Genome-wide association studies of coffee intake in UK/US participants of European ancestry uncover cohort-specific genetic associations

Abstract

Coffee is one of the most widely consumed beverages. We performed a genome-wide association study (GWAS) of coffee intake in US-based 23andMe participants (N = 130,153) and identified 7 significant loci, with many replicating in three multi-ancestral cohorts. We examined genetic correlations and performed a phenome-wide association study across hundreds of biomarkers, health, and lifestyle traits, then compared our results to the largest available GWAS of coffee intake from the UK Biobank (UKB; N = 334,659). We observed consistent positive genetic correlations with substance use and obesity in both cohorts. Other genetic correlations were discrepant, including positive genetic correlations between coffee intake and psychiatric illnesses, pain, and gastrointestinal traits in 23andMe that were absent or negative in the UKB, and genetic correlations with cognition that were negative in 23andMe but positive in the UKB. Phenome-wide association study using polygenic scores of coffee intake derived from 23andMe or UKB summary statistics also revealed consistent associations with increased odds of obesity- and red blood cell-related traits, but all other associations were cohort-specific. Our study shows that the genetics of coffee intake associate with substance use and obesity across cohorts, but also that GWAS performed in different populations could capture cultural differences in the relationship between behavior and genetics.

Introduction

Coffee is a leading global food commodity that has psychoactive properties largely due to the presence of caffeine [1]. While rates of use vary widely by geographic region, it is estimated that 60–85% of adults in Europe and the United States consume between 0.6 to 5.5 cups of coffee daily [2–4]. Intake of coffee and its bioactive constituents is associated with benefits on cognition [5], and a lower risk of liver disease [6, 7] (but see [8]), Parkinson’s and other neurodegenerative diseases [6, 7, 9], cardiovascular disease [6, 7], type II diabetes [6, 7], and certain cancers [6, 7, 10]. However, coffee intake is also associated with higher risks for some adverse outcomes, including other substance use and misuse [11–14], some cancers [7, 10, 15], poor lipid profile [6, 7], pregnancy loss [6, 7], gastrointestinal maladies [16], and worse cardiovascular outcomes following excessive intake [17]. Addressing the full spectrum of coffee’s correlations with health and disease is therefore an important but challenging task.

Genetic studies offer a compelling avenue to investigate the relationships between coffee intake and other complex traits. Twin studies estimate daily coffee intake to be 36–56% heritable [18], suggesting that coffee intake is amenable to genetic analysis. Whereas phenotypic correlations, which depend on measuring traits in the same cohort, can arise from genetic and environmental factors, genetic correlations use the results from genome-wide association studies (GWAS) and can examine correlations between traits measured in non-overlapping cohorts. Genetic correlations can indicate genetically driven relationships, but may also be subject to environment, social influences, phenotyping differences, and other factors such as selection bias [19, 20]. Several coffee GWAS [21–38] have found associations with single nucleotide polymorphisms (SNPs) within or near genes that metabolize caffeine (Supplementary Table 1), such as CYP1A1 and CYP1A2 [21–23, 26–29, 33, 35, 36]. Some of these loci are also associated with other complex traits, including liver disease [39–41], cancers [42–45], and alcohol consumption [46–48]. This pleiotropy could suggest that these associations are mediated by coffee intake or that loci also influence these traits via alternative independent mechanisms. Several genetic correlations with coffee intake have also been reported, such as positive genetic correlations with other substance use [49, 50], negative genetic correlations with major depressive disorder and neuroticism [49], osteoarthritis [51], and migraine [52], and mixed genetic correlations with sleep [49, 53], body mass index (BMI) [54], and type II diabetes [54]. However, these investigations were conducted under a priori justification and may fail to capture the full scope of genetic correlations between coffee intake with other traits. Thus, a data-driven examination of trait associations with coffee intake remains unexplored.

Because coffee is a primary source of caffeine for many globally, coffee intake is often used as a proxy of caffeine intake for its relative ease of assessment. However, coffee and caffeine should not be conflated. Coffee contains a multitude of other, less investigated bioactive chemicals that may affect human health [9] and caffeine can be consumed through other foods including tea, soft drinks, and chocolate. Thus, when we refer to coffee intake, we mean explicit measures of coffee and not caffeine intake unless otherwise specified. Intake of different caffeine sources also varies by geographic region, such as tea preference over coffee in the United Kingdom (UK; tea vs. coffee: ~50% vs. ~20%) compared to the United States (US; ~10% vs. ~30%) [2]. As some previous genetic studies only used data from the UK Biobank (UKB) [21, 37, 51, 53, 55–57] or combined cohorts from regions with different patterns of caffeinated beverage intake (Supplementary Table 1) [33, 35, 36], this distinction may introduce environmental and cultural confounds affecting the genetic associations between coffee intake and other traits.

In this study, we used survey responses from US-based 23andMe, Inc. research participants of European ancestry (N = 130,153) and performed a GWAS of a single item “How many 5-ounce (cup-sized) servings of caffeinated coffee do you consume each day?”. Using genetic correlations and phenome- and laboratory-wide association studies (PheWAS, LabWAS), we explored the relationships between coffee intake and thousands of biomarkers, health features, and lifestyle traits to provide a fuller inventory of genetic associations with coffee intake. We compared our findings from the 23andMe cohort to UKB using GWAS summary statistics of coffee intake (“How many cups of coffee do you drink each day? (Include decaffeinated coffee)”, N = 334,659, http://www.nealelab.is/uk-biobank/). Our results revealed a lower-than-expected genetic correlation between coffee intake in the two cohorts; therefore, we used these datasets to explore cohort differences in the genetic associations with coffee intake across two distinct populations (Supplementary Fig. 1).

Methods

Study cohorts, coffee intake, and univariate GWAS

23andMe

Univariate GWAS was conducted in 130,153 male and female research participants of the genetics testing company 23andMe, Inc., as previously described [58]. Participants provided informed consent and volunteered to participate in research online under a protocol approved by the external AAHRPP-accredited Institutional Review Board (IRB), Ethical & Independent (E&I) Review Services. As of 2022, E&I Review Services is part of Salus IRB (https://www.versiticlinicaltrials.org/salusirb). During 4 months in 2015 and 14 months between 2018–2020, participant responses to the question “How many 5-ounce (cup-sized) servings of caffeinated coffee do you consume each day?” were collected as part of a larger survey [58]. We excluded SNPs of low genotyping quality, including those that failed a Mendelian transmission test in trios or with large allele frequency discrepancies compared to European 1000 Genomes reference data, failed Hardy-Weinberg testing, failed batch effects testing, or had a call rate <90%, as well as imputed variants with low imputation quality or with evidence of batch effects. The 23andMe GWAS pipeline performs linear regression and assumes an additive model for allelic effects. Ancestry was determined by a genetic ancestry classification algorithm [59] using principal component (PC) analysis and genotyping data to define population structures and different ancestry clines. Unrelated participants categorized as of European ancestry from genotyping data were included in the GWAS with age (inverse-normal transformed), sex, the top 5 genetic PCs, and indicator variables for genotype platforms as covariates [59]. We conducted replication using three multi-ancestral cohorts from 23andMe (European N = 689,661; African American N = 32,312; Latin American N = 124,155). Novel SNPs were determined as those not in linkage disequilibrium (LD; r² > 0.1) or within 1 Mb of SNPs uncovered by other GWAS of coffee and caffeine traits sourced from the EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/). For full details on genotyping and GWAS, see Supplementary Methods.

UK Biobank

Publicly available summary statistics of coffee intake (N = 334,659; http://www.nealelab.is/uk-biobank/ field:1498, both sexes) and caffeine intake from coffee (N = 373,522; [23]) were derived from unrelated UKB participants. Participants provided informed consent, were of White British descent, and answered the questions “How many cups of coffee do you drink each day? (Include decaffeinated coffee)” and “What type of coffee do you usually drink?”. Coffee intake (cups/day) GWAS covariates included age, age², sex, age by sex, age² by sex, and the first 20 genetic PCs. Caffeine intake was estimated from the amount and type of coffee consumed, and covariates included age, sex, genotyping array, and the first 30 genetic PCs [23]. Other publicly available summary statistics were not included in our analysis due to differences in coffee intake measurements (Supplementary Table 1).

Gene-based analyses (MAGMA, H-MAGMA, S-PrediXcan/S-MultiXcan)

Significant lead SNPs from UKB were identified using Functional Mapping and Annotation of Genome-Wide Association Studies (FUMA v1.3.8). 23andMe SNPs were annotated in FUMA based on ANNOVAR categories, Combined Annotation Dependent Depletion scores, RegulomeDB scores, expression quantitative trait loci (eQTLs), and chromatin state predicted by ChromHMM. Novel candidate genes were those not previously identified or with start/stop positions within 1 Mb of GWAS-significant SNPs uncovered by other coffee and caffeine GWAS.

MAGMA gene-based and pathway analyses

We used Multi-marker Analysis of GenoMic Annotation (MAGMA, v1.08) to conduct gene-based associations on 23andMe summary statistics. SNPs were annotated to protein-coding genes using FUMA and Ensembl build v92. The default settings (GRCh38; MAGMA Analysis parameters: SNP-wise mean model, MAF = 1%, 0 kb gene window, sex as a covariate, Genome-Tissue Expression [GTEx] v8: 54 tissue types and 30 general tissue types; Positional mapping parameters: 10 kb distance) were used, LD was estimated using the 1000 Genomes European reference sample, and significance determined following Bonferroni correction (p < 2.53E−06). Gene-set analysis was conducted on 10,678 gene-sets and Gene Ontology terms curated from the Molecular Signatures Database (MsigDB v7.0). Tissue-specific gene expression profiles were assessed with average gene expression in each tissue used as a covariate. Using GTEx v8 RNA-seq data, gene expression values were log₂ transformed from the average Reads Per Kilobase Million (max value = 50) per tissue. Significance was determined following Bonferroni correction (p < 9.26E−04 for 54 tissue types; p < 1.67E−03 for 30 general tissue types).

H-MAGMA

We incorporated coffee intake GWAS data with chromatin interaction profiles from human brain tissue using Hi-C coupled MAGMA (H-MAGMA) [60]. H-MAGMA assigns non-coding SNPs to genes based on chromatin interactions from fetal brain, adult brain, iPSC-derived neurons, and iPSC-derived astrocytes datasets (https://github.com/thewonlab/H-MAGMA) [60]. Exonic and promoter SNPs were assigned to genes based on physical position. We applied a Bonferroni correction based on the total number of gene-tissue pairs tested (p < 9.52E−07).

S-PrediXcan and S-MultiXcan

We performed a transcriptome-wide association study using the MetaXcan package (ver0.7.5) [61, 62] consisting of S-PrediXcan and S-MultiXcan to identify eQTL-linked genes associated with coffee intake. This approach uses genetic information to predict gene expression levels in various brain tissues and tests if predicted expression correlates with coffee intake. S-PrediXcan uses precomputed tissue weights from the GTEx project database (https://www.gtexportal.org/) as the reference transcriptome dataset via Elastic net models. S-MultiXcan, an extension of S-PrediXcan, aggregates and jointly examines eQTL-phenotype associations identified across 49 bodily tissues. As input data, we included summary statistics, transcriptome tissue data, and covariance matrices of the SNPs within each gene model (HapMap SNP set available at the PredictDB Data Repository) from all available tissues. We applied Bonferroni correction across all tissues (N = 21,565).

LDSC heritability and genetic correlations

Linkage Disequilibrium Score regression (LDSC; https://github.com/bulik/ldsc) was used to calculate heritability (h2_SNP) and genetic correlations (r_g) [63]. h2_SNP was calculated from pre-computed LD scores (“eur_w_ld_chr/”). r_g, including for conditioned summary statistics (i.e., dietary sugar, cigarettes per day, and Alcohol Use Disorder Identification (AUDIT) Consumption scores; Supplementary Methods), were calculated between coffee intake and health, psychiatric, and anthropologic traits. Genetic correlation unity between coffee intake summary statistics was tested with a one-sample t-test (µ = 1.00) on simulated data. We applied false discovery rate (FDR) correction across all genetic correlations performed (N = 277–282).

Phenome- and laboratory-wide association studies

We tested associations between polygenic scores (PGS) for coffee intake and medical condition liability from hospital-based cohorts using data from the Vanderbilt University Medical Center (VUMC; IRB #160302, #172020, #190418). The BioVU cohort, a subset of VUMC biobank participants (N = 72,821), provided genotyping data and electronic health records (EHR) containing clinical data and laboratory-assessed biomarkers [64–66]. For each unrelated BioVU participant of European ancestry, we computed coffee intake PGS using the PRS-CS “auto” version [66].

Phenome-wide association analyses

We fitted a logistic regression model to each case/control disease phenotype (“phecodes”) to estimate the odds of each diagnosis given coffee intake PGS, while adjusting for sex, median age of the longitudinal EHR, and the first 10 genetic PCs with the PheWAS v0.12 R package [67]. At least two International Disease Classification codes mapping to a PheWAS disease category (Phecode Map 1.2; https://phewascatalog.org/phecodes) and a minimum of 100 cases were required for phecode inclusion (Supplementary Methods). We applied FDR correction across all associations performed (N = 1380).

Laboratory-wide association analyses

We implemented the pipeline established by Dennis et al. [65]. Broadly, LabWAS uses the median, inverse normal quantile transformed age-adjusted values from the QualityLab pipeline in a linear regression to determine the association between coffee intake PGS and 318 phenotypes (Supplementary Methods). We controlled for the same covariates as for the PheWAS analyses, excluding median age because the pipeline corrects for age using cubic splines with 4 knots. We applied FDR correction across all associations performed (N = 318).

All results are presented as the mean ± standard error, unless otherwise specified.

Results

GWAS in the 23andMe US-based cohort replicated seven loci implicated in coffee intake

23andMe participant demographics are described in Supplementary Table 2. The cohort was 65% female, had a mean age of 52.8 ± 0.04 years old, and an average BMI of 28.38 ± 0.07 (range: 14.05–69.10), similar to the US average of 27.5 (95% confidence interval: 25.5–29.4) [68]. The average coffee intake was 1.98 ± 5.49E−03 cups per day, similar to UKB (2.14 ± 2.99E−03; Supplementary Fig. 2 and Supplementary Table 3).

We conducted a GWAS of 14,137,232 imputed genetic variants assuming an additive genetic model (Supplementary Table 4). The genomic control inflation factor (λ = 1.09) suggested no substantial inflation due to population stratification. h2_SNP of coffee intake was 7.57% ± 0.59 (Supplementary Table 5).

We identified seven genome-wide significant (p < 5.00E−08) independent (r² < 0.1) loci associated with coffee intake (Fig. 1, Table 1 and Supplementary Figs. 3–9). These replicated prior coffee or caffeine GWAS findings (Supplementary Table 6) [21, 23, 27, 29, 31, 33, 35–37]. We used three multi-ancestral cohorts from 23andMe participants to replicate these findings (Table 1 and Supplementary Table 3). Of the SNPs that passed QC, all but one replicated (p < 5.0E−08) in a larger European ancestry sample (N = 689,661), one replicated in the African American ancestry sample (N = 32,312), and one replicated in the Latin American ancestry sample (N = 124,155).

Fig. 1. GWAS and secondary analyses of coffee intake from the 23andMe cohort.

A Manhattan plot displays seven genome-wide significant loci associated with coffee intake in the 23andMe cohort (N = 130,153). The horizontal line represents the threshold for significance (p = 5.00E−08). Nearest protein-coding genes (<1 Mb) to significant loci are labeled. Quantile-quantile plot shown in upper left corner. For more details, see Table 1 and Supplementary Table 6. B Overlap of genes identified by MAGMA, H-MAGMA, S-PrediXcan, and S-MultiXcan. Genes identified by all four methods are displayed. C Genes implicated in coffee intake by S-PrediXcan according to brain regions. Upregulated genes are shown in red, downregulated shown in blue.

Table 1.

Significant (p < 5.00E−08) GWAS results for coffee intake from 23andMe research participants (N = 130,153) of European ancestry (EA).

SNP	BP	Alleles	Cytoband	p value	EA EAF	Effect	EA Rep p value	EA Rep EAF	AA p value	AA EAF	LA p value	LA EAF	Nearest gene(s)
rs2472297	75027880	C/T	chr15q24.1	3.60E−65	0.23	0.08	1.28E−234*	0.24	3.09E−05	0.07	5.47E−47*	0.14	CYP1A1, CYP1A2
rs4410790	17284577	C/T	chr7p21.1	5.20E−55	0.38	−0.06	7.58E−212*	0.38	4.52E−15	0.52	9.89E−60*	0.55	AGR3, AHR
rs199612805	24843991	D/I	chr22q11.23	1.80E−10	0.01	−0.10	2.48E−29	0.02	1.28E−07	0.08	6.32E−13	0.02	ADORA2A, UPB1
rs28634426	75675594	G/T	chr7q11.23	2.10E−10	0.24	0.03	3.08E−16	0.24	0.08	0.32	8.17E−06	0.27	STYXL1
rs34645063	98591075	D/I	chr6q16.1	3.30E−09	0.48	−0.02	5.23E−16	0.48	0.05	0.65	4.35E−06*	0.58	MMS22L, POU3F2
rs11474881	62892956	D/I	chr20q13.33	1.10E−08	0.55	−0.02	2.36E−12	0.55	0.21	0.45	0.02*	0.62	PCMTD2
rs117824460	41371480	A/G	chr19q13.2	1.70E−08	0.03	−0.06	2.00E−07	0.03	0.08	0.01	0.08	0.02	CTC-490E21.12

Replication (EA Rep) was conducted in an additional cohort of 23andMe participants of EA (N = 689,661), and two additional cohorts of African American (AA; N = 32,312) and Latin American (LA; N = 124,155) ancestries; *SNPs that did not pass QC in replication. See Supplementary Table 6 for additional information.

Gene-based and tissue enrichment analyses suggest coffee intake is primarily associated with gene expression in the brain

We used gene- and transcriptome-based analyses (MAGMA, H-MAGMA, S-MultiXcan/S-PrediXcan) to identify 165 candidate genes most relevant to coffee intake. MAGMA identified 31 genes in physical proximity to GWAS loci (Supplementary Table 7). H-MAGMA, which maps SNPs to genes via chromatin interaction derived from human brain tissue or iPSC-derived cultures, implicated 143 unique genes across cell types (23.53% cortical neurons, 26.80% midbrain dopamine neurons, 24.83% iPSC-derived neurons, 24.83% iPSC-derived astrocytes) and developmental stages (48.00% fetal, 52.00% adult; Supplementary Table 8). S-PrediXcan (Fig. 1C and Supplementary Table 9) showed that SNPs most frequently correlated with predicted gene expression in the cortex, frontal cortex, and putamen. S-MultiXcan predicted transcriptional regulation of 40 genes implicated in coffee intake (Supplementary Table 10). Overall, four genes, SCAMP2, SCAMP5, MPI, and FAM219B, were identified by all four methods, and six genes, FBXO28, NEIL2, HAUS4, IGDCC4, RP11-298I3.5, and RP11-298I3.5, were novel in their coffee associations (Supplementary Table 11 and Fig. 1B).

MAGMA gene-set analysis revealed coffee intake genetics most strongly associated with the metabolism of xenobiotics or foreign substances (p = 4.75E−07; Supplementary Table 12). MAGMA tissue-based enrichment analyses suggested that coffee intake was only associated with brain tissue, specifically in frontal cortex, cortex, cerebellum, and cerebellar hemispheres (Supplementary Fig. 10 and Supplementary Table 13).

Genetic correlation and polygenic score analyses of coffee intake in US- and UK-based cohorts reveal consistent positive associations with obesity and substance use

To boost statistical power, we sought to meta-analyze our data (metaGWAS) with that from UKB [69]. See Supplementary Figs. 11 and 12 and Supplementary Tables 5, 14–16 for UKB and metaGWAS results. Surprisingly, we found the 23andMe and UKB cohorts were only moderately genetically correlated (r_g = 0.63 ± 0.05, p = 3.54E−43) and not in unity (i.e., r_g < 1.00, t[759,910] = −76.16, p < 2.20E−16), although all top loci (p < 5.00E−05) shared direction and strength of effect (Supplementary Fig. 12B). In addition, heritability of coffee intake of our metaGWAS was lower than for each univariate GWAS (metaGWAS h2_SNP = 4.06% ± 0.25 vs. 23andMe h2_SNP = 7.57% ± 0.59 vs. UKB h2_SNP = 4.85% ± 0.33; Supplementary Table 5). We interpreted these results as an indication of cohort heterogeneity and proceeded to analyze each cohort independently.

We performed a series of genetic correlation and polygenic analyses and compared results from 23andMe and UKB. In the 23andMe cohort, 75 traits were genetically correlated with coffee intake and 74 traits with the UKB cohort (Fig. 2A and Supplementary Table 16). Genetic correlations largely persisted after conditioning on potential confounders (i.e., smoking, alcohol, sugar intake [70]; Supplementary Figs. 13, 14 and Supplementary Tables 17, 18). Of the traits significant in at least one cohort, 29.57% were significant in both datasets and 58.82% shared the same direction of correlation.

Fig. 2. Genetic and phenotypic associations with genetic disposition to coffee intake in US and UK cohorts.

A Comparison of genetic correlations across psychiatric (light gray), anthropologic (medium gray), and health (dark gray) traits between 23andMe (lanes 1 and 2) and UKB (lanes 3 and 4). Lanes 1 and 3 show r_g values calculated by LDSC, and lanes 2 and 4 show FDR-corrected p values. Only traits for which at least one cohort was FDR-significant are displayed. For a full list of correlations and trait names, see Supplementary Table 16. Most signals persisted after conditioning for dietary sugar, cigarettes per day, and Alcohol Use Disorder Identification (AUDIT) Consumption scores using mtCOJO [113] (Supplementary Tables 17, 18 and Supplementary Figs. 13, 14). Genetic correlations that could not be calculated are grayed out; * denotes reverse coding. B Phenomic associations (panel 1: PheWAS [p < 3.62E−05], panel 2: LabWAS [p < 1.57E−04]) identified from PGS of coffee intake from 23andMe and UKB summary statistics. Only traits for which at least one cohort was FDR-significant are displayed (saturated bars = FDR significant; desaturated bars = FDR non-significant). Category abbreviations: neurological (neuro.), genitourinary (gen.), neoplasms (neopl.), sense organs (sense), dermatologic (derma.), immune (imm.). For full trait names and more detail, see Supplementary Tables 19 and 20.

Among traits consistent in both cohorts, we observed positive genetic correlations between coffee intake and substance use traits, such as smoking initiation, drinks per week, and cannabis initiation (23andMe: r_g = 0.21–0.50, p = 4.74E−47 to 1.34E−08; UKB: r_g = 0.09–0.21, p = 1.39E−14 to 5.61E−03). The strength of genetic correlations for substance use and misuse traits (including externalizing behavior [71]) was stronger in 23andMe compared to UKB (0.30 ± 0.03 vs. 0.09 ± 0.02; Welch’s t[51.97] = 5.96, p = 2.23E−07). Associations with substance use disorder (i.e., tobacco, opioid, cannabis use disorders) and dependence (i.e., alcohol, nicotine) traits were observed with the 23andMe cohort (r_g = 0.24–0.43, p = 2.59E−19 to 2.12E−03) and were weaker or not observed with UKB (r_g = −0.15–0.11, p = 7.58E−05 to 0.60), though cluster analysis showed that genetic correlations for coffee intake aligned more with substance use than misuse in both cohorts (Supplementary Fig. 15). Positive genetic correlations with metabolic traits (e.g., BMI, waist-to-hip ratio) were also congruent in both cohorts. Also consistent across cohorts were the lack of significant genetic correlations with most cardiovascular and cancer traits.

The majority of traits were only significant in one cohort (70.43%) or showed discrepant directions of association (58.82%). For example, we identified positive genetic correlations with anxiety-related traits, cross-disorder, attention deficit hyperactivity disorder, schizophrenia, and anorexia nervosa that were exclusive to 23andMe (r_g = 0.12–0.44, p = 1.00E−07 to 0.01). These associations were not apparent or negatively genetically correlated with UKB (r_g = −0.33–0.02, p = 5.49E−06 to 0.55), except for anxiety (r_g = 0.17, p = 1.39E−05). Positive genetic correlations with cognitive variables, such as executive function and intelligence, were found with UKB (r_g = 0.13–0.23, p = 8.04E−23 to 4.55E−08), which were negatively genetically correlated with 23andMe (r_g = −0.17–−0.10, p = 7.83E−08 to 2.06E−03). Other divergences included positive genetic correlations with gastrointestinal and pain traits with 23andMe that were negative or absent in UKB. Across all health and psychiatric traits that were significant within each cohort, all traits showed a positive genetic correlation with 23andMe coffee intake and 41.30% of correlations were positive with UKB.

Next, we conducted PheWAS and LabWAS using PGS for coffee intake derived from either 23andMe or UKB summary statistics and identified 31 traits associated with 23andMe PGS and 24 with UKB PGS (Fig. 2B and Supplementary Tables 19 and 20). Obesity, morbid obesity, and two biomarkers related to red blood cells were consistent in significance and direction of association between both cohorts. Otherwise, all significant associations were observed in one cohort but not the other. For 23andMe coffee intake PGS, the top positive PheWAS and LabWAS associations were substance use disorders, certain respiratory conditions (e.g., chronic airway obstruction, emphysema, and respiratory failure), and absolute monocyte count. Among the top negative associations derived from 23andMe PGS were those with sense organs, neoplasms, other respiratory conditions (i.e., allergic rhinitis, tonsillitis, and adenoiditis), and urea nitrogen serum/plasma. For UKB coffee intake PGS, the top positive PheWAS and LabWAS associations were with endocrine and musculoskeletal disorders, and two metabolic biomarkers, glycated hemoglobin A1c and glucose. The only negative associations from UKB-derived PGS were with anxiety disorders and two biomarkers related to blood and metabolic traits.

Discussion

In this study, we contributed to the existing GWAS literature of coffee intake by analyzing a US population of 130,153 participants. We replicated seven loci associated with coffee intake, mostly in genes implicated in metabolic processes. Coffee-related genes were significantly enriched in the central nervous system. Despite prior phenotypic evidence that coffee intake confers health benefits, we found genetic associations mostly with adverse outcomes in US and UK cohorts, particularly with substance use and obesity-related traits. Relationships with other traits were inconsistent between both cohorts, suggesting that societal, environmental, or analytical differences between populations influence genetic relationships with coffee intake.

Our GWAS replicated prior associations with genes and variants implicated in coffee and caffeine intake [21–38], as well as other metabolic and xenobiotic processes [31], including rs2472297 near CYP1A1/CYP1A2 [21, 27, 29, 33, 36] and rs4410790 near AHR [21, 26, 27, 29, 30, 33, 38]. Gene-based analyses uncovered 165 candidate genes, including four genes previously implicated in coffee intake [21] that overlapped across all four analyses (MPI, SCAMP2, SCAMP5, and FAM219B) and six novel candidates (FBXO28, NEIL2, HAUS4, IGDCC4, RP11-298I3.5, and RP11-298I3.5). We further identified gene enrichment in brain tissues, consistent with prior GWAS [21, 23, 29]. This is further supported by brain imaging studies across cortical and subcortical areas showing morphological and functional differences between those who do and do not habitually drink coffee [72–77].

One of the most striking observations of this study is the breadth and magnitude of positive associations between coffee intake with substance use. It is widely believed that use of one substance heightens risk for use of another and that there are common genetic factors for any substance use [78, 79]; coffee does not appear to be exempt from this. We observed that the genetics of coffee intake aligned with substance consumption phenotypes, corroborating prior GWAS and twin studies of coffee and caffeine intake [50, 80–82], but not with substance misuse. This is perhaps unsurprising because the phenotypes probed focus on quantity rather than clinically-defined dependence, and the genetic architectures of other substance intake versus problematic use are unique [47, 78, 83–86]. This is likely also true for coffee.

We found consistent positive genetic correlations with BMI and obesity with 23andMe and UKB, in contrast to meta-analyses of randomized control trials and epidemiological studies that found a modest inverse relationship between coffee intake and BMI, or unclear effects of coffee on waist circumference and obesity [87, 88]. Highly heterogeneous results are likely due to interindividual variability in other food and substance use habits surrounding daily coffee intake (e.g., sugary additives, concurrent nicotine use [89]). This contentious relationship may also be explained by the amount of coffee intake, as greater coffee intake seems to attenuate coffee’s genetic associations with BMI and obesity [54], possibly due to the appetite suppressant effects of caffeine [90]. Alongside accounting for other dietary intake and consumption habits, future subgroup analyses may help explain discrepant associations between the genetics and prevalence of coffee intake with BMI-related traits.

We did not recapitulate the beneficial phenotypic relationships between coffee intake and a variety of health outcomes that are generally reported by health association studies [6–8, 10, 91–103]. Although these phenotypic and genetic relationships may seem contradictory, a recent meta-analysis of over 100 phenotypic studies on coffee intake health outcomes suggest high levels of cohort heterogeneity [6, 7], especially across geographically-separated populations [6]. This is consistent with our observations of opposing genetic correlations between coffee intake and pain, anxiety, cognitive, psychiatric illnesses, and gastrointestinal traits across 23andMe and UKB cohorts, and few consistently significant PheWAS associations. Also of note is that the number of positive associations between 23andMe coffee intake and other traits was greater compared to UKB, and the strength of these associations was usually stronger. Partially consistent with this, one meta-analysis of mortality found an inverse relationship between coffee intake and all-cause mortality in European but not US studies [104].

Geography has an observable influence on GWAS results [20]. As we observed no effects by subtle geographic differences on coffee intake genetic correlations using location data available for UKB participants (Supplementary Fig. 16), cultural variability in coffee-related habits between the US and UK may be more important to the inconsistencies we observed with coffee’s genetic relationships. For example, tea is the preferred source of caffeine in the UK and may modify coffee intake (Supplementary Fig. 17) or associate with socioeconomic status differences, though we observe minimal deviations in socioeconomic factors between those in UKB who drink tea or coffee (Supplementary Table 21). Comparatively higher levels of coffee intake or caffeine intake from high caloric beverages in the US [2] may partially explain the greater number and magnitude of adverse health associations observed in the 23andMe analysis. Even across coffee beverage subtypes, a recent investigation revealed the volume of ground or instant coffee is important to the potential health effects of its intake [105]; instant coffee is more commonplace in the UK, whereas fresh brewed coffee is preferred in the US [2]. Cultural differences in coffee intake could help explain the divergent patterns of associations between UK and US participants, though the relative contributions of culture, geography, and their interactions will need further exploration.

There are multiple caveats to consider when interpreting our findings. Our coffee intake measure was self-reported and captured a broad phenotype; interindividual differences in how coffee is cultivated and brewed, habits surrounding daily intake (e.g., eating, smoking), and environmental and social norms surrounding coffee drinking may introduce noise to this phenotype, contributing to the cohort discordance we observed. It is likely that only associations with large effect sizes overcome this heterogeneity, which is likely true for other complex phenotypes. Secondly, our study does not address causality between coffee intake and other traits. Mendelian randomization (MR) studies have attempted to address the exposure-outcome relationships between two traits by using genetic instruments (i.e., SNPs identified by GWAS) as proxies for exposure and associating them with an outcome of interest. MR suggests that coffee intake has no causal effect on obesity and endocrine disorders despite observational studies suggesting protective effects [106], nor is there MR evidence of a causal relationship between coffee intake with other substance use [50, 107], although some studies suggests that the causal relationships between of coffee intake and smoking heaviness vary by cohort [108, 109]. There are also several differences between the 23andMe and UKB coffee intake GWAS that should be considered. The phenotype examined by 23andMe was the intake of 5-ounce cups of caffeinated coffee whereas UKB included decaffeinated coffee and did not explicitly define cup volume. The bioactive compounds of coffee, including caffeine, were not directly compared. Although we observed similar patterns of genetic correlations between cups per day and estimated caffeine intake based on primary coffee subtype consumed in UKB [23] (Supplementary Fig. 18 and Supplementary Tables 16 and 22), we cannot discount the possibility that discrepant genetic associations arise from cohort differences in decaffeinated coffee intake, overall caffeine intake, or the intake of other coffee compounds. These compounds, which are under-investigated relative to caffeine [9, 31], may also have health effects and genetic variation could mediate their metabolism, but this was not measured by our study. The UKB dataset also includes more than double the participants of the 23andMe dataset. While cohort discrepancies in genetic correlations could be the result of study power, we do not find this reflected in the strength or significance of genetic correlations in 23andMe versus UKB coffee intake. Cohort discrepancies could also be explained by differences in GWAS covariates or study participation/selection bias, though the covariates used and demographics (e.g., age, socioeconomic status) of these populations are notably similar [110]. Furthermore, the cohorts used here are primarily of European descent, skew older, and are above average socioeconomic status to the general population [110], limiting generalizability of our findings. Some studies also show sex-dependent differences in coffee and caffeine metabolism and health associations with intake [91, 111, 112], which was also not examined.

Overall, we present evidence from two large cohorts of European ancestry that coffee intake genetically associates with other substance use and obesity-related traits. However, we also find cohort differences in genetic associations of coffee intake with other health and psychiatric traits. While robust genetic signals may replicate across diverse cohorts, other associations could be obscured by cohort or cultural differences related to the phenotype in question. Our study provides a cautionary perspective on combining large cohort datasets gathered from unique geo-cultural populations.

Author contributions

SSR and AAP conceived the idea. PF and SLE contributed formal analyses and curation of 23andMe data. HHAT contributed to formal analyses, investigation, and data visualization. BP, AA, and NSCK contributed to formal data analysis and data visualization. JJM and LVR contributed to formal analyses. JM contributed to data visualization. HHAT and SSR wrote the manuscript. All authors reviewed and edited the manuscript.

Funding

MVJ, SBB, and SSR are supported by funds from the California Tobacco-Related Disease Research Program (TRDRP; Grant Number T29KT0526 and T32IR5226). SBB and AAP were also supported by P50DA037844. BKP, JJM, and SSR are supported by NIH/NIDA DP1DA054394. HHAT is funded through a Natural Science and Engineering Research Council PGS-D scholarship and Canadian Institutes of Health Research (CIHR) Fellowship (#491556). JYK is supported by a CIHR Canada Research Chair in Translational Neuropsychopharmacology. LKD is supported by R01 MH113362. NSCK is funded through an Interdisciplinary Research Fellowship in NeuroAIDs (Grant Number R25MH081482) and an NIH/NIAAA Loan Repayment Program (L40AA031140). JM is funded by the National Institute for Health and Care Research (NIHR) Maudsley Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The datasets used for the PheWAS and LabWAS analyses described were obtained from Vanderbilt University Medical Center’s BioVU which is supported by numerous sources: institutional funding, private agencies, and federal grants. These include the NIH funded Shared Instrumentation Grant S10RR025141; and CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. Genomic data are also supported by investigator-led projects that include U01HG004798, R01NS032830, RC2GM092618, P50GM115305, U01HG006378, U19HL065962, R01HD074711; and additional funding sources listed at https://victr.vumc.org/biovu-funding/. PheWAS and LabWAS analyses used CTSA (SD, Vanderbilt Resources). This project was supported by the National Center for Research Resources, Grant UL1 RR024975-01, and is now at the National Center for Advancing Translational Sciences, Grant 2 UL1 TR000445-06.

Data availability

We will provide 23andMe summary statistics for the top 10,000 SNPs upon publication. 23andMe GWAS and metaGWAS summary statistics will be made available through 23andMe to qualified researchers under an agreement with 23andMe that protects the privacy of the 23andMe research participants. Please visit (https://research.23andme.com/collaborate/#dataset-access/) for more information and to apply to access the data.

Competing interests

HHAT is on the Neuropsychopharmacology Special Projects Team. PF, the 23andMe Research Team, and SLE are employees of 23andMe, Inc., and PF and SLE hold stock or stock options in 23andMe. AAP is on the scientific advisory board of Vivid Genomics for which he receives stock options. The remaining authors have nothing to disclose.

Supplementary information

The online version contains supplementary material available at 10.1038/s41386-024-01870-x.