Genetic determinants of obesity in Korean populations: exploring genome-wide associations and polygenic risk scores

Abstract

East Asian populations exhibit a genetic predisposition to obesity, yet comprehensive research on these traits is limited. We conducted a genome-wide association study (GWAS) with 93,673 Korean subjects to uncover novel genetic loci linked to obesity, examining metrics such as body mass index, waist circumference, body fat ratio, and abdominal fat ratio. Participants were categorized into non-obese, metabolically healthy obese (MHO), and metabolically unhealthy obese (MUO) groups. Using advanced computational methods, we developed a multifaceted polygenic risk scores (PRS) model to predict obesity. Our GWAS identified significant genetic effects with distinct sizes and directions within the MHO and MUO groups compared with the non-obese group. Gene-based and gene-set analyses, along with cluster analysis, revealed heterogeneous patterns of significant genes on chromosomes 3 (MUO group) and 11 (MHO group). In analyses targeting genetic predisposition differences based on metabolic health, odds ratios of high PRS compared with medium PRS showed significant differences between non-obese and MUO, and non-obese and MHO. Similar patterns were seen for low PRS compared with medium PRS. These findings were supported by the estimated genetic correlation (0.89 from bivariate GREML). Regional analyses highlighted significant local genetic correlations on chromosome 11, while single variant approaches suggested widespread pleiotropic effects, especially on chromosome 11. In conclusion, our study identifies specific genetic loci and risks associated with obesity in the Korean population, emphasizing the heterogeneous genetic factors contributing to MHO and MUO.

Introduction

Obesity presents a complex global health challenge that predisposes individuals to a myriad of comorbidities such as type 2 diabetes, cardiovascular disease (CVD), and certain types of cancers [1–7]. The worldwide prevalence of obesity has increased in recent decades, prompting extensive research and public health initiatives. However, obesity characteristics exhibit marked variability among different populations. Notably, East Asians (EAS) tend to manifest obesity-related complications at lower body mass index (BMI) levels than their non-Hispanic white (NHW) counterparts [8]. EAS populations also show inconsistent obesity indicators, with low BMI and high waist circumference (WC) [9, 10]. This divergence underscores the need for a distinct obesity classification standard tailored to EAS populations that differs from the NHW criteria.

High-throughput genotyping and genome-wide association studies (GWASs) have successfully identified specific genetic risk factors for obesity-related traits, treating them as polygenic conditions. Since the publication of the initial GWAS on polygenic obesity in 2007, more than 60 GWAS reports have emerged, identifying more than 1100 single nucleotide polymorphisms (SNPs) associated with obesity-related traits [11–17]. Despite the successful identification of disease susceptibility loci for obesity through GWAS, most efforts have focused on NHW populations, primarily BMI, with limited results for EAS [18–21]. Although some progress has been made with the emergence of large-scale biobank data [22–25], comprehensive GWAS specific to EAS and diverse obesity traits or definitions remains limited, except for meta-analyses conducted by the Asian Genetic Epidemiology Network consortium for WC (N = 53,052) and waist to hip ratio (N = 48,312) [26–32].

Furthermore, despite the established consensus linking obesity with comorbidities, observational studies have identified a subset of obese individuals with markedly reduced risk. These investigations have highlighted that the concept of metabolically healthy obesity (MHO), which denotes individuals despite being categorized as obese based on BMI, lacks the usual metabolic disturbances often linked to obesity, such as insulin resistance or dyslipidemia [33, 34]. This distinction underscores the necessity of viewing obesity beyond weight and size, emphasizing the significance of metabolic health in a comprehensive assessment [35–39]. Epidemiological studies have consistently indicated the existence of MHO, highlighting the contrasting risks between MHO and metabolically unhealthy obesity (MUO), with a reduced risk in MHO [40–46]. A few genetic investigations have supported the concept of MHO, underscoring the need to explore genetic associations that may vary based on different definitions of obesity [47–49]. However, research in this area remains limited.

The primary objective of this study was to identify novel loci predisposing individuals to obesity by focusing on diverse obesity indicators and obesity itself, within the context of GWAS. A substantial portion of obese individuals likely experience genetic susceptibility resulting from the cumulative impacts of numerous variants, each exerting modest individual effects—a ‘polygenic’ model akin to other intricate diseases, where most of the inherited susceptibility are attributed to polygenic inheritance involving numerous common genetic variants. The polygenic risk score (PRS) has emerged as a promising tool to address this complexity. To distinguish the genetic heterogeneity between MHO and MUO, we harnessed state-of-the-art computational algorithms for PRS with a cohort of 93,673 Korean subjects. Additionally, we explored the genetic rationale behind heterogeneity factors in obesity.

Results

Descriptive statistics of study populations

Our discovery dataset comprised 85,947 Koreans (51,317 females and 34,630 males). The descriptive statistics for this dataset are presented in Table 1. Table 1 illustrates their respective proportions of these groups in the Korean population: 66.77% non-obese, 15.98% MHO, and 17.25% MUO. The obese group was older and had a higher proportion of males. Additionally, this group exhibited higher BMI, WC, body fat (BF) ratio, and abdominal fat (AF) ratio than the non-obese group. In contrast, in the classification of the obese group based on metabolic health and obesity, the MUO group showed higher levels of obesity than the MHO group and demonstrated unhealthier biochemical levels. This suggests an increasing trend in obesity and metabolic health, in the following order: non-obese, MHO, and MUO. Notably, some indices of metabolic health were healthier in the MHO group. Consistent trends were observed across the three replication datasets, KoGES_Affy, UKB_Chi, and UKB_NHW, as shown in Supplementary Tables 1–3. The analysis scheme for study populations is illustrated in Fig. 1.

Genome-wide association studies

GWAS was conducted using the discovery dataset, and QQ and Manhattan plots depicted in Figs 2 and 3. The estimated genetic inflation factors (⁠${\lambda }_{GC}$⁠) were notably larger than 1, but the LD score intercepts (⁠${\lambda }_{\mathrm{LDSC}}$⁠) were close to 1, which indicated no evidence of inflation. From the GWAS of the continuous obesity traits, we identified 20 genome-wide significant SNPs using linear regression, as presented in Table 2. Rs574367 in the SEC16B gene, known to be associated with BMI in the GWAS catalog, met the genome-wide significance level for WC (BETA = 0.04, P = 7.15 × 10⁻¹⁴). We also identified novel disease susceptibility loci associated with obesity. Among the 13 replicated SNPs in Table 3, rs486394, located in the LINC02702 gene, has been reported to be associated with lipids [50], and our analyses showed its association with obesity (P = 2.67 × 10⁻⁹). The remaining 12 SNPs were located within or near the gene region and have been reported to be associated with BMI. They are also involved in the metabolic syndrome.

We also conducted meta-analyses that combined Korean, Japanese, and Chinese populations. QQ and Manhattan plots, illustrated in Supplementary Fig. 1, were generated from the meta-analysis. We identified rs11199833 (BETA = 0.06, P = 4.03 × 10⁻⁸) within the LOC105378523 gene on chromosome 10 as a potential candidate SNP that reached significance in BF. Unfortunately, we were unable to assess the replicability because of the absence of replication datasets.

Gene-based analyses and gene cluster identification

Supplementary Table 4 presents significantly associated genes at the Bonferroni-corrected significance level α = 7.71 × 10⁻⁶. Most genes associated with BMI or WC have been previously reported in the GWAS catalog. Table 4 displays the significant genes associated with obesity at α = 3.47 × 10⁻⁶. ADCY3 and BDNF showed significant differences in all comparisons between the non-obese and obese groups. Furthermore, the significant genes identified in the comparison between non-obese individuals and those in the combined MHO and MUO groups largely overlapped with those identified in the comparison between the non-obese and MUO groups. In addition to the ZNF259 gene, the MHO and MUO groups exhibited distinct and significant gene sets. The MUO group was primarily associated with genes located on chromosome 3, whereas the MHO group showed significant expression of genes mainly on chromosome 11.

In the context of the obesity gene-set consisting of 24 genes identified through gene-based analysis, our gene-set analysis from STRING revealed that these proteins exhibit significant connectivity as a group (P = 6.23 × 10⁻¹²). This suggests a higher degree of interaction among these genes compared with that expected for a randomly selected set of proteins of the same size and degree distribution from the genome. Additionally, we identified 5 distinct clusters containing 17 genes within the obesity gene set, as shown in Fig. 4, whereas the remaining genes remained un-clustered.

SNP heritability and genetic correlation estimation

SNP heritabilites and genetic correlations for obesity-related measures were estimated using the results presented in Supplementary Table 5. It provides heritability estimates for both Koreans and EAS and illustrates the intra- and inter-class genetic correlations among Koreans and EAS and intra-class trans-ethnic genetic correlations between EAS and NHW. The highest SNP heritability was observed for BMI with h² = 0.21 (P = 7.61 × 10⁻⁷²) for Koreans and h² = 0.16 (P = 9.29 × 10⁻¹⁴⁰) for EAS. All inter-class genetic correlations among the obesity-related traits were significant. Moreover, the intra-class transethnic genetic correlations between EAS and NHW were statistically significant for BMI, WC, and BF. All inter-class transethnic genetic correlations were greater than 0.5, with significant correlations observed between BMI and WC (TGC_{BMI, WC} = 0.68; P_{BMI, WC} = 1.11 × 10⁻¹⁶), BMI and BF (TGC_{BMI, BF} = 0.66; P_{BF, BMI} = 2.08 × 10⁻³) as well as WC and BF (TGC_{WC, BF} = 0.55; P_{WC, BF} = 2.00 × 10⁻⁵).

We calculated the SNP heritability for the obesity trait for each group. Obesity comprises three levels, and we estimated SNP heritability using summary statistics. These estimates were obtained by comparing the MHO and MUO groups with the non-obese group. When comparing the MHO and MUO groups with the non-obese group, we found a genetic correlation of GC_{MHO, MUO} = 0.99 (P = 0). For the non-obese versus MHO comparison, SNP heritability is h² = 0.12 (P = 2.06 × 10⁻³⁰); for the non-obese versus MUO, it is h² = 0.11 (P = 1.94 × 10⁻¹⁹).

Effects of PRS for BMI on obesity

For PRS analyses, the results from the validation dataset, as detailed in Supplementary Table 6, demonstrate that the PRS for BMI derived from LDpred auto yielded the best model performance. This model was employed to construct the PRS for BMI and to assess its association with obesity in the test data. As indicated in Table 5, for obesity based on metabolic health, the analysis compared the non-obese group to the obese groups comprising MHO and MUO. The PRS associations of a decrease in risk were observed in the L group compared with the M group, and an increase in risk was observed in the H group. However, it is noteworthy that a single PRS for BMI failed to differentiate risk associations between the obese groups. The effects of PRS on obesity are illustrated in Fig. 5A.

Multiple trait PRS association

We constructed a multi-PRS for measuring metabolic syndrome. The best models were selected using the correlation between the trait and PRS based on the results from the validation data presented in Supplementary Table 6. Subsequently, the selected models were applied to the test data. As demonstrated in Supplementary Table 7, the multiple trait PRS model (M2) outperformed the BMI PRS model (M1) on comparisons for obesity in terms of both AIC and area under the ROC curve (AUC). However, when comparing MHO + MUO group (case), which consists of the obese group based on BMI with non-missing metabolic traits, to the non-obese group (control), we did not observe a significant improvement in AUC.

Supplementary Figure 2 visualizes the distributions of group-wise PRS for each trait and PRSsum. For respective PRSs, the BMI PRS was well separated between non-obese and obese groups, but no significant mean difference was observed between MHO and MUO. The remaining PRSs showed slightly increasing genetic risk in the order of non-obese, MHO, and MUO, with statistically significant mean differences between the obese groups. In the case of PRSsum, the trend continued, showing higher genetic risk in MUO compared with MHO. The relationships between PRSsum and obesity are summarized in Table 5. This table reveals the consistency in risk direction across different groups based on the summed multiple trait PRS. The group denoted as H and categorized based on summed multiple trait PRS showed persistent associations across group comparisons with decreased risk levels in obesity. This finding contrasts with the limitations of a single PRS for BMI, which failed to distinguish risk association differences between obese groups despite the reported risk increase in metabolic health in the order of non-obese, MHO, and MUO. However, multiple trait PRS in the context of comparing the MHO group (control) with the MUO group (case) yielded nuanced results. It showed a decreased risk in the L group and an increased risk in the H group compared with the M group, shedding light on the complex interplay of genetic factors in different obesity subgroups. The effects of PRS on obesity are illustrated in Fig. 5B.

Heterogeneity of genetic effect on MUO and MHO

We analyzed the overall, regional, and single genetic effects for MUO and MHO (Table 6). To assess the overall genetic effect, we used the trinomial logistic regression to estimate the regression coefficient of PRS across non-obese, MHO, and MUO. The hypothesis test, which compared models with and without the constraint ${\mathrm{H}}_0:{\beta }_{\mathrm{MHO}}={\beta }_{\mathrm{MUO}}$⁠, resulted in P < 2.16 × 10⁻¹⁶, thereby rejecting the null hypothesis. For the regression coefficients for the obese groups, the PRSs for BMI, fasting plasma glucose (FPG), triglycerides (TG), and high density lipoprotein (HDL) were significant for both MHO and MUO compared with the non-obese group, with differing directions except for the PRS for BMI. The PRSs for SBP and DBP were partially significant. We calculated the correlations among PRSs, which were presented in Supplementary Table 8. Using bivariate GREML, we calculated genetic correlations (ρ = 0.89, P = 2.14 × 10⁻³³), indicating genetic similarity with minor differences.

To confirm the regional genetic effect on MUO and MHO, we conducted ρ-HESS analysis, visualized in Supplementary Fig. 3. The region CHR11:116383543-117901740 showed a significant local genetic correlation (⁠$\rho$ = 1.36 × 10⁻³, P = 1.94 × 10⁻⁵), containing 22 genes: LINC00900, LOC101929011, BUD13, ZPR1, APOA5, APOA4, APOA1, APOA1-AS, SIK3, PAFAH1B2, SIDT2, LOC100652768, PCSK7, RNF214, BACE1, CEP164, DSCAML1, FXYD2, FXYD6, TMPRSS13, IL10RA, and TMPRSS4-AS1. Among these, three genes (ZPR1, APOA1, SIK3) and five genes (BUD13, APOA1, APOA4, SIK3, PAFAH1B2) overlapped with significant findings in GWAS and gene-based analyses.

We then examined the single SNP effects on obese groups. We tested whether the 13 genome-wide significant SNPs identified from GWAS have equal effects on MHO and MUO groups using trimodal logistic regression (Table 3). Eight SNPs, including four SNPs located on chromosome 11, showed P < 0.05 for the null hypothesis ${\mathrm{H}}_0:{\beta }_{\mathrm{MHO}}={\beta }_{\mathrm{MUO}}$⁠. Rs9368219 and rs4776987 showed the same direction of effects, while the rest showed opposite direction, suggesting the presence of a subgroup within the obesity groups.

Discussion

In this study, we aimed to address the genetic components implicated in obesity by conducting GWASs and PRS analyses in the Korean population, which is a representative sample of EAS. We assumed that the genetic effect on obesity differs according to metabolic health conditions and considered four obesity-related traits and obesity based on metabolic health as outcomes to comprehensively evaluate the genetic effect on anthropometric adiposity.

We first conducted GWASs on obesity traits and identified 20 genome-wide significant SNPs in BMI and WC that were also found to be significant in NHW populations. We observed that the SNP heritabilities estimated using GWAS summary statistics ranged from 0.12 to 0.21 and were significantly larger than 0 for all traits, indicating a substantial genetic contribution to their phenotypic variability. When comparing the estimates between the EAS and NHW groups, a portion of the genetic correlations remained significant, albeit at reduced magnitudes, even after accounting for the differences in genetic architecture between these geographical genetic ancestries. Despite the disparate nature of the obesity-related traits investigated in this study (for instance, BMI serving as an anthropometric indicator, and WC as a measure of visceral fat distribution), they appear to have shared genetic influences. Consequently, it is reasonable that SNPs and genes that are significant for one obesity trait may also be significant for other obesity traits. We observed the significance of the SMIM4 and SPCS1 gene associations reported in Supplementary Table 4 for WHR [51, 52] in the GWAS catalog to extend to BMI in our analyses.

Contrary to the conventional assumption of homogeneity in obese participants, we observed heterogeneous SNP and gene effects between the MHO and MUO groups. This observation was consistent with epidemiological evidence [33, 53]. Regarding the SNP effects, we specifically identified that rs486394 in LINC02702 (Long Intergenic Non-Protein Coding RNA 2702) exhibited effects of −0.12 for MHO and 0.04 for MUO. This finding is consistent with recent research indicating that lncRNAs play a role in obesity via their involvement in adipogenesis and lipid metabolism [54]. A genetic correlation of −0.99 between these groups indicates that the same causal SNPs affect both phenotypes, but in opposite directions. Intriguingly, despite these divergent SNP effects, SNP heritability estimates were comparably close in both groups: 0.12 for MHO and 0.11 for MUO, both of which were statistically significant.

On the other hand, significant genes such as ADCY3 and BDNF in Table 4 were observed in all comparisons of the non-obese and obese groups, which are known to function in metabolism or energy balance, as well as the regulation of BMI and body weight. Although the mechanisms of ZNF259 have not been extensively studied, they have been reported to be associated with the risk of CVD, which serves as the main association that distinguishes MHO from MUO (the MHO group is known to have a lower risk of CVD than the MUO group) [33, 41, 47]. The remaining genes, which are mostly significant SNPs found in GWASs, have been reported to be associated with metabolism and obesity. Considering the limitations that LDSC do not distinguish among the various scenarios depending on the direct/indirect SNP effects or direction of effects (e.g. the SNPs on phenotype 1 affect phenotype 2, or vice versa), and the unknown mechanisms of genes, we cannot ascertain the detailed mechanisms underlying the genetic factors determining obesity classifications based on metabolic health.

However, the obesity gene set listed in Table 4 displayed statistically significant enrichment in interactions compared with a randomly selected gene set. Furthermore, among the five distinct clusters identified in the cluster analysis, two clusters consisted of genes on chromosome 3 that were significantly different in both the obese and MUO groups compared with the non-obese group. The remaining three clusters mainly consisted of genes on chromosome 11 and the three genes APOA1, APOA4, and LPL in the yellow cluster in Fig. 4 were reported to be involved in cholesterol metabolism in the KEGG pathways (hsa04979; FDR = 0.01). While this result does not fully describe the genetic predisposition to obesity based on metabolic health, it suggests the existence of heterogeneous genetic components between the obese groups using pleiotropic effects on metabolic traits. Given that the obesity criteria for metabolic health support the association between both obesity and metabolic traits, this appears to be a case of pleiotropy.

The inherited genetic risk of BMI was evaluated using the PRS for BMI, and we found that individuals with a high PRS for BMI were more likely to be obese. The performance of the PRS for BMI improved with multiple trait PRS. This observation suggests that the shared genetic components among the various obesity-related traits identified in this study contribute to a better understanding of obesity across different criteria. However, when it comes to subgroups within the obese category, although a single PRS for BMI may offer insights into genetic predisposition, it may not fully capture the polygenicity. The complexity of these conditions likely involves a combination of multiple genetic factors beyond BMI alone. In conclusion, the PRS for BMI proves to be valuable in assessing the inherited genetic risk of obesity; however, for subgroups within obese groups, a more comprehensive approach that incorporates multiple trait PRS and accounts for other genetic components may be necessary to gain a deeper understanding of their genetic underpinnings.

The heterogeneity in obese groups identified through GWAS, gene-based analyses, and the PRS trend in obesity classification was quantified by comparing the overall, regional, and single SNP effects between MHO and MUO groups. For the overall genetic effect comparison, we found significant differences in the effects of each PRS on MHO and MUO with small differences in regression coefficients. These differences were further confirmed with genetic correlations (0.89 from bivariate GREML). The regional genetic effect comparison revealed significant local genetic correlations primarily in chromosome 11, with genetic correlation of 1.36 × 10⁻³. Single variant approaches indicated more widespread minor heterogeneity across the entire genome. These observations underscore some genetic differences between MHO and MUO, highlighting the complex interplay of genetic factors in obesity. Our study emphasizes the importance of considering both common and distinct genetic components in understanding MHO and MUO.

Although we unveiled novel associations between GWASs and PRS, along with post-hoc analyses, there are certain limitations that should be considered. First, our findings pertained primarily to the Korean population. Although our summary statistics were from the Japanese population, it is important to note that the Japanese and Korean populations are genetically close (GC over 0.9 on BMI). It is possible that some bias may have been introduced because of genetic differences between these populations for some phenotypes. Secondly, the absence of a universally accepted gold standard for defining MHO poses a challenge. In our study, we used moderate criteria for MHO. However, there are no established validation methods for determining the optimal definition. Despite these limitations, our research offers valuable insights into obesity by studying a large Korean population, which is the largest genetic study conducted on this population to date. Additionally, we provide evidence of heterogeneous genetic effects within obese groups, highlighting the need to consider subpopulations within this category. Overall, these findings contribute to our understanding of genetic predisposition to obesity.

Materials and methods

Study population

We considered several prospective cohorts drawn from patients registered at our hospital. For discovery dataset, we utilized portions of the following four cohorts: the Korean Genome Epidemiology Study (KoGES) cohort (N = 211,725) [24] (https://biobank.nih.go.kr/cmm/main/mainPage.do), Gene-Environment of Interaction and phenotype cohort [55] (GENIE, N = 10,300), the YonSei University Hospital medical center (YSUH, N = 6679), and Veterans Health Service Medical Cent (VHSMC, N = 2598). The KOGES cohort is a population-based cohort comprising three subcohorts from the Korean Association REsource (N = 10,030), Cardiovascular Disease Association (N = 28,338), and Health EXAminee (N = 173,357) studies. This cohort was segmented into two groups based on the genotyping platform used: the Korean Biobank Array (KoreanChip) [56] and Affymetrix Genome-Wide Human SNP Array 5.0 or 6.0 (Affymetrix, Santa Clara, CA, USA). For the discovery analyses, we specifically used subjects genotyped using KoreanChip. For the GENIE, YSUH, and VHSMC cohorts, health-related information was collected from patients who visited the hospital and provided informed consent. After excluding participants with missing obesity outcomes, the discovery dataset comprised 85,947 participants and 4,736,957 SNPs.

For replication, we used three datasets: a fraction of the KoGES cohort genotyped with Affymetrix (N = 8856), as well as two ethnic groups of Chinese (N = 1503) and NHW (N = 459,259) from the United Kingdom Biobank (UKB, N = 502,413) cohort. These were denoted as KoGES_Affy, UKB_Chi, and UKB_NHW, respectively. The UKB cohort [57] (https://www.ukbiobank.ac.uk) was a large-scale prospective cohort study that included subjects residing in the UK. The baseline survey for this cohort commenced in 2006, and follow-up is still in progress. All participants were aged 40–69 years at baseline and provided electronically signed consent. Each cohort was designed to examine both genetic and phenotypic data in order to investigate a wide range of risk factors associated with common complex diseases. The final replication datasets for KoGES_Affy, UKB_Chi, and UKB_NHW included 7726 Koreans with 2,221,602 SNPs, 1496 Chinese individuals with 2,076,717 SNPs, and 392,160 NHW with 4,936,389 SNPs.

Obesity-related measures and operational definition of metabolic health and obesity

Several obesity metrics including BMI, WC, BF, and AF were evaluated. For GWASs, continuous measurements were used for BMI and WC. In the context of PRS analyses, BMI was dichotomized to cases (obese group) if BMI ≥ 25 kg/m², while those with lower BMIs were defined as controls. For WC, females with a measurement ≥85 cm and males with a measurement ≥90 cm were designated as cases [58].

Participants with no missing biochemical observations were further divided into three groups based on obesity as defined by BMI and metabolic health status: non-obese, MHO, and MUO. The EAS with a BMI less than 25 kg/m² and NHW with a BMI less than 30 kg/m² were classified as non-obese, whereas the others were classified as obese. An individual was considered MHO if they were obese but had fewer than or equal to two of the following metabolic risk conditions: (i) elevated blood pressure [either systolic blood pressure (SBP)/diastolic blood pressure (DBP) ≥ 130/85 mm Hg or taking anti-hypertensive medication], (ii) impaired FPG (≥100 mg/dL or a diagnosis of diabetes mellitus, or prescription for antidiabetic medication), (iii) high plasma TG (≥150 mg/dL), and (iv) low HDL cholesterol (<40 mg/dL in men or <50 mg/dL in women) [59]. Based on these criteria, participants were either classified as MHO (obese with fewer than or equal to two metabolic risk factors) or MUO (obese with at least three metabolic risk factors).

Genotyping, quality control, and imputation

KoGES included 81,153 participants. Of these, 72,297 were genotyped using KoreanChip and the remaining 8856 participants were genotyped using Affymetrix. Participants in the GENIE, YSUH, and VHSMC cohorts were genotyped using KoreanChip. The genotypes generated by KoreanChip were called using the K-medoid algorithm [60] to minimize heterogeneities between batches and studies.

For downstream quality control, SNPs were removed if the missing genotype call rates were >0.05 or the Hardy Weinberg Equilibrium (HWE) was P < 10⁻⁵. Participants were excluded if they had missing genotype call rates >0.05 or there was a sex inconsistency. The remaining SNPs were prephased with Eagle v2.4 [61] and untyped SNPs were imputed using the Northeast Asian Reference Database [62] imputation server (https://nard.macrogen.com/). Imputed SNPs were removed if R² < 0.3 or the number of alleles ≠ 2, missing genotype rate > 0.05, minor allele frequency (MAF) < 0.005 (KoreanChip), MAF < 0.05 (Affymetrix), or HWE P < 10⁻⁵. Participants were excluded if SNP heterozygosity was >5 × IQR or if the mean difference between the top two principal components (PCs) was >5 × IQR.

The imputed genotypes were downloaded from the UKB. SNPs were removed if the missing genotype rate > 0.01, P for HWE < 10⁻⁶, or the number of alleles ≠ 2. We considered participants of Chinese and European ancestry, with NHW participants identified as having a White, British, Irish, and any other White background. For the Chinese group, we eliminated SNPs with an MAF < 0.05. In the NHW group, we excluded participants if MAF < 0.005 and if SNP heterozygosity >5 × IQR, estimated genetic relationship >0.125 or mean difference of top two PCs > 5 × IQR.

All data management and quality control processes were performed using PLINK [63], PLINK2 [64], GCTA [65], and ONETOOL [66].

Genome-wide association studies

In order to identify the loci associated with obesity, we conducted a GWAS on obesity-related variables using the discovery data. Linear regression was used for continuous outcomes, incorporating age, sex, and the first 10 PCs as covariates. For categorical outcomes, binary or multinomial logistic regression was performed using the same covariates as those used in the linear regression. For the multinomial logistic regression, association analyses were conducted using a likelihood ratio test (LRT). Linear and logistic regressions were conducted using PLINK and Rex [67] and the VGAM package in R version 1.1.7. The significance level was set at α = 5 × 10⁻⁸. Significant genome-wide SNPs were annotated using ANNOVAR [68]. Furthermore, we performed two meta-analyses. The first involved GWASs for the discovery dataset and replication dataset of KoGES_Affy by using METAL [69]. The second incorporated all the GWAS results for EAS, including the publicly accessible Biobank Japan (BBJ) summary statistics for BMI (N = 158,284).

Gene-based analyses and gene cluster identification

We conducted gene-based analyses of 18,432 genes within the discovery datasets, KoGES_Affy, and UKB_Chi using FUMA [70] with 1000G Phase 3 EAS as a reference population. We performed the same analyses on 14,411 genes using the logistic regression results from the discovery dataset. This allowed us to compare the obese group comprising both the MHO and MUO groups and the MHO and MUO groups with the non-obese group. For UKB_NHW, on the other hand, we utilized the 1000G Phase 3 EUR.

Specifically for the analysis comparing MHO, MUO, or both groups with non-obese group, we expanded our analysis by performing gene-set and gene cluster analyses using the Search Tool for the Retrieval of Interacting Genes (STRING version 12.0) [71]. For the gene set enrichment analysis, the whole genome was used as the statistical background. For the gene network clustering analysis, we applied the Markov Cluster Algorithm with the inflation parameter set to three.

SNP heritability and genetic correlation

GWAS results were used in conjunction with LDSC [72] to estimate SNP heritability and genetic correlations within EAS populations. The transient genetic correlation between the EAS and NHW populations was calculated using POPCORN [73]. In cases where there was an overlap among participants used to calculate summary statistics, for instance, regression coefficients between MHO compared with non-obese individuals and between MHO compared with non-obese individuals were adjusted for correlations between summary statistics using Erase Sample Overlap and Relatedness (EraSOR) [74], resulting in 595,220 SNPs and then incorporated into LDSC to adjust correlations between summary statistics.

PRS and prediction model building

The construction of PRS requires summary statistics from the GWAS. We used GWAS data on BMI from the BBJ [75] (http://jenger.riken.jp/en), focusing on SNPs with MAF > 0.005. The remaining 5,925,388 SNPs were used for the PRS calculation. The PRS model has been constructed using various methods, including clumping + thresholding (CT) [76], LDpred with infinitesimal, grid, and auto models [77, 78], lassosum [79], and PRScs [80]. The EAS subpopulation from 1000 Genomes Phase 3 was used as the linkage disequilibrium reference panel. For LDpred grid model, proportions of causal SNP $\rho$ were set at 1%, 3%, 10%, 30%, and 100% with the default value used for the other options.

PRS analyses require both validation and testing. For validation, we randomly selected 15,000 participants from all Koreans genotyped using KoreanChip, while the remaining 70,947 Koreans were used as test data. The validation data underwent 10-fold cross-validation for BMI, and we used a logistic regression model to create a prediction model incorporating baseline age, sex, BMI RPS, and PC1–5 as covariates.

To enhance the accuracy of the prediction model, we considered a multiple trait PRS approach [81]. This approach included PRSs for BMI, SBP, DBP, FPG, TG, and HDL as covariates. The PRS models were constructed using summary statistics downloaded from the BBJ. These multiple trait PRS models were validated using the same method used for the BMI. These multiple trait PRS were used to develop a predictive model for obesity. The best method for the PRS was selected using the correlation between the trait and PRS. The prediction accuracy of the selected model was evaluated using the test data, and the AUCs between the models based on BMI PRS and multiple trait PRS were compared using the DeLong test [82].

Association analyses of PRS

The association between the PRS and obesity was investigated using logistic regression analysis. We considered two distinct PRSs: one for BMI and another PRSsum [83] that combined multiple trait PRS for BMI, SBP, DBP, FPG, TG, and HDL with equal weights. From the constructed PRSs, we compared the PRS among the groups using pairwise T-tests to determine the mean differences between MHO and MUO. PRSs were categorized into three groups: low (L), medium (M), and high (H). Here ‘L’ represents the lower 10%, ‘M’ encapsulates the range of 10%–90%, and ‘H’ denotes the upper 10%.

Heterogeneity of genetic effect on MHO and MUO

We compared the overall, regional, and single genetic effects between MHO and MUO groups. To compare the overall genetic effect, we tested the hypothesis that the effects of PRSs for BMI, SBP, DBP, FPG, TG, and HDL on MHO and MUO were the same compared with the non-obese group (i.e. ${\beta }_{\mathrm{MHO}}={\beta }_{\mathrm{MUO}}$for BMI PRS, etc.) using trinomial regression (VGAM package in R version 1.1.7). Additionally, genetic correlations were measured using bivariate GREML analysis [84] implemented in GCTA for individual-level data. Regional genetic effects were compared using local genetic correlation using ρ-HESS [85], which quantifies the correlation between traits in specific genomic regions, with significance determined using Bonferroni-corrected $\alpha$ = 3.48 × 10⁻⁵. Lastly, single SNP genetic effects were compared using trinomial regression with restriction ${\beta }_{\mathrm{MHO}}={\beta }_{\mathrm{MUO}}$⁠. SNPs with P < 0.05 were identified as having heterogenous SNP effects.

Author contributions

Jang Won Son and Sungho Won contributed to the design of the study and revised the manuscript. Jinyeon Jo analyzed the data and wrote the manuscript. Nayoung Ha, Yunmi Ji, Ahra Do, Je Hyun Seo, Bumjo Oh, Sungkyung Choi, Eun Kyung Choe, and Woojoo Lee reviewed the manuscript.

Conflict of interest: The authors declare that they have no competing interest.

Funding

This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (HI22C0154), the National Research Foundation (NRF) grants (NRF-2021R1A5A1033157) and (RS-2024-00346850) funded by the Korean government (MSIT). Statistical analyses were supported by the National Supercomputing Center with supercomputing resources including technical support (KSC-2022-CRE-0319 and KSC-2023-CRE-0117). This study was conducted with bioresources from National Biobank of Korea, the Korea Disease Control and Prevention Agency, Republic of Korea (NBK-2020-101). The committee of VHS Biobank (VBP-2020-03) approved the use of bioresources for this study. This study was supported by Research Grant (Grant No. KSSO-D-2021002) from Korean Society for the Study of Obesity.

Data availability

The data used in this study are sourced from various repositories, including KoGES (Korean Genome Epidemiology Study), GENIE (Gene–Environment of Interaction and Phenotype Cohort), YSUH (YonSei University Hospital Medical Center), VHSMC (Veterans Health Service Medical Center), and UK Biobank. For researchers interested in accessing the KoGES and UK Biobank data, these datasets are available through institutional data access and can be requested by eligible researchers. Detailed information and access criteria for KoGES can be found at KoGES Data Access (https://biobank.nih.go.kr/cmm/main/mainPage.do), and UK Biobank data access can be requested via UK Biobank (https://www.ukbiobank.ac.uk/). Access to GENIE, YSUH, and WHSMC data is facilitated through collaboration with researchers in the respective hospitals. Additionally, summary statistics from the BioBank Japan (BBJ) project can be downloaded from https://pheweb.jp/. We encourage interested researchers to follow the provided links for more information on data availability and access procedures.

References