2.1 Study population

The present study utilized data from the UK Biobank, a large-scale prospective cohort study conducted among middle-to-older adults across the United Kingdom. The design and the protocol of this project have been previously published.^{14, 15} Approximately 500 000 participants between 40 and 69 years were recruited from the general population at baseline during 2006–2010. Comprehensive assessments were undertaken at 22 centers throughout the UK to collect extensive data on participant demographics, socioeconomic status, medical history, lifestyle factors, health outcomes, and biological measures. Moreover, blood specimens were collected for biochemical assays and genotyping. All UK Biobank participants were genotyped using either the Affymetrix UK BiLEVE Axiom or Affymetrix UK Biobank Axiom array. Ethical approval was obtained from the North West Multi-Centre Research Ethics Committee (reference number: 11/NW/0382), and written informed consent was secured from all participants before data collection. The present analyses were conducted under UK Biobank application number 44407.

2.2 Exposure assessment

Habitual coffee consumption was ascertained using both a food frequency questionnaire (FFQ) and repeated 24-h dietary recalls collected at different time points. On the FFQ, participants reported their average daily coffee intake over the past year, measured in cups per day, with each cup defined as 200–250 mL. The FFQ also inquired about the intake of specific coffee subtypes including instant, ground, and decaffeinated, among others.

Additionally, subsets of participants completed up to five 24-h dietary recalls. On each recall, participants reported their coffee consumption on the previous day, including intake in cups of various subtypes such as instant, filtered, and espresso. Further, the recalls captured additions to coffee such as milk, sugar, and artificial sweeteners.

2.3 Outcomes

The primary outcome was defined as metabolic syndrome, in accordance with the harmonized criteria set by the International Diabetes Federation.¹ Specifically, participants meeting the threshold for central obesity based on population-specific waist circumference cutoffs, along with any two of the following, were classified as having metabolic syndrome: elevated fasting glucose, blood pressure, triglycerides, and reduced high density lipoprotein (HDL)-cholesterol. Detailed criteria for each condition were outlined in Table S1. Waist circumference was measured to the nearest centimeter during the physical assessment utilizing a SECA tape measure. Blood pressure was assessed twice via an automated Omron monitor, with the average of the two readings utilized in analyses. Fasting blood biomarkers were assayed using a Beckman Coulter AU5800 analyzer. Self-reported medication use for treating dysglycemia, dyslipidemia, and hypertension was ascertained through touchscreen questionnaires. The corresponding medication codes were shown in Table S2.

2.4 Case and control ascertainment of outcomes

For each outcome, cases were defined as meeting the ascertainment criteria, while controls were defined as not meeting the ascertainment criteria without missing data. Handling of missing data was addressed in the next section. For metabolic syndrome, individuals meeting the criteria for cases in three or more of the component conditions were considered cases, while those meeting the criteria for controls in three or more of the component conditions were labeled as controls. Participants not meeting the criteria for either case or control were excluded from the analysis.

2.5 Handling of missing data

Since the UK Biobank collected random blood samples during the baseline recruitment, only 6% of the total participants provided fasting blood samples, defined as blood samples taken after at least 7 h of fasting for this study. For those who did not provide a fasting blood sample, a case of high fasting glucose level was defined for individuals taking medication to manage high fasting glucose levels, while those not taking medication were treated as missing since confirmation of their fasting levels was not feasible. This same approach was applied when identifying cases for high triglycerides.

In cases where participants did not provide data on medication use, a case of high fasting glucose levels was established if fasting glucose levels exceeded the threshold specified in Table S1. Conversely, the rest were treated as missing, considering uncertainty regarding whether observed fasting glucose levels were influenced by medication use. Similar logic was extended to ascertain cases for high triglycerides, high blood pressure, and low HDL-cholesterol levels.

2.6 Measurement of confounding variables

Socioeconomic position was characterized using the Townsend Deprivation index.¹⁶ Self-reported smoking status, alcohol intake frequency, and highest qualification attained were subsequently recoded as follows: smoking status (never/previous/current), alcohol intake frequency (monthly or less/weekly/daily), and highest qualification obtained (below General Certificate of Secondary Education [GCSE], GCSE, A-level, or degree holder). Physical activity level was evaluated using the validated International Physical Activity Questionnaire (IPAQ) and categorized as low, moderate, or high based on IPAQ research committee guidelines.¹⁷ Intakes of fruit, vegetable, and tea were self-reported by the participants. Fruit intake was calculated by summing the intake of fresh and dried fruit while vegetable intake was calculated by summing the cooked and raw vegetable intake. Tea intake was measured in cups per day.

2.7 Mendelian randomization analysis

To complement the observational analysis, a two-sample Mendelian randomization study was undertaken to estimate the causal effect of habitual coffee intake on metabolic syndrome risk. This instrumental variable analysis is less susceptible to residual confounding and reverse causation and relies on three instrumental variable assumptions.¹³ First, the instruments should be strongly associated with the exposure of interest. Second, there should be no unmeasured confounding of instruments on outcome. Third, the instruments should be independent of the outcome given the exposure and the confounders.

Single-nucleotide polymorphisms (SNPs) associated with coffee consumption were obtained from a large genome-wide association study conducted by the Coffee and Caffeine Genetics Consortium, which did not include data from the UK Biobank.¹⁸ These included (i) coffee consumption in cups per day among coffee drinkers (phenotype 1, n = 129 488, of which 93.9% were Europeans) and (ii) no/low versus high coffee consumption (phenotype 2, n = 65 842, of which 98.2% were Europeans). Only independent SNPs (r² < 0.001) that achieved genome-wide significance (p < 5 × 10⁻⁸) in the trans-ethnic meta-analysis were selected (Table S3).

Genetic associations between the identified SNPs for cups of coffee consumed (phenotype 1) and being a coffee drinker (phenotype 2) and metabolic syndromes and its components, including high fasting glucose, high triglycerides, central obesity, high blood pressure, and low HDL, were derived using individual-level data in the UK Biobank. Effect estimates were obtained using multivariable logistic regression. Adjustments were made for age, sex, top 20 principal components, genotype array, and assessment center. Exclusions encompassed participants with subpar genotyping quality (missing rate > 0.01), ≥10 putative third-degree relatives, sex chromosome aneuploidy, mismatch between reported and genetic sex, and inconsistency in reported and genetic ethnicity. Genetic associations involving outcomes were observed among up to 360 903 participants.

2.8 Statistical analyses

All analyses were done using R version 4.1.1.¹⁹ mendelian randomization (MR) analyses were performed using the “TwoSampleMR” package^{20, 21} and “MVMR” package.²² To address the issue of multiple testing, we established a p-value threshold for statistical significance across all analyses at 0.0083 (derived from dividing 0.05 by the six distinct outcomes examined). All scripts used for data cleaning and analyses were included in the Supporting Information.

We employed multivariable logistic regression to scrutinize the cross-sectional associations between coffee consumption and metabolic syndrome, along with its constituent components. In this analysis, individuals who did not consume coffee served as the reference group. Effect estimates were meticulously adjusted for potential confounding variables, including age, sex, Townsend deprivation index, highest educational qualification achieved, smoking status, frequency of alcohol intake, level of physical activity, and consumption of fruits, vegetables, and tea. Furthermore, when evaluating data derived from 24-h recalls, we introduced coffee type as a covariate. To explore the interplay between the use of milk, sugar, and artificial sweeteners with coffee and their collective impact on outcomes, we undertook subanalyses stratified by these factors whenever feasible. Specifically, for instant and filtered coffee, we conducted stratification based on milk use, and for all coffee types except espresso, we executed stratification based on the use of sugar and artificial sweeteners.

We calculated the R² (only phenotype 1) and F statistics of the instruments, serving as indicators of instrument strength. These calculations adhered to the methodologies outlined in a prior Mendelian randomization study,²³ where a higher F statistic indicated diminished evidence of weak instrument bias. Moreover, we harmonized the effect alleles of all instruments with the allele associated with increased coffee consumption. To assess the causal link between coffee consumption and metabolic syndrome, in addition to its constituent conditions, we employed the inverse variance weighted (IVW) method, with multiplicative random effect models. The IVW estimator operates under the assumption of balanced horizontal pleiotropy.²⁴ Notably, a high level of heterogeneity within instrument-specific Wald Ratios, as determined by Cochran's Q test, may suggest the potential inclusion of invalid instruments. To fortify the robustness of our findings, a series of sensitivity analyses were conducted, grounded in diverse assumptions. The MR-Egger estimator, designed to account for overall directional pleiotropic effects, was employed. It assumes that the instrument's strength is independent of the direct effect (InSIDE assumption). The presence of a statistically significant MR-Egger intercept (p < 0.05) signifies the existence of a horizontal pleiotropic effect.²⁴ Additionally, the weighted median method, predicated on at least 50% of the weight arising from valid SNPs, was employed.²⁵

Since smoking status and alcohol consumption were well-recognized confounding factors in the associations between coffee consumption and various health outcomes²⁶ and genetic predictors of coffee consumption were found to be associated with both traits as well,²⁶ we conducted multivariable MR to obtain effect estimates adjusted for alcohol consumption and liability to smoking initiation. Genetic variants linked to alcohol consumption and smoking initiation liabilities were identified with genome-wide significance (p < 5 × 10⁻⁸) from the genome-wide association study (GWAS) and Sequencing Consortium of Alcohol and Nicotine use study,²⁷ featuring participants exclusively of European ancestry, distinct from those in the UK Biobank cohort (n = 249 171 for smoking initiation; n = 226 223 for alcohol consumption). For the multivariable MR analysis, genetic predictors of coffee consumption were based on stage one results from the CCGC meta-analysis, limited to European participants. The lists of SNPs used in the multivariable MR for both phenotypes are provided in Tables S4 and S5. We excluded SNPs in linkage disequilibrium with other instruments using a more lenient cutoff (r² < 0.1) so that we have a sufficient number of instruments and maintain statistical power for multivariable MR. Genetic associations for exposure and outcome were aligned to the same effect allele, and additionally effect allele frequency for palindromic SNPs. In cases where the target SNPs were absent from any dataset, proxy SNPs (r² ≥ 0.8) were used. The causal effect estimates of coffee consumption on metabolic syndrome and its individual components were subsequently computed through multivariable IVW and multivariable MR-Egger approaches.^{28, 29}

2.9 Use of large language models

During the preparation of this work, the authors used ChatGPT 3.5 to improve the flow, clarity, coherence, and grammar of the text drafted by the authors. After using this tool/service, the senior author (JCYL) reviewed and edited the content as needed and takes full responsibility for the content of the publication.

3.1 Participant characteristics

After excluding individuals who withdrew consent, were non-British, lacked coffee intake data, or had missing covariate data, 351 805 participants were included (Figure 1). Table 1 displays participant characteristics by daily coffee consumption. Coffee drinkers tended to have a higher proportion of males, smokers, frequent alcohol consumption, education attainment, and lower tea intake compared to nondrinkers.

3.2 Coffee intake and metabolic syndrome

Figure 2 depicts the association between daily coffee intake and metabolic syndrome prevalence among 142 594 participants after adjustment for sociodemographic and lifestyle factors. Consuming up to three cups per day was associated with lower metabolic syndrome odds compared with abstainers (1 cup/day odds ratio [OR]: 0.88, 95% confidence interval [CI]: 0.85–0.92; 2 cups/day OR: 0.90, 95% CI: 0.86–0.93; 3 cups/day OR: 0.94, 95% CI: 0.90–0.99). However, intakes of four cups or more daily were potentially associated with higher metabolic syndrome odds (4 cups/day OR: 1.03, 95% CI: 0.98–1.08; 5 cups/day OR: 1.06, 95% CI: 0.99–1.12). Coffee intake displayed positive associations with high fasting glucose and triglycerides, while inverse associations were observed for high blood pressure and low HDL-cholesterol. For central obesity, a J-shaped relationship was found (Figure 2).

We identified eight instruments for phenotype 1 (cups of coffee consumed, R²: 0.4%, range of F statistics: 9.0–196.0) and three instruments for phenotype 2 (liability to high coffee consumption, range of F statistics: 44.4–100.0) for the MR analysis (Figure S1). Results did not support a causal effect of genetically predicted coffee intake on metabolic syndrome risk (Figure 3 and Tables S6 and S7). Despite heterogeneous instruments based on Cochran's Q test, MR-Egger regression suggested did not provide strong evidence for overall horizontal pleiotropy. Additional multivariable Mendelian randomization adjusting for smoking and alcohol attenuation yielded consistent results (Figure S2 and Tables S8 and S9). However, coffee intake was causally associated with higher central obesity risk based on analyses with both phenotypes. There was also suggestive evidence for a causal effect on high fasting glucose (Figure 3).

3.3 Coffee type, additives, and metabolic outcomes

Stratified analyses by coffee subtype aligned mostly with the overall findings, with some exceptions (Figure 2). For example, instant coffee intake displayed null associations with metabolic syndrome and high blood pressure, while decaffeinated coffee was not associated with low HDL-cholesterol. Intakes of latte/cappuccino were inversely linked to metabolic syndrome and high blood pressure only, whereas espresso intake showed no associations (Figures S3–S6). Regarding additives, milk and sugar did not substantially modify the associations. However, artificial sweetener use was positively associated with all outcomes among instant coffee drinkers. Similarly, filtered coffee with sweeteners displayed positive associations except for glucose and triglycerides. Avoiding sweeteners was associated with lower metabolic syndrome, central obesity, and low HDL-cholesterol among filtered and latte/cappuccino consumers.