Genetic risk score predicts risk for overweight and obesity in Finnish preadolescents

Summary Common genetic variants predispose to obesity with varying contribution by age. We incorporated known genetic variants into genetic risk scores (GRSs) and investigated their associations with overweight/obesity and central obesity in preadolescents. Furthermore, we compared GRSs with lifestyle factors, and tested if they predict the change in body size and shape in a 4‐year follow‐up. We utilized 1142 subjects from the Finnish Health in Teens (Fin‐HIT) cohort. Overweight and obesity were defined with age‐ and gender‐specific body mass index (BMI) z‐score (BMIz), while central obesity by the waist‐to‐height ratio (WHtR). Background data on parental language, eating habits, leisure‐time physical activity (LTPA) and sleep duration were included. Genotyping was performed with the Metabochip platform. Weighted, standardized GRSs were derived. Of the11‐year‐old children, 25.5% were at least overweight and 90.8% had Finnish speaking background. BMI‐GRS was associated with higher risk for overweight with odds ratio (95% confidence interval) of 1.39 (1.20; 1.60) and obesity 1.41 (1.08; 1.83), but not with central obesity. BMI‐GRS was weakly and inversely associated with the changes in BMIz and WHtR in the 4‐year follow‐up. Waist‐to‐hip ratio‐GRS was not related to any obesity measures at baseline nor in the follow‐up. The effect of BMI‐GRS is similar to that of low LTPA on overweight. An interaction between parental language and BMI‐GRS was noted (P = .019): BMI‐GRS associated more strongly with overweight in Swedish than in Finnish speakers. We further identified two suggestive genetic variants near LOC101926977 and LOC105369677 associated with BMIz in preadolescents which were replicated in the adult population. In preadolescents, known genetic predisposing factors induce a risk for overweight comparable to low LTPA. However, the GRS was poor in predicting short‐term changes in BMI or WHtR.


| INTRODUCTION
The world is in the midst of an epidemic of obesity, and the number of obese children worldwide has increased to 124 million in 2016. 1 Data from the prospective Finnish Health in Teens (Fin-HIT) cohort in 2011 to 2014 showed that about 12.7% of 9-to 12-year-old Finnish children were overweight, while 2.6% were obese 2 and similar prevalence was maintained in the follow-up 3 to 5 years later. Children with overweight have a higher risk for non-communicable diseases that might also appear at a younger age than in normal-weight peers. 3,4 Since early prevention of weight gain is of great importance, 5 any tools for early identification of persons at risk are valued.
The aetiology of obesity is complex, with many factors involved in its development, for example, lifestyle factors, other environmental factors, genetic susceptibility, and likely their interactions. The era of genome-wide association studies (GWASs) has increased our knowledge of common genetic variants together explaining about 6% of the variation in adult body mass index (BMI). 6 Critically considered, the GWASs have not added to our knowledge of predicting BMI beyond maternal BMI, 7 or other well-characterized conventional factors, for example, family income, birth weight, high-sensitive c-reactive protein. 8 In turn, several other studies have witnessed that many of the same genetic variants contribute to the variation in child BMI and BMI trajectories from childhood. [9][10][11] However, their contribution may appear and vary with age and in different stages of life. 11,12 BMI is a measure of body size reflecting fat accumulation and overall adiposity, while body shape in terms of waist circumference or waist-to-hip ratio (WHR) distinguishes fat distribution and serves as markers of central obesity. WHR is independently of BMI associated with an elevated risk for chronic diseases including type 2 diabetes and cardiovascular diseases. 13 Despite the crudeness of these anthropometric measures, these traits and some of their associated genetic loci have been validated against more sophisticated phenotypes (body fat content or local fat distribution by magnetic resonance imaging). 14,15 During puberty, body size and fat distribution change dramatically. It is shown that overweight, especially during puberty, modifies the risk of type 2 diabetes in adulthood, while normalization of BMI before puberty may reduce this risk. 16 In the present study we investigated associations of genetic loci with body size and shape in 1142 Finnish preadolescents and created genetic risk scores (GRSs) for BMI and WHR based on the literature. 17,18 We hypothesize that the GRSs identify persons with elevated risk for obesity and central obesity. In addition, we compared GRSs with other lifestyle factors and tested if the GRSs predict change in body size and shape in a 4-year follow-up study.

| Study design and population
Material collected in the Fin-HIT cohort was utilized in the present study. The cohort is described elsewhere in detail. 2 Shortly, the Fin-HIT is a prospective cohort consisting of over 11 000 9-to 12-yearold (born between 1998 and 2006) 22 BMIz cut-offs for underweight, normal-weight, overweight and obesity were less than −2, between −2 and +1, over +1 and over +2, respectively.
Waist circumference was assessed midway between the hip bone and the ribs to the nearest 0.1 cm with a measuring tape as previously described. 2 Waist-to-height-ratio (WHtR) was calculated by dividing the waist circumference (cm) by the height (cm), and it illustrates body shape. In children, WHtR mirrors cardiovascular risk factors more accurately than a combination of BMI and waist. 23 A cut-off of 0.5 is used for central obesity. 24,25 In the follow-up, the child's height, weight and waist were selfmeasured and -reported by a parent. We have previously reported the validity of the self-reported anthropometry. 26 2.4 | Questionnaire data

| Eating habits
Subjects filled in a 16-item food frequency questionnaire covering the preceding month. The selected food items were adopted from Health Behaviour in School-Aged Children Study protocol 27 and these included fruits, fresh or cooked vegetables, sugary soft drinks, dark grain bread, milk or soured milk, fresh juice, water, pizza, hamburger or hot dog, biscuits/cookies, ice cream, chocolate or sweets, salty snacks and sugary juice drinks. Children reported the frequency of consumption for each item on a 7-point scale ranging from 0 (not consumed) to 6 (consumed several times per day).
Based on food item frequencies, eating habit variable was created utilizing a factor and cluster analyses as explained more in depth elsewhere. 28 The cluster analysis confirmed three eating habits in the cohort: unhealthy eating habit (loaded with pizza, hamburger or hot dog, baked goods, salty snacks and sugary drinks), fruit and vegetable avoider (avoided unhealthy food items, fruit, berries and vegetables in all forms) and healthy eating habit (loaded with dark grain bread, milk, fruits, berries and vegetables in all forms).

| Leisure-time physical activity
Duration of leisure-time physical activity (LTPA) was assessed with the question: "How many hours a week do you normally exercise or do sports during your free time?" with 10 response options ranging from "An hour a week or less" to "Around ten hours a week." We categorized the responses into two categories, less than 7 hours per week or at least 7 hours per week, to mimic the adherence to current guidelines. 29

| Sleep duration
The subjects replied the time asleep on school nights with the question: "When do you usually fall asleep in the evenings on a school night?" with 12 response options. Correspondingly, the waking up on school days was asked: "When do you usually wake up on school days?" with seven response options. Sleep duration on school nights was calculated and categorized into three groups: (a) less than recommended, (b) recommended and (c) more than recommended according to the age-specific Childhood Sleep Guidelines by the American Academy of Pediatrics 30 : children between 6 and 12 years should sleep daily 9 to 12 hours on a regular basis to promote optimal health. Since we had only three observations in the third category, these subjects were combined into the recommended category.

| Creation of genetic risk scores
We selected 32 SNPs for BMI and 32 SNPs for WHR based on the large GWASs performed on these traits at the time of study design in 2010, reflecting also the selection of SNPs to the Metabochip genotyping platform. 17,18 Of the 32 BMI-SNPs, 21 were directly available on the Metabochip platform, and a good proxy SNP was available for nine SNPs (r 2 > 0.80, MAF > 5% or r 2 = 1.00 if MAF < 5%). Proxy SNPs were derived using LDlink's LDproxy tool, and Finnish population data. 33 Two BMI SNPs (rs4771122 and rs4836133) were not available and had no good proxies. Thus, the final BMI-GRS consisted of 30 SNPs (Table S1). Of the 32 WHR SNPs, two were not available on the Metabochip platform (rs9687846 and rs12608504) and had no proxies, and thus, 30 SNPs were included in the WHR-GRS.
We created weighted GRSs using the score function in Plink v1.09, which calculates an average score per non-missing SNP. This score is calculated by multiplying the weight of each SNP with the number of risk alleles for that SNP. As weights, we used the effect sizes from the original publications for BMI (Speliotes et al,17 Table S1) and WHR (Heid et al, 18 Table S2). Finally, we standardized the GRS using the mean and SD in the sample. Since there were more recent and larger GWASs available 6,7,34 than the one of Speliotes et al, 17 we reconstructed BMI-GRSs based on the different sources and compared their performance in our sample (Table S3) Bonferroni-corrected threshold for 68 183 independent SNPs. We defined a suggestive hit as P < 1.0 × 10 −4 .

| Other statistical analyses
The normal distribution of variables was visually inspected, and transformations applied when needed. Comparison of baseline characteristics between two groups was performed with independent samples t test in case of a continuous variable, and with chi-square test for categorical variables.
Logistic regression was used to study the association of GRSs with overweight (BMIz > +1), obesity (BMIz > +2) and central obesity (WHtR > 0.5) in two models: first GRS alone adjusted for PC1 and PC2, and a second multivariate model including GRS with all covariates, and adjusted for age, PC1 and PC2. Covariates in the analysis were gender, parental language, LTPA, sleep duration, and eating habits. Female gender, Finnish language, LTPA >7 hours per week, recommended sleep duration and healthy eating habit were considered as reference categories in the analysis. The associations were reported as odds ratio (OR) with 95% confidence interval (CI).
We had missing values in several covariates: parental language (n = 33), LTPA (n = 11), eating habit (n = 109) and sleep (n = 59). The missing values were replaced in logistic regression analysis using the multiple imputation procedure in SPSS, in order to maintain the full sample size. The imputation method was "fully conditional specification," which suits for arbitrary missing data. Multiple variables, for example, all covariates, GRSs and outcome measures in original scale, were included in the imputation process.
Associations of covariates (gender, parental language, LTPA, sleep duration, eating habit) with the two GRSs were tested with t test or analysis of variance. Interactions between covariates and GRSs were tested with loglikelihood ratio comparing models with and without interaction terms.

Association of GRSs with changes in BMIz and
All statistical analyses were conducted using the IBM SPSS program for Windows, version 22 (IBM, Chicago, Illinois). The statistical significance level was set at 5%.

| Baseline characteristics
Of the total of 1142 subjects, 8 (0.7%) were categorized as underweight, 843 (73.8%) normal weight, 230 (20.1%) overweight and 61 (5.3%) obese based on IOTF age-and gender-specific BMIz. A combined prevalence of overweight and obesity was more common in boys than in girls (29.7% vs 21.3%, P = 0.001), while the prevalence did not differ by parental language: 25.2%, 30.8% and 20.8% with Finnish, Swedish and other language family background, respectively (P = .732). The subjects were recategorized into two groups combining under-and normal-weight into "UW/NW" (BMIz ≤ +1) and overweight and obese into "OW/OB" (BMIz > +1) ( Table 1). As expected, weight, waist, BMI, WHtR and parental BMI were higher in OW/OB than in UW/NW group, while no differences were observed in height or age. Male subjects and lower physical activity level were more commonly seen in OW/OB than UW/NW group.
Follow-up measurements were available from 727 subjects, and their baseline characteristics were similar to those of 1142 subjects (Table S3). During the follow-up, weight increased similarly in the two groups, but a higher increment in height (22.1 vs 20.4 cm, P = .005) and waist circumference (8.5 vs 5.6 cm, P < .001) were observed in UW/NW than in OW/OB group. BMIz increased among those in UW/NW group, while decreased in OW/OB group. In both groups, the WHtR decreased, but more so in OW/OB group.

| Genetic risk scores
BMI-GRS was higher in the OW/OB than in the UW/NW group (P < .001), while no group difference was observed for WHR-GRS (Table 1). With our sample size we had 81% power to detect an explained variance (R 2 ) of 1.0% and 95% power to detect R 2 of 1.5%.

| Associations of GRSs with overweight, obesity and central obesity at baseline
Associations of GRSs with overweight (BMIz > +1), obesity (BMIz > +2) and central obesity (WHtR ≥ 0.5) were tested with logistic regression in two models (Tables 2 and 3). An SD-increase in weighted BMI-GRS increased the risk for overweight with an OR of 1.39 (95% CI:   baseline BMIz or WHtR overruled the effects of BMI-GRS. The WHR-GRS was not associated with changes in BMIz or WHtR (Table S4).

| Interaction with parental language
Interactions were tested between GRSs and sex, parental language, LTPA, sleep, and eating habits in all models regarding overweight, obesity and central obesity. An interaction appeared between parental language and BMI-GRS concerning overweight/obesity (P = .019) ( Figure 1).
None of the variants were associated with BMI in young children in the EGG Consortium 34 (Table 4).
We further tested the correlation between the effect size estimates of the suggestive SNPs found for BMIz in our study and in previous GWASs on BMI. 6,7,34 Borderline significant (r = 0.6, P = .08) positive correlations were observed with SNP effects in the two adult population studies, while the effect sizes did not correlate with estimates for BMI in young children from the EGG Consortium (P = .43; Figure 2).
Of the 16 suggestive SNPs associated with WHtR, variants in or near LTBP1, PDE1C, GALNT17, DPP6, MAP4K5 and EYA2 were nominally associated with BMI or WHR in adult population 17,18,35,36 (  38 We created BMI-GRSs based on data of Speliotes et al, 17 and 30 of these 32 SNPs were incorporated in Metabochip. The BMI-GRS was associated with risk of overweight and obesity, and explained as much as 3.7% of BMIz at the baseline, but was not informative regarding the risk of central obesity. Previously, GRS applied to paediatric data has been reported to predict adulthood obesity efficiently 20 to 30 years later. 8 However, different genetic factors may affect the short-term changes in BMI, especially during rapid growth period. 11 Besides body size, we were interested in body shape in terms of central obesity, which has been linked to cardiovascular diseases and diabetes more strongly than BMI. 13 While WHR has been used as a marker of central obesity in adult GWA studies, we utilized WHtR as a tool for determining central obesity in children in the present   39,40 : the prevalence of overweight and obesity are reported to be higher in boys than in girls in Finland. A higher prevalence of overweight and obesity in boys than in girls at this age seems a common observation, 1 although not consistent across the globe. Differences in the prevalence of excess weight appear after childhood, suggesting that growth pattern and hormonal changes play a role, which is also supported by the fact that similar or even higher prevalence is reported in girls at younger age groups. 41 Sex differences may also be F I G U R E 2 Effect size comparison of the Fin-HIT suggestive SNPs for BMIz in adult population and in young children related to life-style factors and beliefs in children and their parents. 42 Previously, we have shown that unhealthy eating habit was more common in boys than in girls, 28 while typically boys are more physically active at this age 43,44 providing no simple explanation for the phenomenon.
Exercising less than 7 hours per week led to a higher risk for overweight and central obesity in our children, which is in line with current physical activity guidelines. 29  We observed an interaction between GRSs and parental language.
In the present study, parental language mirrors socioeconomic status as previously suggested. 21 8,10 However, in our study, the BMI-GRS was poor in predicting short-term changes in BMIz or WHtR. Surprisingly, higher BMI-GRS was associated with smaller changes in BMI and WHtR, but not after further adjustment for baseline measures. This is further supported by our data pointing that during the 4-year follow-up the BMI was to some extent normalized in overweight/obese group, while this continued to increase in under-/normal-weight group. This implies that growth speed varies by age. It seems the overweight/obese had grown at earlier state compared with under-/normal-weight group, as suggested before. 11,51 Metabochip was utilized here to discover potential novel genetic variants for BMIz in preadolescents. Even though our metabochipwide association study was underpowered, we identified 12 + 16 suggestive SNPs associated with P < 10 −4 with BMIz and WHtR, respectively. Two BMI-related SNPs (rs12680842 and rs10840674) were robustly replicated in adult population, 6 while several other SNPs reached nominal replication in the adult population as well. Of note, none of the suggestive SNPs were associated with BMI in the EGG consortium GWAS in children aged between 2 and 11 years. 34 Furthermore, the effect size estimates of the suggestive SNPs for BMIz correlated with the corresponding ones in the adult population, 6,7 but not with effect size estimates in young children, 34 which suggest that genetic factors related to body size in preadolescents are more similar to adults than to younger children.
The study had several limitations. The sample size limited the power of metabochip-wide association analysis, and only suggestive associations were identified. The Metabochip contained over 65 000 SNPs with MAF < 1% in this Finnish study, and those were excluded as rare SNPs were not of interest here. The Metabochip platform is not a fully genome-wide genotyping platform as it contains markers only in some regions of the genome, for example, regions selected for fine-mapping and replication based on GWASs on metabolic traits performed before 2012. Thus, it provides less data required for constructing the LD patterns, which is important for genotype imputation. 52 Regarding the phenotype, the secondary trait in our study was WHtR, which differs from WHR that has been used as a marker of central adiposity in adults, 18 and a discrepancy between the traits may cause some inaccuracy. An additional limitation was that we had follow-up data only on 63% of the participants. At follow-up, the anthropometric measures were self-assessed, which might cause some bias. Previously, we have validated the self-reported measures 26 : home-measured mean height, weight and waist circumference were slightly higher, but derived BMI lower than measured by the fieldworker. However, the differences were so small that they had no impact on weight status. The sampling for Metabochip array was random, but the participants in the present study showed a somewhat higher prevalence of overweight/obesity than seen in the entire cohort, but still comparable with the Finnish population at this age. 39,40 We have demonstrated that BMI-GRS is a significant risk factor for overweight and obesity in preadolescents, but is poor in predicting short-term changes in body size and shape during puberty. The associations of BMI-GRS and low physical activity with overweight/ obesity were similar, but somewhat weaker than being a male. An interaction with socioeconomic status was observed, suggesting that