Introduction

Obesity is a major risk factor for type 2 diabetes, dyslipidaemia, hypertension and cardiovascular disease, and has a strong genetic component [1]. Twin studies have generally found heritability estimates of 0.75 to 0.85 for BMI and approximately 0.70 for weight [2]. Genome-wide association (GWA) studies have provided evidence that several loci are associated with common obesity mostly in populations of European descent [315]. The first such loci reported was the fat-mass and obesity-associated gene (FTO) [3, 16, 17]. A common variant in the FTO locus, rs9939609, was originally identified as part of a GWA study for type 2 diabetes [3]; people with homozygous risk alleles weighed approximately 3 kg more than those without a risk allele. It has been suggested that this variant can predispose individuals to type 2 diabetes and metabolic disorders through its primary effect on BMI in Europeans [18].

Ethnic differences have been assumed to exist between Europeans and Asians in terms of the components and impacts of genetic factors for obesity, and traits or disorders related to obesity (e.g. type 2 diabetes) [2, 19, 20]. For instance, the relationship between BMI and body fat per cent differs between these populations, with Asians in general having a higher body fat per cent at a lower BMI than Europeans [20]. It has also been estimated that the absolute genetic variances for BMI and weight are greater in Europeans than in East Asians, according to an adolescent twin study [2]. From an epidemiological viewpoint, moreover, it has been hypothesised that the overall impact of obesity on type 2 diabetes is greater in Asians than in Europeans [21]. Accordingly, it is of interest to compare the genetic associations between populations of European descent and Japanese populations.

To date, several studies on non-European populations have replicated the associations for a number of novel obesity loci previously identified by GWA studies in Europeans [13, 2225]; nevertheless, statistical power has not been sufficient to make strong conclusions. Therefore to test associations between obesity measures (BMI and weight) and type 2 diabetes for 14 single nucleotide polymorphisms (SNPs) from 13 candidate loci recently reported by two GWA meta-analyses [10, 11], we performed a replication study in the general Japanese populations.

Methods

Study populations

We performed a replication study of previously identified variants in the general Japanese populations (Table 1, electronic supplementary material [ESM] Study samples for BMI association analysis). Specifically, 5,695 Japanese participants (referred to hereafter as the Amagasaki panel) were consecutively enrolled in the population-based setting as described previously [26] and 12,569 other Japanese participants (referred to hereafter as the Fukuoka panel) were randomly selected from residents aged 50 to 74 years in the general population [27]. Among the candidate loci tested in the two panels, those showing a replication of association (p < 0.05 one-tailed, i.e. p < 0.1 two-tailed, in concordant direction with previous reports) were subjected to tests to examine type 2 diabetes associations in a Japanese case–control study panel (ESM Table 1, ESM Study samples for type 2 diabetes case–control studies). For the purpose of uniformity, two tailed p values are shown throughout the text, unless otherwise indicated. Of 2,041 cases and 2,418 controls enrolled from the Cardiovascular Genome Epidemiology (CAGE) Network [28], 931 cases and 1,404 controls were included in stage 1 and 1,110 cases and 1,014 controls in stage 2. In addition, participants in the Fukuoka panel were included in stage 1 and Biobank Japan (http://biobankjp.org/ [in Japanese], accessed 1 February 2011) cases were included in stage 2 (ESM Table 1). From the CAGE Network, type 2 diabetes cases were enrolled according to the 1999 WHO criteria, while unaffected controls were enrolled according to the following criteria: (1) no past history of urinary glucose or glucose intolerance; (2) HbA1c < 5.6% or a normal result from 75 g glucose tolerance test; and (3) age at examination ≥55 years. From the population-based participants in the Fukuoka panel, diabetic participants (n = 740) and unaffected controls (n = 4,889) were selected for case–control analysis; among these samples, there was some overlap between the BMI/weight and type 2 diabetes studies. Here, diabetes was defined as HbA1c ≥ 7.0 or under treatment for type 2 diabetes; a relatively stringent criterion for HbA1c (≥7.0%) was adopted because of the lack of fasting plasma glucose data for participants in the Fukuoka panel. The controls were chosen as non-diabetic participants who met the following conditions: age ≥55 years; HbA1c ≤ 5.0%; no previous and/or current treatment for diabetes; and absence of renal failure (serum creatinine <265.2 μmol/l), as previously described [27]. In total, 6,781 cases and 7,307 controls were used for the type 2 diabetes case–control study in Japanese. The CAGE Network samples were categorised, according to the stages of a previous GWA study [28], into two panels: CAGE–GWAS (stage 1); and CAGE–replication (stage 2) (ESM Table 1); genotyping in the CAGE-GWAS panel was performed with Infinium assay (Illumina, San Diego, CA). All participants enrolled in these different studies provided written informed consent. Local Ethics Committees approved the protocols used.

Table 1 Clinical characteristics of study participants for BMI

Height and body weight were measured by trained personnel using standard anthropometric techniques for all participants other than Biobank Japan type 2 diabetes participants, for whom the relevant data were self-reported from questionnaire.

SNP genotyping and quality control

Samples (except for the CAGE–GWAS panel) were genotyped using the TaqMan assay (Applied Biosystems by Life Technologies, Carlsbad, CA, USA) for 14 SNPs from 13 unique obesity loci previously identified in populations of European descent [10, 11]. These SNPs included: rs2815752 (NEGR1), rs10913469 (SEC16B), rs4854344 (TMEM18), rs7647305 (ETV5), rs10938397 (GNPDA2), rs2844479 (NCR3AIF1), rs6265 (BDNF), rs10838738 (MTCH2), rs7138803 (BCDIN3DFAIM2), rs4788102 (SH2B1ATP2A1), rs6499640 (FTO), rs9939609 (FTO), rs12970134 (MC4R) and rs29941 (KCTD15). The genotype distribution of all tested SNPs was in Hardy–Weinberg equilibrium (p > 10−4). We obtained successful genotyping call rates of >99.6% for all SNPs and >99.8% for all included samples (across 14 SNPs).

Statistical analysis

SNP association analysis

BMI and weight were inverse-normal transformed separately by sex in each panel before association analysis. In addition, in the Fukuoka study panel, we examined the WHR as a variable of fat distribution, which was also inverse-normal transformed. We tested SNPs for the trait association using linear regression analysis in an additive genotype model after adjustment for age classes separately by sex. Age classes were defined according to age distribution in the individual panels, and included ≤40, 41−50, 51−60 and >60 years for the Amagasaki panel, and ≤55, 56−60, 61−65, 66−70 and >70 years for the Fukuoka panel. A one tailed value of p < 0.05 (p < 0.1 two-tailed) was considered statistically significant. We combined association results for the two Japanese panels by using the inverse variance method. We used PLINK (version 1.06; http://pngu.mgh.harvard.edu/~purcell/plink/) [29], R software (version 2.8.1; www.r-project.org) and rmeta (version 2.16; http://cran.r-project.org) for association test and meta-analysis (websites accessed 6 February 2011).

Assessment of genetic effect of obesity variants

To assess the proportion of variance for BMI that was explained by each SNP, we calculated a coefficient of determination R 2 as: 2f(1 − f2, where f is the minor allele frequency and β is the per-allele effect on the standardised values of BMI. We measured the cumulative effect of multiple SNPs by summing the R 2 values for individual SNPs.

Test of ethnic diversity and sex specificity

We compared the per-allele effect size of each SNP on inverse-normal transformed BMI between the ethnic groups (Japanese vs Europeans) and between sexes. Taking into account the well-known male–female differences in body composition (e.g. fat distribution, deposition and accumulation) [30], we tested the potential sex specificity in the genetic associations with obesity. We examined the heterogeneity of the effect size with Cochran’s Q-test [31].

Association analysis of type 2 diabetes

While we tested the primary association with obesity measures, we tested type 2 diabetes association as a secondary analysis in the present study. Thus, we performed two-staged analysis for diabetes case–control study. That is, all SNPs were tested for association with type 2 diabetes in stage 1 samples and only SNPs with p < 0.1 for BMI or weight were then tested in stage 2 samples (ESM Table 1). Using logistic regression analysis, we tested association of candidate SNPs with type 2 diabetes, with and without adjustment for BMI. We adjusted the diabetes trait for sex and BMI, but not for age because age distribution differed between case and control groups; cases were younger than controls, who were defined as being ≥55 years of age.

Results

BMI and weight association at reported SNPs

In this study, we tested the association of BMI and weight with 14 SNP loci that had attained genome-wide significance levels (p < 5 × 10−8) in previously performed GWA meta-analyses for obesity in Europeans [10, 11]. In the combined sample (n = 18,264), we found that 11 of the 14 SNPs had nominally significant (p < 0.1 two-tailed) associations with BMI and/or weight; these were rs10913469 (SEC16B), rs4854344 (TMEM18), rs10938397 (GNPDA2), rs6265 (BDNF), rs10838738 (MTCH2), rs7138803 (BCDIN3DFAIM2), rs4788102 (SH2B1ATP2A1), rs6499640 (FTO), rs9939609 (FTO), rs12970134 (MC4R) and rs29941 (KCTD15) (Table 2). Furthermore, significant associations between BMI and two SNPs at FTO were determined to be independent (i.e. p = 0.016 for rs6499640 and p = 3.9 × 10−6 for rs9939609, both two-tailed, when these were simultaneously included in the linear regression model). Given the much stronger association for rs9939609 than rs6499640 (Table 2), rs9939609 was considered to be the key SNP for FTO, although rs6499640 still appears to have some role. The strength of association did not appear to significantly differ between the study panels (Amagasaki vs Fukuoka), except at BDNF (p < 0.01 for heterogeneity; ESM Table 2). The presence of sexual dimorphism was indicated for two loci, NEGR1 and GNPDA2 (p < 0.05 for heterogeneity; ESM Table 3).

Table 2 Association of previously reported SNPs with BMI and type 2 diabetes in Japanese individuals

Ethnic heterogeneity in effect sizes

To test for the potential presence of ethnic heterogeneity, we compared the effect sizes (β for BMI) between the Japanese and European populations (Fig. 1a). The sample size was smaller in the Japanese group, and consequently the 95% CI was wider in this group, although the direction of association was concordant between the ethnic groups for all SNPs. However, we did find significant (p < 0.05) inter-ethnic heterogeneity in per-allele effect at rs2844479 (NCR3AIF1) and rs29941 (KCTD15) for standardised BMI, and at rs9939609 (FTO) for non-standardised BMI (in kg/m2; ESM Table 4). We calculated R 2 as the proportion of phenotypic variance explained by a SNP (see Methods). Based on these R 2 measurements, the association of three replicated loci, rs4854344 (TMEM18), rs6265 (BDNF) and rs10838738 (MTCH2), was stronger in the Japanese population than in the population of European descent. We detected no significant gene × gene interactions by regression analysis for BMI associations with 14 SNPs (data not shown); the genotypes of each SNP were concomitantly included assuming an additive effect when combined. For the 14 SNPs, the cumulative R 2 totalled 0.65% and 1.2% in the Japanese and Europeans, respectively. Thus, the explained variance tended to be smaller in Japanese than in Europeans.

Fig. 1
figure 1

Effect size for BMI and type 2 diabetes at SNPs previously reported to be associated with BMI in Europeans. a Cross-population comparison of per-allele effect of 14 BMI-associated SNPs between the Japanese and European populations. The per-allele effects (β in %SD) of each variant on BMI are shown by squares (proportional in size to tested sample size) and vertical lines (representing 95% CI) for the Japanese (black squares) and Europeans (white squares). Curves (in grey) indicate R 2 = 0.003, 0.002, 0.001 and 0.0005 (from top to bottom). b Comparison of genetic impacts on BMI (β in x-axis) and type 2 diabetes adjusted for BMI (OR in y-axis) for the 14 SNPs. Those showing significant (black circles) and non-significant (white circles) association with BMI in Japanese are depicted. SNP rs numbers of individual loci are as follows: rs2815752 for NEGR1, rs10913469 for SEC16B, rs4854344 for TMEM18, rs7647305 for ETV5, rs10938397 for GNPDA2, rs2844479 for NCR3, rs6265 for BDNF, rs10838738 for MTCH2, rs7138803 for BCDIN3D, rs4788102 for SH2B1, rs6499640 for FTO(1), rs9939609 for FTO(2), rs12970134 for MC4R and rs29941 for KCTD15

Association of obesity susceptibility SNPs with type 2 diabetes

For the case–control study of type 2 diabetes, we genotyped 11 obesity-associated SNPs in the stage 1 and stage 2 panels; the remaining three SNPs were genotyped only in the stage 1 panel (ESM Table 1). Some evidence of association with type 2 diabetes in a direction consistent with the BMI–SNP associations (p = 0.02−4.3 × 10−10 and OR 1.06–1.20 before adjustment for BMI; p = 0.04−2.4 × 10−7 and OR 1.05–1.17 after adjustment for BMI) was provided for six (of 11) SNPs in the Japanese population (Fig. 1b, Table 2). After adjusting for BMI, most of the observed associations between obesity-associated SNPs and diabetes were slightly attenuated, although nominal significance remained (Table 2, ESM Table 5). Considering the possibility of some selection bias in the case–control study design, we also tested the diabetes association by nested case–control comparison within the same population (i.e. Fukuoka panel) and verified a fair consistency in the OR for type 2 diabetes (ESM Table 6). Further, to examine the possibility that the BMI-independent diabetes associations derive from genetic susceptibility to fat distribution, we tested the association with WHR among the control participants in the Fukuoka panel (n = 4,889), where none of the six SNPs showed significant association with WHR in a direction consistent with the SNP–diabetes associations (ESM Table 7).

Assuming the potential presence of ‘unadjusted’ confounding influences of BMI, we plotted (Fig. 2) the data from the genetic studies examining type 2 diabetes association at the FTO locus [3, 11, 24, 3242], our aim being to evaluate the relationship between BMI in the control group (or derived population) and the OR for type 2 diabetes. We detected no apparent relationship between the tested variables (p = 0.15 for East Asians and p = 0.41 for Europeans; Fig. 2, ESM Table 8). While there was a significant cross-population difference in the OR for type 2 diabetes (e.g. OR 1.13 and 1.04 for East Asians and Europeans, respectively; I 2 = 89%; p = 0.003), the BMI-adjusted association with type 2 diabetes proved to be highly significant (OR 1.09, p = 3.0 × 10−11) in the meta-analysis involving 36,064 cases and 61,234 controls of various ethnic origins.

Fig. 2
figure 2

Relationship between BMI in the control group (or derived population) and the OR for type 2 diabetes in genetic studies examining type 2 diabetes association at the FTO locus (by using rs9939609 or its proxies). The studies are categorised into three ethnic groups: East Asians (diamonds), Europeans (squares) and others (triangles). For further details, see ESM Table 8

Discussion

The present study investigated genetic susceptibility to common forms of obesity and its relevance to type 2 diabetes in Japanese populations. Replicating a study of candidate loci previously identified by GWA meta-analyses of Europeans [10, 11], we found some ethnic diversity in obesity variants (Fig. 1a). The top hit associations were consistently reported at FTO and MC4R in populations of European descent [5, 6, 1012, 14, 15], whereas equivalent or more pronounced associations with obesity localised to TMEM18 and BDNF, together with modest association at SEC16B, GNPDA2, MTCH2, BCDIN3DFAIM2, SH2B1ATP2A1 and KCTD15, were found in the Japanese populations (Table 2, ESM Table 4). In addition, we found that the direction of BMI association was concordant between the ethnic groups at all tested loci, although the cumulative effect of the associated SNPs was smaller in Japanese than in Europeans. Of particular note is the fact that, in the present study, we highlighted the genetic impact of six loci on type 2 diabetes, this impact being independent of obesity susceptibility in all six loci.

Recently, two GWA meta-analyses were performed for examination of obesity in populations of European descent; the studies involved >32,000 and >25,000 individuals, with follow-up analysis using genotypes from large cohorts (>50,000 samples in total) and computer-generated association results [10, 11]. Collectively, these studies revealed 13 unique obesity-associated loci in Europeans. Four studies in total, including our present study, have attempted to replicate the obesity association in East Asians; of these, two studies involved Japanese individuals and two involved Chinese individuals [2325]. While the design of these studies was not identical (two case–control studies, one quantitative trait analysis and one study with both-types of analytical approaches combined), all four studies consistently reported significant association with obesity at the GNPDA2 (rs10938397) and FTO (rs9939609 or its proxies) loci in East Asians. The present study, which is the largest of the four studies and the only one that involved quantitative trait analysis using the general population sample, has enabled us to validate nine other SNP loci. Eight of these SNP loci had previously shown obesity association in one or two studies [2325], whereas replication at MTCH2 (rs10838738) has not been previously reported in East Asians.

Compared with the data for Europeans [10, 11], the variance for BMI explained by individual SNP loci appeared to be modest as a whole in the Japanese population, except for TMEM18 rs4854344 (R 2 = 0.13%), BDNF rs6265 (R 2 = 0.1%) and MTCH2 rs10838738 (R 2 = 0.06%; Fig. 1b, ESM Table 4). It should be kept in mind that the larger cumulative effect on BMI in Europeans than in Japanese is partly due to the ‘winner’s curse effect’ [43], where the original meta-analysis tends to overestimate the true population effect, in addition to ethnic difference in at-risk alleles. We found that the effect size (β) was larger at TMEM18 rs4854344 than at FTO rs9939609 in the independent panels of Japanese samples (ESM Table 2). Nevertheless, there seems to be an overall consistency of genetic variants for susceptibility to obesity between the examined ethnic groups.

A remaining issue of interest is whether the correlations between obesity-associated SNPs and type 2 diabetes can be explained by the SNP–trait and trait–type 2 diabetes associations [18]. A simple way to assess the causal direction of associations (i.e. whether increased adiposity resulting from the variant under investigation is causally related to type 2 diabetes) is to test the inconsistency of the SNP–diabetes association both with and without adjustment for BMI. In this regard, our results, although not conclusive at some loci, indicated that six obesity variants could lead to altered fat mass and altered susceptibility to type 2 diabetes, at least in part, through separate mechanisms, based on the statistical evidence for independence following analysis of either the whole sample or separately by sex (Table 2, ESM Table 5). Interestingly, a previous study in Europeans [18] demonstrated that the associations between FTO variants and obesity-related metabolic traits (including type 2 diabetes) are likely to be entirely mediated by BMI; this study also emphasised that the use of appropriately powered studies was important in making such assessments. Our meta-analysis for the FTO locus, involving 36,064 cases and 61,234 controls, has shown that the BMI-adjusted diabetes association is highly significant (OR 1.09, p = 3.0 × 10−11). Based on analysis of stratification of Japanese samples by extent of obesity (underweight, normal weight, overweight and obese), sex (men, women) and age (<60 and ≥60 years), we determined that the strength of type 2 diabetes association tended to be more pronounced among underweight people aged <60 years in both sexes, although it did not reach statistical significance due to relatively small sample size (n < 100 for cases and controls) in the corresponding (age- and obesity-stratified) group (p = 0.18 for a group of underweight people aged <60 years vs the others; Fig. 3, ESM Fig. 1). Previous epidemiological studies have reported that being underweight is likely to be associated with the risk of type 2 diabetes in the Japanese population [44, 45] and that Japanese people have low insulin secretory capacity and a high risk of type 2 diabetes at BMIs lower than the existing WHO cut-off point for being overweight, 25 kg/m2 [46, 47]. Thus, together with other, unidentified confounding factors, the extent of obesity (>2/3 of the Japanese participants were underweight or normal weight) may account for cross- and within-population differences in type 2 diabetes susceptibility at obesity loci such as FTO. Despite limited data availability at present on the topic, the direction of BMI-adjusted diabetes association seems to be concordant in some five studies [3, 4, 10, 24, 42] that have investigated the 14 obesity variants (ESM Table 9).

Fig. 3
figure 3

Stratification of Japanese sample by extent of obesity, sex (men and women), and age (<60 and ≥60 years) in relation with genetic impacts at FTO rs9939609. Obesity was graded as: underweight (BMI <18.5 kg/m2), normal weight (18.50–24.99 kg/m2), overweight (25.00–29.99 kg/m2) and obese (≥30 kg/m2). The samples were analysed separately by sex. a Risk allele (A in the plus strand) frequencies of rs9939609, compared among four subgroups of Fukuoka panel (n = 12,569) stratified by extent of obesity. b ORs for type 2 diabetes, compared among four subgroups of case–control sample (6,781 cases, 7,307 controls) stratified by extent of obesity. c ORs for type 2 diabetes, compared between two subgroups of case–control sample (6,781 cases, 7,307 controls) stratified by age. Whiskers indicate 95% CIs. For detailed stratification, see ESM Fig. 1

Sex specificity is another issue of interest in the genetics of obesity. Some epidemiological studies have suggested that different genes influence variation of BMI in men and women [30]. However, little evidence has been provided to support this notion. In the present study, we found nominally significant sex-related differences in the effect sizes for two obesity variant loci, NEGR1 and GNPDA2 (ESM Table 3). Although heterogeneity was not statistically significant (p = 0.09), obesity association only in women appeared to be replicated for MC4R. Notably, a previous study has reported the equivalent type of sexual dimorphism for the association between a MC4R variant and obesity measures in the Swedish population [42]. Despite no prior evidence of sex specificity at this locus in Europeans [4, 48], these findings warrant further investigation with reference to ethnic homogeneity and/or the specificity of target populations.

We acknowledge several limitations inherent in the present study. Specifically, the sample size used to screen obesity association was smaller for the Japanese group (n = 18,264) than that used for the European GWA meta-analyses (n > 25,000 in the discovery stage). Therefore, we assessed statistical power for each of the tested loci, assuming effect sizes equivalent to those reported in the original GWA studies for the allele frequencies in the Japanese (ESM Table 10). Apart from the replicated associations, sufficient power (i.e. >0.8) was not attainable for two (of 14) SNP loci in the Japanese samples. In addition to the issue of power, differences in linkage disequilibrium patterns between the ethnic groups could also have contributed to the lack of replicated association at a given locus. We found some cross-population differences in linkage disequilibrium relations at several SNP loci, which could also attenuate (or strengthen) the association signals accordingly. Regional examination of SNP–obesity association, which is ongoing in Asians, but not point-wise studies, will resolve this issue.

In summary, we confirmed that 11 SNPs from ten candidate loci were significantly associated with BMI in Japanese individuals, as was previously reported in Europeans. We also found some cross-population differences in the effect sizes of individual obesity variants. These studies, moreover, highlight the genetic influences on type 2 diabetes that appear to be independent of BMI, as well as the potential presence of sex specificity in the genetics of common obesity.