New genetic loci link adipose and insulin biology to body fat distribution.

Our study included 224,459 individuals of European, East Asian, South Asian, and African American ancestry. The European ancestry arm included 142,762 individuals from 57 cohorts genotyped with genome-wide SNP arrays and 67,326 individuals from 44 cohorts genotyped with the Metabochip 11 ( Extended Data Fig. 1 , Supplementary Table 1 ). The non-European ancestry arm comprised ~1,700 individuals from one cohort of East Asian ancestry, ~3,400 individuals from one cohort of South Asian ancestry, and ~9,200 individuals from six cohorts of African American ancestry, all genotyped with the Metabochip. There was no overlap between individuals genotyped with genome-wide SNP arrays and Metabochip. For each study, local institutional committees approved study protocols and confirmed that informed consent was obtained. Our primary trait was WHRadjBMI, the ratio of waist and hip circumferences adjusted for age, age 2 , study-specific covariates if necessary, and BMI. For each cohort, residuals were calculated for men and women separately and then transformed by the inverse standard normal function. Cohorts with related men and women provided inverse standard normal transformed sex-combined residuals. For each cohort, the same transformations were applied to other traits: (i) WHR without adjustment for BMI (WHR); (ii) waist circumference with (WCadjBMI) and without (WC) adjustment for BMI; and (iii) hip circumference with (HIPadjBMI) and without (HIP) adjustment for BMI. Sample and SNP quality control (QC) were undertaken within each cohort ( Supplementary Table 3 ) 44 . The GWAS scaffold in each cohort was imputed up to CEU haplotypes from HapMap resulting in ~2.5 million SNPs. Each directly typed and imputed SNP passing QC was tested for association with each trait under an additive model in a linear regression framework ( Supplementary Table 3 ). SNP positions are reported based on NCBI Build 36. For each cohort, sex-specific association summary statistics were corrected for residual population structure using the genomic control inflation factor 45 (median λ GC =1.01, range=0.99 – 1.08). SNPs were removed prior to meta-analysis if they had a minor allele count ≤ 3, deviation from Hardy-Weinberg equilibrium exact P <10 −6 , directly genotyped SNP call rate<95%, or low imputation quality (below 0.3 for MACH, 0.4 for IMPUTE, and 0.8 for PLINK). Association summary statistics for each trait were combined via inverse-variance weighted fixed-effects meta-analysis and corrected for a second round of genomic control to account for structure between cohorts ( Extended Data Fig. 1 , Supplementary Fig. 1 ). Sample and SNP QC analyses were undertaken in each cohort ( Supplementary Table 3 ). Each SNP passing QC was tested for association with each trait under an additive model using linear regression. The Metabochip array 11 is enriched, by design, for loci associated with anthropometric and cardiometabolic traits, thus, we based our correction on 4,425 SNPs selected for inclusion based on associations with QT-interval that were not expected to be associated with anthropometric traits (>500 kb from variants on Metabochip 46 for these traits). These study-specific inflation factors had a median λ GC =1.01(range 0.93–1.11), with only one study exceeding 1.10. After removing SNPs for QC as described in the previous section, association summary statistics were combined via inverse-variance weighted fixed-effects meta-analysis and corrected for a second round of genomic control on the basis of QT-interval SNPs to account for structure between cohorts. Association summary statistics from the two parts of the European ancestry arm were combined via inverse-variance weighted fixed-effects meta-analysis using METAL 47 with no further genomic control correction. Results were reported for SNPs with a sex-combined sample size≥50,000. The meta-analyses were repeated for men and women separately for each trait. Analyses were corrected for population structure within each sex. The meta-analysis of WHRadjBMI in men included up to 93,480 individuals, and in women up to 116,742 individuals. Sample and SNP QC, tests of association, genomic control correction (median λ GC =1.01, range=0.90–1.17, with only one study exceeding 1.10), and meta-analyses were performed as described above. Association summary statistics from the European and non-European ancestry meta-analyses were combined via inverse-variance weighted fixed-effects meta-analysis without further genomic control correction. For each lead SNP, we tested for sex differences based on the sex-specific beta estimates and standard errors, while accounting for potential correlation between estimates as previously used in Randall et al 10 . Similarly, we tested for potential differences in effects between European and non-European samples, comparing the effects from GWAS+Metabochip data for Europeans and Metabochip data for non-Europeans, and we tested for differences between population-based studies and samples ascertained on diabetes status, and cardiovascular disease, or both. In assessing effects of ascertainment overall, we compared effects in seven subsets of our study sample using population-based studies (i.e., those not ascertained on any phenotype) as the referent population: 1) all studies ascertained on any phenotype, 2) T2D cases, 3) T2D controls, 4) T2D cases+controls, 5) CAD cases, 6) CAD controls, and 7) CAD cases+controls. We evaluated significance for heterogeneity tests within each comparison using a Bonferroni-corrected p-value of 0.05/49=0.05/49=1.02×10 −3 as well as an FDR threshold 48 of <5% ( Supplementary Table 28 ). Between-study heterogeneity in all meta-analyses was assessed using I 2 statistics 49 . We calculated the heritability and genetic correlations of several central obesity traits using variance component models 50 , 51 in the Framingham Heart Study (FHS) and TWINGENE study. In this approach, the phenotypic variance is decomposed into additive genetic, non-additive genetic, and environmental sources of variation (including model error), and for sets of traits, the covariances between traits. We report narrow sense heritability ( h 2 ), the ratio of the additive genetic variance to the total phenotypic variance. Sex-specific inverse normal trait residuals, adjusted for age (and cohort in FHS), were used to estimate heritability separately in men and women, using variance components analysis in SOLARv.4.2.7 52 (FHS) or M×1.703 53 (TWINGENE). Additionally, the sex-specific residuals were used to conduct bivariate quantitative variance component genetic analyses that calculate genetic and environmental correlations between traits. The genetic correlations obtained are estimates of the additive effects of shared genes, and a genetic correlation significantly different from zero suggests a direct influence of the same genes on more than one trait. Similarly, significant environmental correlations suggest shared environmental effects. We estimated sex-stratified correlations between all waist traits, as well as BMI, height, and weight in TWINGENE, FHS, KORA, and EGCUT. In TWINGENE and FHS, age-adjusted Pearson correlations were used; in EGCUT and KORA, correlations were adjusted for age and age 2 . To evaluate the evidence for multiple association signals within identified loci, we performed approximate conditional analyses of sex-combined, women-specific, and men-specific data as implemented in the GCTA software 14 , 54 . This approach makes use of association summary statistics from the combined European ancestry meta-analysis and a reference dataset of individual-level genotype data to estimate LD between variants and hence also the approximate correlation between allelic effect estimates in a joint association model. To evaluate robustness of the GCTA results, we performed analyses using two reference datasets: Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS) consisting of 949 individuals from Uppsala County, Sweden with both GWAS and Metabochip genotype data; and Atherosclerosis Risk in Communities (ARIC) consisting of 6,654 individuals of European descent from four communities in the USA with GWAS data. Both GWAS datasets were imputed using data from Phase II of the International HapMap Project 55 . Results shown use the PIVUS reference dataset because Metabochip genotypes are available (see a comparison in the Supplementary Note ) . Assuming that the LD correlations between SNPs more than 10 Mb away are zero, and using each reference dataset in turn, we performed a genome-wide stepwise selection procedure to select associated SNPs one-by-one at a P value<5×10 −8 . For each locus at which multiple association signals were observed in the sex-combined, women-, and/or men-specific data, the SNPs selected by GCTA as independently associated with WHRadjBMI in any of the three meta-analyses are reported, with the SNP identified in the sex-combined analysis taken by default when proxies are identified in the women- and/or men-specific analyses. For SNPs not selected by a particular joint conditional analysis, but identified by either of the other two analyses, summary statistics were calculated for association analysis of the SNP conditioned on the GCTA-selected SNP(s). We calculated a genetic risk score for each individual in the population-based KORA study (1,670 men and 1,750 women) using the 49 identified variants, weighted by the allelic effect from the European ancestry meta-analyses of WHRadjBMI. Sex-combined scores were computed on the basis of the sex-combined meta-analysis. Sex-stratified scores were calculated on the basis of men- and women-specific meta-analyses, where SNPs not achieving nominal significance in the respective sex ( P ≥0.05) were excluded. For each individual, the sex-combined and sex-stratified risk scores were rounded to the nearest integer for plotting. Risk scores were then tested for association with WHRadjBMI using linear regression. We calculated the variance explained by a single SNP as: 2 • M A F • ( 1 − M A F ) • β 2 V a r ( Y ) where MAF is the minor allele frequency, β is the SNP effect estimate computed by meta-analysis, and Var(Y) is the variance of the phenotype Y as it went into the study-specific association testing. To derive the total variance explained by a set of independent SNPs, we computed the sum of single-SNP explained variances under the assumption of independent contributions. To estimate the polygenic variance explained by all HapMap SNPs, we used the all-SNP estimation approach implemented in GCTA and analysed individuals in the ARIC and TwinGene cohorts, including the first 20 principal components as fixed covariates. After removing one of each pair of individuals with estimated genetic relatedness>0.025, 11,898 unrelated individuals with WHRadjBMI were available. We considered each identified locus, defined as 500 kb upstream and downstream of the lead SNP, and computed 95% credible intervals using a Bayesian approach. On the basis of association summary statistics from the European ancestry, non-European ancestry, or all ancestries sex-combined meta-analyses, we calculated an approximate Bayes’ factor 56 in favor of association, given by: B F j = 1 − R j exp ( − R j β j 2 2 σ j 2 ) where β j is the allelic effect of the j th SNP, with corresponding standard error σ j , and R j = 0.04 ∕ ( σ j 2 + 0.04 ) , which incorporates a N(0,0.2 2 ) prior for β j . This prior gives high probability to small effect sizes, and only small probability to large effect sizes. We then calculated the posterior probability that the j th SNP is causal by: φ j = B F j Σ k B F k where the summation in the denominator is over all SNPs passing QC across the locus. We compared the meta-analysis results and credible sets of SNPs likely to contain the causal variant as described 57 . Assuming a single causal variant at each locus, a 95% credible set of variants was then constructed by: (i) ranking all SNPs according to their Bayes’ factor; and (ii) combining ranked SNPs until their cumulative posterior probability exceeded 0.95. For each locus, we calculated the number of SNPs contained within the 95% credible sets, and the length of the genomic interval covered by these SNPs. To determine whether the identified loci were also associated with any of 22 cardio-metabolic traits, we obtained association data from meta-analysis consortia DIAGRAM (T2D) 58 , CARDIoGRAM-C4D (CAD) 59 , ICBP (SBP, DBP) 60 , GIANT (BMI, height) 36 , 37 , GLGC (HDL, LDL, and TG) 61 , MAGIC (fasting glucose, fasting insulin, fasting insulin adjusted for BMI, and two-hour glucose) 62 - 64 , ADIPOGen (BMI-adjusted adiponectin) 65 , CKDgen (urine albumin-to-creatinine ratio (UACR), estimated glomerular filtration rate (eGFR), and overall CKD) 66 , 67 , ReproGen (age at menarche, age at menopause) 68 , 69 , and GEFOS (bone mineral density) 70 ; others provided association data for IgA nephropathy 71 (also Kiryluk K, Choi M, Lifton RP, Gharavi AG, unpublished data) and for endometriosis (stage B cases only) 72 . Proxies ( r 2 >0.80 in CEU) were used when an index SNP was unavailable. We also searched the National Human Genome Research Institute (NHGRI) GWAS Catalog for previous SNP-trait associations near our lead SNPs 73 . We supplemented the catalog with additional genome-wide significant SNP-trait associations from the literature 13 , 70 , 74 - 80 . We used PLINK to identify SNPs within 500 kb of lead SNPs using 1000 Genomes Project Pilot I genotype data and LD ( r 2 ) values from CEU 81 , 82 ; for rs7759742, HapMap release 22 CEU data 81 , 83 were used. All SNPs within the specified regions were compared with the NHGRI GWAS Catalog 16 . To evaluate whether the alleles associated with increased WHRadjBMI at the 49 identified SNPs convey effects for any of the 22 cardiometabolic traits, we conducted meta-regression analyses of the beta-estimates on these metabolic outcomes from other consortia with the beta-estimates for WHRadjBMI in our data 65 . Based on the association data across traits, we generated a matrix of Z -scores by dividing the association betas for each of the 49 WHRadjBMI SNPs for each of 22 traits by their respective standard errors. The traits did not include WHRadjBMI or nephropathy in Chinese subjects, but did include HIPadjBMI and WCadjBMI. Each Z -score was made positive if the original trait-increasing allele also increased the look-up trait and negative if not. Missing associations with were assigned a value of zero. We performed unsupervised hierarchical clustering of the Z score matrix in R using the default settings of the “heatplot” function from the made4 library (version 1.20.0), agglomerating clusters using average linkage and Pearson correlation metric distance. The rows and columns of matrix values were each automatically scaled to range from 3 to −3. Confidence in the hierarchical clustering was assessed by bootstrap analysis (10,000 resamplings) using the R package “pvclust” 84 . The 1000 Genomes CEU pilot data were queried for SNPs within 500 kb and in LD ( r 2 >0.7, an arbitrary threshold) with any index SNP. All identified variants were then annotated based on RefSeq transcripts using Annovar to identify potential nonsynonymous variants near identified association signals. The distance between each variant and the nearest transcription start site were calculated using gene annotations from GENCODE (v.12). To investigate whether SNPs in LD with index SNPs are also in LD with common copy number variants (CNVs), we extracted waist trait association results for a list of SNP proxies that are in high LD (r 2 >0.8, CEU) with CNVs in European populations as described previously 7 . Altogether 6,200 CNV-tagging SNPs were used, which are estimated collectively to capture>40% of CNVs>1 kb in size. We examined our lead SNPs in eQTL datasets from several sources ( Supplementary Note ) for cis effects significant at P< 10 −5 . We then checked if the trait-associated SNP also had the strongest association with the expression level of its corresponding transcript. If not, we identified a nearby SNP that had a stronger association with expression (peak transcript SNP) of that transcript. To check whether effects of the peak transcript SNP and waist trait-associated SNP overlapped, we conducted conditional analyses to estimate associations between the waist-associated SNP and transcript level when the peak transcript-associated SNP was also included in the model, and vice versa. If the association for the expression-associated SNP was not significant ( P >0.05) when conditioned on the waist-associated SNP, we concluded that the waist-associated SNP is likely to explain a substantial proportion of the variance in gene transcript levels in the region. For SNPs that passed these criteria in either women or men eQTL datasets from deCODE, we investigated sex heterogeneity in gene transcript levels for whole blood (312 men, 435 women) and subcutaneous adipose tissue (252 men, 351 women) based on the sex-specific beta estimates and standard errors, while accounting for potential correlation between the sex-specific associations 8 . We examined overlap of regulatory elements with the 68 trait-associated variants and variants in LD with them ( r 2 >0.7, 1000 Genomes Phase 1 version 2 EUR 85 ), totaling 1,547 variants. We obtained regulatory element data sets from the ENCODE Consortium 24 and Roadmap Epigenomics Project 25 corresponding to eight tissues selected based on a current understanding of WHRadjBMI pathways. The 226 regulatory element datasets included experimentally identified regions of open chromatin (DNase-seq, FAIRE-seq), histone modification (H3K4me1, H3K27ac, H3K4me3, H3K9ac, and H3K4me2), and transcription factor binding ( Supplementary Table 17 ). When available, we downloaded data processed during the ENCODE Integrative Analysis 24 . We processed Roadmap Epigenomics sequencing data with multiple biological replicates using MACS2 86 and the same Irreproducible Discovery Rate pipeline used in the ENCODE Integrative Analysis. Roadmap Epigenomics data with only a single replicate was processed using MACS2 alone. We performed permutation-based tests in a subset of 60 open chromatin (DNase-seq) and histone modification (H3K27ac, H3K4me1, H3K4me3, H3K9ac) datasets to identify global enrichment of the WHRadjBMI-associated loci. We matched the index SNP at each locus with 500 variants having no evidence of association ( P> 0.5, ~1.2 million total variants) with a similar distance to the nearest gene (±11,655 bp), number of variants in LD (±8 variants), and minor allele frequency. Using these pools, we created 10,000 sets of control variants for each of the 49 loci and identified variants in LD ( r 2 >0.7) and within 1 Mb. For each SNP set, we calculated the number of loci with at least one variant located in a regulatory region under the assumption that one regulatory variant is responsible for each association signal. We initially calculated an enrichment P value by finding the proportion of control sets for which as many or more loci overlap a regulatory element than the set of associated loci. For increased P value accuracy, we estimated the P value assuming a sum of binomial distributions to represent the number of index SNPs or their LD proxies that overlap a regulatory dataset compared to the 500 matched control sets. Gene Relationships Among Implicated Loci (GRAIL) 19 is a text-mining algorithm that evaluates the degree of relatedness among genes within trait regions. Using PubMed abstracts, a subset of genes enriched for relatedness and a set of keywords that suggest putative pathways are identified. To avoid potential bias from papers investigating candidate genes stimulated by GWAS, we restricted our search to abstracts published prior to 2006. We tested for enrichment of connectivity in the independent SNPs that were significant in our study at P <10 −5 . To investigate if pathways including predefined sets of genes were enriched in the lower part of the gene P value distribution for WHRadjBMI, we performed a pathway analysis using Magenta 2.4 20 and SNPs present in both the Metabochip and GWAS meta-analyses. SNPs were assigned to a gene if within 110 kb upstream or 40 kb downstream of the transcript’s boundaries. The most significant SNP P value within this interval was adjusted for putative confounders (gene size, number of SNPs in a gene, LD pattern) using stepwise linear regression, creating a gene association score. If the same SNP was assigned to multiple genes, only the gene with the lowest gene score was kept. The HLA region was removed from further analyses due to its high LD structure and gene density. Each gene was then assigned pathway terms using Gene Ontology (GO), PANTHER, Ingenuity and Kyoto Encyclopedia of Genes and Genomes (KEGG) 87 - 90 . Finally, the genes were ranked based on their gene association score, and a modified gene-set enrichment analysis (GSEA) using MAGENTA was performed. This analysis tested for enrichment of gene association score ranks above a given rank cutoff (including 5% of all genes) in a gene-set belonging to a predefined pathway term, compared to multiple, equally sized gene-sets that were randomly sampled from all genes in the genome. 10,000-1,000,000 gene-set permutations were performed. This method is described in detail elsewhere 23 , 36 . Briefly, DEPICT uses gene expression data derived from a panel of 77,840 expression arrays 91 , 5,984 molecular pathways (based on 169,810 high-confidence experimentally-derived protein-protein interactions 92 ), 2,473 phenotypic gene sets (based on 211,882 gene-phenotype pairs from the Mouse Genetics Initiative 93 ), 737 Reactome pathways 94 , 184 KEGG pathways 95 , and 5,083 GO terms 19 . DEPICT uses the expression data to reconstitute the protein-protein interaction gene sets, mouse phenotype gene sets, Reactome pathway gene sets, KEGG pathway gene sets, and GO term gene sets. To avoid biasing the identification of genes and pathways covered by SNPs on the Metabochip, analyses were restricted to GWAS cohort data and included 226 WHRadjBMI SNPs in 78 non-overlapping loci with sex-combined P <10 −5 . We used DEPICT to map genes to associated WHRadjBMI loci, which then allowed us to (1) systematically identify the most likely causal gene(s) in a given associated region, (2) identify reconstituted gene sets that were enriched in genes from associated regions, and (3) identify tissue and cell type annotations in which genes from associated regions were highly expressed. Associated regions were defined by all genes residing within LD ( r 2 >0.5) distance of the WHRadjBMI-associated index SNPs. Overlapping regions were merged, and SNPs that mapped near to or within the HLA region were excluded. The 93 WHRadjBMI SNPs with P <10 −5 (clumping thresholds: HapMap release 27 CEU r 2 =0.01, 500 kb) resulted in 78 non-overlapping regions. GWAS+Metabochip index SNPs were annotated with DEPICT-prioritized genes if the DEPICT (GWAS-only) SNP was located within 500 kb. To mark related gene sets, we first quantified significant gene sets’ pairwise overlap using a non-probabilistic version of the reconstituted gene sets and the Jaccard index measure. Groups of gene sets with mutual Jaccard indices >0.25 were subsequently referred to as meta gene sets and named by the most significant gene set in the group ( Supplementary Table 18 and Fig. 2a ). In Figures 2a-b , gene sets with similarities between 0.1-0.25 were connected by an edge that was scaled according to degree of similarity. The Cytoscape tool was used to construct parts of Figure 2 96 . In Figure 2c , we show the significance of all cell type annotations and annotations that were categorized as “Tissues” at the outermost level of the Medical Subject Heading ontology.

We performed meta-analyses of GWAS of WHRadjBMI in up to 142,762 individuals of European ancestry from 57 new or previously described GWAS 7 , and separately in up to an additional 67,326 European ancestry individuals from 44 Metabochip studies ( Extended Data Fig. 1 ; Supplementary Tables 1-3 ). The combination of these two meta-analyses included up to 2,542,447 autosomal SNPs in up to 210,088 European ancestry individuals. We defined new loci based on genome-wide significant association ( P500 kb) from previously established loci 7 , 8 . We identified 49 loci for WHRadjBMI, 33 of which were new and 16 previously described 7 , 8 . Of these, a European ancestry sex-combined analysis identified 39 loci, 24 of which were new ( Table 1 , Supplementary Table 4, and Supplementary Figs. 1-3 ) 7 , 8 . European ancestry sex-specific analyses identified nine additional loci, eight of which were new and significant in women but not in men (all P men >0.05; Table 1 , Supplementary Fig. 4 ). The addition of 14,371 individuals of non-European ancestry genotyped on Metabochip identified one additional locus in women (rs1534696, near SNX10, P women =2.1×10 −8 , P men =0.26, Table 1 , Supplementary Tables 1-3 ), with no evidence of heterogeneity across ancestries ( P het =0.86, Supplementary Note ). To evaluate sexual dimorphism, we compared sex-specific effect size estimates of the 49 WHRadjBMI lead SNPs. The effect estimates were significantly different ( P difference <0.05/49=0.001) at 20 SNPs, 19 of which showed larger effects in women ( Table 1 , Extended Data Fig. 2a ), similar to previous findings 7 , 8 . The only SNP that showed a larger effect in men mapped near GDF5 (rs224333, β men =0.036 and P= 9.0×10 −12 , β women =0.009 and P =0.074, P difference = 6.4 × 10 −5 ), a locus previously associated with height (rs6060369, r 2 =0.96 and rs143384, r 2 =0.96, 1000 Genomes Project CEU), though without significant differences between sexes 12 , 13 . Consistent with the larger number of loci identified in women, variance component analyses demonstrated a significantly larger heritability ( h 2 ) of WHRadjBMI in women than men in the Framingham Heart ( h 2 women =0.46, h 2 men =0.19, P difference =0.0037) and TwinGene studies ( h 2 women =0.56, h 2 men =0.32, P difference =0.001, Supplementary Table 5 , Extended Data Fig. 2b ). To identify multiple association signals within observed loci, we performed approximate conditional analyses of the sex-combined and sex-specific summary statistics using GCTA 14 ( Supplementary Note ). Multiple signals ( P< 5×10 −8 ) were identified at nine loci ( Extended Data Table 1 ). Fitting SNPs jointly identified different lead SNPs in the sex-specific and sex-combined analyses. For example, the MAP3K1-ANKRD55 locus showed near-independent (linkage disequilibrium (LD) r 2 <0.06) SNPs 54 kb apart that were significant only in women (rs3936510) or only in men (rs459193, Extended Data Table 1 , Supplementary Table 4 ). Other signals are more complex. The TBX15-WARS2 locus showed different but correlated lead SNPs in men and women near WARS2 ( r 2 =0.43), an independent signal near TBX15, and a distant independent signal near SPAG17 ( Fig. 1 ). At the HOXC gene cluster, conditional analyses identified independent ( r 2< 0.01) SNPs ~80 kb apart near HOXC12-HOXC13-HOTAIR and near HOXC4-HOXC6 ( Fig. 1 ). These results suggest that association signals mapping to the same locus might act on different underlying genes and may not be relevant to the same sex. We assessed the aggregate effects of the primary association signals at the 49 WHRadjBMI loci by calculating sex-combined and sex-specific risk based on genotypes of the lead SNPs. In a linear regression model, the risk scores were associated with WHRadjBMI, with a stronger effect in women than in men (overall effect per allele β=0.001, P =6.7×10 −4 , women β=0.002, P =1.0×10 −11 , men β=7.0×10 −4 , P =0.02, Extended Data Fig. 3 , Supplementary Note ). The 49 SNPs explained 1.4% of the variance in WHRadjBMI overall, and more in women (2.4%) than in men (0.8%) ( Supplementary Table 6 ). Compared to the 16 previously reported loci 7 , 8 , the new loci almost doubled the explained variance in women and tripled that in men. We further estimated that the sex-combined variance explained by all HapMap SNPs 15 (h 2 G ) is 12.1% (SE=2.9%). At 17 loci with high-density coverage on the Metabochip 11 , we used association summary statistics to define credible sets of SNPs with a high probability of containing a likely functional variant. The 99% credible sets at seven loci spanned <20 kb, and at HOXC13 included only a single noncoding SNP ( Supplementary Table 7, Supplementary Fig. 5 ). Imputation up to higher density reference panels will provide greater coverage and may have more potential to localize functional variants. Given the epidemiological correlations between central obesity and other anthropometric and cardiometabolic measures and diseases, we evaluated lead WHRadjBMI variants in association data from GWAS consortia for 22 traits. Seventeen of the 49 variants were associated ( P <5×10 −8 ) with at least one of the traits: high-density lipoprotein cholesterol (HDL-C; n =7 SNPs), triglycerides (TG; n= 5), low-density lipoprotein cholesterol (LDL-C; n= 2), adiponectin adjusted for BMI ( n= 3), fasting insulin adjusted for BMI ( n= 2), T2D ( n= 1), and height ( n= 7) ( Supplementary Tables 8-9 ). WHRadjBMI SNPs also showed enrichment for directional consistency among nominally significant ( P <0.05) associations with these traits and also with fasting and 2-hour glucose, diastolic and systolic blood pressure (DBP, SBP), BMI and coronary artery disease (CAD) ( P binomial <0.05/23=0.0022, Extended Data Table 2 ); these results were generally supported by meta-regression analysis of the regression coefficient-estimates ( Supplementary Table 10 ). Furthermore, our WHRadjBMI loci overlap with associations reported in the NHGRI GWAS Catalog ( Table 2 , Supplementary Table 11 ) 16 , the strongest of which is the locus near LEKR1, which is associated ( P =2.0×10 −35 ) with birthweight 17 . Unsupervised hierarchical clustering of the corresponding matrix of association Z -scores showed three major clusters characterized by patterns of anthropometric and metabolic traits ( Extended Data Fig. 4 ). These data extend knowledge about genetic links between WHRadjBMI and insulin resistance-related traits; whether this reflects underlying causal relations between WHRadjBMI and these traits, or pleiotropic loci, cannot be inferred from our data. We next examined variants in LD with the WHRadjBMI lead SNPs ( r 2 >0.7) for predicted effects on protein sequence, copy number, and cis -regulatory effects on expression ( Table 2 , Supplementary Tables 12-15, Supplementary Note ). At 11 of the new loci, lead WHRadjBMI SNPs were in LD with cis -expression quantitative trait loci (eQTLs) for transcripts in subcutaneous adipose tissue, omental adipose tissue, liver, or blood cell types ( Table 2 , Supplementary Table 15 ). No additional sex-specific eQTLs were identified, perhaps reflecting limited power ( Supplementary Table 16 ) . At the 11 WHRadjBMI loci harboring eQTLs, we compared the location of the candidate variants to regions of open chromatin (DNase I hypersensitivity and formaldehyde-assisted isolation of regulatory elements [FAIRE]) and histone modification enrichment (H3K4me1, H3K4me2, H3K4me3, H3K27ac, and H3K9ac) in adipose, liver, skeletal muscle, bone, brain, blood, and pancreatic islet tissues or cell lines ( Supplementary Table 17 ). At seven of these 11 loci, at least one variant was located in a putative regulatory element in two or more datasets from the same tissue as the eQTL, suggesting that these elements may influence transcriptional activity ( Supplementary Table 18 ). For example, at LEKR1 , five variants in LD with the WHRadjBMI lead SNP are located in a 1.1 kb region with evidence of enhancer activity (H3K4me1 and H3K27ac) in adipose tissue ( Extended Data Fig. 5a ). We also examined whether any variants overlapped with open chromatin or histone modifications from only one of the tested tissues, possibly reflecting tissue-specific regulatory elements ( Supplementary Table 18 ). For example, five variants in a 2.2 kb region, located 77 kb upstream from a CALCRL transcription start site, overlapped with peaks in at least five datasets in endothelial cells ( Extended Data Fig. 5b ), suggesting that one or more of these variants may influence transcriptional activity. CALCRL , which is expressed in endothelial cells, is required for lipid absorption in the small intestine, and influences body weight in mice 18 . Other variants located in tissue-specific regulatory elements were detected at NMU for endothelial cells, at KLF13 and MEIS1 for liver, and at GORAB and MSC for bone ( Supplementary Table 18 ). To identify potential functional connections between genes mapping to the 49 WHRadjBMI loci, we used three approaches ( Supplementary Note ). A survey of literature using GRAIL 19 identified 15 genes with nominal significance ( P< 0.05) for potential functional connectivity ( Table 2 , Supplementary Table 19 ). The predefined gene set relationships across loci identified using MAGENTA 20 highlighted signaling pathways involving vascular endothelial growth factor (VEGF), phosphatase and tensin (PTEN) homolog, the insulin receptor, and peroxisome proliferator-activated receptors ( Supplementary Table 20 ). VEGF signaling plays a central, complex role in angiogenesis, insulin resistance, and obesity 21 , and PTEN signaling promotes insulin resistance 22 . Analyses using DEPICT 23 facilitated prioritization of genes at associated loci, analyses of tissue specificity, and enrichment of reconstituted gene sets through integration of association results with expression data, protein-protein interactions, phenotypic data from gene knockout studies in mice, and predefined gene sets. DEPICT identified at least one prioritized gene (false discovery rate (FDR)<5%) at nine loci ( Table 2 , Supplementary Table 21 ) and identified 234 reconstituted gene sets (161 after pruning of overlapping gene sets) enriched for genes at WHRadjBMI loci. Among these we highlight biologically plausible gene sets suggesting roles in body fat regulation (including adiponectin signaling, insulin sensitivity, and regulation of glucose levels), skeletal growth, transcriptional regulation, and development ( Fig. 2 , Supplementary Table 22 ). We also note gene sets that are specific for abundance or development of metabolically active tissues including adipose, heart, liver, and muscle. Specific genes at the loci were significantly enriched (FDR<5%) for expression in adipocyte-related tissues, including abdominal subcutaneous fat ( Fig. 2 , Supplementary Table 23 ). Together, these analyses identified processes related to insulin and adipose biology and highlight mesenchymal tissues, especially adipose tissue, as important to WHRadjBMI. We also tested variants at the 49 WHRadjBMI loci for overlap with elements from 60 selected regulatory datasets from the ENCODE 24 and Epigenomic RoadMap 25 data and found evidence of enrichment in 12 datasets ( P< 0.05/60=8.3×10 −4 , Extended Data Table 3 ). The strongest enrichments were detected for datasets typically attributed to enhancer activity (H3K4me1 and H3K27ac) in adipose, muscle, endothelial cells, and bone, suggesting that variants may regulate transcription in these tissues. These analyses point to mechanisms linking WHRadjBMI loci to regulation of gene expression in tissues highly relevant for adipocyte metabolism and insulin resistance. We also reviewed functions of candidate genes located near new and previously established WHRadjBMI loci 7 , 8 , identifying genes involved in adipogenesis, angiogenesis, and transcriptional regulation ( Table 2 , literature review in the Supplementary Note ). Adipogenesis candidate genes include CEBPA, PPARG, BMP2, HOXC/miR196, SPRY1, TBX15, and PEMT . Of these, CEBPA and PPARG are essential for white adipose tissue differentiation 26 , BMP2 induces differentiation of mesenchymal stem cells toward adipogenesis or osteogenesis 27 , and HOXC8 is a repressor of brown adipogenesis in mice that is regulated by miR-196a 28 , also located within the HOXC region ( Fig. 1 ). Angiogenesis genes may influence expansion and loss of adipose tissue 29 ; they include VEGFA , VEGFB , RSPO3 , STAB1 , WARS2 , PLXND1 , MEIS1 , FGF2 , SMAD6 , and CALCRL . VEGFB is involved in endothelial targeting of lipids to peripheral tissues 30 , and PLXND1 limits blood vessel branching, antagonizes VEGF, and affects adipose inflammation 31 , 32 . Transcriptional regulators at WHRadjBMI loci include CEBPA , PPARG , MSC , SMAD6 , HOXA , HOXC , ZBTB7B , JUND , KLF13 , MEIS1 , RFX7 , NKX2-6 , and HMGA1 . Other candidate genes include NMU, FGFR4 , and HMGA1 , for which mice deficient for the corresponding genes exhibit obesity, glucose intolerance, and/or insulin resistance 33 - 35 . To determine whether the WHRadjBMI variants exert their effects primarily through WC or HIP and to identify loci that are not reported for WHRadjBMI, BMI, or height 36 , 37 , we performed association analyses for five additional traits: WCadjBMI, HIPadjBMI, WHR, WC, and HIP. Based on phenotypic data alone, WC and HIP are highly correlated with BMI ( r =0.59 - 0.92), and WHR is highly correlated with WHRadjBMI ( r =0.82 - 0.95), while WCadjBMI and HIPadjBMI are moderately correlated with height ( r =0.24 - 0.63, Supplementary Table 24 ). In contrast to WHRadjBMI, which has almost no genetic correlation (see Methods) with height ( r G <0.04, Extended Data Fig. 2c ), WCadjBMI ( r G =0.42) and HIPadjBMI ( r G = 0.82) have moderate genetic correlations with height. These data suggest that some, but not all, WCadjBMI and HIPadjBMI loci would be associated with height. Across all meta-analyses, we identified an additional 19 loci associated with one of the five traits ( P< 5×10 −8 ), nine of which showed significantly larger effects ( P difference <0.05/19=0.003) in one sex than in the other ( Table 3 , Supplementary Figs. 1-4, Supplementary Table 25 ). Three of four new loci with larger effects in women were associated with HIPadjBMI and three of five new loci with larger effects in men were associated with WCadjBMI. Most of the 19 loci showed some evidence of association with WHRadjBMI in sex-combined or sex-specific analyses, but four loci showed no association ( P >0.01) with WHRadjBMI, BMI, or height ( Supplementary Tables 8, 26 ). We next asked whether the genes and pathways influencing these five traits are shared with WHRadjBMI or are distinct. Candidate genes were identified based on association with other traits, eQTLs, GRAIL, and literature review ( Extended Data Table 4 , Supplementary Tables 8, 11-13, 15-16, 19 ). Candidate variants identified based on LD ( r 2 >0.7) included coding variants in NTAN1 and HMGXB4 , and six loci showed significant eQTLs in subcutaneous adipose tissue. Based on the literature, several candidate genes are involved in adipogenesis and insulin resistance. For example, delayed induction of preadipocyte transcription factor ZNF423 in fibroblasts results in delayed adipogenesis 38 , and NLRP3 is part of inflammasome and pro-inflammatory T-cell populations in adipose tissue that contribute to inflammation and insulin resistance 39 . GRAIL analyses identified connections that partially overlap with those identified for WHRadjBMI ( Supplementary Table 19 ). Taken together, the additional loci appear to function in processes similar to the WHRadjBMI loci. The identification of loci that are more strongly associated with WCadjBMI or HIPadjBMI than the other anthropometric traits suggests that the additional traits characterize aspects of central obesity and fat distribution that are not captured by WHRadjBMI or BMI alone.

These meta-analyses of GWAS and Metabochip data in up to 224,459 individuals identified additional loci associated with waist and hip circumference measures and help elucidate the role of common genetic variation in body fat distribution that is distinct from BMI and height. Our results emphasize the strong sexual dimorphism in the genetic regulation of fat distribution traits, a characteristic not observed for overall obesity as assessed by BMI 36 . Differences in body fat distribution between the sexes emerge in childhood, become more apparent during puberty 40 , and change with menopause, generally attributed to the influence of sex hormones 41 , 42 . At loci with stronger effects in one sex than the other, these hormones may interact with transcription factors to regulate gene activity. Annotation of the loci emphasized the role for mesenchymally-derived tissues, especially adipose tissue, in fat distribution and central obesity. The development and regulation of adipose tissue deposition is closely associated with angiogenesis 29 , a process highlighted by candidate genes at several WHRadjBMI loci. These tissues are implicated in insulin resistance, consistent with the enrichment of shared GWAS signals with lipids, T2D, and glycemic traits. The identification of skeletal growth processes suggests that the underlying genes affect early development and/or differentiation of adipocytes from mesenchymal stem cells. In contrast, BMI has a significant neuronal component, involving processes such as appetite regulation 36 . Our results provide a foundation for future biological research in the regulation of body fat distribution and its connections with cardiometabolic traits, and offer potential target mechanisms for interventions in the risks associated with abdominal fat accumulation.

Data (dashed lines) and analyses (solid lines) related to the genome-wide association study (GWAS) cohorts for waist-hip ratio adjusted for body mass index (WHRadjBMI) are colored red and those related to the Metabochip (MC) cohorts are colored blue. The two genomic control (λ GC ) corrections (within-study and among-studies) performed on associations from each dataset are represented by gray-outlined circles. The λ GC corrections for the GWAS meta-analysis were based on all SNPs and the λ GC corrections for the Metabochip meta-analysis were based on a null set of 4,319 SNPs previously associated with QT interval. The joint meta-analysis of the GWAS and MC datasets is colored purple. All SNP counts reflect a sample size filter of N ≥ 50,000 subjects. Additional WHRadjBMI meta-analyses included Metabochip data from up to 14,371 subjects of East Asian, South Asian, or African American ancestry from eight cohorts. Counts for the meta-analyses of waist circumference (WC), hip circumference (HIP), and their BMI-adjusted counterparts (WCadjBMI and HIPadjBMI) differ from those of WHRadjBMI because some cohorts only had phenotype data available for one type of body circumference measurement (see Supplementary Table 2 ). a, Figure showing effect beta estimates for the 20 WHRadjBMI SNPs showing significant evidence of sexual dimorphism. Sex-specific effect betas and 95% confidence intervals for SNPs associated with waist-hip ratio adjusted for body mass index (WHRadjBMI) are shown as red circles and blue squares for women and men, respectively. The SNPs are classified into three categories: ( i ) those showing a female-specific effect (“Women SSE”), namely a significant effect in women and no effect in men ( P women < 5 × 10 −8 , P men ≥ 0.05), ( ii ) those showing a pronounced female effect (“Women CED”), namely a significant effect in women and a less significant but directionally consistent effect in men ( P women < 5 × 10 −8 , 5 × 10 −8 < P men ≤ 0.05); and ( iii ) those showing a male-specific effect (“Men SSE”), namely a significant effect in men and no effect in women ( P men < 5 × 10 −8 , P women ≥ 0.05). Within each of the three categories, the loci were sorted by increasing P value of sex-based heterogeneity in the effect betas. b, Figure showing standardized sex-specific phenotypic variance components for six waist-related traits. Values are shown in men (M) and women (W) from the Swedish Twin Registry ( N = 11,875). The ACE models are decomposed into additive genetic components (A) shown in black, common environmental components (C) in gray, and non-shared environmental components (E) in white. Components are shown for waist circumference (WC), hip circumference (HIP), waist-hip ratio (WHR), and their body mass index (BMI)-adjusted counterparts (WCadjBMI, HIPadjBMI, and WHRadjBMI). When the A component is different in men and women with P < 0.05 for a given trait, its name is marked with an asterisk. c, Table showing genetic correlations of waist-related traits with height, adjusted for age and body mass index. Genetic correlations of three traits with height were based on variance component models in the Framingham Heart Study and TWINGENE study (see Online Methods). WCadjBMI, waist circumference adjusted for BMI; WHRadjBMI, waist-hip ratio adjusted for BMI; HIPadjBMI, hip circumference adjusted for BMI. a, All subjects ( N = 3,440, P trend = 6.7 × 10 −4 ). b, Only women ( N = 1,750, P trend = 1.0 × 10 −11 ). c, Only men ( N = 1,690, P trend = 0.02). Each genetic risk score (GRS) illustrates the joint effect of the waist-hip ratio adjusted for body mass index (WHRadjBMI)-increasing alleles of the 49 identified variants from Table 1 weighted by the relative effect sizes from the applicable sex-combined or sex-specific meta-analysis. The mean WHRadjBMI residual and 95% confidence interval is plotted for each GRS category (red dots). The histograms show each GRS is normally distributed in KORA (gray bars). The matrix of Z -scores representing the set of associations was scaled by row (locus name) and by column (trait) to range from −3 to 3. Negative values (blue) indicate that the waist-hip ratio adjusted for body mass index (WHRadjBMI)-increasing allele was associated with decreased values of the trait and positive values (red) indicate that this allele was associated with increased values of the trait. Dendrograms indicating the clustering relationships are shown to the left and above the heat map. The WHRadjBMI-increasing alleles at the 49 lead SNPs segregate into three major clusters comprised of alleles that associate with: 1) larger waist circumference adjusted for BMI (WCadjBMI) and smaller hip circumference adjusted for BMI (HIPadjBMI) ( n = 30 SNPs); 2) taller stature and larger WCadjBMI ( n = 8 SNPs); and 3) shorter stature and smaller HIPadjBMI ( n = 11 SNPs). The three visually identified SNP clusters could be statistically distinguished with >90% confidence. Alleles of the first cluster were predominantly associated with lower high density lipoprotein (HDL) cholesterol and with higher triglycerides and fasting insulin adjusted for BMI (FIadjBMI). eGFRcrea, estimated glomerular filtration rate based on creatinine; LDL cholesterol, low-density lipoprotein cholesterol; UACR, urine albumin-to-creatinine ratio; BMD, bone mineral density. a, Five variants associated with waist-hip ratio adjusted for body mass index (WHRadjBMI) and located ~77 kb upstream of the first CALCRL transcription start site overlap regions with genomic evidence of regulatory activity in endothelial cells. b, Five WHRadjBMI variants, including rs8817452, in a 1.1 kb region (box) ~250 kb downstream of the first LEKR1 transcription start site overlap evidence of active enhancer activity in adipose nuclei. Signal enrichment tracks are from the ENCODE Integrative Analysis and the Roadmap Epigenomics track hubs on the UCSC Genome Browser. Transcripts are from the GENCODE basic annotation. P values and β coefficients for the association with WHRadjBMI from the joint model in the approximate conditional analysis of combined GWAS and Metabochip studies. SNPs selected by conditional analyses as independently associated with WHRadjBMI in a meta-analysis (sex-combined, women- or men-specific) have their respective summary statistics for these analyses marked in black and bold. SNPs not selected by a particular conditional analysis as independently associated are marked in gray and show the association analysis results for the SNP conditioned on the locus SNPs selected by GCTA. Sample sizes are from the unconditioned meta-analysis. Locus and lead SNPs are defined by Table 1 The effect allele is the WHRadjBMI-increasing allele in the sex-combined analysis. Test for sex difference in conditional analysis based on the effect correlation estimate from primary analyses; values significant at the table-wise Bonferroni threshold of 0.05 / 25=2×10 −3 are marked in bold. SNPs selected by conditional analysis in the sex-combined analysis; proxies were selected by joint conditional analysis in the women- and/or men-specific analyses. SNP not present in the sex-specific meta-analyses due to sample size filter requiring N ≥ 50,000; sample size from GCTA. At NFE2L3-SNX10 , different lead SNPs were identified in the European and all-ancestry analyses but LD is reported with respect to rs10245353. Chr, chromosome; EA, effect allele; EAF, effect allele frequency. The 49 waist-hip ratio adjusted for body mass index (WHRadjBMI) SNPs were tested for association with other traits by GWAS meta-analyses performed by other groups (see Online Methods). The maximum sample size available is shown overall or separately for 61 cases/controls. N indicates the number of the total SNPs for which the WHRadjBMI-increasing allele is associated with the trait in the concordant direction (increased levels, except for HDL-C, adiponectin, and BMI). One-sided binomial P values test whether this number is greater than expected by chance (null P = 0.5 and null P = 0.025, respectively). The tests do not account for correlation between WHRadjBMI and the tested traits. P values representing significant column-wise enrichment ( P < 0.05 / 23 tests) are marked in red and bold. Enrichment of waist-hip ratio adjusted for BMI (WHRadjBMI)-associated loci in regulatory elements from selected WHRadjBMI-relevant tissues. P values are derived using a sum of binomial distributions (see Methods). P values below a Bonferroni-corrected threshold for 60 tests of 8.3 × 10 −4 are indicated in bold font. The binomial based P values are similar to P values generated from 10,000 permutation tests. Dashes indicate that datasets were not available. Candidate genes for loci shown on Table 3 based on secondary analyses or literature review. Further details are provided in other Supplementary Tables and the Supplementary Note . Loci are shown in order of chromosome and position. Gene transcript levels associated with SNP genotype (expression QTL) in the indicated tissue(s). Genes in pathways identified as enriched by GRAIL analysis. Strongest candidate genes identified based on manual literature review. Traits associated at P 0.7), and the genes(s) named in those reports. Nonsynonymous variants (nsSNPs) and copy number variants (CNVs) with tag SNPs in high LD with index SNP based on a 1000 Genomes CEU reference panel. DEPICT analysis was not performed for loci associated with these traits. Chr, Chromosome; WCadjBMI, waist circumference adjusted for body mass index (BMI); HIPadjBMI, hip circumference adjusted for BMI; WHR, waist-to-hip ratio.