Gene-environment interaction study for BMI reveals interactions between genetic factors and physical activity, alcohol consumption and socioeconomic status

1Gene-environment interactions result from individuals responding differently to environmental stimuli depending on their genotype, or from genetic effects that vary between groups of individuals depending on their lifestyles. In humans, the most famous examples include skin color and risk of melanoma in response to ultra-violet rays, and phenylketonuria (PKU) in response to foods containing phenylalanine in individuals who carry mutations that lead to phenylalanine hydroxylase deficiency. Gene-environmental interactions are likely to exist for complex human traits and identifying gene-environment interactions can potentially improve risk-assessment for disease and help unravel underlying biological pathways.

2Obesity and being overweight are serious public health issues due to their strong associations with diseases such as cardiovascular disease, type 2 diabetes and cancer. In addition, their global prevalence has increased dramatically over the latter part of the 20th century and up to the present day. Body mass index (BMI) is a standardised measure of human body size that is calculated from weight and height. Twin studies have demonstrated a heritable component of BMI and genome-wide association studies (GWAS) have shown that BMI is influenced by hundreds of common genetic variants. Recently, a GWAS for BMI on 339,224 individuals, reported 97 genetic loci to be associated with variation in BMI. However, only a few studies have investigated the effect of gene-environment interactions on BMI. Previous studies have reported physical activity to attenuate the effect of genetic factors on BMI, including the effects of genetic variants within the FTO locus. Identification of gene-environment interactions for complex human traits poses several challenges. For instance, most GWAS of complex traits have been performed by large-scale meta-analyses of multiple cohorts, which complicate a harmonised collection of lifestyle and environmental data. Also, the effects of genetic variants identified through GWAS are generally small, and differences in the effects of genetic variants between groups exposed to different lifestyle factors may be difficult to detect in smaller cohorts due to lack of statistical power.

3Initiatives such as the UK Biobank provide a unique opportunity to study interactions between genetic and lifestyle factors. Data collection in UK Biobank has been performed in a standardised manner and data include a large number of lifestyle and environmental factors collected from approximately half a million UK citizens, as well as comprehensive, genome-wide genotyping. Recent studies on the UK Biobank found that the effect of the FTO locus variant, rs1421085, interacts with several lifestyle risk factors such as alcohol consumption, sleep patterns, diet and physical activity. Another recent study on the UK Biobank determined that the effect of a genetic risk score for BMI is modified by socioeconomic status.

4Here, we study the effects of gene-environment interactions on BMI, by analysing 131 lifestyle factors assessed by touchscreen questionnaires. These factors include diet, smoking, alcohol consumption habits, physical activity, socioeconomic status, mental health, sleep, general health as well as factors that are specific to females such as number of live births. For the purpose of our analyses, we constructed a genetic score for BMI (GSBMI) which was composed of 94 single-nucleotide polymorphisms (SNPs) that have previously been associated with BMI in a GWAS.

5We utilised data from the UK Biobank Resource (http://www.ukbiobank.ac.uk/about-biobank-uk/) for all analyses. UK Biobank has recruited more than 500,000 individuals aged 40–69 from the United Kingdom during the years 2006–2010. Participants underwent standardised measurements of anthropomorphic traits, and additionally provided biological samples and detailed information about themselves via touchscreen questionnaires. Genotyping had been performed using two custom-designed UK Biobank Axiom Arrays with 820,967 and 807,411 SNPs respectively (BiLEVE and Axiom). Genotypes that were not directly assayed had been imputed using a combined set consisting of the UK10K haplotype reference panel and the 1000 Genomes phase 3 reference panel. We utilized the initial release of genotype data (data accessed January 2016) as a discovery cohort, and the remaining participants with genotype data available in the second release as a replication cohort (data accessed July 2017). In the initial release, data were available for 73,355,667 SNPs in 152,249 UK Biobank participants. To identify related individuals, we used information provided by UK Biobank (Data-Field: 22011—Genetic relatedness pairing). Briefly, kinship coefficients had been calculated for each pair of participants in the cohort using the genetic data and pairs of related individuals had been identified (at least 3rd degree relatives = kinship coefficient > 0.044). In addition, only people who self-identified as white British (Data-Field 21000) and that were classified as Caucasians based on the genetic principal components (Data-Field 22006) were included. After filtering, 116,138 individuals remained for the analysis in the discovery cohort. The same filtering was applied to the replication cohort, leaving 246,358 participants for the replication.

6All participants had provided signed consent to participate in UK Biobank. UK Biobank has been given ethical approval to collect participant data by the North West Multicentre Research Ethics Committee, which covers the UK; the National Information Governance Board for Health & Social Care, which covers England and Wales, and the Community Health Index Advisory Group, which covers Scotland. UK Biobank possesses a generic Research Tissue Bank approval granted by the National Research Ethics Service (http://www.hra.nhs.uk/), which lets applicants conduct research on UK Biobank data without obtaining separate ethical approvals. Access to UK Biobank genetic and phenotypic data was granted under application no. 15152: “Interaction between diet, food preference and lifestyle with genetic factors influencing body mass, body adiposity and obesity”. Written consent was obtained from all participants.

7Participants’ weights were assessed by a variety of means during the initial UK Biobank assessment centre visit. For weight, we utilised data-field 21002, which is an amalgate of all weight values into a single item. Standing height was measured on a SECA 240 Height Measure. BMI was constructed from height and weight measurements during participants’ initial visit to assessment centers. For most analyses, BMI was transformed using rank-based inverse normal transformation.

8Lifestyle factors have primarily been collected via self-report touchscreen questionnaire. All lifestyle variables that had been assessed in more than 20,000 of the participants were included for analyses. This resulted in 131 quantitative, ordinal and categorical measurements of lifestyle factors representing dietary habits, general health, sleep, smoking and alcohol consumption, physical activity, mental health and socioeconomic status (S1 Table).

9We aimed to use linear regression models to test for interaction between lifestyle factors and genetic factors on BMI. To this end, “Prefer not to answer” and “I don’t know” were set to “missing” in our analyses. We removed the 99th percentile of quantitative phenotypic variables, such as, for example: “Average weekly red wine intake” in number of glasses, and “Duration of moderate physical activity” in minutes, to reduce the effect of outliers. We analysed ordinal phenotypic data as quantitative variables. For example, data-field 1558—frequency of alcohol intake: which is coded as: 1 = “Daily or almost daily”, 2 = “Three or four times a week”, 3 = “Once or twice a week”, 4 = “One to three times a month”, 5 = “special occasions only”, 6 = “Never”.

10Data field 20126 represents bipolar and major depression status among participants. This variable was derived from self-report questionnaire data. Very few patients were assessed to have bipolar disorder type I and II (n = 808 & 807 respectively) and these were designated missing. Severity of depression was assessed as 0 = “No depression”, 1 = “Single probable major depression episode”, 2 = “Probable recurrent major depression (moderate)”, 3 = “Probable recurrent major depression (severe)”. Categories 1 to 3 were combined and this field was recoded as “No depression” = 0, and “Probable depression” = 1. ‘Had menopause’ (Data-field 2724) was recoded to better represent linearity: participants who were uncertain due to having undergone a hysterectomy were designated “missing”. Data field 680: “Own or rent accommodation lived in”, was recoded to better represent linearity with regard to socioeconomic status: 1 = “Own outright”, 2 = “Own with mortgage”, 3 = “Rent, from local authority”, 4 = “Rent, from private landlord or letting agency”. Categories 5: “Pay rent and part mortgage (shared ownership)” and 6: “Live in accommodation rent free”, were set to missing due to the low number of participants in these groups (N = 303 and N = 735, respectively). Self-reported drinking habits were converted to amounts in ml alcohol per week using standard sizes for serving and percentages: red wine—125 ml per glass, 13.5% alcohol; white wine– 125 ml per glass, 12.0% alcohol; beer and cider– 570 ml per pint, 5.5% alcohol; spirits– 30 ml per measure, 41.5% alcohol; fortified wine– 58 ml per glass, 19% alcohol. Amounts of exercise per week for specific exercise-types, e.g., “10+ minute walks”, “walking for pleasure”, “moderate physical activity”, and “vigorous physical activity”; were calculated by multiplying the exercise frequency per week with the duration of activity in minutes.

11Genotype data for 97 SNPs that have previously been identified to be associated with BMI were considered for our analyses (S2 Table). One SNP, rs2033529 was not part of the UK biobank dataset and was replaced by another linked SNP rs751414 (r2 = 0.99, D’ = 1). Three SNPs were removed due to deviation form Hardy Weinberg equilibrium, which left 94 SNPs for the analyses. Since many of these variants have not been replicated in an independent cohort, we first tested for association in the initial release of genotype data from the UK Biobank cohort. This was done using linear regression models with BMI as a response variable. BMI was first transformed using a rank-based inverse normal transformation similar to the discovery study. The UK Biobank participants were genotyped on two different genotyping arrays: (BiLEVE and Axiom), and a variable to adjust for this was included as a covariate in addition to sex, age, age2 and the first 15 genetic principal components (PCs). Some of the SNPs were identified as being associated with BMI in females or males separately in the previous study. We therefore also tested for association in males and females separately and compared whether there was a significant difference in the estimates between males and females.

12Genotype data was used in dosage format, where SNP genotypes were represented by the number of copies of the effective allele. To calculate GSBMI, regression coefficients (β-estimates) were retrieved from the GIANT consortium meta-analysis for BMI for the European populations with males and females combined. Weighted GSBMI were then calculated for each individual by multiplying the number of effective alleles for each of the 94 SNPs (all SNPs in HWE in UK biobank) with the respective β-estimates (i.e., ) and calculating the sum over all SNPs (S2 Table):

13Linear regression modeling was used to estimate the effect of gene-environment interaction (GSBMI × E) on BMI, for 131 lifestyle factors (E) separately. In addition to the GSBMI × E interaction term, each of the 131 models was adjusted for covariates: age, age2, sex, PCs, and genotyping array (batch). Interaction terms for GSBMI with age, age2, and sex as well as interaction terms for the lifestyle factor with age, age2, and sex were also included in order to properly control for possible confounding effects of these interactions, in accordance with previously published recommendations, such that:

14We assume that the error term ε ~ NID(0, σ2). The models also included 15 principal components (PCs) to account for effects of population stratification in UK Biobank. In the primary analyses, models for each of the 131 lifestyle factors were analysed separately. The aim of this study was to investigate the effect of the interaction term GSBMI×E on BMI. For this purpose, we focused our attention on the estimate of the coefficient β3 in (2), and more specifically whether this estimate significantly deviated from zero. The null hypothesis H0: β3 = 0 was either accepted or rejected, depending on the outcome of a two-sided marginal student’s t-test, which in this case (i.e., one degree-of-freedom difference between the nested models and normal regularity conditions) is equivalent to a likelihood-ratio test of the hypothesis H0: β3 = 0. P-values lower than the significance level α = 0.05/131 ≈ 3.82*10−4 were considered significant to account for the family-wise error rate using the Bonferroni method. Interaction effects that were considered significant in the discovery cohort were then tested in the replication cohort using the same covariates as well as 15 PCs. Calculations were performed in R version 3.3.0 using the “lm” function included in the stats package.

15In order to visualise and make it easier to interpret the significant interactions, we also estimated the effect of GSBMI and of individual SNPs on BMI in different subgroups with regards to lifestyles, e.g., the genetic effect in participants with different frequencies of alcohol consumption. In these analyses, linear regression was performed in the subgroups using the same covariates as above, but with untransformed BMI values as a response variable, for easier interpretability of the regression coefficients (presented in kg/m2). Here the differences in effect between the subgroups reflect the interaction term from the previous analyses. Interactions were visualised with bar graphs using the ggplot2 package in R. We also used the plot3D R package to visualise interactions in 3D-plots.

16Genetic variants within intron one of FTO have consistently been shown to be the strongest genetic factors associated with BMI. We therefore constructed a genetic score that excluded the FTO-linked SNP rs1558902 (GSBMI’), and performed linear regression modeling in exactly the same manner as previously. We also performed additional analyses to assess how individual SNPs interacted with lifestyle factors. SNP-interactions that were considered significant in the discovery cohort were further tested in the replication cohort. We also performed sensitivity analyses by including TDI and its interactions with age, age2, sex, GSBMI, as well as each of the lifestyle factors in the model, in addition to the variables described in Eq (2).

17For calculating GSBMI, we used the regression coefficients from the GIANT consortium. However, in a discovery GWAS, the regression coefficients are often overestimated, which will introduce a bias in the GSBMI. For this reason, we also performed additional analyses using the regression coefficients estimated in UK Biobank, for the sake of comparison.

18To determine which of the interacting lifestyle factors had an independent contribution in the regression model for BMI, we performed stepwise linear regression (SLR) using the ‘step’ function included in the ‘stats’ package in R. This function uses the Akaike information criterion (AIC) to select variables for the model. A base-model for BMI was constructed that included GSBMI, age, age2, sex, a batch variable to control for the two genotyping platforms used in UK Biobank, as well as 15 principal components. Variables that were significant after replication were included in SLR. SLR was performed using ‘both’ directions so that variables were either added or dropped depending on how they improved AIC. The process is repeated until no improvement in AIC can be made. Individuals with any missing data were excluded from the analyses, and in order to maintain a large sample size for the analyses, we performed SLR on a combined set of the discovery and replication cohorts. Individuals with missing data in any of the tested factors were excluded before running SLR, which resulted in 290,441 participants remaining after filtering. All secondary interactions between variables were included in the analysis to control for potential confounding, in accordance with recommendations by Keller.

19Genotype and phenotype data were available for 487,409 participants, of which 362,496 passed the QC and were included in the analyses (116,138 in the discovery and 246,358 in the replication). Basic characteristics are provided in Table 1. The distribution of BMI was slightly skewed (S2 Fig), and we therefore applied a rank-based inverse normal transformation of BMI prior to the analyses (Fig 1), in agreement with previous GWAS on BMI. Out of the 94 SNPs that were in HWE in UK Biobank, the association with BMI replicated for 63 of them (S1 Table). However, the effect sizes for the 94 SNPs were consistent with previous data from the GIANT consortium (S1 Fig). We therefore included all 94 SNPs in the construction of GSBMI. No differences in effect size between males and females were observed for any of the 94 SNPs (S3 Table) and we therefore used the same regression coefficients for calculating the GSBMI in males and females. GSBMI was calculated so that a one-unit increase in GSBMI was associated with a one-standard deviation increase in BMI, which in our data equals 4.83 kg/m2 (Fig 1). GSBMI explained 1.85% of the variation in BMI in the studied subset of the UK Biobank participants.

20Linear regression modeling in the initial release of genotype data from the UK Biobank cohort (discovery cohort) revealed GSBMI to interact with 19 lifestyle factors related to physical activity, alcohol consumption, smoking, socioeconomic status, sleep, mental health and number of live births (Table 2, S4–S9 Tables), when we applied Bonferroni adjustment for multiple testing (p < 3.8*10−4). Of these, interactions with 15 factors were replicated using the second release of genotype data from the UK Biobank cohort (Table 3). If we instead apply the false discovery rate to adjust for multiple testing in the discovery cohort, 38 interacting lifestyle factors were identified. The additional FDR-significant factors fall into the same categories as previously mentioned with the addition of variables related to variation in diet, and intake of bread, processed meat, and salad or raw vegetables (S4–S9 Tables).

21Strong evidence for interaction with GSBMI was seen for alcohol intake frequency (p = 1.45*10−29) with larger effect of GSBMI in infrequent drinkers (Fig 2A, Tables 2 and 3, S5 Table). The effect of GSBMI decreased, in a dose-dependent manner, as alcohol consumption frequency increased and the effect of GSBMI was less than half the effect in everyday drinkers compared to infrequent drinkers (Fig 2A). The interaction between GSBMI and alcohol intake frequency means that the increase in BMI per GSBMI unit is higher in infrequent drinkers compared to more frequent drinkers (Fig 3A). In addition to the interaction between GSBMI and alcohol intake frequency, we also observe a highly significant inverse association between alcohol intake frequency and BMI (S3 Fig, p < 2.2*10−308).

22In contrast to alcohol frequency, we were unable to observe significant interactions with the number of alcoholic beverages per week or total weekly alcohol intake (gram/week). An interaction was observed with total weekly intake of red wine in the discovery cohort, where higher genetic effects were associated with low weekly consumption of red wine (Table 2, S5 Table). However, this effect was not observed in the replication cohort (Table 3).

23Interactions were also identified for several factors related to physical activity, such as: usual walking pace, stair climbing, and TV watching, as well as frequencies of light- (Fig 3B), moderate-, as well as vigorous exercise (Tables 2 and 3, S6 Table). Particularly strong evidence was found for an interaction with walking pace, (p = 3.38*10−26, Table 2). While the frequencies of physical activity were interacting with GSBMI, no significant interactions were identified for durations of physical activity (S6 Table).

24Several markers of socioeconomic status were observed to interact with GSBMI in the discovery cohort including: Townsend deprivation index (TDI), as well as number of vehicles in household and total household income (Tables 2 and 3, S7 Table). TDI is a composite score for socioeconomic status that is generated for each national census output area and incorporates area inhabitants’ unemployment rates, car- and house-ownership as well as the number of people in a household. Higher TDI corresponds to a larger degree of social deprivation and was associated with an increased effect of GSBMI (Tables 2 and 3, S7 Table).

25We also observed significant interactions with factors related to depression including: fed-up feelings and frequency of depressed mood (Tables 2 and 3, S8 Table). Higher effects of GSBMI were observed in participants who reported often feeling fed-up, and in participants who reported higher frequency of feeling down, depressed or hopeless (data field 2050). Interactions were also observed for factors related to sleeping patterns with higher genetic effects in the group that often take a nap during the day and in the group that often felt tired or reported having low energy. Significant interactions were also identified for smoking status (Tables 2 and 3, S5 Table) and number of treatments/medications taken (Tables 2 and 3, S9 Table), where smokers had a higher effect of GSBMI compared to non-smokers and where the genetic effects increased with number of treatments/medications taken (Table 3).

26We followed up the findings from the main analyses by performing linear regression modeling on individual SNPs that were included in GSBMI to study their interactions with lifestyle factors (supplementary data). These analyses resulted in large number of statistical tests (94 SNPs times 131 lifestyle factors = 12,314 tests), which reduces the statistical power. Adjusting for all tests performed using Bonferroni or FDR resulted in only one significant interaction between rs1558902 at the FTO locus and usual walking pace (Table 4). However, using the same p-value cut-off as for GSBMI, (p = 3.82 *10−4) interactions were found for eleven lifestyle factors in the discovery cohort, of which three interactions with the FTO SNP rs1558902 could be replicated (Table 4). Rs1558902 was observed to interact with total household income, usual walking pace and alcohol intake frequency (Table 4). Previous GWAS has estimated the effect of the FTO variant, rs1558902, to 0.39 kg/m2 per A-allele. Stratification by frequency of alcohol intake revealed the effect of rs1558902 to be 0.64 kg/m2 per A-allele in non-drinkers, and to be attenuated to 0.25 kg/m2 per A-allele in participants that drink daily or almost daily. The interaction between rs1558902 and alcohol intake frequency means that, the effect of rs1558902, i.e. the increase in BMI per allele, is lower in high-frequency alcohol consumers, and higher in low-frequency consumers (Fig 3B).

27To test whether interactions were driven primarily by the FTO-SNP rs1558902, we constructed a genetic score for BMI with rs1558902 excluded (GSBMI’). In the discovery analysis, all Bonferroni significant interaction terms from the previous analyses remained significant when we used GSBMI’, except for smoking status and number of treatments/medications taken (Table 2, S4–S9 Tables). Effect estimates for all significant interactions were in the same direction and within 15% of the interaction effects with the previous GSBMI. We also utilized GSBMI’ in the replication cohort, and replication was successful for all 15 interactions observed with GSBMI, except for frequency of depressed mood and smoking status (Table 3).

28We also performed sensitivity analyses by including TDI and its interaction terms as covariates in linear regression models. These results were highly correlated with previous results (Pearson’s r = 0.98 for effect estimates of interaction terms). Including TDI as a covariate led to a general slight decrease in effect sizes of interaction terms. The largest decreases were seen for factors related to socioeconomic status (S11 Table) and smoking status, which is consistent with the highly significant correlation between these variables and TDI (S12 Table).

29In the present study, we utilised SNP effect estimates from a previous GWAS by the GIANT consortium to calculate the genetic score for BMI. These estimates may be somewhat overestimated due to the “winner’s curse”. Using overestimated effect sizes results in a GSBMI that is associated with a slightly lower BMI-increase in UK Biobank, compared to when using the correct effect size estimates. This can clearly be seen in our data since a one-unit increase in the GSBMI results in a 0.82 unit increase in the rank transformed BMI in UK Biobank. We therefore also tested interaction effects using a genetic score composed of SNP effect estimates calculated in the UK Biobank cohort (GSBMI_UKBB), so that a single-standard-unit increase in the GSBMI_UKBB results in an exactly 1.00 unit increase in the rank-transformed BMI, which is 22% higher compared to 0.82 for GSBMI. The interaction results were also strongly correlated between using GSBMI_UKBB or GSBMI (Pearson’s r = 0.99 for effect estimates of interaction terms; S11 Table). However, the effect estimates for FDR-significant interaction terms were, not surprisingly, on average 16% larger when we utilised GSBMI_UKBB (S11 Table).

30Several of the exposures that were found to interact with GSBMI in our primary analysis showed highly significant evidence for correlation with one another (S12 Table). In order to identify the most informative interacting variables and interaction terms for a predictive model for BMI, we performed SLR on a combined set of the discovery and replication cohort. SLR was performed in both directions using the 15 lifestyle factors whose interactions with GSBMI were replicated (Table 3). This resulted in inclusion of 290,441 participants with non-missing data when combining the discovery and replication cohorts. The final model generated by SRL included gene-environment interaction terms for ten lifestyle factors (S10 Table, supplementary Data), of which eight were nominally significant: alcohol intake frequency (p = 2.82*10−15), usual walking pace (p = 1.55*10−14), frequency of 10+ minute walks (p = 6.26*10−4), smoking status (p = 1.45*10−3), frequency of vigorous exercise (p = 2.28*10−3), number of vehicles in household (p = 1.03*10−2), TDI (p = 3.02*10−2) and frequency of tiredness/lethargy (p = 3.87*10−2) and. The adjusted R2 value for the final model was 0.1957.

31In this study, we performed a gene-environment interaction study using genetic variants and self-reported lifestyle data. We identified several lifestyle factors that influence the effect of genetic variants on BMI. Interactions were observed for factors related to alcohol intake, physical activity, socioeconomic status, mental health and sleeping patterns. Interactions were seen for factors related to physical activity, where a more active lifestyle attenuated the genetic effects, which is consistent with previous reports. Interactions were observed for light, moderate intensity, and vigorous physical activity. However, we observed that the interaction between physical activity and the genetic score was strong for frequencies of physical activity, in contrast to durations in minutes/day.

32Strong evidence was also observed for an interaction with frequency of alcohol intake. The genetic effect was attenuated with higher frequency of alcohol intake in an almost dose-dependent manner with twice as large effects in non-drinkers compared to daily drinkers. Alcohol consumption is common in western societies, where also most previous GWAS have been performed. Our results indicate that the interaction associated with alcohol intake frequency may have partially attenuated the full effect of BMI-associated genetic variants observed in previous association studies.

33Alcohol intake frequency was also associated with lower average BMI. This is consistent with clinical reports of lower BMI and fat mass in severely alcoholic patients. In vitro and in vivo experiments have also shown ethanol exposure to increase lipolysis and reduce white adipose tissue mass. This can also be compared to data from the National Institute on Alcohol Abuse and Alcoholism (NIAAA), which suggests that moderate daily consumption of alcoholic beverages, 1–2 drinks per day, reduces the risk of myocardial infarction as well as all-cause mortality. In addition, a cohort study on 38,077 male health professionals reported that alcohol consumption frequency, rather than total amounts, was the primary determinant of the inverse association between alcohol consumption and risk of myocardial infarction. Unfortunately, we do not have data on UK Biobank participants’ daily consumption amounts and we are unable to determine how this factors into the association between alcohol consumption frequency, BMI and the interaction between alcohol intake frequency and GSBMI.

34In a previous study on gene-environment interactions in the UK Biobank, Tyrrell et al. used a genetic score composed of 69 BMI-associated variants to study interactions with measurements of the obesogenic environment, with focus on physical activity, diet and socioeconomic status. Interactions were identified with measurements of physical activity and socioeconomic status (TDI) which were consistent with the current study. TDI serves as a proxy for environmental and lifestyle factors that are correlated with income and social position. The study by Tyrell et al. study contrasts the current in the selection of twelve obesogenic factors. The current study instead utilised a hypothesis-free approach to test interactions between GSBMI and 131 factors, which allows us to contrast aspects of the same behaviour, e.g., between amounts of physical activity and frequency and also gives us the potential to identify new gene-environment interactions. A drawback to this approach is the increased statistical power required in order to correct for the family-wise error rate.

35In our primary analyses, we have investigated interactions between lifestyle factors and a genetic score composed by 94 independent SNPs, located in different loci. SNPs were combined into genetic scores that explain a greater amount of the variation in BMI compared to the individual SNPs, in order to gain statistical power (S1 Supporting Information). These SNPs have previously been shown to influence BMI. However, combining them into a genetic score before testing for interactions with lifestyle factors assumes that the interaction effect of the BMI-increasing alleles are all in the same direction: e.g., that alcohol intake frequency decreases the genetic effects of all SNPs rather than some genetic effects being larger among frequent drinkers and others larger among non-drinkers. For alcohol intake frequency, we have the statistical power to detect an interaction if interaction effects in the same direction are present for at least 37 of the SNPs (S1 Supporting Information). However, if some SNPs are interacting in the opposite direction, our power will decrease dramatically (S1 Supporting Information). It is therefore possible that there are gene-environment interactions that are masked by SNPs having interaction effects with the same environmental factor, but in opposite directions. For this reason, we also performed follow-up analyses of individual SNPs. These analyses revealed that the FTO-linked SNP rs1558902, in addition to interacting with alcohol intake frequency, also interacts with average total household income and physical activity.

36For BMI, as well as for other complex traits, knowledge on the biological implications of associated genetic variation is limited, which impedes deduction of causal mechanisms underlying gene-environment interactions. The FTO variant, rs1558902, is associated with the expression of two upstream genes (IRX3 and IRX5) which affect adipocyte “browning”, i.e. the occurrence of thermogenic ‘beige’ adipocytes in white adipose tissue depots. This may partly explain the observed interaction between rs1558902 and frequency of alcohol intake, as in vitro experiments have shown that ethanol exposure interferes with mobilization of glucose transporters to the adipocyte cellular membrane in response to insulin. Beige adipocytes, on the other hand, are able to take up glucose from the circulation in an insulin-independent manner. The altered lipolysis in white adipose tissue due to ethanol exposure, in combination with an altered rate of thermogenesis due to differential propensity for adipocyte browning between individuals with different rs1558902 genotypes may explain the interaction between this SNP and frequency of alcohol intake.

37Enrichment analyses from previous GWAS have also implicated central nervous system processes to play an important role in BMI. The central nervous system contains regions that regulate several functions related to BMI, such as appetite, homeostasis, reward, and motivation. Ethanol confers several well-known behavioural effects on humans, but also acts in a bi-phasic manner as a central nervous system stimulant at low doses, and a general depressant at higher doses. BMI-associated genetic variants that affect BMI-related central nervous system function may also factor into the observed interaction between alcohol consumption frequency and GSBMI.

38A possible limitation to our study is responder bias in the self-report questionnaire data. This may be more likely for factors pertaining to self-image such as alcohol, tobacco use and physical activity. The lack of an interviewing person, and assuring participants of the confidentiality and anonymity of their data aim to reduce the likelihood of responder bias. We tested the validity of factors related to alcohol consumption and physical activity by comparing these to data collected through a 24-hour recall questionnaire. We observe that both frequency and amounts of alcohol consumption, as well as measurements of frequency and duration of physical exercise, agreed well with 24-hour recall data, which supports the validity of these measurements (S2 Supporting Information).

39In this study, we primarily investigated associations, and the underlying causal mechanisms behind gene-environment interactions are difficult to deduce from cross-sectional data alone. We constructed separate models for each environmental exposure or lifestyle factor. As such, it is important to be aware that confounding effects of factors that are not included in the models, or that are unknown, may be present in the results from these tests. In order to fully correct for confounding factors and correctly characterise causal factors, controlled experiments such as clinical trials in controlled settings serve as the gold standard. We have attempted to correct for confounding by including TDI and the interaction term TDI*GSBMI as covariates in all analyses, which resulted in very little effect on the main results. To identify factors with the highest predictive value, we also performed SLR, which showed evidence for interactions between GSBMI and alcohol consumption frequency, physical activity, smoking, and socioeconomic status all contributed independently to a predictive model for BMI.

40In conclusion, the standardised collection of genetic and lifestyle data in UK Biobank has enabled us to identify several factors that modify the effect of BMI-associated genetic variants. The most significant interactions were observed between GSBMI and frequency of alcohol intake, frequency of physical activity and socioeconomic status. Previous studies have reported interactions between genetic variants at the FTO locus and environmental factors. However, most interactions were still observed even when the FTO locus was excluded from the genetic score, which indicates that the individual interactions are not solely dependent on FTO variants. We can therefore conclude that the presence of genetic interactions is more general and will be identified to a higher degree for individual SNPs once the sample size increases even more and reaches sufficient power.