (cDC). Detailed results are reported in Supplementary Table 8.

eQTL analysis

After applying a FDR 5% correction for the variants, genes, and tissues tested, we observed 212 significant eQTLs out of 2,855 tests conducted with respect to the GWS loci observed in the trans-ancestry meta-analysis. Considering the top CNS tissue for each eQTL surviving multiple testing correction (Table 3), we observed 37 significant results. Thirty-four relate to rs61667602 (chromosome 17), associated with the expression of multiple genes, where the strongest significance was mostly observed in the cerebellum transcriptomic profile (22/34). Additionally, we identified significant eQTLs with respect to rs1360983 on chromosome 13 (FGF14-AS2, top CNS tissue: spinal cord) and rs2291317 on chromosome 8 (BIN3, top CNS tissue: nucleus accumbens; FAM160B2, top CNS tissue: substantia nigra). Consistent with the strong linkage disequilibrium with the loci identified in the trans-ancestry analysis, similar eQTL results were observed with respect to the variants identified in the EUR analysis.

Table 3.

Significant eQTLs observed with respect to the GWS variants identified in trans-ancestry meta-analysis considering 13 CNS tissues. TSS: transcription start site.

rsID	Gene	TSS distance	slope	se	P value	FDR Q	Top CNS Tissue
rs1360983	FGF14-AS2	−178872	−0.361	0.078	2.00E-05	1.72E-04	Spinal cord (cervical c-1)
rs61667602	LRRC37A2	−803528	1.169	0.063	2.29E-36	4.69E-34	Cerebellum
	LRRC37A4P	157648	−1.096	0.061	4.20E-34	7.89E-32	Nucleus accumbens (basal ganglia)
	AC005829.1	−559054	1.132	0.068	1.67E-32	2.18E-30	Cerebellum
	KANSL1-AS1	−485593	1.202	0.072	3.39E-32	4.05E-30	Cortex
	AC005829.2	−552623	1.219	0.076	2.65E-31	2.45E-29	Cerebellum
	PLEKHM1	217234	−0.951	0.062	7.95E-30	5.43E-28	Cerebellum
	ARL17A	−871739	1.130	0.076	5.96E-29	3.80E-27	Cerebellum
	MAPK8IP1P2	105643	1.093	0.083	5.77E-25	2.30E-23	Cerebellum
	MAPK8IP1P1	−535623	1.157	0.085	1.25E-24	4.78E-23	Cerebellar Hemisphere
	DND1P1	122112	1.166	0.096	3.14E-22	9.89E-21	Cortex
	LINC02210	87655	0.690	0.059	3.02E-21	8.66E-20	Cortex
	SPPL2C	−136907	0.734	0.067	8.67E-20	2.18E-18	Cerebellum
	LRRC37A	−584750	1.004	0.093	3.42E-19	7.91E-18	Cerebellum
	AC091132.1	204723	−0.922	0.098	4.67E-16	7.83E-15	Cerebellum
	FAM215B	−854812	0.844	0.090	6.89E-16	1.14E-14	Cerebellum
	FMNL1	485759	−0.672	0.074	2.54E-15	3.87E-14	Cerebellum
	AC091132.3	176406	0.757	0.120	6.44E-09	6.76E-08	Cerebellar Hemisphere
	ARHGAP27	273562	0.459	0.078	4.94E-08	4.93E-07	Nucleus accumbens (basal ganglia)
	AC091132.2	255139	−0.573	0.101	1.03E-07	1.00E-06	Cerebellum
	AC008105.3	486194	−0.469	0.086	3.15E-07	2.99E-06	Cerebellum
	MAPT-AS1	−187617	0.508	0.096	6.23E-07	5.72E-06	Cerebellum
	MAPT	−186399	−0.316	0.065	3.62E-06	3.22E-05	Cerebellum
	CRHR1	86082	−0.399	0.098	1.01E-04	8.31E-04	Putamen (basal ganglia)
	RPS26P8	99440	0.698	0.169	1.11E-04	9.12E-04	Spinal cord (cervical c-1)
	KANSL1	−517384	0.355	0.090	1.42E-04	0.001	Cerebellum
	NMT1	656319	−0.211	0.057	3.44E-04	0.003	Cerebellum
	NSF	−882686	−0.158	0.043	3.93E-04	0.003	Cerebellum
	AC008105.1	466248	−0.247	0.072	7.87E-04	0.006	Cerebellum
	AC091132.4	162179	−0.387	0.118	0.001	0.010	Cerebellum
	AC015936.1	760069	0.575	0.173	0.002	0.011	Spinal cord (cervical c-1)
	CR936218.1	−327330	−0.393	0.122	0.002	0.013	Putamen (basal ganglia)
	ACBD4	575382	−0.175	0.057	0.003	0.020	Cerebellum
	PLCD3	574628	0.189	0.067	0.006	0.041	Hypothalamus
	ARL17B	−653781	0.351	0.127	0.007	0.043	Cerebellum
rs2291317	BIN3	−699499	0.212	0.074	0.005	0.034	Nucleus accumbens (basal ganglia)
	FAM160B2	−119533	0.251	0.089	0.007	0.043	Substantia nigra

Discussion

We report here findings from a GWAS of maximum habitual alcohol use from the US MVP sample, in EUR and AFR. In EUR, we observed 7.8% SNP-based heritability that is consistent with other large-GWAS of alcohol-related traits which also range from 5% to 10% (14). These SNP-based heritability estimates account for about 15–25% of the heritability reported by twin studies (40). The phenotype tested (i.e., “In a typical month, what is/was the largest number of drinks of alcohol (beer, wine, and/or liquor) you may have had in one day?”) has genetic overlap with both alcohol consumption (UK Biobank: “Amount of alcohol drunk on a typical drinking day” and “Frequency of consuming six or more units of alcohol”) and with alcohol misuse (PGC: DSM-IV Alcohol Dependence; UK Biobank: “Ever been injured or injured someone else through drinking alcohol” and “Ever had known person concerned about, or recommend reduction of, alcohol consumption”).

Our findings provide strong support for association in the chromosome 4 ADH region, for ADH1B*rs1229984, as has been reported multiple times previously (here with p=4.9×10⁻⁴⁷), spanning a lengthy chromosomal region (Figure 2a). A different signal at the same locus, rs2066702 (2.29×10⁻¹²; Figure 2b), was the only GWS result in AFR. We also report three additional regions in EUR, with prior varying, but never GWS, support: a region on chromosome 17 including CRHR1*rs77804065 (p=1.5×10⁻¹²; Figure 2c); chromosome 8, lead SNP rs7821592 (3.6×10⁻⁸), closest to XPO7; and at chromosome 10, lead SNP rs1577857 (4.2×10⁻⁰⁸), LOC105378478, which has unknown function. The trans-population meta-analysis added two additional novel GWS regions, FGF14 (p=9.86×10⁻⁰⁹) and LOC105376602 (p=4.63×10⁻⁰⁸)(19).

Lead SNP ADH1B*rs1229984 is a long-established risk locus from the pre-GWAS era that has been strongly confirmed by GWAS (2, 6). To identify rs1229984 as the lead variant, we needed to address a data-cleaning dilemma, as this variant was initially excluded from analysis on HWE criteria. Knowing the importance of the variant, we investigated the situation further, and discovered two subpopulations within the EUR, one with higher rs1229984 MAF (that clusters with Ashkenazi Jews(41)), and another much larger subpopulation with lower MAF. HWE criteria for this key variant were met within both of these individual subpopulations. In the initial quality control investigation, the violation of HWE expectations was, we conclude, attributable to this demonstrable violation of the random mating assumption (and not to a problem with data quality). We recommend that studies that may have excluded ADH1B*rs1229984 on HWE grounds examine this same issue.

This variant has many orders of magnitude greater support for association than the next-best-supported independent region on chromosome 17, lead SNP rs77804065, which maps to CRHR1, observed p=1.5×10⁻¹² in EUR. CHRH1 variants were previously implicated in candidate gene studies of alcohol use phenotypes(42, 43) and in an animal study regarding sensitivity to relapse into alcohol seeking induced by environmental stress (44). This GWS association signal maps to a well-known 900 kb inversion region (45) containing numerous other genes, some of which could also be considered MaxAlc candidate loci. The inversion is much less common in Africans (45, 46) consistent with the complex evolutionary history at this locus (47), so meta-analysis between EUR and AFR could potentially narrow the associated region greatly, if there is association information in that population as well, even if non-significant taken only in AFR. Indeed, the transpopulation meta-analysis showed that statistical significance increased by over an order of magnitude (to 1.02×10⁻¹³) with improved evidence for localization of the lead SNP at CHRH1 (Figure 2c). A similar phenomenon has been observed in narrowing associated regions for schizophrenia when meta-analyzing EUR and Asian GWAS results (48). Gene-based analyses and the replication in the UK Biobank provided additional evidence supporting CRHR1 as a risk locus.

On chromosome 8, rs7821592, the implicated locus is XPO7. Although this locus was identified as being of interest in a prior sparse “pooled GWAS” study of AD (49), and was identified in a study of AD comorbid with bipolar disorder (50), it has never previously been identified for these traits at anything approaching GWS. Finally, in EUR, rs1577857 (LOC105378478) on chromosome 10 has apparently not been reported previously. Although this variant is located in a non-coding RNA gene not previously associated with any human phenotype, the association in the MVP cohort was also replicated in PGC and UK Biobank cohorts. Additionally, the regulatory functional significance of this locus is supported by the fact that the variant is in a DNase I hypersensitivity site detected in twelve different cell types(51).

In AFR, we identified a single region led by ADH1B rs2066702 which, like rs1229984 in EUR, is well replicated (2).

The trans-population meta-analysis added two novel GWS loci, i.e., six in the meta-analysis, vs. only four in the European-only analysis (and one, overlapping with an EUR-associated region, GWS in AFR taken individually, albeit with a different SNP). These were FGF14 - Fibroblast Growth Factor 14 – at p=9.86×10⁻⁰⁹: a gene implicated in inherited cerebellar ataxias (52), among other traits, which regulates KCNQ2/3 potassium channels (53); and an uncharacterized RNA gene locus, LOC105376602 (at p=4.63×10⁻⁰⁸). FGF14 is particularly relevant because KLB, a locus previously identified as associated to alcohol consumption (4, 6, 8, 54) and replicated in MVP (min p=5.54×10⁻⁶), is a receptor that acts as a targeting signal for several FGF genes (55), suggesting the strong possibility of wider involvement of the FGF family in predisposition to alcohol consumption.

Thus, although our AFR sample was too small for novel locus identification when taken individually, trans-population meta-analysis was very valuable because of the differences in local LD (allowing improved CRHR1 region mapping) and additional association information for risk regions apparently in common between these populations.

The phenome-wide genetic-correlation analyses identified correlations with numerous traits including tobacco smoking behaviors, socioeconomic status, physical activity, reproductive behaviors, fat mass, personality traits, and, to a lesser extent, certain psychiatric disorders. Similar findings have been reported previously, even with small numbers of markers or with ADH1B*rs1229984 taken individually (56). These genetic relationships of MaxAlc are consistent with the pervasive role of alcohol use and abuse on human morbidity and mortality (57). Gene-based analysis, besides supporting CRHR1 as noted above, supported other genes associated in SNP-based analysis such as XPO7 and FGF14, as well as, for example, KANSL1, which maps to the same inversion region as CRHR1, and PDE4B (Phosphodiesterase 4B), previously implicated in other neuropsychiatric disorders. Tissue and human cell-type expression enrichments were noted for cerebellar hemisphere and cerebellum; dopaminergic and GABAergic neurons in human midbrain; and delta cells in pancreas. The cerebellar enrichment is particularly relevant with respect to the known effects of alcohol on this brain region: ethanol is the most common injurious agent to Purkinje cells (58, 59). In this context, inter-individual variability in the genetic regulation of cerebellum may be linked to the ability to drink large amounts of alcohol. Additionally, alcohol affects the type-A γ-aminobutyric acid (GABA_A) receptor, which mediates autocrine signaling mechanisms in pancreatic cells (60). Individuals with high resistance to the effects of ethanol on this system may be able to drink larger amounts of alcohol; subjects at risk for alcohol dependence tend to have lower levels of response to measures including body sway (61) which is presumably at least in part cerebellar in origin. In a mouse model, it was demonstrated that genetically-influenced differences in cerebellar alcohol response affect alcohol consumption (62). eQTL analysis provide further evidence for functional effects of risk loci, particularly those mapped to chromosome 17, in central nervous system, particularly cerebellum.

In summary, we mapped a) four risk loci for MaxAlc in EUR, of which only one (ADH1B) was previously known; b) one in AFR, which was previously known (a different marker in ADH1B than in EUR); and c) an additional two loci, both novel, in the transpopulation meta-analysis. MaxAlc is a clinically meaningful trait that differs from, but is genetically correlated with, DSM diagnosis of AUD. It is unclear to what extent the novel findings are due to the phenotype definition, or to the size and other characteristics of the clinical sample. MaxAlc, relating not merely to habitual alcohol use but to maximal habitual use, is more strongly related to the pathological range of alcohol use than some other measures such as the Alcohol Use Disorders Identification Test - Consumption (AUDIT-C) or MAXDRINKS (63). The negative correlation with “healthy” alcohol use behaviors, such as “alcohol usually taken with meals,” supports this interpretation.

Although our study is based on a large sample, we are still underpowered to conduct additional analyses to dissect the differences in the polygenic architecture of excessive drinking behaviors between sexes and age classes. MaxAlc, although a valid and useful phenotype, has previously been used only rarely. The high correlation with AUD per se may encourage more use in future studies, in the context of the results we report here. Additionally, the MVP uses an array that, while adequate for studies of EUR, is sparse for AFR and accordingly leaves much of the genome unstudied (64). This is the case because AFR are a genetically older population than EUR and have lower linkage disequilibrium genomewide; hence each SNP tends to query a shorter genomic region. For studies including large AFR populations, a more informative array would, ideally, be employed.

Finally, this study demonstrates the tremendous utility of the MVP sample for locus discovery. The large sample and informative set of surveys (combined with electronic health record data, which were not used here) will permit powerful and virtually unprecedented association studies of a vast array of traits and diseases. Furthermore, the inclusion of a sizeable sample of individuals of African descent contributes to additional locus identification opportunities.

Figure 4.

Manhattan plot, gene-based association results