A genetic sum score of risk alleles associated with body mass index interacts with socioeconomic position in the Heinz Nixdorf Recall Study

Body mass index (BMI) is influenced by genetic, behavioral and environmental factors, while interactions between genetic and socioeconomic factors have been suggested. Aim of the study was to investigate whether socioeconomic position (SEP) interacts with a BMI-related genetic sum score (GRSBMI) to affect BMI in a population-based cohort. SEP-related health behaviors and a GRS associated with educational attainment (GRSEdu) were included in the analysis to explore potential interactions underlying the GRSBMIxSEP effect. Baseline information on SEP indicators (education, income), BMI, smoking, physical activity, alcohol consumption and genetic risk factors were available for 4,493 participants of the Heinz Nixdorf Recall Study. Interaction analysis was based on linear regression as well as on stratified analyses. In SEP-stratified analyses, the highest genetic effects were observed in the lowest educational group with a 0.24 kg/m2 higher BMI (95%CI: 0.16; 0.31) and in the lowest income quartile with a 0.14 kg/m2 higher BMI (95%CI: 0.09; 0.18) per additional risk allele. Indication for a GRSBMIxSEP interaction was observed for education (ßGRSbmixeducation = -0.02 [95%CI:-0.03; -0.01]) and income (ßGRSbmixincome = -0.05 [95%CI: -0.08; -0.02]). When adjusting for interactions with the GRSEdu and SEP-related health behaviors, effect size estimates of the GRSBMIxSEP interaction remained virtually unchanged. Results gave indication for an interaction of BMI-related genetic risk factors with SEP indicators, showing substantially stronger genetic effects in low SEP groups. This supports the hypothesis that expression of genetic risks is higher in socioeconomically disadvantaged environments. No indication was observed that the GRSBMIxSEP interaction was affected by other SEP-related factors included in the analysis.


Introduction
Obesity is a complex condition caused by various genetic, environmental and behavioral factors [1,2]. Research has shown that indicators of socioeconomic position (SEP), such as income and education, are also strongly related to obesity, indicating the highest prevalence in groups of low SEP, especially in western countries [3][4][5]. Genome-wide association studies (GWAS) have detected numerous single nucleotide polymorphisms (SNPs) robustly associated with body mass index (BMI), indicating a polygenetic contribution to obesity risk. The effects of single genetic variants are modest to small, as well as the variance explained by the combined effect of all genetic loci detected [1,6]. It has been suggested that interactions between genes and the environment (GxE) may partly account for the unexplained variance of BMI, meaning that for some genetic variants, the genetic effect depends on environmental exposures [7][8][9]. In the study of GxE interaction, SEP indicators can serve as proxy markers for a wide range of socially unequally distributed environmental, psychosocial and, in particular, behavioral health risks (e.g., low physical activity, unhealthy diet, smoking) [10] which are usually more prevalent among lower socioeconomic groups [11][12][13]. Twin studies have suggested an interaction of genetic factors with education on BMI [14] and with education and income on overall physical health [15,16]. However, previous studies on GxSEP interactions have rarely included underlying risk factors potentially responsible for the interactions detected.
The aims of the present study were to investigate whether (1) the SEP indicators income and education interact with a genetic sum score of BMI-related risk alleles (GRS BMI ) to affect BMI in a population-based cohort and (2) whether these GRS BMI xSEP interactions might be mediated by physical activity, smoking or alcohol consumption as SEP associated health behaviors. Further, (3) a GRS associated with educational attainment (GRS Edu ) was included to investigate potential gene by-gene interaction effects (GRS BMI xGRS Edu ).

Study population
Baseline data of the Heinz Nixdorf Recall Study, a population-based prospective cohort, was used. The rationale of the study has been described in detail elsewhere [17]. In brief, 4,814 participants aged 45-74 years were randomly selected from mandatory registries of residence of the cities Bochum, Essen and Mülheim/Ruhr within the largest metropolitan region in the western part of Germany. The baseline response proportion (December 2000 to June 2003) was 55.8% [18]. The study was approved by the ethics committees of the University Hospital Essen and included extended quality management procedures and certification according to DIN ISO 9001:2000. Informed consent was obtained from all participants.

Data collection
At study baseline, BMI (kg/m 2 ) was calculated based on standardized weight (in underclothes) and height measures. Information on educational attainment, household income, physical activity and smoking was assessed in standardized computer-assisted face-to-face interviews. Education was defined by combining school and vocational training as total years of formal education according to the International Standard Classification of Education [19]. Years of education was used as a continuous variable or categorized into four groups with �10 years, 11-13 years, 14-17 years and �18 years of education. The lowest educational group is equivalent to a minimum compulsory school attendance and no additional vocational degree and the highest educational group is comparable to a vocational training including additional qualification or a university degree. Income was measured as the monthly household equivalent income calculated by dividing the participants' household net income by a weighting factor for each household member [20]. Income was included as a continuous variable or divided into four groups, using sex-specific quartiles. In order to take account for their different mechanisms in causing health inequalities, both SEP indicators were analyzed separately [21,22].
Smoking status was dichotomized for analyses as current smoker (smoking cigarettes during the past year) versus former and never smoker. Physical activity was defined as no regular engagement in physical exercise versus exercising one and more times per week. Amount of alcohol intake is given in gram per week and was estimated from information on the total number of alcoholic drinks by type of drink (beer, wine, sparkling wine, and spirits) usually consumed in a week and was included as a continuous variable into analyses.
Using the meta-analyses of genome-wide association studies (GWAS) by Locke et al., 2015 [6], 97 SNPs or suitable proxy SNPs (r 2 >0.90), representing independent genetic loci robustly associated with BMI, were identified (S1 Table and S2 Table). In addition, 74 education-associated SNPs, or their proxy SNPs (r 2 > 0.94), were selected from GWAS meta-analysis by Okbay et al., 2016 [24] (S3 Table). For two loci associated with educational attainment, neither the original SNP nor their proxy SNP was available within the study population. Unweighted genetic risk scores associated with BMI (GRS BMI ) and with educational attainment (GRS Edu ) were calculated by aggregating the total number of risk alleles for each individual across the selected SNPs (for detailed information see S2 Fig and S3 Fig). The GRS BMI was additionally weighted by the effect size estimates of each SNP reported by Locke et al. [6]. As the results using the weighted score did not differ to those resulting from the unweighted GRS BMI , only the latter were reported.

Statistical analyses
Overall, 4,493 participants with non-missing information on BMI, gender, age and genetic variants were included in the analyses (S1 Fig). As participants had missing information on education (n = 13) and income (n = 279), smoking status (n = 76), physical activity (n = 71) and alcohol consumption (n = 109), analysis populations differed in the respective analyses. Participants with missing information on income or education did not differ substantially in age, BMI or in GRS BMI compared to the analysis population. As most of the SNPs associated with educational attainment were found on the GWAS microarrays, which were available for n = 4,147 participants only, analyses including the GRS Edu were carried out in this sub-sample.
Linear regression models were fitted, adjusted for sex and age, to estimate effects and 95% confidence intervals (95% CIs) on BMI for education and income (model 1), the GRS BMI (model 2) and SEP-related health behaviors smoking, physical activity and alcohol consumption (models 3-5). To assess GRS BMI xSEP interactions, the GRS BMI and SEP main effects as well as GRS BMI xSEP interaction terms were included (model 6). The SEP indicators education and income were used separately as continuous variables. Interaction analysis was also repeated for each BMI-associated SNP and SEP indicator. The genetic effect on BMI was calculated stratified by education groups and income quartiles; the effect of SEP indicators on BMI was stratified by tertiles of the GRS BMI .
All possible combinations of GRS BMI tertiles and SEP groups were entered as dummy variables into regression models to calculate the joint effects of the GRS BMI and SEP indicators on BMI, separately for income and education, using the group with the highest SEP and lowest GRS tertile as single reference [25].
To analyze whether the GRS BMI xSEP interaction may be affected by underlying interactions between GRS BMI and SEP-related health behaviors, smoking (S), physical activity (PA) or alcohol (A) main effects and the respective GRS BMI xS/PA/A interaction terms, in addition to an SEPxPA/S/A interaction term, were included in model 6, separately for each health behavior and SEP indicator (models 7-9) [26]. Moreover, a model was fitted including GRS BMI , GRS Edu , SEP main effects, GRS BMI xGRS Edu and GRS BMI xEducation interaction terms to investigate whether there is indication for overall gene-by-gene interaction (model 10). For all analyses the statistical computing software R v3.1.1 [27] was used. For single SNP analyses Plink v1.07 software package for Windows was used [28].

Results
The mean BMI (± standard deviation) in the study population was 27.9 ± 4.6 ( Table 1). Men showed a higher BMI than women (28.2 ± 4.0 vs. 27.7 ± 5.2). Sex differences were also observed in the distribution of the two SEP indicators with women reporting lower formal education and a lower median income. One fourth of the study population was current smoker, while almost half of the study population stated to be physically inactive. The median alcohol consumption was 13.9 gram per week. Socioeconomic inequalities in BMI were found in the study population with a lower BMI observed with higher household income and years of education (Table 2, model 1). On average, a 0.10 kg/m 2 higher BMI was seen per additional risk allele (Table 2, model 2). Adjusted for SEP, reporting no physical activity resulted in a higher BMI, while current smoking and higher alcohol consumption showed a negative effect on BMI (Table 2, model [3][4][5]. In stratified analyses, the genetic effect on BMI was strongest in groups with low SEP (Fig 1). Participants with �10 years of education showed the strongest genetic effect which almost disappeared in the highest educational group. Similar results were observed for income, although the trend across income quartiles was less pronounced. Results for the association of SEP indicators with BMI stratified by GRS BMI tertiles revealed the strongest effect sizes estimates in the highest GRS BMI stratum (S4 Table). In the analysis of joint effects with a single reference group, beta estimates showed a clear trend within and between groups: with increasing years of education and decreasing number of risk alleles, the weaker the effect size estimates. Compared to the reference group with the highest education and lowest GRS BMI , participants with a high GRS BMI and less than 10 years of education showed on average a 4.54 units higher BMI. Less clear patterns were observed for the joint effects of income and GRS BMI on BMI (Table 3).

Table 2. Sex-and age-adjusted effects and corresponding 95% confidence intervals (95% CI) on body mass index (BMI) in linear regression models including main effects of income (per 100€/month) and education (per year) as indicators of socioeconomic position (SEP), a BMI-associated genetic risk score (GRS BMI ), and factors of SEPrelated health behavior (no physical inactivity, current smoking, alcohol consumption [per 100g/week]
For the GRS BMI by income interaction term (ß = -0.05 [95% CI: -0.08; -0.02] per 1000 €/month increase and additional risk allele) as well as for the GRS BMI by education interaction term (ß = -0.02 [95% CI: -0.03; -0.01] per year of education and additional risk allele) negative effect size estimates were observed (Table 4, model 6).
After including physical activity, smoking and alcohol consumption, effect size estimates of the GRS BMI xSEP interaction virtually did not change (Table 4, models 7-9). Moreover, an indication for GRS BMI xPA interaction was found in regression models adjusted for income and for education (Table 4, model 7). However, no changes in interaction effects were observed after excluding income main effects and respective interaction effects from models 7-9. The GRS Edu showed an association with education (ß = 0.03 [95%-CI: 0.02, 0.05] per additional risk allele), but no indication for GRS BMI xGRS Edu interaction (Table 4, model 10).
Interaction effect size estimates between individual BMI-associated SNPs and SEP indicators were close to each other in the range of -0.20 to 0.15 for education and -0.33 to 0.38 for income (S1 Table and S2 Table). Interaction effect size estimates for some single SNPs pointed in the opposite direction than the respective interaction effect of the GRS. Effect sizes of the interaction effect between rs1558902 of the FTO gene and both SEP indicators were rather small compared to other SNPs of the GRS BMI . Interaction of a BMI-associated genetic risk score with socioeconomic position

Discussion
The aim of the present study was to investigate whether a sum score of genetic variants associated with BMI interacts with indicators of SEP in a population-based cohort and to explore SEP-related health behaviors and genetic factors which may explain such GRS BMI xSEP interactions. Results gave indication for a negative interaction between the GRS BMI and the SEP indicators income and education. This was supported by results of stratified analysis in which the strongest genetic effects on BMI were seen in groups of low SEP. Joint effects of all possible combinations of GRS BMI tertiles and education groups showed the strongest effect on BMI for Table 3 Table 4    participants with highest genetic risk within the lowest education group. Effect size estimates for the GRS BMI xSEP interactions remained virtually unchanged after including SEP-related health behaviors (i.e., smoking, physical activity and alcohol consumption) and educationrelated genetic factors into regression analysis, suggesting that these factors do not explain the observed GRS BMI xSEP interaction. Independent of the SEP effect, a GRS BMI by physical activity interaction was observed.

. Sex-and age-adjusted effects and corresponding 95% confidence intervals (95% CI) on body mass index (BMI) in linear regression models including main effects and interaction terms of a BMI-associated genetic risk score (GRS BMI ), indicators of socioeconomic position (SEP; years of education and 1000€ income/ month), SEP-related health behaviors (no physical activity [PA], current smoking [S], alcohol consumption [A; per 100g/week]) and a genetic risk score related to educational attainment (GRS Edu
The direction of the GRS by SEP interaction effect was consistent with results of previous studies: Liu et al. have investigated interaction between a GRS of BMI-associated risk alleles and education, using data of the Health and Retirement Study. While the GRSxEducation interaction on BMI revealed non-significant results, the direction of interaction effect size estimates has been consistent to the present results [29]. Likewise, another study that has examined interactions between BMI-associated GRSs and SEP reported smaller genetic effects within groups of higher educational attainment in two different study samples, albeit interaction effects were again statistically non-significant [30]. Within study samples of the UK Biobank, GRS BMI xSEP interaction has been observed using the area-related Townsend deprivation index (TDI) as SEP indicator [31,32]. Additionally, Tyrrell et al. have assessed interactions of GRS BMI with job class and education, measured by the ISCED, which, however, yield non-significant results. Rask-Andersen et al. have evaluated interactions between GRS BMI and 131 environmental factors and reported an GRS BMI xIncome interaction [31], which is comparable to the present results.
Even though education and income are correlated SEP indicators, each of them represents certain aspects of SEP related to different health behaviors and risks. Regarding the impact of SEP on BMI, several causal pathways have been described ranging from an improved knowledge about promoting health skills, increased social support in groups of higher education, to lower levels of stress and stress reactivity and better access to recreational facilities and healthy food with higher income [14,15,33]. Moreover, higher SEP leads to greater control over important determinants of health due to a better ability to engage in healthier behaviors. Via these pathways, SEP may also have an effect on the genetic susceptibility to higher BMI, including the possibility of altered gene expression by epigenetic modifications [34,35]. However, including SEP-related health behaviors physical activity, smoking and alcohol consumption in the present analysis no substantial changes in the GRS BMI xSEP interaction effect size estimates were revealed, indicating that other SEP-related risk factors potentially mediate the found GRS BMI xSEP interaction. In a recent study by Komulainen et al. within a Finnish study  sample, a GRS BMI xEducation interaction has also been reported [36]. Similar to the present results, the observed interaction effect size estimates attenuated only slightly after further adjusting for main effects of physical activity and their respective interaction effects. While in the present study no indication for a GRS BMI xS or GRS BMI xA interaction was observed, some indication for positive GRS BMI xPA interaction was obtained, supporting results of previous studies [31,37,38].
The present study adds further knowledge to the results of previous studies with the investigation of gene by gene interactions potentially underlying the observed GRS BMI xSEP interaction. However, the GRS Edu showed no indication for interaction with GRS BMI . This suggested that strong GenexGene interactions in the present study population are unlikely, albeit interactions of single BMI-with education-related genetic loci cannot be ruled out. In using a GRS the cumulative genetic risk of a person was being considered and thus, results of interaction analyses have to be interpreted rather global. The power of the study was not sufficient for robustly analysing single SNPxSNP interactions, indicators of SEP or SEP-related health behaviors. Effect size estimates of SNPxSEP interactions, however, were found to be rather similar in magnitude, indicating that no single genetic variant by itself noticeably triggered the observed GRS BMI xSEP interaction, which is in line with the results of a previous study [32]. Different to other studies, which have found interactions especially between variants of the fatmass and obesity-associated FTO gene [37,[39][40][41], the variant of the FTO gene used in the present study did not belong to the SNPs with strongest interaction effect.
Strengths of the present study include its population-based study sample and the use of two different individual SEP indicators in the analyses. Further, evidence on interaction was not only based on testing GRS BMI xSEP interaction terms, but also on stratified analyses and analysis of joint effects. Another advantage of the present study was that several health behaviors and a GRS of genetic variants related to educational attainment were included in the analyses. However, following limitations need to be considered: Next to the sample size and the limited statistical power for single SNP analyses, the age range of study participants did not enable to make statements about interactions between childhood/youth BMI and SEP during childhood/youth. As the detection of BMI-related genetic loci is still ongoing, the genetic information included in the analysis does not represent all genetic susceptibility to higher BMI.
In conclusion, the results of the present study provide indication for an interaction of genetic factors related to BMI with indicators of SEP in a population-based study sample, showing stronger genetic effects in groups of low SEP. This gives support to the hypothesis that SEP influences the expression of genetic susceptibility related to BMI: Higher SEP groups seem to be better enabled to reduce their genetic effect on BMI. However, physical activity, smoking status or alcohol consumption did not seem to explain the observed SEP-related differences. GRS BMI xSEP interaction has thus to be explained by other factors underlying the interaction effect. To fully understand the complex relationship of genetic variants and environmental factors on BMI, further research is needed that takes account of other factors, such as dietary patterns or stress levels, which might modify the effect of BMI-associated genetic variants on BMI.  (6), 97 SNPs representing independent genetic loci robustly associated with BMI were identified. Of these, 95 SNPs (incl. 2 SNPs were proxy SNPs [r 2 > 0.9]) were found in genotyped data of the Metabochip available for all study participants included in the analysis. Two SNPs (incl. 1 SNP was a proxy SNP [r 2 = 1]) were found in imputed and combined data of different Illumina genome-wide chips (Omni1-Quad n = 779, Omni1S n = 1,348, OmniExpressv1.0 n = 457, HumanCoreExome n = 1,747 with partly overlapping information available for 3,874 participants of the study population) with an imputation quality of 0.98 for both SNPs. As for these two SNPs information for 718 individuals was not available in the study population, this information was imputed based on the sample allele frequency according to the PLINK scoring routine (Purcell et al., 2007). The GRS BMI was then calculated for n = 4,493. (TIF)

S3 Fig. Selection of SNPs from different Illumina BeadArrays for calculating the genetic risk score associated with educational attainment (GRS Edu ) in the HNR study population.
Using the meta-analyses of genome-wide association studies by Okbay et al., 2016 (24), 74 independent genetic loci robustly associated with educational attainment were selected. Of these, 12 SNPs were found in data of the Metabochip (incl. proxies for 6 SNPs [r 2 between 0.94 and 1.00] and 3 imputed SNPs with an imputation quality between 0.94 and 0.99). Further, 60 SNPs were found in combined data of different Illumina genome-wide chips (Omni1-Quad n = 779, Omni1S n = 1,348, OmniExpressv1.0 n = 457, HumanCoreExome n = 1,747 with partly overlapping information available for 4,147 participants of the study population). Of these, 12 SNPs were selected from genotyped data and 48 SNPs (incl. 1 proxy SNP with r 2 = 1.0) were imputed with an imputation quality between 0.92 and 1.00. For 2 loci, neither the original SNP nor the proxy SNP was available. Thus, the GRS Edu was calculated including 72 SNPs associated with educational attainment for n = 4,147. (TIF)