A genome-wide association study of body mass index across early life and childhood

Background: Several studies have investigated the effect of known adult body mass index (BMI) associated single nucleotide polymorphisms (SNPs) on BMI in childhood. There has been no genome-wide association study (GWAS) of BMI trajectories over childhood. Methods: We conducted a GWAS meta-analysis of BMI trajectories from 1 to 17 years of age in 9377 children (77 967 measurements) from the Avon Longitudinal Study of Parents and Children (ALSPAC) and the Western Australian Pregnancy Cohort (Raine) Study. Genome-wide significant loci were examined in a further 3918 individuals (48 530 measurements) from Northern Finland. Linear mixed effects models with smoothing splines were used in each cohort for longitudinal modelling of BMI. Results: A novel SNP, downstream from the FAM120AOS gene on chromosome 9, was detected in the meta-analysis of ALSPAC and Raine. This association was driven by a difference in BMI at 8 years (T allele of rs944990 increased BMI; PSNP = 1.52 × 10−8), with a modest association with change in BMI over time (PWald(Change) = 0.006). Three known adult BMI-associated loci (FTO, MC4R and ADCY3) and one childhood obesity locus (OLFM4) reached genome-wide significance (PWald < 1.13 × 10−8) with BMI at 8 years and/or change over time. Conclusions: This GWAS of BMI trajectories over childhood identified a novel locus that warrants further investigation. We also observed genome-wide significance with previously established obesity loci, making the novel observation that these loci affected both the level and the rate of change in BMI. We have demonstrated that the use of repeated measures data can increase power to allow detection of genetic loci with smaller sample sizes.


Introduction
Obesity, defined by high body mass index (BMI), is a complex condition associated with an increased risk of many chronic diseases. 1 Evidence suggests that the adverse health consequences of obesity begin in early life. [2][3][4] BMI in adults primarily reflects weight independent of height; however, changes in BMI throughout childhood are influenced by changes in both height and weight.
Choh et al. 5 estimated that the narrow-sense heritability of BMI over infancy, childhood and adolescence ranges from 47% to 76%, with similar estimates reported in adults. 6,7 Genome-wide association studies (GWAS) have identified and replicated 34 adult BMI-associated loci (P-values < 5 Â 10 À8 from discovery and replication meta-analysis) [8][9][10] that explain approximately 1.5% of the variance in adult BMI. Several studies have investigated the association between these adult BMI SNPs and childhood BMI, including BMI trajectories. [11][12][13][14][15] These publications have shown that SNPs associated with adult BMI are not associated with birthweight or length, but have effects that begin from early infancy and strengthen throughout childhood and adolescence. The effects of some variants appear to change over time, for example the A allele at the FTO rs9939609 SNP is associated with a lower BMI in infancy but a higher BMI from 5.5 years onwards. 16 In contrast, the ADCY3 rs11676272 SNP has been shown to have consistent effects across ages. 17 A GWAS investigating childhood obesity (i.e. BMI > 95th percentile) in populations of European descent 18 identified two novel loci, one near the olfactomedin 4 (OLFM4) gene and the other in the homeobox B5 (HOXB5) gene. The results of these studies suggest that the expression of genetic variants may change across the life course.
In a recent review of obesity genetics, Day and Loos highlight the importance of conducting GWAS in children and adolescents to identify loci that may have effects early in life rather than adulthood. 19 Therefore, the aim of the current study was to investigate the genetic basis of BMI and BMI trajectories across childhood and adolescence.

Key Messages
• We performed a genome-wide association study of body mass index trajectories over childhood and adolescence using repeated measures data from two large cohort studies.
• A novel association between body mass index at 8 years of age and single nucleotide polymorphisms near the FAM120AOS gene on chromosome 9 was identified. The top SNP in this region begins to have an observable effect on BMI around 2 years of age, and the effect appears to be driven by changes in weight rather than height. This association warrants further investigation.
• Novel associations between several known adult body mass index-associated loci and body mass index at 8 years of age and/or change in body mass index over childhood were observed. In addition, a known childhood obesityassociated locus was associated with body mass index at 8 years of age.
• Through the use of repeated measures data, genetic associations can be detected with smaller sample sizes.
• Our results highlight that the genetic determinants of susceptibility to obesity in adulthood begin acting in early life and develop over the life course.

Genotyping
Imputed genotypic data used in both cohorts has been described previously. 24,25 Briefly, genotype data in both cohorts were cleaned using standard thresholds (HWE P > 5.7 Â 10 À7 , call rate >95% and minor allele frequency >1%). Imputation for chromosomes 1 to 22 was performed with the MACH software 26  Within-study GWAS analysis and quality control BMI was skewed, so a natural log transformation was applied before analysis. A semi-parametric linear mixed model was fitted to the BMI measures with mean centred age (centred at age 8); smoothing splines, with knot points at two, eight and 12 years and a cubic slope for each spline, were used to produce a smooth growth curve estimate. 27 A continuous autoregressive of order 1 correlation structure was assumed. Full details of the statistical methodology are in Supplementary material, available as Supplementary data at IJE online.
The fixed effects in the ALSPAC model included a binary indicator of measurement source (questionnaire vs clinic or health visitor measurement) to allow for differential measurement error. The fixed effects in the Raine model included the first five principal components for population stratification, calculated in EIGENSTRAT; 28 no adjustment for population stratification was made in ALSPAC, as previous analyses have shown that there is no obvious stratification.
A recent GWAS meta-analysis of adults from the Genetic Investigation of ANthropometric Traits (GIANT) consortium has shown that there were no genome-wide significant gender difference in SNP-BMI associations. 29 Therefore, the sexes were combined with the inclusion of a main sex effect and an interaction between sex and the spline function for age. This allowed the average BMI trajectory to differ between males and females, but with genetic effects assumed to be consistent across sexes.
The semi-parametric linear mixed model was fitted for each individual SNP. Genetic differences in the trajectories were estimated by including both a main effect for the imputed dosage and an interaction between the spline function for age and the imputed dosage for each genetic variant (i.e. an additive genetic model). We have previously shown that the type 1 error of the genetic effect over time in linear mixed effects models may be inflated if the function involving age terms in the fixed and random effects differs. 30

Meta-analysis
The meta-analysis for the SNP effect at age 8 years was conducted in Metal 32 [http://www.sph.umich.edu/csg/abecasis/ metal/], using inverse variance weighting. The analysis in Metal included an adjustment for genomic control in both cohorts, and a test of heterogeneity of the effect sizes was carried out.
Stouffer's method for combining P-values was used for the global Wald test and the Wald test for the SNP by age interaction. 33 Genomic inflation k was estimated by dividing the chi-square statistics, with 7 degrees of freedom for the global Wald test and 6 degrees of freedom for the SNP by age interaction Wald test, by the median of the central v 2 7 and v 2 6 distributions, respectively. 34 P-values from ALSPAC and Raine were adjusted for the estimated k values and weighted by the square root of the respective samples sizes in the meta-analysis.

Follow-up
The Northern Finland Birth Cohort of 1966 (NFBC1966) was used to replicate regions that reached genome-wide significance (P < 5 Â 10 À8 ) on any one of the three tests conducted. We investigated the SNP with the lowest P-value in each region and any potentially functional SNPs that were in high linkage disequilibrium (LD) with that SNP; LD was determined from SNAP 35

Characterization of significant loci
For loci reaching genome-wide significance, we plotted the trajectory for each of the genotypes and estimated the earliest age at which the effect of the SNP could be detected.
For novel SNPs, we also investigated their association with height and weight trajectories using a spline model similar to that used for BMI (see ref. 11 for description of methods used). These additional analyses were conducted in the ALSPAC study only, being the largest discovery cohort.

Results
Each cohort has a similar proportion of males and females; NFBC1966 has a lower average weight, and consequently BMI, than the two discovery cohorts from approximately 6 years of age, whereas their average height is similar ( Table 1,  Five regions reached genome-wide significance at 5 Â 10 À8 (Table 2, and Supplementary Table 1 available as Supplementary data at IJE online); three of these loci have previously been identified through GWAS of adult BMI (FTO, 38 MC4R 39 and ADCY3 8 ) and one locus has previously been shown to be associated with paediatric obesity (OLFM4. 18 ). Manhattan plots for each of the three tests are in Supplementary Figures 4-6, (available as Supplementary data at IJE online).
A novel genome-wide significant locus was found downstream from FAM120AOS, which has not previously been reported to be associated with any adiposityrelated traits. The most statistically significant SNP, rs944990, was associated with BMI intercept at 8 years (T allele: b SNP ¼ 0.012, P SNP ¼ 1.52 Â 10 À8 ), and showed modest evidence of association with change in BMI (P Wald(Change) ¼ 0.006; Supplementary Figure 9, available as Supplementary data at IJE online). A consistent direction of effect was seen in NFBC1966 for rs944990; however, P > 0.05 for all three tests ( Table 2). Figure 1 shows the BMI trajectories in the ALSPAC cohort from 1 to 17 years of age in males and females for individuals who have zero, one or two BMI decreasing alleles (major allele) at the rs944990 locus. A male homozygous for the T allele would have an average BMI of 18.16 kg/m 2 at 1 year of age, which would increase to an average of 21.88 kg/m 2 by age 16. In contrast, a male homozygous for the C allele would on average have a BMI of 18.12 kg/m 2 at 1 year of age and by age 16 it would be 21.48 kg/m 2 .  In ALSPAC, the effect is detectable from 2 years of age ( Figure 2). Consistent results were observed after including additional fixed effects in the model to account for change in height (Supplementary Table 2, available as Supplementary data at IJE online). rs944990 showed a stronger association with weight than height (Supplementary Table 2) and had the strongest influence on both weight and height over the pre-pubertal years (7-10 years; Figures 3 and 4). We investigated the association between energy intake and rs944990 from 3 to 14 years of age in the ALSPAC cohort (Supplementary  Table 3, available as Supplementary data at IJE online).
Assuming an additive effect of the SNP, there was a marginal association at 39 months of age (b ¼ 14.28 kCal per T allele, P < 0.001) and only a slight increase in energy intake was observed around the peak of the BMI association (10 years of age; b ¼ 12.80 kCal per T allele, The most statistically significant SNP at this novel FAM120AOS locus in our meta-analysis, rs944990, is in LD (r 2 ¼ 0.72, D' ¼ 0.96 35 ) with a non-synonymous variant that may disrupt splicing activity 36 , rs10821128. In the meta-analysis, the C allele of rs10821128 was associated with an increase in BMI intercept at age 8 years (b SNP ¼ 0.011, P SNP ¼ 5.29 Â 10 À8 ) and showed weak evidence for an increased change in BMI (P Wald ¼ 1.14 Â 10 À4 ; P Wald(Change) ¼ 0.014). Given the potential functionality of this SNP, we analysed rs10821128 in the NFBC1966 cohort. Consistent with the results from the meta-analysis, the C allele at rs10821128 was associated with BMI intercept at 8 years of age in NFBC1966 (b SNP ¼ 0.005, P SNP ¼ 0.041) and the global Wald test (P Wald ¼ 0.041), but not with change in BMI (P Wald(Change) ¼ 0.139). Neither this SNP nor rs944990 was associated with mean BMI (rs944990: P ¼ 0.887; rs10821128: P ¼ 0.931) in the GIANT consortium metaanalysis 8 but showed weak evidence of association with phenotypic variability of BMI 40 (rs944990: P ¼ 0.063; rs10821128: P ¼ 0.092). We used publicly available results from GWAS in various consortia to investigate whether rs944990 or rs10821128 was associated with any other growth-related endpoint (including BMI, height, waist to hip ratio, obesity, glucose, insulin, homeostatic model assessment-insulin resistance (HOMA-IR) and type 2 diabetes in adults, age of menarche, pubertal growth and obesity in children and birthweight, birth length and head circumference at birth). Both rs944990 (P ¼ 4.0 Â 10 À4 ) and rs10821128 (P ¼ 5.9x10 À4 ) were associated with age of menarche, and rs10821128 was marginally associated with total pubertal growth (P ¼ 0.044), consistent with the observation that rs944990 influences pre-pubertal weight and height growth (Figures 3 and 4) Figures 12 and 13, available as Supplementary data at IJE online) and MC4R ( Supplementary Figures 14 and 15, available as Supplementary data at IJE online) reached genome-wide significance for all three tests, indicating that these loci influence both BMI intercept at 8 years and change in BMI trajectory over childhood (FTO rs1558902: b SNP ¼ 0.013, P SNP ¼ 1.54 Â 10 À10 ; P Wald ¼ 1.48 Â 10 À21 ; P Wald(Change) ¼ 4.99 Â 10 À22 ; MC4R rs571312: b SNP ¼ 0.013, P SNP ¼ 4.94 Â 10 À8 ; P Wald ¼ 1.27 Â 10 À9 ; P Wald(Change) ¼ 2.81 Â 10 À9 ). The A allele at rs1558902 in FTO was associated with a decrease in BMI until 2 years of age, but an increase in BMI from 6 years of age (Figure 2 and Supplementary Figure 13).
The OLFM4 rs12429545 SNP reached genome-wide significance for the SNP effect at age 8 (b SNP ¼ 0.016, P SNP ¼ 1.90 Â 10 À8 ) and global Wald tests (P Wald ¼ 1.13 Â 10 À8 ) and showed weak evidence for association with change in BMI over time ( Supplementary Figures 10 and 11, available as Supplementary data at IJE online).
To investigate whether known adult BMI-associated loci influenced BMI trajectory through childhood and adolescence, we investigated the relationship between the 33 top adult BMI-associated SNPs and the SNP effect at age 8 in our meta-analysis. One SNP from the adult BMI GWAS, rs11847697, was not available in our study as it had a minor allele frequency less than 5%. We observed 15 loci with directionally consistent associations at nominal significance (P < 0.05), 12 of which reached P < 0.01. Figure 5 indicates that the majority of the loci show the same direction of effect for both the SNP effect at age 8 and the adult BMI meta-analyses. Results from the metaanalysis of the three tests for the 33 adult BMI-associated SNPs are presented in Supplementary Figure 16 and Supplementary Table 5 (available as Supplementary data at IJE online). We also looked up the HOXB5 and OLFM4 SNPs, previously associated with childhood obesity. 18 The obesity risk T allele at rs9299 in HOXB5 increased BMI intercept at 8 years (b SNP ¼ 0.004, Figure 2. Associations from the ALSPAC cohort between the genome-wide significant SNPs and BMI from age one to 16 years. Error bars represent the regression coefficient of BMI on the natural log scale and 95% confidence intervals derived from the longitudinal additive genetic models. The SNPs are aligned to the minor allele. P SNP ¼ 0.068), but was not associated with change in BMI over time (P Wald ¼ 0.591; P Wald(Change) ¼ 0.728). The obesity risk A allele at rs9568856 near OLFM4 was associated with increased BMI intercept at 8 years (b SNP ¼ 0.014, P SNP ¼ 1.68 Â 10 À6 ) and increased change in BMI (P Wald ¼ 3.24 Â 10 À6 ; P Wald(Change) ¼ 0.002).

Discussion
Previous studies have shown that some genetic variants for BMI have varying effects across the life course, suggesting potential age-specific gene expression. [14][15][16][17] We conducted a genome-wide meta-analysis of BMI trajectories from 1 to 17 years of age and identified a novel association with rs944990, downstream from FAM120AOS on chromosome 9. This SNP was associated with BMI intercept at 8 years, an effect that was independent of height. The effect size for the T allele, 0.20 kg/m 2 in ALSPAC, is comparable to the effect sizes reported for adult BMI-associated SNPs. 8 In addition, rs944990 showed a stronger association with adiposity than skeletal growth and had the greatest influence on pre-pubertal growth. This association did not replicate in the NFBC1966 cohort; however, a potentially functional non-synonymous SNP in LD was associated with BMI intercept in NFBC1966. The lack of convincing evidence for association in NFBC1966 could be due to the observed difference in BMI trajectories to the ALSPAC and Raine cohorts (as seen in Supplementary Figure 1); with fewer individuals in the right-hand tail of the BMI distribution, it is more difficult to detect an association. This difference in trajectories could be due to several causes, including different genetic profiles or generational effects.
Little is known about the function of FAM120AOS. However, there are two genes in the region surrounding FAM120AOS that have shown evidence to support a role for this region in growth. An SNP in the first of those genes, NINJ1, has been shown to be associated with severe obesity in children. 41 The second gene, PHF2, has been shown to influence bone development in newborn mice 42 and adipocyte differentiation. 43 In addition to these two plausible candidate genes in the region surrounding FAM120AOS, we identified that the T allele at rs944990 was associated with increased BMI, increased change in BMI and earlier age of menarche. The direction of effect in this study is consistent with the observed phenotypic correlation, 44,45 where girls who reach menarche earlier tend to have a higher BMI in later life than girls who reach menarche later. The effect seen by rs944990 is similar to that reported for the LIN28B locus, whereby it influences BMI and weight from adolescence to early/mid adulthood, and is associated with age of menarche. 46 Therefore, we believe this region of chromosome 9 warrants further investigation.
Four loci previously associated with either adult BMI (FTO, 38 MC4R 39 and ADCY3 8 ) or childhood obesity (OLFM4 18 ) reached genome-wide significance for BMI level and/or slope. These loci were detected with a smaller sample size than the original studies (>20 000 subjects in ref. 18 and 249 796 subjects in ref. 8), consistent with the increased statistical power gained through using repeated measures. 47 A small increase in BMI may not necessarily lead to obesity; however, our results demonstrate for the first time that the obesity risk allele of the OLFM4 rs9568856 SNP 18 is associated with increased BMI at 8 years and increased change in BMI over childhood, rather than having an effect on obesity only. This SNP, along with the top SNPs from the FTO and MC4R loci, had an effect at baseline as well as showing an increasing effect over childhood. In contrast, the top SNP at ADCY3 was associated with baseline and its effect over time was relatively constant. These results are consistent with previous studies on adult BMI loci. 16,17,39 Furthermore, we detected a decrease in BMI in infancy for the FTO SNP, rs1558902, followed by an increase from early childhood: a result which is consistent with our previous work. 16 Approximately half of the known adult BMI SNPs showed nominal effects on BMI at age 8, suggesting that these SNPs begin having an effect in childhood. The SNP effect sizes on BMI at age 8 are larger than the effects on adult BMI for some SNPs; for example, the RBJ/ADCY3 locus had an effect size of 0.14 kg/m 2 in adults 8 and 0.18 kg/m 2 at age 8. Those SNPs that did not have an effect on BMI in childhood may indicate early-onset vs adult-onset SNPs or may be due to lower power in our study.
There are several limitations to this study. These analyses were much more computationally intensive than the cross-sectional models commonly used in GWAS. 27 For example, both the ALSPAC and Raine GWAS took approximately 2 months (approximately 1440 h) to analyse on high performance clusters, i.e. BlueCrystal Phase 2 cluster: [https://www.acrc.bris.ac.uk/acrc/phase2.htm] and [http://www.ivec.org/], respectively. This computational burden has limited the current analyses to genetic data on the 22 autosomes using HapMap2 imputation, rather than the more recent 1000 genomes, and to two cohorts with this capacity. Replication is also challenging because of the need for detailed repeated measures data from across childhood and adolescence. Furthermore, there has been some criticism over recent years of the use of BMI as a measure of adiposity throughout childhood because at some ages BMI remains correlated with height and this correlation changes with age. 48 Stergiakouli et al. 17 showed that a different power of height was required at different ages throughout childhood, ranging from 1.5 at 12 months to 3.1 at 8-12 years. However, their results suggest that genetic influences on weight/height are similar to those for BMI adjusted for height. We therefore conducted a sensitivity analysis of our novel SNP by including a fixed effect for change in height and showed that the association between rs944990 and BMI remained.
Identifying genetic variants that have age-specific effects has the potential to shed light on the life-course aetiology of health and disease as well as potentially providing clues to gene function. Our results are consistent with the hypothesis that the genetic determinants of adult susceptibility to obesity act on both BMI intercept at 8 years and trajectory from childhood, change with age and develop over the life course. As with adult loci, the genetic variants associated with BMI in children only explain a small proportion of the estimated heritability of childhood BMI. Hence, considerable opportunity exists for new insights into the biology of childhood obesity.

Supplementary Data
Supplementary data are available at IJE online.  . Association results of SNP effect at age 8 years for known adult BMI-associated loci. The y-axis displays the effect size from published meta-analyses [8][9][10] and the x-axis displays the effect size from the meta-analysis presented in this paper of the BMI intercept at age 8. Effect sizes are aligned to the adult BMI-increasing allele. Colour indicates association P-values from the longitudinal model over childhood: dark grey ¼ P < 5 Â 10 À8 , medium grey ¼ 5 Â 10 À8 P < 0.001, light grey ¼ 0.001 P < 0.01, white ¼ P ! 0.01. The dotted grey line indicates the mean effect size for the 33 SNPs in the longitudinal model parameters. The effect sizes from the meta-analysis were transformed from the log scale 49 using an intercept of 2.80 and standard deviation of 0.057, for comparison with the adult effect sizes.