The BCAMS study, which has been described in detail elsewhere (22, 23), is an ongoing cohort study of obesity and related metabolic abnormalities in a representative sample of school-age children (n = 19,593, aged 6-18 y, 50% boys) recruited from the Beijing area between April and October 2004. Within this cohort, 4,500 subjects were identified as being at risk of metabolic syndrome (MS), based on at least one of the following criteria: 1) overweight, as defined by body mass index (BMI) percentile; 2) increased total cholesterol (TC) ≥ 5.2 mmol/L; 3) triglycerides (TG) ≥ 1.7 mmol/L; and 4) fasting blood glucose (FBG) ≥ 5.6 mmol/L based on finger capillary blood tests. Next, all children at risk of MS, together with a parallel normal sample of 1,024 schoolchildren, were invited to participate in a further medical examination including venipuncture-based blood tests. In total, 3,506 subjects, including 2,525 subjects with MS risk, ultimately completed the further clinical examination. Thus, the presence of pediatric MS based on clinical examination was defined by the presence of three or more of the following five components (22, 23) (1): central obesity defined as ≥90th percentile for age and gender (2); elevated systolic and/or diastolic blood pressure ≥90th percentile for age, sex and height (3); hypertriglyceridemia defined as TG ≥1.24 mmol/L (4); low high-density lipoprotein (HDL) cholesterol defined as <1.03 mmol/L; and (5) hyperglycemia defined as FBG ≥ 5.6 mmol/L. Accordingly, in the current study, we used the cross-sectional data of 3,402 participants (including 2,112 subjects with MS risk and 1,290 subjects without MS risk based on clinical evaluation), who completed the examination of adiponectin levels, genotype, and lifestyle factors in 2004 ( Supplementary Figure S1 and Supplementary Table S1 ).

The subjects’ height and weight were measured according to our standard protocol (22, 24). Height in centimeters was measured without shoes to the nearest 0.1 cm. Bodyweight was measured to the nearest 0.1 kg (light indoor clothing, without shoes) using a calibrated electronic scale. BMI was calculated as weight (kg) divided by height squared (m²). Age- and sex-specific BMI percentiles were used to define overweight (85^th) and obesity (95^th) following the Working Group for Obesity in China (25). Puberty stages were assessed by two pediatricians of the same gender, based on Tanner’s stages of breast development in girls and testicular volume in boys, in line with Marshall and Tanner (26). The categories of puberty were defined as pre-puberty (Tanner stage I), mid-puberty (Tanner stage II-III), and post-puberty (Tanner stage ≥ IV) (26).

Fourteen lifestyle factors (27), including walking to school, frequency of exercise, duration of habitual sleep, and eleven dietary habits were selected for examination in this study ( Table 1 ). And the comparison of lifestyle factors between subjects with risk of MS and without risk of MS was listed in Supplementary Table S1 . Surveys were conducted retrospectively and participants were asked how often on average they had consumed each food in the previous month. Each dietary item was scored on a 5-point scale according to the frequency categories from “seldom or never” to “> 5 times per week “, with ascending values for favorable foods (1 for “seldom or never” and 5 for “> 5 times per week”) and vice versa, according to the direction of their linear associations with adiponectin levels. The diet score was generated by summing all the selected item scores, with a higher diet score indexing predisposition to higher adiponectin levels. Effective exercise was deemed as exercise lasting longer than 30 minutes for extracurricular physical activities such as cycling, running, swimming, dancing, and team sports. Transportation refers to “walking to school” which was recorded and reclassified into two modes, ‘walking’ and ‘non-walking’.

Venous blood samples were collected after 12 h of fasting. The adiponectin concentration was measured using a monoclonal antibody-based enzyme-linked immunosorbent assay (28). The intra- and interassay coefficients of variation were < 5.4% and 8.5%, respectively.

Genomic DNA was isolated from peripheral white blood cells using QIAamp DNA blood midikits (Qiagen). Genotyping was carried out on a Sequenom Mass Array iPLEX genotyping platform by BioMiao Biological Technology Co., Ltd (29). All these SNPs had genotyping efficiency >0.95 and were in Hardy-Weinberg equilibrium with p-value >0.008 (0.05/6). We selected six SNPs showing strong associations with adiponectin levels in previous genome-wide association studies (GWASs) (12, 13). Among them, five SNPs (CDH13-rs4783244, ADIPOQ-rs10937273, PEPD-rs889140, CMIP-rs2925979, and WDR11-FGFR2-rs3943077) were identified in an East Asian adult population (12) to have the five strongest associations with adiponectin levels (P < 10^-10 for each of the five SNPs); while another SNP, ADIPOQ-rs6773957, was identified in a European population (13) and was included because it is located at the 3’ UTR of ADIPOQ, a very important regulatory area of the gene. Details of the SNPs are listed in Supplementary Table S2 .

As the effects of the SNPs were dramatically different, we used weighted rather than unweighted genetic predisposition scores (wGPSs) to evaluate the genetic structure of the children. wGPSs were calculated by summing the scores of the six SNPs, each of which was weighted using the mean of linear regression β from our study (wGPS_all-_BCAMS) or that of published GWASs (wGPS_all-_GWAS) for adiponectin levels (12, 13). CDH13-rs4783244 was found to be the strongest modulator of adiponectin levels in the current study, explaining at least 10-fold higher adiponectin levels than the others ( Supplementary Table S2 ). We generated two GPSs that excluded CDH13-rs4783244, namely, wGPS_{no CDH13-GWAS} and wGPS_{no CDH13}-_BCAMS, to evaluate the association of genetic structure with adiponectin expression. For each participant j (j = 1, 2,…, 3405), we calculated wGPSs using the following equation:

where n is 6 (i.e., wGPS_all; including all six SNPs) or 5 (i.e., wGPS_{no CDH13}; all SNPs but CDH13-rs4783244), β_i is the β coefficient of each SNP for adiponectin levels (natural logarithm-transformed) adjusted for age, sex, and BMI, and EA_i is the number of effect alleles (0, 1 or 2) in each SNP. Thus, wGPS_all ranged from 0 to 12, while wGPS_{no CDH13} ranged from 0 to 10. However, there were no significant differences between the results of wGPS_all-_BCAMS and wGPS_all-_GWAS or between wGPS_{no CDH13}-_GWAS and wGPS_{no CDH13}-_BCAMS in the current study. Considering that SNP-adiponectin associations could be different between children and adults, we only report the results for wGPS_all-_BCAMS and wGPS_{no CDH13}-_BCAMS in the current pediatric study, in the form of wGPS_all and wGPS_{no CDH13} for brevity.

All analyses were performed using SPSS version 22.0 software for Windows (SPSS Inc.) (30). Adiponectin levels were natural logarithm transformed for analysis. We assigned a score of 1-5 (“seldom or never” = 1; “1 time every 2 weeks” = 2; “1-2 times per week” = 3; “3-5 times per week” = 4; “> 5 times per week” = 5) to each lifestyle factor to facilitate the analyses. The exception was transportation (non-walking = 0, walking = 1). The results are expressed as the mean ± SD or mean (95% CI) unless otherwise stated. We used ANOVA, ANCOVA, and t tests to compare the values of factors between different puberty groups. We performed a linear regression adjusted for confounders including age, sex, and BMI, to evaluate the associations of adiponectin with SNPs, GPSs, and lifestyle factors. Furthermore, gene-lifestyle interactions on adiponectin were tested using linear regression models by including the interaction terms (e.g. diet*genotype) in these models. The β coefficient, which reflects the change in the serum adiponectin concentration, was used to report the effects of genetic variants and lifestyle on adiponectin levels. Associations between SNPs and adiponectin levels were assessed using an additive model in which a score of 0, 1, or 2 was assigned to genotypes according to the number of effect alleles. An SNP association was considered statistically significant if the resulting P-value was less than the Bonferroni-corrected significance threshold of 0.05/6 = 0.008. We only tested the interactions between gene variants and lifestyle factors that were statistically significantly associated with adiponectin levels, and then stratified analyses were conducted to observe effect modification. The gene-by-lifestyle interaction was considered to be significant if the resulting P-value was less than the Bonferroni-corrected significance threshold of 0.05/9 = 0.006 and considered to be “nominally significant” if the P-value between 0.006 and 0.05.

Among the 3,402 children in our study, 32% were prepubertal, 32% were midpubertal, and 36% were postpubertal ( Table 1 ). Except for fruit intake, the examined lifestyle factors differed significantly among children at different pubertal stages. Children at an advanced pubertal stage generally exhibited a higher BMI and lower adiponectin levels. Figure 1 highlights the adiponectin levels among the different groups according to sex and puberty status. Compared with prepuberty, adiponectin levels significantly decreased after the onset of puberty, and adiponectin levels were higher in girls after adjusting for age.

Supplementary Table S2 shows the genetic information for the six selected SNPs. After controlling for age, sex, and BMI, five of these loci significantly influenced adiponectin levels: CDH13-rs4783244, ADIPOQ-rs10937273, PEPD-rs889140, ADIPOQ-rs6773957, and WDR11-FGFR2-rs3943077, explaining 5.2%, 0.7%, 0.5%, 0.2%, and 0.2% of the total variance, respectively. As for the two wGPSs, both wGPS_all and wGPS_{no CDH13} were strongly positively associated with adiponectin levels and explained 7.1% and 1.9% of the total variance, respectively. Table 2 outlines the associations between the SNPs and adiponectin levels according to pubertal status in the current study. The changes in the strength of these associations during puberty after adjustment for age, sex and BMI are graphically highlighted in Figure 2A . The association of ADIPOQ-rs6773957 with adiponectin levels was only significant at postpuberty [β = 0.058 (0.017, 0.099), P = 0.005], while WDR11-FGFR2-rs3943077 presented a significant effect on adiponectin levels at midpuberty [β = 0.076 (0.027, 0.125), P = 0.002]. The other three SNPs, namely, ADIPOQ-rs10937273, CDH13-rs4783244, and PEPD-rs889140, together with wGPS_all and wGPS_{no CDH13}, were associated with adiponectin levels throughout puberty. Most notably, among the six SNPs, CDH13-rs4783244 showed the strongest association with adiponectin levels at all pubertal stages.

Controlling for age and sex, both exercise (P = 0.049) and walking to school (P = 4.47×10^-12) was associated with increased adiponectin levels; adjustment for BMI attenuated the association of adiponectin levels with both exercise (P = 0.750) and walking to school (P = 1.07×10^-4). Six diet items (breakfast, meat, dairy, soft drink, fried food, and snack) were significantly associated with adiponectin levels in the model adjusted for age and sex (soft drink, fried food, and snack), or in the model further adjusted for BMI (breakfast, meat, dairy, and soft drink) ( Table 3 ). To better analyze the association of adiponectin with dietary structure, we used the six dietary items shown to be associated with adiponectin in either adjusted model to construct a new diet score. The diet score showed a stronger association with increased adiponectin levels than any single diet factor in both models with (P = 1.25 × 10^-8) or without (P = 6.32 × 10^-6) adjustment for BMI ( Supplementary Table S3 ). Further analyses of lifestyle factors and gene-by-environment interactions will only focus on nine lifestyle factors that were associated with adiponectin levels in either adjustment model, including walking to school, exercise, diet score, and consumption of breakfast, meat, dairy, soft drink, fried food, and snacks. Further analyses (depicted in Figure 2B and summarized in Supplementary Table S3 ) adjusted for age and sex revealed different effects of the nine lifestyle factors on adiponectin levels among the three pubertal stages. However, none of the dietary items nor the diet score was associated with adiponectin levels at midpuberty. Walking to school was the only factor that influenced adiponectin levels throughout puberty [prepuberty: β = 0.114 (0.043, 0.185), P = 0.002; midpuberty: β = 0.077 (0.003, 0.151), P = 0.044; postpuberty: β = 0.203 (0.142, 0.264), P < 0.001]. In addition, exercise frequency was only associated with adiponectin levels after the onset of puberty [midpuberty: β = 0.031 (0.002, 0.060), P = 0.036; postpuberty: β = 0.025 (0.001, 0.049), P = 0.031]. After further adjustment for BMI, most of the effects of lifestyle factors on adiponectin levels were weakened, except for that of breakfast frequency, which decreased adiponectin levels significantly at postpuberty [β = -0.039 (-0.059 to -0.019), P = 0.003] ( Supplementary Table S3 ). Additionally, none of the lifestyle factors, including walking to school, significantly affected adiponectin levels at midpuberty independent of BMI.

After controlling for age and sex ( Supplementary Table S4 ), we found two statistically significant interactions in the entire population: one between WDR11-FGFR2-rs3943077 and walking to school [β = -0.079 (-0.132 to -0.026), P = 0.004] and the other between ADIPOQ-rs6773957 and exercise [β = 0.038 (0.018 to 0.058), P = 1.71 × 10^-4], as well as five nominally significant interactions. Regarding the interactions between GPS and lifestyle factors, as depicted in Figure 3 and Supplementary Table S5 , we identified nominally significant negative interactions of wGPS_all with diet score [β = -0.003 (-0.005 to 0.000), P = 0.022] and wGPS_{no CDH13} with exercise [β = -0.009 (-0.017 to 0.000), P = 0.049] and snacks [β = -0.006 (-0.012 to 0.000), P = 0.038] when age, sex and other lifestyle factors were controlled. The interaction between wGPS_all and diet score was more prominent at postpuberty than at other puberty stages. Further adjustment for BMI did not change the significance of the interactions between wGPS_{no CDH13} and exercise and snacks in the entire population and between wGPS_all and diet score in children at postpuberty ( Supplementary Tables S4, S5 ).

As we identified an interaction between wGPS_all and diet score, and an interaction of wGPS_{no CDH13} with exercise, further stratified analyses were undertaken to observe effect modification. We firstly analyzed the associations of adiponectin levels with diet score and exercise according to the categories of genetic risk ( Figure 4 ). We found that diet score and exercise had greater effects in children with a higher genetic risk for low adiponectin ( Figure 4 ).

Second, we compared adiponectin concentrations between different categories of genetic risk and lifestyle levels ( Figure 5 ). The difference in adiponectin levels between children at high genetic risk and those at intermediate or low genetic risk was more prominent among children with low and intermediate exercise levels than among those with high exercise levels ( Figure 5A ), whereas no significant difference in adiponectin levels was found among the genetic risk groups when the exercise frequency was high. A similar pattern was observed for diet levels, while an increased diet score significantly attenuated the difference in adiponectin levels between the high genetic risk group and intermediate genetic risk group ( Figure 5B ).