A phenome-wide association and Mendelian randomisation study of alcohol use variants in a diverse cohort comprising over 3 million individuals.

Data collection was not performed as part of this study. Summary statistics from the PheWAS and MR analyses are presented in Supplementary Tables S11–S15 .

The funders, as listed in the Acknowledgements section, did not have any role in study design, data collection, data analyses, interpretation, or writing of the report.

Our cohort consisted of research participants from 23andMe, Inc., a direct-to-consumer genetics company. Demographic information about this sample is presented in Supplementary Table S2 . Ancestry was determined through an analysis of local ancestry, 39 as described in detail in the Supplementary Material . Only individuals who were categorised as being of European, Latin American, or African American ancestries (which were the largest available ancestral groups at the time of the analysis) based on empirical genotype data 39 were retained. Individuals of African American and Latin American ancestries are admixed with broadly varying contributions from Europe, Africa and the Americas, and sometimes other populations as well. The distributions of the length of segments of European and African American ancestry are very different between individuals of African American and Latin American ancestry, because of distinct admixture timing between the three ancestral populations in the two ethnic groups. Therefore, we trained a logistic classifier that takes one individual's length histogram of segments of African, European, and American ancestries, and predicts whether the individual is likely of African American or Latin American descent. Additional genotyping information is included in the Supplementary Material . Randomisation and blinding did not apply to the current study. An overview of the data collection process has been previously described. 40 All participants provided answers to online surveys available on the 23andMe platform. The traits were self-reported and pertained to 28 phenotypic categories (e.g., substance-use related, mental health, sleep, immune, cardiovascular, metabolic) that we identified. All 23andMe research participants included in the analyses provided informed consent and volunteered to participate in the research online, under a protocol approved by the external AAHRPP-accredited IRB, Ethical & Independent (E&I) Review Services. As of 2022, E&I Review Services is part of Salus IRB ( https://www.versiticlinicaltrials.org/salusirb ). We selected three protective SNPs for alcohol consumption ( ADH1B : rs1229984-T; rs2066702-A; ADH1C : rs698-T; Supplementary Figures S1–S3 ) based on their roles in alcohol metabolism and results from diverse large-scale GWAS ( 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ; see 3 , 4 for comprehensive reviews), that were available in the 23andMe cohort and passed the quality control criteria ( Table 1 , Supplementary Tables S3–S7 ). The functional SNP ALDH2 (rs671-A, Supplementary Figure S4 ), which is heavily studied in individuals of East Asian ancestry, could not be tested in this study due to near zero (<0.002) allele frequency in the populations studied, which did not include East Asians. Imputed variant statistics are shown in Supplementary Table S4 . Table 1 Alcohol variants included in the PheWAS and Mendelian randomisation analyses. rsID chr:pos:alleles Protective allele Gene Effect predicted by VEP eQTL in GTEx eQTL tissues eQTL p values Amino acid freq_EUR freq_LA freq_AA rs2066702 4:100229017:G:A A ADH1B missense variant Y ( ADH1A ) Adipose-Subcutaneous; Breast—Mammary Tissues; Skin—Sun Exposed (Lower Leg) 8.00E-07; 5.70E-06; 8.60E-05 R (Arg) > C (Cys) 0.002 0.017 0.189 rs1229984 4:100239319:T:C T ADH1B missense variant N – – H (His) > R (Arg), H (His) > P (Pro) 0.035 0.068 0.012 rs698 4:100260789:T:C T ADH1C intron variant ( ADH1B ), missense variant, downstream gene variant ( ADH1C ) Y ( ADH1C ) I (Ile) > F (Phe), I (Ile) > V (Val) 0.588 0.707 0.856 rs3114045 4:100252560:T:C T ADH1B intron variant ( ADH1B ) Y (RP11-571L19.8; METAP1 ) Adipose—Subcutaneous; Skin–Sun Exposed (Lower leg); Nerve—Tibial; Thyroid 3.00E-07; 2.70E-06; 1.80E-04; 2.70E-04 0.135 0.299 0.071 Shown are the alcohol-protective allele frequencies. VEP , variant effect predictor. Additional eQTL information is presented in Supplementary Tables S3 and S9 . Alcohol variants included in the PheWAS and Mendelian randomisation analyses. Shown are the alcohol-protective allele frequencies. VEP , variant effect predictor. Additional eQTL information is presented in Supplementary Tables S3 and S9 . We employed single-SNP PheWASs to examine associations between the three selected SNPs and 1316 traits in the European cohort, 962 traits in the Latin American cohort and 785 traits in the African American cohort from the 23andMe database ( Fig. 1 ). We excluded traits with <1,000 responses (15.73%, 8.84% and 11.46% from the European, Latin American and African American cohorts respectively), based on a prior simulation study for power analysis of PheWAS. 41 The final number of traits tested was 1109 in the European cohort, 883 in the Latin American cohort and 695 in the African American cohort, with the sample size ranging from 1,012 to 2,619,939 in the European cohort, 1,006–446,646 in the Latin American cohort, and 1,000–146,776 in the African American cohort ( Fig. 1 and Supplementary Table S8 ). Most traits had sample sizes >10,000, which provided power (80%) to detect modest associations (e.g., beta = 0.02; Supplementary Figure S5 ). For case–control comparisons, we computed association test results by logistic regression. For quantitative traits, association tests were performed by linear regression. We assumed additive allelic effects and included covariates for age at the time of the analysis (as determined by participant date of birth), sex, and the top five ancestry-specific principal components, to account for residual population structure within each group. Within each cohort, we used a 5% FDR Benjamini & Hochberg correction for multiple testing. 42 Fig. 1 Overview of the study . Four variants (rs1229984-T, rs2066702-A, rs698-T, rs671-A) were selected based on a systematic literature review of genes ( ADH1B, ADH1C, ALDH2 ) implicated in alcohol use disorder and other alcohol-related traits. rs671-A could not be tested due to extremely low allele frequency in the populations studied. We tested the association of the three remaining SNPs and 1109 traits available in 23andMe via a phenome-wide association study ( PheWAS ). We included traits with >1000 responses per cohort (EUR N = 1012–2,619,939; LA N = 1006–446,646; AA N = 1000–146,776). To assess evidence of a causal effect of more alcohol consumption by ethanol metabolising enzyme SNPs and PheWAS significant traits, a two-sample MR analysis and a one-sample replication Mendelian randomisation ( MR ) analysis were conducted. SNP effects associated with AUDIT Consumption and number of drinks consumed over the past 2 weeks were treated as exposures for the two- and one-sample MR analyses respectively. ∗In order to have a minimum of three SNPs for MR analyses, an additional ADH1B SNP (rs3114045) was added to replace replaced rs2066702, as this SNP was not available in the AUDIT Consumption GWAS. Overview of the study . Four variants (rs1229984-T, rs2066702-A, rs698-T, rs671-A) were selected based on a systematic literature review of genes ( ADH1B, ADH1C, ALDH2 ) implicated in alcohol use disorder and other alcohol-related traits. rs671-A could not be tested due to extremely low allele frequency in the populations studied. We tested the association of the three remaining SNPs and 1109 traits available in 23andMe via a phenome-wide association study ( PheWAS ). We included traits with >1000 responses per cohort (EUR N = 1012–2,619,939; LA N = 1006–446,646; AA N = 1000–146,776). To assess evidence of a causal effect of more alcohol consumption by ethanol metabolising enzyme SNPs and PheWAS significant traits, a two-sample MR analysis and a one-sample replication Mendelian randomisation ( MR ) analysis were conducted. SNP effects associated with AUDIT Consumption and number of drinks consumed over the past 2 weeks were treated as exposures for the two- and one-sample MR analyses respectively. ∗In order to have a minimum of three SNPs for MR analyses, an additional ADH1B SNP (rs3114045) was added to replace replaced rs2066702, as this SNP was not available in the AUDIT Consumption GWAS. In the PheWAS, we examined the association between three SNPs in ethanol metabolising enzyme genes, well-studied for their protective effects on alcohol consumption, 3 , 4 and various outcomes. Some of the associations could be the result of horizontal pleiotropy (i.e., a SNP is independently associated with both alcohol consumption and these other traits 43 ) or vertical pleiotropy, also known as causality (i.e., a SNP is associated with these other traits as a consequence of alcohol consumption 38 , 44 ). To explore whether the associations between alcohol consumption and these outcomes are causal, we conducted univariable two-sample MR analysis 37 , 38 , 45 , 46 , 47 (for a flowchart see Supplementary Figure S6 ), following the STROBE guidelines ( Supplementary Material ). We extracted the summary statistics (beta and SE) for the same SNPs used in the PheWAS from a previously published GWAS of alcohol consumption measured in an independent cohort (UK Biobank, N = 160,824 9 ;). Alcohol consumption was measured by the Alcohol Use Disorder Identification Test ( AUDIT ) Consumption scores (items 1–3 48 ), which we treated as our exposure. Summary statistics (beta and SE) for the PheWAS significant outcomes were extracted from the European cohort results. Exposure and outcomes were thus derived from similar underlying populations (i.e., European ancestry, middle-aged adults). We restricted the MR analysis to the European cohort because there are currently no well-powered GWAS of alcohol consumption in individuals of Latin American or African American populations that we could use for the analysis. The SNPs extracted were each independently ( r 2 < 0.27) associated with alcohol consumption at a genome-wide significant level ( p  < 5.00E-08). rs2066702-A was not available in the alcohol consumption GWAS. 9 Because some MR methods require a minimum of three SNPs, 49 we expanded our PheWAS results to include another ADH1B variant (rs3114045; Supplementary Table S9 , Supplementary Figure S6 ), which was well genotyped and imputed in the 23andMe cohort, and was significant in the alcohol consumption GWAS. 9 We evaluated instrument strength by assessing the proportion of variance in the phenotype explained by each instrument and the F-statistics, where a value under 10 may indicate a weak instrument. 50 Next, we harmonised the exposure-outcome datasets (i.e., consistent direction of effect between exposure and outcome associations, where the effect alleles for all SNPs were associated with more alcohol consumption), and applied an inverse variance weighted ( IVW ) analysis to give an overall estimate of causal effect. 51 The causal estimate of the IVW analysis indicates the causal increase in the outcome (or log odds of the outcome for a binary outcome) per unit change in the exposure. Because IVW assumes that all variants are valid instrumental variables in the absence of horizontal pleiotropy ( Supplementary Figure S10 ), any deviation from this assumption can introduce bias in MR analysis estimates if the SNPs affect the outcome via a different pathway. To evaluate the robustness of our interpretations, we compared the IVW results with two other methods known to be more robust to horizontal pleiotropy. Specifically, we used weighted median MR, which allows for 50% of the instrumental variables to be invalid, and MR-Egger regression, which relaxes the assumption that the effects of the variants on the outcome are entirely mediated via the exposure. 52 , 53 , 54 This is achieved by not constraining the intercept, where a non-zero intercept indicates directional horizontal pleiotropy. To test the suitability of the MR-Egger method, we calculated the I 2 GX statistic to quantify the degree of regression dilution bias due to measurement error of SNP-exposure effects. 50 Where there was evidence of violation of negligible measurement error ( NOME , I 2 0.6–0.9), MR-Egger was performed with simulation extrapolation ( SIMEX ) correction. 50 , 55 SIMEX relies on simulation to estimate or reduce bias due to measurement error, considering additional data sets with increasing measurement error variance. Lastly, we tested for heterogeneity between the individual SNPs included in the genetic instrument using Cochran's Q test of heterogeneity. To assess whether any single SNP was driving effect estimates, we conducted leave one SNP out MR analyses. All analyses were conducted using the TwoSampleMR package (v0.5.6) 56 in R (v 4.2.1). Triangulating results across multiple MR methods provides the strongest evidence of causal inference. 57 In our case, the IVW method was the main analysis and the other methods served as sensitivity analyses. We thus present the estimates from the IVW method in the manuscript; results for all three methods are shown in Supplementary Table S14 . Significance was determined after multiple testing correction based on the number of outcomes tested ( Supplementary Figure S6 ). Finally, we performed one-sample MR as a form of replication using the MendelianRandomization package (v. 0.6.0) 49 in R (v. 4.2.1). We included the summary statistics for alcohol consumption (i.e., number of drinks consumed over the past 2-weeks) as the exposure and any significant outcomes obtained via two-sample MR. Our analyses were not pre-registered, and therefore our results should be considered exploratory. 58

The European cohort included up to 2,619,939 individuals (56.20% female; mean age = 51.77 [SD = 18.71]); the Latin American cohort included 446,646 individuals (57.30% female; mean age = 42.89 [SD = 17.74]); and the African American cohort included 146,776 individuals (58.30% female; mean age = 44.27 [SD = 17.93]). Alcohol consumption (≥1 drink) in the past two weeks was reported by 65.29% (European cohort), 61.25% (Latin American cohort), and 57.16% (African American cohort). Additional sociodemographic and health-related outcome results can be found in Supplementary Table S2 . The list of PheWAS association results after 5% FDR correction is available in Supplementary Table S11 (European), Supplementary Table S12 (Latin American), and Supplementary Table S13 (African American). The standardised effect sizes were generally small ( Supplementary Figures S7–S9 ; European, β = −6.25 to 3.65; Latin American, β = −2.70 to 2.32; African American, β = −1.26 to 1.85), as is expected for a single SNP's effect on a complex trait. Overall, all findings that were observed in two or more cohort groups showed the same direction of effect and similar effect sizes ( Supplementary Figures S11–S14 ). rs1229984-T (protective allele (T) frequency: 0.06 EUR, 0.08 LA, 0.04 AA) showed the highest number of FDR-significant PheWAS associations ( Fig. 2 , Supplementary Table S11 ) in the populations studied, despite its low frequency in those populations. Fig. 2 FDR (5%)-significant PheWAS associations between rs1229984-T and a selection of traits from multiple categories in the European (EUR), Latin American (LA), and African American (AA) cohorts. The full list of FDR-significant associations is shown in Supplementary Tables S11–S13 . FDR (5%)-significant PheWAS associations between rs1229984-T and a selection of traits from multiple categories in the European (EUR), Latin American (LA), and African American (AA) cohorts. The full list of FDR-significant associations is shown in Supplementary Tables S11–S13 . In the European cohort, rs1229984-T was associated with 232 traits from each of the 28 categories we considered; we describe some of the most prominent findings below: Alcohol-related behaviours. The alcohol-protective rs1229984-T allele was most strongly associated, as expected, with alcohol-related traits [e.g., less alcohol consumption (β = −0.25, p 3 daily servings; β = −0.60, p  = 2.97E-44), less concern regarding a need to stop drinking alcohol (β = −0.45, p =  1.54E-144)]. We also found associations with conditions that are often comorbid with alcohol consumption, such as gastrointestinal disorders [e.g., less cirrhosis (β = −0.25, p  = 1.19E-09); less gastroesophageal reflux disease ( GERD ; β = −0.03, p  = 1.47E-08)], renal conditions [i.e., more kidney stones (β = 0.07, p  = 1.21E-22)], as well as respiratory conditions that tend to be associated with smoking, which commonly co-occurs with alcohol consumption [e.g., less chronic bronchitis (β = −0.11, p  = 3.61E-05)]. Central nervous system ( CNS ). We found associations between the protective allele (rs1229984-T) and mental health traits [e.g., lower rates of self-reported post-traumatic stress disorder (β = −0.04, p  = 2.21E-05), bipolar disorder (β = −0.03, p  = 4.12E-03), and depression (β = −0.01, p  = 5.43E-03)], and neurological traits [e.g., fewer reports of memory loss (β = −0.10, p  = 1.51E-05), lower rates of epilepsy (β = −0.04, p  = 9.79E-03), and less restless leg syndrome (β = −0.05, p  = 2.30E-07)]. In addition, we observed associations with sleep health [e.g., less insomnia (β = −0.03, p  = 6.05E-07), less sleep paralysis (β = −0.08, p  = 3.15E-16), less REM sleep behaviour disorder (β = −0.10, p  = 1.42E-09), less sleep apnea (β = 0.03, p  = 1.13E-04), higher sleep quality (β = 0.01, p  = 1.01E-09)]. Immune, cardiovascular, endocrine/metabolic. Many of the associations we observed were related to immune, cardio-metabolic, and endocrine/metabolic traits ( Fig. 3 ). For example, we found associations between rs1229984-T and multiple immune traits [e.g., fewer allergies (β = −0.07, p  = 4.80E-11), less Crohn's or ulcerative colitis (β = −0.11, p  = 7.45–16)]; cardiovascular traits [e.g., fewer self-reports of high blood pressure (β = −0.05, p  = 4.93E-23), less coronary artery disease (β = −0.06, p  = 4.60E-07)]; and metabolic/endocrine traits [e.g., less obesity (β = −0.05, p  = 3.11E-25), less hyperglycaemia (β = −0.08, p  = 7.14E-13), less type II diabetes (β = −0.05, p  = 7.68E-09), fewer self-reports of medications to reduce cholesterol (β = −0.06, p  = 1.57E-36), and less heart metabolic disease burden (β = −0.03, p  = 2.69E-42)]. Fig. 3 a ) Integrated bar chart featuring significant traits identified in both the PheWAS and MR analyses, grouped by categories. Each bar represents the number of outcomes for each category. The percentage represents the proportion of significant outcomes from the total possible number of outcomes within that specific category. b ) Integrated network analysis of PheWAS and MR results in the European cohort. Each node in the network denotes one of the 233 PheWAS or 35 MR significant traits (annotated), with colour indicating the category of the association, the width of the line denoting the magnitude (beta) of the effect size, and the size of the shape reflecting the association strength (log10 p values). a ) Integrated bar chart featuring significant traits identified in both the PheWAS and MR analyses, grouped by categories. Each bar represents the number of outcomes for each category. The percentage represents the proportion of significant outcomes from the total possible number of outcomes within that specific category. b ) Integrated network analysis of PheWAS and MR results in the European cohort. Each node in the network denotes one of the 233 PheWAS or 35 MR significant traits (annotated), with colour indicating the category of the association, the width of the line denoting the magnitude (beta) of the effect size, and the size of the shape reflecting the association strength (log10 p values). Other. We also found associations between rs1229984-T and several other traits, such as musculoskeletal [e.g., fewer fractures (β = −0.04, p  = 1.54E-04), less fibromyalgia (β = −0.07, p  = 5.84E-09)] and dermatologic traits [e.g., less rosacea (β = −0.03, p  = 1.57E-03)], eye health [e.g., less age-related macular degeneration (β = −0.07, p  = 1.66E-05)], and reproductive health [e.g., less preeclampsia or high blood pressure during pregnancy (β = −0.06, p  = 4.72E-07) and less endometriosis liability (β = −0.05, p  = 1.34E-06)]. Overall, the alcohol-protective rs1229984-T allele was associated with better health outcomes, including general health (β = 0.03, p  = 9.90E-55), less chronic fatigue (β = −0.08, p  = 2.270E-11), and less daily assistance required (β = −0.08, p  = 2.53E-03). However, we also found that 32 out of the 232 associations (13.79%) were with worse health outcomes, including more self-reports of lifetime tobacco use (β = 0.03, p  = 1.25E-09), a mental health trait [i.e., more emotional eating (β = 0.09, p  = 2.13E-06)], neurological conditions [e.g., more migraine with aura (β = 0.06, p  = 4.37E-11)], immune and endocrine/metabolic conditions [e.g., more Graves’ disease (β = 0.09, p  = 1.09–05), more gout (β = 0.13, p  = 4.74E-37), more hyperthyroidism (β = 0.14, p  = 8.46E-03)], infectious/parasitic diseases [e.g., more malaria (β = 0.30, p  = 3.93E-03)], conditions related to eye health [i.e., more myopia (β = 0.06, p  = 9.74E-09)], and several cancers [e.g., more skin cancer (β = 0.03, p  = 1.41E-04), more lung cancer (β = 0.04, p  = 1.65E-03)]. In the Latin American cohort, we identified similar findings: rs1229984-T was associated with 29 traits across 13 categories ( Supplementary Table S12 ): Alcohol-related behaviours. Again, the strongest associations of the protective allele were observed with alcohol-related traits [e.g., less alcohol consumption (β = −0.24, p  = 7.09E-216), more wine headaches (β = 0.29, p  = 1.68E-22), fewer concerns regarding a need to stop drinking alcohol (β = −0.42, p  = 3.44E-24)], and conditions related to alcohol consumption and smoking, such as respiratory disorders [i.e., less asthma (β = −0.10, p  = 1.37E-03)]. CNS . We found associations with neurological traits [i.e., less restless leg syndrome (β = −0.13 p  = 5.04E-07) and Parkinson's disease (β = −0.34 p  = 1.41E-04)]. Immune, cardiovascular, endocrine/metabolic. Many of the associations were again related to immune, cardio-metabolic, and endocrine/metabolic traits. For example, we found associations with immune traits [e.g., fewer allergies (β = −0.05, p  = 2.72E-07) and less Crohn's disease (β = −0.12, p  = 1.02E-03)]; cardiovascular traits [e.g., lower rates of high blood pressure (β = −0.08, p  = 4.94E-10) and fewer cardiovascular problems at an older age (β = −0.09, p  = 3.91E-04)]; endocrine/metabolic traits [e.g., less obesity (β = −0.07, p  = 1.35E-12), less hyperglycaemia (β = −0.10, p  = 1.30E-04), and less heart metabolic disease burden (β = −0.03, p  = 6.26E-16)]. Other. Again, we found associations with understudied traits such as musculoskeletal conditions [e.g., fewer fractures (β = −0.03, p  = 9.84E-04)]. We also noted that 3 out of the 20 significant associations (10.34%) were unexpected, including a mental health trait [i.e., more emotional eating (β = 0.12, p  = 8.57E-04)], a dermatologic trait [i.e., more keloids (β = 0.06, p  = 2.23E-04)], and a metabolic trait [i.e., more gout (β = 0.14, p  = 6.95E-06)]. In the African American cohort, which was the smallest cohort, rs1229984-T was associated with 7 traits ( Supplementary Table S13 ). The strongest associations were with alcohol-related traits [e.g., less alcohol consumption (β = −0.15, p  = 1.12E-12) and less preference for whisky (β = −0.19, p  = 1.37E-04)]. In addition, we found associations with similar immune [e.g., fewer allergies (β = −0.14, p  = 4.35E-08)] and musculoskeletal traits [i.e., fewer fractures (β = −0.11, p  = 1.43E-04)] observed in the other cohorts. In this cohort only, rs1229984-T was associated with lower height (β = −0.05 p  = 1.87E-05). rs2066702-A (protective allele (A) frequency: 0.002 EUR, 0.017 LA, 0.189 AA) was associated with less alcohol consumption in the European (β = −0.09, p  = 3.21E-07), Latin American (β = −0.09, p  = 1.82E-08), and African American (β = −0.08, p  = 3.78E-19) cohorts, but no other alcohol-related behaviours were found. In the Latin American cohort, rs2066702-A was associated with three additional traits, including a mental health trait [i.e., lower rates of self-reported depression (β = −0.08, p  = 7.04E-05)] and two metabolic traits [i.e., lower BMI (β = −0.05, p  = 2.50E-10) and lower weight (β = −0.04, p  = 1.54E-07)]. No other associations were found in the European and African American cohorts. rs698-T (protective allele (T) frequency: 0.61 EUR, 0.67 LA, 0.82 AA), was associated with less alcohol consumption in the European (β = −0.04, p  = 6.44E-136), Latin American (β = −0.05, p  = 7.39E-33) and African American (β = −0.04, p  = 1.94E-05) cohorts. In the European cohort, rs698-T was associated with 9 additional traits: Alcohol-related behaviours. The protective allele rs698-T was associated with two additional alcohol-related traits [i.e., more wine headaches (β = 0.02, p  = 9.48E-05), and fewer concerns regarding a need to stop drinking alcohol (β = −0.05, p  = 5.55E-12)]. CNS . We found an association with a sleep trait [i.e., less movement during REM sleep (β = −0.02, p  = 1.76E-04)]. Cardiovascular, endocrine/metabolic. We observed associations with the same cardiovascular [i.e., lower rates of high blood pressure (β = −0.01, p  = 1.86E-05)] and metabolic traits [i.e., less heart metabolic disease burden (β = −0.01, p  = 1.55E-04)] that were found to be associated with rs1229984-T. Other . rs698-T was associated with physical characteristic traits [i.e., height (β = −0.003, p  = 1.55E-04) and body type (β = −0.02, p  = 5.83E-05)]. Only one trait was associated with the protective rs698-T allele in the unexpected direction [i.e., more reports of lifetime tobacco use (β = 0.01, p  = 1.11E-04)]. In the Latin American cohort, rs698-T was associated with 2 additional traits, which included an alcohol-related trait [i.e., fewer concerns regarding a need to stop drinking alcohol (β = −0.08, p  = 4.66E-05)], and one metabolic trait [i.e., reduced heart metabolic disease burden (β = −0.01, p  = 1.27E-04)]. We did not observe any significant associations in the African American cohort. Using three SNPs in ethanol metabolising enzyme genes that were robustly and independently associated with alcohol consumption measured via AUDIT scores, two-sample MR in the European cohort provided evidence of a causal effect of genetically determined alcohol consumption on 35 outcomes (15.09% of the 232 PheWAS significant traits) after accounting for multiple testing correction and sensitivity analyses ( Supplementary Figures S10, S15–S18 , Supplementary Table S14 ). The proportion of the phenotypic variance of the exposure explained by each instrument (rs1229984-C, R 2  = 0.04; rs3114045-C, R 2  = 0.02; rs698-C, R 2  = 0.01) and the F-statistics were adequate (F > 32.58), indicative of valid instrument strength. Alcohol-related behaviours. MR results showed strong evidence of a causal effect of increased alcohol consumption on higher preference for whisky (IVW OR = 1.34, 95% CI = 1.25–1.44, p  = 5.39E-17), and increased concern about the need to stop drinking alcohol (IVW OR = 12.09, 95% CI = 10.11–14.46, p  = 1.00E-164). Gastrointestinal conditions that are often associated with alcohol consumption, such as cirrhosis (IVW OR = 3.90, 95% CI = 2.58–5.88, p  = 9.18E-11), elevated aspartate aminotransferase ( AST ), which is a clinical biomarker of liver damage (IVW OR = 1.90, 95% CI = 1.50–2.42, p  = 1.64E-07) and GERD (IVW OR = 1.17, 95% CI = 1.11–1.22, p  = 6.07E-10), showed evidence of being causally associated with alcohol consumption. CNS. We found evidence of a causal effect of increased alcohol consumption on one mental health trait [i.e., fear of speaking (IVW OR = 1.21, 95% CI = 1.11–1.30, p  = 2.88E-06)] and sleep alterations [e.g., REM sleep behaviour disorder (IVW OR = 1.76, 95% CI = 1.50–2.06, p  = 4.53E-12), poor sleep quality (IVW OR = 1.07, 95% CI = 1.05–1.10, p  = 7.14E-13)]. Cardiovascular, endocrine/metabolic traits. We found evidence of a causal effect of increased alcohol consumption on cardiovascular conditions [e.g., coronary artery disease (IVW OR = 1.43, 95% CI = 1.26–1.63, p  = 5.86E-08)], metabolic/endocrine conditions [e.g., obesity (IVW OR = 1.30, 95% CI = 1.24–1.36, p  = 3.32E-28), type 2 diabetes ( T2D , IVW OR = 1.35, 95% CI = 1.23–1.47, p  = 2.59E-11), high plasma glucose (IVW OR = 1.20, 95% CI = 1.11–1.29, p  = 6.17E-07), and heart metabolic disease burden (IVW OR = 1.17, 95% CI = 1.15–1.20, p  = 2.08E-51)]. Other. While the majority of the associations were positive (i.e., higher alcohol consumption increasing the odds for these outcomes), 28.57% of the associations were negative [e.g., lower self-reports of: ever using tobacco (IVW OR = 0.85, 95% CI = 0.79–0.91, p  = 1.49E-05), emotional eating (IVW OR = 0.61, 95% CI = 0.52–0.72, p  = 5.58E-09), migraines (IVW OR = 0.75, 95% CI = 0.69–0.81, p  = 2.84E-12), kidney stones (IVW OR = 0.67, 95% CI = 0.60–0.74, p  = 1.13E-13), gout (IVW OR = 0.48, 95% CI = 0.43–0.54, p  = 1.26E-09), and Grave's disease (IVW OR = 0.57, 95% CI = 0.46–0.70, p  = 2.47E-07)]. We identified evidence of causal effect of increased alcohol consumption on other domains, such as physical characteristics [e.g., heavy freckles (IVW OR = 0.74, 95% CI = 0.67–0.82, p  = 8.58E-09)], that may be indicative of residual biases not fully accounted for by sensitivity analyses. For all outcomes, the direction of effect was consistent across the three MR methods tested (IVW, weighted median, MR-Egger), and the high I 2 GX values of the SNP-exposure associations indicated low bias from regression dilution (i.e., attenuation towards the null of an association as a result of measurement error). There was some heterogeneity in the SNP effects, funnel plots identifying an influential potential outlier (rs1229984-C; Supplementary Figure S16 ); nonetheless, all results were still significant following leave-one-out analyses ( Supplementary Table S14 ). All of these associations were replicated in the one-sample MR ( Supplementary Table S15 ).

The societal and economic consequences of alcohol consumption are extensive, 59 , 60 and the importance of the role of ethanol metabolising enzyme variants on alcohol-related behaviours is undeniable. 3 , 4 This is the largest study investigating the role of three of these variants in a diverse population comprising over 3 million individuals. Through a hypothesis-free approach aimed at investigating phenome-wide associations across numerous traits and categories, our study reveals three key insights. First, we provide further evidence that the alleles that encode the more active alcohol metabolising enzymes 3 of these functionally important variants offer protection against excessive alcohol consumption across three ancestries. 61 We replicated associations between three variants in alcohol metabolising enzyme genes and alcohol-related phenotypes–from alcohol consumption to wine headache to heavy drinking. Second, our PheWAS provides evidence of even further-reaching effects of these variants on human health, including replicating associations with mental, sleep, and cardio-metabolic health from other large datasets like the UK Biobank, and revealing associations of these SNPs with numerous understudied traits with no prior GWAS data. Third, MR analyses suggest that some of these associations may be mediated, at least in part, by genetically determined alcohol consumption, while others may reflect common genetic factors across the traits studied. Identifying shared pathways of risk between alcohol consumption and these outcomes could have high translational potential, allowing tailoring of treatments for multiple medical conditions. It is well-established that alcohol consumption is associated with poor mental health. 62 PheWAS associations with mental health traits (e.g., PTSD, depression and earlier onset of depression, bipolar disorder) were primarily identified with rs1229984-T in the European cohort, which is largely due to greater sample size and therefore discovery power compared to the other ancestries. This aligns with GWAS in individuals of European ancestry showing strong genetic correlations between alcohol-related behaviours and mental health traits 7 , 8 , 9 , 13 , 14 ; our PheWAS results show that this individual SNP contributes to those genetic correlations. MR, which could only be performed in the European ancestry cohort due to the lack of well-powered alcohol consumption GWAS in other populations, provided limited evidence that alcohol consumption contributes to mental disorders, consistent with a comprehensive systematic review (including 63 studies) identifying causal links between mental disorders and increased alcohol consumption without evidence for the reverse. 63 Another cluster of associations related to two of the alcohol-protective alleles we studied (rs1229984-T and rs698-T) were with a variety of sleep conditions. Prior large-scale GWAS of alcohol consumption behaviours in individuals of European ancestry revealed significant genetic correlations with sleep alterations, 7 , 8 , 9 , 13 , 14 implying that there are common underlying biological pathways for alcohol consumption and sleep health. The MR analyses in our study support these associations by identifying a potential effect of more alcohol consumption on 29% of the PheWAS significant sleep traits, consistent with prior work, 63 , 64 , 65 , 66 , 67 and extending to other sleep behaviours, such as REM sleep disorder. Our MR results (based only on the SNPs we studied) are somewhat contradictory with a prior MR study (based on polygenic scores) suggesting a causal effect of another sleep behaviour (insomnia) on alcohol consumption and dependence but not vice versa 64 ; however, we did not test for bi-directional associations in the current study. While collectively this line of work suggests the potential for simultaneously treating both excessive alcohol consumption and sleep disturbances, 68 future research should explore bi-directional associations in larger and more diverse samples and consider potential confounding factors such as psychiatric and metabolic disorders not accounted for in this or prior studies. 69 , 70 Beyond mental health, the breadth of information available in our 23andMe cohort revealed PheWAS associations with other neurological conditions that often co-occur with alcohol use, such as migraines. The association between alcohol-protective alleles and increased risk for migraines was specifically for rs122998-T. Our results add to previous observational studies reporting that migraines can be triggered by alcohol consumption, 71 , 72 and prior MR studies linking genetically determined alcohol consumption with lower migraine risk, 73 as well as providing evidence for the reverse. 74 Alcohol consumption also tends to be associated with a range of other medical conditions, such as gastrointestinal or respiratory health. In our study, nearly half of the gastrointestinal conditions that were found to be significant in the European PheWAS, such as cirrhosis and GERD, were suggested as a direct result of alcohol consumption in our MR analysis. This is consistent with a recent MR study based on polygenic scores showing a causal effect of alcohol consumption on gastrointestinal diseases, such as duodenal ulcer, cirrhosis, and chronic pancreatitis, 75 and supporting observational studies showing that chronic alcohol consumption, especially in excess, can lead to a range of gastrointestinal symptoms (e.g., 76 , 77 , 78 ). We also observed associations with respiratory conditions, such as self-reported diagnosis of asthma and, for the first time, the efficacy of beta steroids in the treatment of asthma. Although prior prospective studies have found a U-shaped relationship between alcohol consumption and incidents of asthma, 79 our MR analyses were discrepant, either because we were unable to stratify by alcohol consumption levels, or because comorbid smoking consumption, which is heavily associated with alcohol consumption and is a strong contributor to respiratory conditions, was not included in this analysis. In addition, we identified that carriers of the alcohol-protective alleles we studied had a more favourable cardiovascular (e.g., high blood pressure, coronary artery disease) and endocrine-metabolic (e.g., obesity, hyperglycaemia, hyperthyroidism, T2D, cholesterol medication) profile, consistent with early candidate gene studies 80 , 81 and prospective cohort studies. 82 Although these loci are known for their contribution to ethanol metabolism in hepatic tissues, transcripts of ADH1B are also highly expressed in multiple other tissues, including cardiac and adipose tissues. 3 , 4 , 83 , 84 Therefore, it is not surprising that variants in ADH1B have been implicated in lipid metabolism in humans 84 , 85 , 86 , 87 , 88 , 89 , 90 , 91 and in cholesterol and triglyceride absorption in animal models, regardless of alcohol consumption. 92 While our PheWAS results suggest that these associations may partially arise from common genetic underpinnings, our MR results contribute to the growing body of literature showing that the association with cardiometabolic outcomes may also be, at least in part, a direct consequence of alcohol drinking, 46 , 93 , 94 , 95 and suggest that decreasing alcohol use could have widespread effects on cardiometabolic health. This is in contrast to prior large-scale GWAS (incorporating genome-wide variants) reporting negative genetic associations between alcohol consumption and better cardiometabolic health, 8 , 9 , 11 , 33 which may be partially attributed to biases, such as differences in socioeconomic status, measurement errors, or changes in alcohol consumption over time in the different cohorts. 9 , 19 , 96 Multivariable MR including additional variables, such as BMI and high blood pressure, or experimental analyses, such as those using animal models, may assist in further disentangling the role of alcohol consumption in cardiometabolic health. 81 , 84 , 93 Our study also revealed broad implications of variations in ethanol metabolising enzyme genes for domains of health that are relatively unexplored. For example, across all three ancestries, a large number of the PheWAS associations were with immune traits, particularly between rs1229984-T and various types of allergies. The nature of these associations is still unclear. While some epidemiological studies have found positive associations between alcohol consumption and allergies, 97 , 98 , 99 others have failed to identify significant associations, 99 , 100 or have only found associations with lifetime alcohol consumption, but not the quantity consumed, 97 or with levels of Immunoglobulin E (a type of antibody that plays a crucial role in the immune system), but not with allergic disease. Consistent with these findings, our MR analysis did not establish these associations as causal, indicating other common mechanisms whereby these variants influence the development and course of allergies. Beyond allergies, our PheWAS results revealed an association between rs1229984-T (protective allele) and other immune conditions, including lower incidence of rheumatoid arthritis, lupus, and many others. MR did not identify a causal relationship between alcohol consumption and most of these outcomes, consistent with prior studies presenting conflicting findings and challenging prior observational studies that may be more highly subjected to residual confounders, such as socioeconomic and lifestyle factors. Furthermore, our PheWAS identified associations with eye health (e.g., myopia, cataracts, age related macular degeneration), which is consistent with previous prospective observations 101 , 102 , 103 and MR studies identifying an association between genetically predicted alcohol consumption and risk of developing age-related macular degeneration. However, these results are discrepant with our and other MR studies showing inconsistent results between alcohol consumption and eye health, 104 , 105 emphasising the need for more in-depth studies exploring the effects of alcohol consumption on the visual system. Finally, we also identified associations between rs1229984-T and fractures in all three ancestries. This is perhaps unsurprising, as persistent alcohol consumption (along with many other risk factors such as smoking, low BMI) has been linked to higher odds of developing osteoporosis and bone fractures, 106 , 107 , 108 and to increased injuries as a result of heavy drinking. 109 Although a previous MR study identified an association between alcohol dependence (based on a single SNP, rs1229984) and fractures, our MR results were primarily null, suggesting that additional studies, particularly those using a broader range of SNPs and examining more severe alcohol-related phenotypes, are needed to draw more robust conclusions. Variants in genes encoding ethanol metabolising enzymes are also known to affect the risk of developing cancer. 3 , 4 , 6 , 110 , 111 , 112 , 113 We identified unexpected positive associations between the protective alleles of ethanol metabolising enzyme variants that encode more active enzymes and higher rates for several types of cancer (e.g., skin and lung cancer) in the European cohort. It is not clear whether this association is mediated by alcohol use, by alcohol metabolism in non-hepatic tissues, by both mechanisms, or due to other issues. Our MR results identified one association between increased alcohol consumption and one type of cancer (skin cancer) but not others. These results suggest an indirect pathway explaining the effect of rs1229984 on cancer other than drinking, and replicate the mixed findings observed in prior MR studies for other cancer types. 110 , 114 , 115 , 116 , 117 Future studies exploring additional confounders (smoking, poor diet), increasing granularity in alcohol-behaviours (e.g., only including individuals drinking amounts higher than the recommended thresholds), and using higher numbers of cancer cases, can contribute to clarifying the pathogenic pathways by which alcohol contributes to carcinogenesis. Lastly, our study discovered several ancestry-specific associations (e.g., endocrine/metabolic conditions only being reported in individuals of European and Latin American ancestry; mental health conditions primarily reported in individuals of European ancestry), many of which may simply reflect the difference in power. Another possible explanation for these differences could be attributed to environmental differences among participants from the European, Latin American, and African American cohorts, 118 , 119 , 120 such as education and socioeconomic factors ( Supplementary Table S2 ) that may affect both drinking patterns and disease ascertainment. Several limitations are also worth mentioning. First, there are hundreds to thousands of additional SNPs and other classes of genetic variation beyond the three SNPs that we studied–however, the three SNPs we tested have by far the highest effect sizes on alcohol consumption in the cohorts that we studied. Although we included diverse cohorts, our study lacked many major ancestry groups such as East Asians and South Asians, which were not available at the time of the analysis. Including those of East Asian ancestry would have allowed testing of rs671, a SNP in ALDH2 that has the most profound effect of any studied SNP in alcohol metabolising enzymes. 3 , 21 , 31 , 32 , 35 Another important caveat is that our results may be biased by ascertainment and related characteristics of the sample. 23andMe participants are volunteers, and they tend to be older than the general population, with higher education and socioeconomic status, and low levels of alcohol consumption ( Supplementary Table S2 ). Indeed, approximately 30% of individuals in this study reported no current alcohol drinking. Thus, this cohort includes both lifetime abstainers and former drinkers who may have reduced or quit drinking because of other detrimental outcomes, which are known to introduce biases in genetic studies. 19 This cohort composition may explain some of our paradoxical inverse relationships between alcohol-protective variants and increased liability for some conditions (e.g., more lifetime tobacco use, cancer, Graves’ disease). The mediating effects of age and alcohol exposure on the reported associations could be assessed in future studies, and therefore replication using younger populations, 121 longitudinal data 122 and cohorts with lower socioeconomic status and higher levels of alcohol consumption is warranted. Moreover, our PheWAS used minimal (self-reported) phenotyping to allow a large number of diverse phenotypes to be surveyed within a single study; in particular, alcohol phenotypes pertain to recent use at the time of the analyses and not lifetime use, which would be preferable. Although we have studied the largest diverse cohort to date, some traits were still underpowered for some SNPs ( Supplementary Figure S5 ). Similarly, while we examined them separately, most conditions we studied are correlated. For example, the comorbidity between alcohol consumption and high blood pressure is necessary to consider when assessing the elevated likelihood of cardio-metabolic disease. Thus, despite our attempt to disentangle the mechanisms underlying pairs of comorbidities with problematic alcohol consumption, the risk may be better represented by a matrix of comorbidities with many shared genetic and non-genetic pathways. In this study, the genetic liability for alcohol consumption was defined by using three specific SNPs in the ADH region. MR is less prone to the bias of observational studies, particularly in relation to reverse causation and confounding, providing a more robust estimate of causal relationships between exposure and outcomes. Incorporating established instruments, particularly rs1229984, 3 , 4 provides an additional layer of support for the validity of our MR analysis; however, relying solely on three variants with modest effects on the magnitude of alcohol consumption can diminish predictive accuracy. Furthermore, although our MR results were robust to different assumptions and biases and to sensitivity analyses accounting for horizontal pleiotropy, 46 some confounding may remain (e.g., associations with physical characteristics). Because we are only testing one pathway related to ethanol metabolism, it is possible that other molecular causal pathways could play a significant role in these associations. Similarly, there may be other unknown or unmeasured confounding factors we have not accounted for in MR analyses, such as dynastic effects (e.g., parents with high alcohol consumption genetic propensity conferring some unrelated advantage/disadvantage on these outcomes) or assortative mating (genetic resemblance between spouses due to selecting each other based on genetically influenced traits). 123 Our results should therefore be replicated with even larger sample sizes (or by using multiple SNPs in combination to increase the variance explained in alcohol consumption) and employing additional study designs, such as within families, to counteract those potential confounders. This study performs an in-depth exploration of the relationships between variants in alcohol metabolising enzyme genes and thousands of other traits across three large ancestral groups. We provide evidence that these loci contribute directly or indirectly to several leading causes of disability. Interventions that decrease alcohol consumption may be a promising therapeutic target for a constellation of other disorders. Future research using larger sample sizes and/or observational longitudinal studies in multiple ancestries may allow for further investigation into these complex associations.

SSR conceived of the idea for the paper, with input from AAP. PF, MVJ, JJMM. and the 23andMe team contributed analyses and data. PF and SLE accessed and verified the original data. All contributing authors wrote and edited the paper, and approved the final version.

Heavy alcohol consumption is a heritable trait associated with numerous negative social and health outcomes. 1 Individual differences in how people metabolise alcohol alter the risk for excessive alcohol consumption and development of alcohol use disorders ( AUD ). 2 , 3 The primary enzymes responsible for alcohol metabolism are alcohol dehydrogenases ( ADHs ) and aldehyde dehydrogenases ( ALDHs ). 2 , 3 Some individuals have certain single nucleotide polymorphisms ( SNPs ) of the alcohol dehydrogenase 1B ( ADH1B ) and 1C ( ADH1C ) genes that change their amino acid sequence such that they metabolise alcohol to acetaldehyde at a faster rate. Similarly, individuals with coding polymorphisms in aldehyde dehydrogenase 2 ( ALDH2 ) convert acetaldehyde to acetate more slowly. Both of these polymorphisms influence the unpleasant effects of acetaldehyde concentrations 3 , 4 , 5 thereby reducing alcohol consumption and serving as protective factors against the development of AUD and alcohol-related health problems, such as alcohol-related liver disease and cancer. 3 , 4 , 5 Single nucleotide polymorphisms ( SNPs ) located in genes that encode ethanol metabolising enzymes have been consistently associated with alcohol-related traits both by candidate gene studies and genome-wide association studies ( GWAS ). 3 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 As extensively reviewed by others, 3 , 4 rs1229984-T, a missense variant in ADH1B , is associated with reduced alcohol consumption and is protective against AUD and several similar alcohol phenotypes in individuals of European and Asian ancestries. 5 , 7 , 8 , 11 , 13 , 14 , 16 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 Additionally, rs2066702-A, another missense variant in ADH1B , which is primarily found in individuals of African ancestry, is similar to rs1229984-T. 13 , 15 , 20 The functional variant rs671-A in ALDH2 , which is primarily found in individuals of Asian ancestry, 3 , 13 , 21 , 25 , 31 , 32 is associated with the alcohol flush reaction, and is also associated with reduced alcohol consumption and is strongly protective against AUD. 3 Other variants, including rs698-T in ADH1C , which is known to contribute to alcohol metabolism, have also been implicated in alcohol use behaviours and risk for AUDs. 3 , 9 , 12 , 14 , 33 , 34 Intriguingly, these variants also have effects on other social and health traits beyond those directly related to alcohol ( Supplementary Table S1 ). These associations could reflect common genetic factors (pleiotropy) that confer liability to both alcohol-related behaviours and AUD comorbidities, or could arise as a consequence of the effects of these variants on alcohol consumption. In this study, we focused on three of the best studied protective variants in the ethanol metabolising enzyme genes ( ADH1B : rs1229984-T, rs2066702-A; ADH1C : rs698-T). Different ancestry groups often have differences in allele frequencies and linkage disequilibrium patterns ( LD ; i.e., frequency with which variants are inherited together). Therefore, we studied the role of these variants on outcomes via phenome-wide association analyses ( PheWAS ) across three diverse populations (N ≤ 2,619,939, European; N ≤ 446,646, Latin American; N ≤ 146,776, African American) from the 23andMe research cohort. We included rs2066702-A for its prior association with AUD in individuals of African ancestry, 3 , 7 despite its very low frequency in other populations. ALDH2 rs671-A, which is strongly negatively associated with alcohol consumption in Asian populations, 3 , 35 could not be tested due to extremely low allele frequency in the populations studied, which did not include an East Asian cohort. We employed a hypothesis-free PheWAS approach to examine up to 1109 self-reported traits across 28 phenotypic categories (e.g., substance-use related, mental health, sleep, immune, cardiovascular, metabolic) that had been previously collected by 23andMe, most of which have never been published as GWAS or explored in non-European ancestry populations, meaning that these data are not available from commonly consulted resources such as the EMBL-EBI GWAS catalog. 36 Lastly, we examined causality via two- and one-sample Mendelian randomisation ( MR ) analyses. 37 , 38 We anticipated replication of previously reported associations with alcohol-related behaviours and discovery of additional associations across numerous categories given our increased sample size and more ancestrally diverse dataset.

PF, SLE and members of 23andMe Research Team are employees of 23andMe, Inc., and hold stock or stock options in 23andMe. ASH has filed for a provisional Patent entitled “Methods of Treatments for Multi-Drug or Broad Addiction Liability”, University patent application T-020489. ASH reports consulting fees from Psychiatric Genomics Consortium Suicide working group, speaker fees from Jackson Laboratories and a grant from National Institute on Alcoholism and Alcohol Abuse (NIAAA K01 AA030083-01). CK received a new investigator award from TRDRP and consulting fees and stock options from CARI Health, Inc. JG is paid for editorial work for the journal ‘Complex Psychiatry’ and is named as an inventor on PCT patent application #15/878,640 entitled: “Genotype-guided dosing of opioid agonists” filed January 24, 2018 and issued on January 26, 2021 as U.S. Patent No. 10,900,082. AP received consulting fees from universities and companies and holds a biomedical patent (related to inhibitions of the gene Glo1 ), both unrelated to this work. SSR received consulting fees from the Externalizing Consortium as well as a Royal Society Alcoholism Early Career Award honoraria. EA reports consulting fees from PGC-PTSD and LATINO consortium. All other authors declare no conflicts of interest.