Introduction

Obesity is prevalent in the Western world, posing a serious health risk.1 Genes are known to contribute to its pathogenesis.2 The discovery of genetic interactions with age, other genes or the environment may improve consistency of associations as well as the understanding of the mechanisms leading to or underlying obesity. Here we present longitudinal interaction analyses for two candidate single-nucleotide polymorphisms (SNPs); one located upstream of the insulin-induced gene 2 (INSIG2) on chromosome 2 and the other in the melanocortin-4 receptor gene (MC4R) on chromosome 18. INSIG2 rs7566605 and MC4R rs2229616 were the first SNPs with a replicated polygenic effect on body mass in the general population3, 4 and known functional implications.5, 6

The INSIG2 region has been connected with obesity in human linkage studies7 as well as in one of the first genome-wide association studies (GWAS)3 to address obesity. This GWAS identified common variant rs7566605, located 10-kb upstream of the INSIG2 transcription start site. Elevated body mass index (BMI) or increased obesity risk was found for minor allele homozygotes (∼11% Caucasians) in at least six independent studies.3, 8 INSIG2 rs7566605 interaction analyses are of particular interest, as its main effect on BMI or obesity status was questioned following equal numbers of replications and non-replications of association in >20 well-powered data sets.3, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17 Combining all study results on INSIG2 suggest age as the most probable interaction candidate. A large meta-analysis18 (34 studies, 74 345 individuals) confirmed association of INSIG2 with obesity across study designs when comparing extremes (eg, BMI ≥32.5 kg/m2 versus BMI <25 kg/m2). No heterogeneity in INSIG2 main effect estimates was found between studies with higher compared with lower mean subject age; however, contrasted studies overlapped considerably in age range, with mean ages lying between 41 and 58 years.18 Nevertheless, this meta-analysis revealed an increased obesity risk for rs7566605 minor allele homozygotes in general population studies (odds ratio 1.092, P=0.035, 48 844 subjects from 16 studies), but a decreased obesity risk in population-based studies with subjects selected for better health status (odds ratio 0.796, P=0.028, 7640 subjects from 5 studies). Six other studies also indicated a tendency toward a protective effect of rs7566605 risk genotype CC on BMI or obesity status,3, 9, 10, 12, 14 or on waist-to-hip ratio.16 The effect strength of rs7566605 association varied between the first six longitudinal examinations on unrelated Framingham-Offspring subjects.8 Tests for rs7566605–age interaction were not significant for early childhood to middle age (4–50 years).19, 20 However, a lower gain in weight-for-length was reported for genotype CC babies between birth and the age of 6 months.21 These findings suggest an INSIG2–age interaction with weak marginal effect because of crossing BMI-age trends for different INSIG2 genotypes at a mid-life age (Figure 1, left). In this case, contrasting BMI extremes is robust but the heterogeneity test will not detect this, and power and sign of marginal effect estimates depend on the age range studied.22, 23

Figure 1
figure 1

INSIG2–age interaction on longitudinal BMI. Left: growth curves of raw BMI for INSIG2 genotypes in men (joint analysis: FHS-Off, KORA3, KORA4). Displayed is a corridor of ±1 SD of the random subject-effect on intercept. Random subject-effects on the regression coefficient were small. Similar results were obtained for women and in analyses of rank-normalized BMI (Table 2). Right: INSIG2–age interaction on residual BMI (joint model: FHS-Off examinations 3 (1983–1987) and 5 (1991–1995); KORA3, KORA4). Displayed are the median and interquartile of residual BMI, after adjusting raw BMI for sex, baseline age, education and cohort. INSIG2 genotype is displayed by box colour (GG=white, GC=light grey, CC=dark grey), baseline age class by figure column. Longitudinal examinations T1 (KORA baseline, FHS-Off examination 3) and T2 (KORA follow-up, FHS-Off examination 5) are presented adjacent to each other.

Full size image

INSIG2 fulfils a role complementary to insulin-induced gene 1 (INSIG1) in the regulation of cholesterol and triglyceride levels.24, 25 However, both genes also have separate additional functions. INSIG2 (but not INSIG1) is predominant in adipogenesis with a marked increase in gene expression during adipocyte differentiation.5 Multi-marker tag-SNP haplotypes in the close neighbourhood of rs7566605 associate with waist-to-hip ratio and computed tomography measures of visceral and subcutaneous adipose tissue.26 This suggests that rs7566605 tags a variant for altered adipogenesis.

This interaction analysis also examined the functional SNP rs2229616 in MC4R, which contributes to hormonal appetite control in response to the amount of adipose tissue.1, 27 MC4R harbours several susceptibility loci for obesity.6, 28, 29 Known functional SNPs in MC4R are all rare, have been found mostly in obesity studies and mostly associate with elevated BMI. The missense variant V103I (rs2229616), however, is relatively frequent and well-studied in population-based samples.4, 30, 31, 32, 33, 34 Minor allele carriers (∼3% Caucasians) have lower BMI,30, 31 lower risk of obesity4, 32, 33 and metabolic syndrome,34 and beneficially altered triglyceride and HDL-cholesterol levels.34, 35 Concurrently, functional studies demonstrated that V103I alters MC4R receptor responsiveness to its endogenous antagonist agouti-related protein.6

In this candidate study, we tested INSIG2 interaction with age and MC4R variant V103I on individual-level BMI data of unrelated adults (25–74 years) from three large longitudinal population-based cohorts. No other SNPs or hypotheses were tested.

Materials and methods

Samples

We analysed independent individuals from the KORA S3/F3 and KORA S4/F4 cohorts30, 36 (from the KORA study, http://www.helmholtz-muenchen.de/en/kora), and from the Framingham Heart Study Offspring cohort (FHS-Off)37 (Genetic Analysis Workshop 16 Framingham data, accession number phs000128.v1.p1 from the Database of Genotypes and Phenotypes (dbGaP), http://www.ncbi.nlm.nih.gov/gap). Data were retrieved and analysed in compliance with the Declaration of Helsinki. All studies were approved by local, regional and national ethics committees as required.

The KORA platform (Cooperative Research in the Region of Augsburg) recruited from the adult general population of the southern German city of Augsburg and its two adjacent districts.30, 36 The aim was to describe the prevalence of common diseases and their risk factors. We analysed baseline survey S3 (1994–1995) with 10-year follow-up F3 (KORA3), and baseline survey S4 (1999–2001) with 7-year follow-up F4 (KORA4).

Furthermore, we analysed independent subjects from the FHS-Off37 (26 years of follow-up: examinations 1 (1971–1975), 3 (1983–1987), 5 (1991–1995), 7 (1998–2001)). Subjects were adults recruited from the town of Framingham, MA, USA, with the identification of common factors in cardiovascular disease as objective. We excluded subjects on cholesterol treatment, as INSIG2 contributes to the regulation of cholesterol synthesis.

BMI is known to exhibit a gain phase up to late mid-life (55–62 years, dependent on study and sex); followed by diminished gain and a plateau or even slight decline at older age.38 We analysed subjects (Table 1) whose examinations occurred between 25 and 74 years of age, and at least half their examinations within the BMI gain phase (age ≤62 years). For FHS-Off with 26-year follow-up and four examinations, this restricted the selected baseline age range to 25–46 years (accounting for any discrepancy between actual age on examination and scheduled age). For KORA3 and KORA4 with 10 or 7-year follow-up and two examinations, this restricted the selected baseline age range to the whole adult BMI gain phase (25–62 years). The cohorts are otherwise comparable, including minor allele frequencies of considered SNPs and duration between examinations.

Table 1 Age-restricted, population-based cohorts analysed
Full size table

Statistical analysis

We performed longitudinal interaction analyses of rs7566605 with age and rs2229616 on BMI and obesity status. An additive model is not powerful for a recessive SNP (INSIG2)3 and its interaction with a dominant SNP (MC4R).4 We therefore used co-dominant SNP coding. All analyses were adjusted for sex, age, study and education (see Appendix for details). All P-values are nominal and two sided.

BMI is distributed right-skewed. Rank normalization of longitudinal BMI (joint Blom-transformation39 of all examinations; all cohorts in the joint analysis) removed this skewness but maintained the longitudinal rank correlation and order of the BMI data. Rank-normalized BMI was analysed using growth curve models.40 The resulting relative estimates of effect size were converted into absolute values (BMI units) by multiplying with the sample SD of BMI (KORA3: 4.5 kg/m2, FHS-Off, KORA4: 4.8 kg/m2, joint analysis: 4.7 kg/m2). Longitudinal obesity status (BMI ≥30 kg/m2) was analysed by logistic regression, accounting for dependence by random subject intercept or by solving generalized estimating equations with an unstructured working correlation matrix. The longitudinal nonparametric rank-sum test LNPT41, 42 validated the parametric analyses. LNPT resembles special repeated-measures ANOVA on rank order data42 of BMI, necessitating conversion of predictors and covariates into factors. Hence, baseline age is categorized for LNPT; cutoffs (Tables 2 and 3) ensure balanced classes for each cohort. LNPT was applied on raw BMI (adjusting for covariates within LNPT) and on residual BMI (after a priori covariate adjustment of raw BMI, see Appendix). FHS-Off and KORA cover the same age range. LNPT models the longitudinal study design by the factor longitudinal time course (examination number) and adjusts for baseline age class. However, power of the parametric models is optimized by using the covariable age on examination directly.42

Table 2 INSIG2 rs7566605–age interaction effect on longitudinal BMI in age-restricted population-based cohorts
Full size table
Table 3 MC4R—INSIG2 interaction effect on longitudinal BMI in age-restricted population-based cohorts
Full size table

Results

INSIG2 rs7566605–age interaction on BMI

We detected SNP–age interaction for INSIG2 rs7566605, but not for MC4R rs2229616. The BMI growth curve (Table 2, upper panel) of INSIG2 minor allele carriers has a lower intercept and significantly increased regression coefficient compared with the GG genotype (FHS-Off, KORA3, joint analysis). KORA3 and KORA4 contributed 46% and 43% of the total sample size to the joint analysis. Longitudinal logistic regression of obesity status (Table 2, bottom panel) confirmed a SNP-age effect. Expectedly, binary trait obesity is less powerful compared with quantitative BMI; however, evidence of SNP–age interaction is consistent between both phenotype models. The initially lowered obesity odd for genotype CC increases significantly with age.

Interestingly, estimates of INSIG2 CC genotype interaction with age were greater in the full KORA4 cohort (0.02 kg/m2 per year more than for GG genotype subjects, P=0.124, with negative intercept estimate), in contrast to the age-restricted KORA4 sample (Table 2), in which 835 subjects with baseline age >62 years were excluded. An INSIG2 main effects model also estimated greater CC genotype effect in the full KORA4 cohort (0.64 kg/m2, P=0.0015, in agreement with Herbert et al3) when compared with the age-restricted KORA4 sample (0.42 kg/m2, P=0.0656). The GC genotype’s main effect, however, was stable (full cohort: 0.29 kg/m2, P=0.0203; age-restricted sample: 0.32 kg/m2, P=0.0258; compared with the GG genotype). This strongly implies that INSIG2 genotype CC also interacts with age in KORA4.

SNP–age effect estimates were stable for different models of longitudinal correlation (comparing unstructured estimation of longitudinal covariance (data not shown) with parametric covariance models (Tables 2 and 3: autoregressive covariance (growth curve), or random subject intercept (logistic regression))). The interaction was also confirmed in raw BMI, albeit with weaker power (Supplementary Table). Further validation comes from the rank-sum test LNPT, which is model-free and free of distributional assumptions.42 LNPT on raw BMI (Table 2, middle panel, single cohorts) agreed strongly with LNPT on residuals of covariate-adjusted raw BMI (data not shown). LNPT tested two-way interactions of INSIG2 with baseline age (averaging over repeated examinations), or with longitudinal time course (averaging over the baseline age classes), and three-way interaction of INSIG2 with longitudinal time course and baseline age class. Power of the interaction tests with longitudinal time course42 is aided by the long-term follow-up (26 years) and relatively strong longitudinal autocorrelation of BMI (ρ=0.38) in FHS-Off. However, the two-way interaction test with baseline age is expected to have better power42 in KORA3 and KORA4, because of short-term follow-up (10, 7 years) and lower autocorrelation (ρ=0.16, 0.27). The three-way interaction tested for differential effects in older compared with younger subjects (which were suspected previously18). LNPT agreed with parametric results. INSIG2–age interaction on BMI was significant for FHS-Off (26 years follow-up, P=0.0079, three-way interaction with longitudinal time course and baseline age) and for KORA3 (10 years follow-up, P=0.0044, two-way interaction with baseline age) but not for KORA4. The three-way interaction in FHS-Off is due to a significant INSIG2 effect on longitudinal time course of BMI in older subjects (P=0.022, baseline age 34–46years), while no significance was reached in younger FHS-Off subjects (P=0.16, baseline age 25–33 years). LNPT joint analysis used FHS-Off examinations 3 and 5 to have balanced age classes (≤46 versus >46 years KORA baseline age, or FHS-Off age on examination 3) in each study. INSIG2 interaction with age on examination (Figure 1, left: growth curve) is consistent with INSIG2 interaction with longitudinal time course and age class (Figure 1, right: residuals of covariate-adjusted raw BMI, as analysed by LNPT). Growth curves of INSIG2 genotypes cross (Figure 1, left). Mean BMI of genotype CC subjects (bold dotted line) is initially lower compared with the other genotypes, but increases more rapidly with age and is elevated for ages ≥40 years. Consistently, LNPT detected no INSIG2 average effect but a tendency toward different marginal effects over the two baseline age classes (≤46 years/other; two-way interaction with age; joint analysis Table 2). P-values were also close to 0.05 for the interaction with examination (Table 2: joint analysis, comparable to KORA3).

MC4R–INSIG2 interaction on BMI

The final genetic interaction model contains the INSIG2–age interaction and tests for INSIG2–MC4R interaction effects on intercept (because of power considerations22). This model was tested in the large KORA cohorts and jointly over all three cohorts (ensuring >10 observations for all two-locus genotypes and all examinations). INSIG2–MC4R interaction was significant (P=0.001, Table 3, upper panel). Compared with additive SNP effects, we found an enhanced protective effect of the MC4R variant in INSIG2 heterozygotes (−1.26 kg/m2 on intercept (95% CI: −2.21, −0.31), P=0.009) and a tendency toward an enhanced detrimental effect in INSIG2 minor allele homozygotes who also have the MC4R variant (0.97 kg/m2 on intercept, P=0.148). A MC4R rs2229616 main effect was found in large, single-locus meta-analyses4, 30, 31, 32, 33 but was not detected by our interaction model. INSIG2 interaction with age was consistent between Tables 2 and 3. Growth curve results again were consistent with LNPT analyses of residuals of covariate-adjusted raw BMI (Table 3, Figure 2). However, LNPT had weaker power. LNPT P-values for the INSIG2–MC4R interaction were close to 0.05 whenever the growth curve model revealed significance (KORA3: P=0.036 interaction test for GC-GA subjects, excluding INSIG2 genotype CC, remaining subgroups have sufficient sample size; joint analysis: P=0.058 global test over all genotypes). Previous literature3, 4 on main effects reported that the INSIG2 SNP is recessive and the MC4R SNP is dominant. A recessive-dominant growth curve model of BMI confirmed the interaction of INSIG2 with MC4R and age (increased intercept of INSIG2–MC4R CC-GA genotype by 1.32 kg/m2 (95% CI: 0.10, 2.54), P=0.034; and increased regression coefficient for INSIG2 CC genotype by 0.02 kg/m2 per year, P=0.028).

Figure 2
figure 2

MC4R–INSIG2 interaction on longitudinal BMI in population-based cohorts (joint model: FHS-Off examinations 3 (1983–1987) and 5 (1991–1995); KORA3, KORA4). Displayed are the median and interquartile of residual BMI, after adjusting raw BMI for sex, baseline age, education and cohort. INSIG2 rs7566605 genotype is displayed by box colour (GG=white, GC=light grey, CC=dark grey) in MC4R rs2229616 homozygotes (left) and heterozygotes (right); stratified by baseline age class (figure columns). Longitudinal examinations T1 (KORA baseline, FHS-Off examination 3) and T2 (KORA follow-up, FHS-Off examination 5) are displayed adjacent to each other. Sample sizes are given at the bottom of the graphs.

Full size image

Discussion

Previous literature (Table 4) reported a recessive main effect3 of 0.60 kg/m2 for INSIG2 rs7566605 genotype CC in KORA4, with replication failures8 in FHS-Off and KORA3. In contrast, we provide consistent evidence of rs7566605–age interaction with both major statistical interaction models, testing interaction on both the additive scale (growth curve, LNPT) and multiplicative scale (logistic regression). MC4R rs2229616 was previously analysed for baseline examinations KORA S3 and S4, demonstrating a protective main effect30 of −0.52 kg/m2 for minor allele carriers. This may be a weaker marginal effect from a stronger INSIG2–MC4R interaction shown here (Table 3).

Table 4 This article compared with previous research: INSIG2 rs7566605 association with BMI and obesity
Full size table

We found that INSIG2 minor allele homozygotes accumulate on average an excess of 1.86 kg/m2 in 62 years but have an MC4R genotype-dependent intercept (−0.05 kg/m2 for CC-GA genotype, −1.17 kg/m2 for CC-GG genotype), compared with GG-GG genotype subjects. The INSIG2–age effect is observed directly in INSIG2–MC4R CC-GA genotype subjects with a mean BMI elevated by 1.81 kg/m2 at the age of 62 years (∼0.4% Caucasians). In contrast, the mean BMI of CC-GG genotype subjects is lowered by −0.42 kg/m2 at the age of 25 years and raised by 0.69 kg/m2 at the age of 62 years (∼10.8% Caucasians). In contrast, mean BMI for GC-GA genotype subjects is lowered by −1.32 kg/m2 with no significant genotype–age effects (∼1.6% Caucasians). Hence, elderly INSIG2–MC4R CC-GA genotype subjects are particularly at risk of obesity. This also may explain previous INSIG2 main effect replication failures, suggesting that the proportion of younger and elderly subjects in a study can influence association outcome when not accounting for SNP–age interaction.

The strengths of this study are the modelling of individual participant data from high quality, well-characterized, population-based studies (no meta-analysis). The presented effects were robust with respect to different modelling approaches. In contrast to previous analyses, we restricted our analysis to subject age, motivated by the general dependence of BMI on age.38 No interaction was found earlier between INSIG2 rs7566605 and the GWAS identified common SNP rs17782313 near MC4R.17 This interaction analysis, however, examined the functional SNP6 rs2229616 in MC4R.

A few study limitations exist. As with every other joint analysis, we had to accommodate differences in study design (baseline age and follow-up duration) and different covariate information. Follow-up in KORA is relatively short and sample size for INSIG2–MC4R interaction analysis is still relatively small because of the rare rs2229616 variant. Our finding of INSIG2–age interaction (altered adipogenesis) and of INSIG2–MC4R interaction (interplay between adipogenesis and hormonal appetite control) would require validation by an independent study. We conclude that gene–gene or gene–age interaction may explain differences in genetic main effects across studies. More elaborate modelling could tackle this important question in future, large, well-characterized studies. Our results also encourage investigation into interactions between other obesity genes and with MC4R.