Association analyses of 249,796 individuals reveal eighteen new loci associated with body mass index

Abstract

Obesity is globally prevalent and highly heritable, but the underlying genetic factors remain largely elusive. To identify genetic loci for obesity-susceptibility, we examined associations between body mass index (BMI) and ~2.8 million SNPs in up to 123,865 individuals, with targeted follow-up of 42 SNPs in up to 125,931 additional individuals. We confirmed 14 known obesity-susceptibility loci and identified 18 new loci associated with BMI (P<5×10⁻⁸), one of which includes a copy number variant near GPRC5B. Some loci (MC4R, POMC, SH2B1, BDNF) map near key hypothalamic regulators of energy balance, and one is near GIPR, an incretin receptor. Furthermore, genes in other newly-associated loci may provide novel insights into human body weight regulation.

Obesity is a major and increasingly prevalent risk factor for multiple disorders, including type 2 diabetes and cardiovascular disease^1,2. While lifestyle changes have driven its prevalence to epidemic proportions, heritability studies provide evidence for a substantial genetic contribution (h²~40–70%) to obesity risk^3,4. BMI is an inexpensive, non-invasive measure of obesity that predicts the risk of related complications⁵. Identifying genetic determinants of BMI could lead to a better understanding of the biological basis of obesity.

Genome-wide association (GWA) studies of BMI have previously identified ten loci with genome-wide significant (P < 5×10⁻⁸) associations in or near FTO, MC4R, TMEM18, GNPDA2, BDNF, NEGR1, SH2B1, ETV5, MTCH2, and KCTD15^6–10. Many of these genes are expressed or known to act in the central nervous system, highlighting a likely neuronal component to the predisposition to obesity⁹. This pattern is consistent with results in animal models and studies of monogenic human obesity, where neuronal genes, particularly those expressed in the hypothalamus and involved in regulation of appetite or energy balance, are known to play a major role in susceptibility to obesity^11–13.

The ten previously identified loci account for only a small fraction of the variation in BMI. Furthermore, power calculations based on the effect sizes of established variants have suggested that increasing the sample size would likely lead to the discovery of additional variants⁹. To identify more loci associated with BMI, we expanded the GIANT (Genetic Investigation of ANtropometric Traits) consortium GWA meta-analysis to include a total of 249,769 individuals of European ancestry.

Results

Stage 1 GWA studies identify novel loci associated with BMI

We first conducted a meta-analysis of GWA studies of BMI and ~2.8 million imputed or genotyped SNPs using data from 46 studies including up to 123,865 individuals (Online Methods, Supplementary Fig. 1 and Supplementary Note). This stage 1 analysis revealed 19 loci associated with BMI at P < 5×10⁻⁸ (Table 1, Fig. 1a and Supplementary Table 1). These 19 loci included all ten loci from previous GWA studies of BMI^6–10, two loci previously associated with body weight¹⁰ (FAIM2 and SEC16B) and one locus previously associated with waist circumference¹⁴ (near TFAP2B). The remaining six loci, near GPRC5B, MAP2K5/LBXCOR1, TNNI3K, LRRN6C, FLJ35779/HMGCR, and PRKD1, have not previously been associated with BMI or other obesity-related traits.

Table 1.

Stage 1 and stage 2 results of the 32 SNPs that were associated with BMI at genome-wide significance (P < 5.10⁻⁸) levels.

SNP	Nearest gene	Other nearby genes*	Chr	Position** (bp)	Alleles**		Frequency effect allele (%)	Per allele change in BMI beta (se)***	Explained variance (%)	Stage 1 P-value	Stage 2 P-value	Stage 1 + 2
					Effect	Other						n	P-value
Previous BMI loci
rs1558902	FTO		16	52,361,075	a	t	42%	0.39 (0.02)	0.34%	2.05E-62	1.007E-60	192,344	4.8E-120
rs2867125	TMEM18		2	612,827	c	t	83%	0.31 (0.03)	0.15%	2.42E-22	4.42E-30	197,806	2.77E-49
rs571312	MC4R B		18	55,990,749	a	c	24%	0.23 (0.03)	0.10%	1.82E-22	3.19E-21	203,600	6.43E-42
rs10938397	GNPDA2		4	44,877,284	g	a	43%	0.18 (0.02)	0.08%	4.35E-17	1.45E-15	197,008	3.78E-31
rs10767664	BDNFB,M		11	27,682,562	a	t	78%	0.19 (0.03)	0.07%	5.53E-13	1.17E-14	204,158	4.69E-26
rs2815752	NEGR1C,Q		1	72,585,028	a	g	61%	0.13 (0.02)	0.04%	1.17E-14	2.29E-09	198,380	1.61E-22
rs7359397	SH2B1Q,B,M	APOB48RQ,M, SULT1A2Q,M, AC138894.2M, ATXN2LM, TUFMQ	16	28,793,160	t	c	40%	0.15 (0.02)	0.05%	1.75E-10	7.89E-12	204,309	1.88E-20
rs9816226	ETV5		3	187,317,193	t	a	82%	0.14 (0.03)	0.03%	7.61E-14	1.15E-06	196,221	1.69E-18
rs3817334	MTCH2Q,M	NDUFS3 Q , CUGBP1 Q	11	47,607,569	t	c	41%	0.06 (0.02)	0.01%	4.79E-11	1.10E-03	191,943	1.59E-12
rs29941	KCTD15		19	39,001,372	g	a	67%	0.06 (0.02)	0.00%	1.31E-09	2.40E-02	192,872	3.01E-09
Previous waist & weight loci
rs543874	SEC16B		1	176,156,103	g	a	19%	0.22 (0.03)	0.07%	1.66E-13	2.41E-11	179,414	3.56E-23
rs987237	TFAP2B		6	50,911,009	g	a	18%	0.13 (0.03)	0.03%	5.97E-16	2.40E-06	195,776	2.90E-20
rs7138803	FAIM2		12	48,533,735	a	g	38%	0.12 (0.02)	0.04%	3.96E-11	7.82E-08	200,064	1.82E-17
rs10150332	NRXN3		14	79,006,717	c	t	21%	0.13 (0.03)	0.02%	2.03E-07	2.86E-05	183,022	2.75E-11
Newly identified BMI loci
rs713586	RBJ	ADCY3Q, M, POMCQ,B	2	25,011,512	c	t	47%	0.14 (0.02)	0.06%	1.80E-07	1.44E-16	230,748	6.17E-22
rs12444979	GPRC5BC,Q	IQCK Q	16	19,841,101	c	t	87%	0.17 (0.03)	0.04%	4.20E-11	8.13E-12	239,715	2.91E-21
rs2241423	MAP2K5	LBXCOR1 M	15	65,873,892	g	a	78%	0.13 (0.02)	0.03%	1.15E-10	1.59E-09	227,950	1.19E-18
rs2287019	QPCTL	GIPRB,M	19	50,894,012	c	t	80%	0.15 (0.03)	0.04%	3.18E-07	1.40E-10	194,564	1.88E-16
rs1514175	TNNI3K		1	74,764,232	a	g	43%	0.07 (0.02)	0.02%	1.36E-09	7.04E-06	227,900	8.16E-14
rs13107325	SLC39A8Q,M		4	103,407,732	t	c	7%	0.19 (0.04)	0.03%	1.37E-07	1.93E-07	245,378	1.50E-13
rs2112347	FLJ35779 M	HMGCR B	5	75,050,998	t	g	63%	0.10 (0.02)	0.02%	4.76E-08	8.29E-07	231,729	2.17E-13
rs10968576	LRRN6C		9	28,404,339	g	a	31%	0.11 (0.02)	0.02%	1.88E-08	3.19E-06	216,916	2.65E-13
rs3810291	TMEM160 Q	ZC3H4 Q	19	52,260,843	a	g	67%	0.09 (0.02)	0.02%	1.04E-07	1.59E-06	233,512	1.64E-12
rs887912	FANCL		2	59,156,381	t	c	29%	0.10 (0.02)	0.03%	2.69E-06	1.72E-07	242,807	1.79E-12
rs13078807	CADM2		3	85,966,840	g	a	20%	0.10 (0.02)	0.02%	9.81E-08	5.32E-05	237,404	3.94E-11
rs11847697	PRKD1		14	29,584,863	t	c	4%	0.17 (0.05)	0.01%	1.11E-08	2.25E-04	241,667	5.76E-11
rs2890652	LRP1B		2	142,676,401	c	t	18%	0.09 (0.03)	0.02%	2.38E-07	9.47E-05	209,068	1.35E-10
rs1555543	PTBP2		1	96,717,385	c	a	59%	0.06 (0.02)	0.01%	7.65E-07	4.48E-05	243,013	3.68E-10
rs4771122	MTIF3	GTF3A Q	13	26,918,180	g	a	24%	0.09 (0.03)	0.02%	1.20E-07	8.24E-04	198,577	9.48E-10
rs4836133	ZNF608		5	124,360,002	a	c	48%	0.07 (0.02)	0.01%	7.04E-07	1.88E-04	241,999	1.97E-09
rs4929949	RPL27A	TUB B	11	8,561,169	c	t	52%	0.06 (0.02)	0.01%	7.57E-08	1.00E-03	249,791	2.80E-09
rs206936	NUDT3	HMGA1 B	6	34,410,847	g	a	21%	0.06 (0.02)	0.01%	2.81E-06	7.39E-04	249,777	3.02E-08

^*Genes within +/− 500 kb of the lead SNP

^**Positions according to Build 36 and allele coding based on the positive strand

^***Effect sizes in kg/m2 obtained from Stage 2 cohorts only

^QAssociation and eQTL data converge to affect gene expression

^BBiological candidate

^MBMI-associated variant is in strong LD (r2 ≥ 0.75) with a missense variant in the indicated gene

^CCNV

Figure 1. Genome-wide association results for the BMI meta-analysis.

(a) Manhattan plot showing the significance of association between all SNPs and BMI in the stage 1 meta-analysis, highlighting SNPs previously reported to show genome-wide significant association with BMI (blue), weight or waist circumference (green), and the 18 new regions described here (red). The 19 SNPs that reached genome-wide significance at Stage 1 (13 previously reported and 6 new) are listed in Table 1). (b) Quantile-quantile (Q-Q) plot of SNPs in stage 1 meta-analysis (black) and after removing any SNPs within 1 Mb of the 10 previously reported genome-wide significant hits for BMI (blue), after additionally excluding SNPs from the four loci for waist/weight (green) and after excluding SNPs from all 32 confirmed loci (red). The plot was abridged at the Y-axis (at P < 10⁻²⁰) to better visualise the excess of small P-values after excluding the 32 confirmed loci (Supplementary Fig. 3 shows full-scale Q-Q plot). The shaded region is the 95% concentration band. (c) Plot of effect size (in inverse normally transformed units (invBMI)) versus effect allele frequency of newly identified and previously identified BMI variants after stage 1 + stage 2 analysis; including the 10 previously identified BMI loci (blue), the four previously identified waist and weight loci (green) and the 18 newly identified BMI loci (blue). The dotted lines represent the minimum effect sizes that could be identified for a given effect-allele frequency with 80% (upper line), 50% (middle line), and 10% (lower line) power, assuming a sample size of 123,000 individuals and a α-level of 5×10⁻⁸.

Stage 2 follow-up leads to additional novel loci for BMI

To identify additional BMI-associated loci and to validate the loci that reached genome-wide significance in stage 1 analyses, we examined SNPs representing 42 independent loci (including the 19 genome-wide significant loci) with stage 1 P < 5×10⁻⁶. Variants were considered to be independent if the pair-wise linkage disequilibrium (LD; r²) was less than 0.1 and if they were separated by at least 1 Mb. In stage 2, we examined these 42 SNPs in up to 125,931 additional individuals (79,561 newly genotyped individuals from 16 different studies and 46,370 individuals from 18 additional studies for which GWA data were available; Table 1, Supplementary Note, and Online Methods). In a joint analysis of stage 1 and stage 2 results, 32 of the 42 SNPs reached P < 5×10⁻⁸. Even after excluding SNPs within these 32 confirmed BMI loci, we still observed an excess of small P-values compared to the distribution expected under the null hypothesis (Fig. 1b), suggesting that more BMI loci remain to be uncovered.

The 32 confirmed associations included all 19 loci with P < 5×10⁻⁸ at stage 1, 12 additional novel loci near RBJ/ADCY3/POMC, QPCTL/GIPR, SLC39A8, TMEM160, FANCL, CADM2, LRP1B, PTBP2, MTIF3/GTF3A, ZNF608, RPL27A/TUB, NUDT3/HMGA1, and one locus (NRXN3) previously associated with waist circumference¹⁵ (Table 1, Supplementary Table 1, Supplementary Fig. 1 and 2). In all, our study increased the number of loci robustly associated with BMI from 10 to 32. Four of the 22 new loci were previously associated with body weight¹⁰ or waist circumference^14,15, whereas 18 loci had not previously associated with any obesity-related trait in the general population. Whilst we confirmed all loci previously established by large-scale GWA studies for BMI^6–10 and waist circumference^14,15, four loci identified by GWA studies for early-onset or adult morbid obesity^16,17 [at NPC1 (rs1805081; P = 0.0025), MAF (rs1424233; P = 0.25), PTER (rs10508503; P = 0.64), and TNKS/MSRA (rs473034; P = 0.23)] showed limited or no evidence of association with BMI in our study.

As expected, the effect sizes of the 18 newly discovered loci are slightly smaller, for a given minor allele frequency, than those of the previously identified variants (Table 1 and Fig. 1c). The increased sample size also brought out more signals with low minor allele frequency. The BMI-increasing allele frequencies for the 18 newly identified variants ranged from 4% to 87%, covering more of the allele frequency spectrum than previous, smaller GWA studies of BMI (24%–83%)^9,10 (Table 1 and Fig. 1c).

We tested for evidence of non-additive (dominant or recessive) effects, SNP×SNP interaction effects and heterogeneity by sex or study among the 32 BMI-associated SNPs (Online Methods). We found no evidence for any such effects (P > 0.001, no significant results after correcting for multiple testing) (Supplementary Tables 1 and Supplementary Note).

Impact of 32 confirmed loci on BMI, obesity, body size, and other metabolic traits

Together, the 32 confirmed BMI loci explained 1.45% of the inter-individual variation in BMI of the stage 2 samples, with the FTO SNP accounting for the largest proportion of the variance (0.34%) (Table 1). To estimate the cumulative effect of the 32 variants on BMI, we constructed a genetic-susceptibility score that sums the number of BMI-increasing alleles weighted by the overall stage 2 effect sizes in the ARIC study (N = 8,120), one of our largest population-based studies (Online Methods). For each unit increase in the genetic-susceptibility score, approximately equivalent to one additional risk allele, BMI increased by 0.17 kg/m², equivalent to a 435–551 g gain in body weight in adults of 160–180 cm in height. The difference in average BMI between individuals with a high genetic-susceptibility score (≥38 BMI-increasing alleles, 1.5% (n=124) of the ARIC sample) and those with a low genetic-susceptibility score (≤21 BMI-increasing alleles, 2.2% (n=175) of the ARIC sample) was 2.73 kg/m², equivalent to a 6.99 to 8.85 kg body weight difference in adults 160–180 cm in height (Fig. 2a). Still, we note that the predictive value for obesity risk and BMI of the 32 variants combined was modest, although statistically significant (Fig. 2b, Supplementary Fig. 4). The area under the receiver operating characteristic (ROC) curve for prediction of risk of obesity (BMI ≥ 30 kg/m²) using age, age² and sex only was 0.515 (P = 0.023 compared to AUC of 0.50), which increased to 0.575 (P < 10⁻⁵) when also the 32 confirmed SNPs were included in the model (Fig. 2b). The area under the ROC for the 32 SNPs only was 0.574 (P < 10⁻⁵).

Figure 2. Combined impact of risk alleles on BMI/obesity.

(a) Combined effect of risk alleles on average BMI in the population-based Atherosclerosis Risk in Communities (ARIC) study (n = 8,120 individuals of European descent). For each individual, the number of “best guess” replicated (n = 32) risk alleles from imputed data (0,1,2) per SNP was weighted for their relative effect sizes estimated from the stage 2 data. Weighted risk alleles were summed for each individual and the overall individual sum was rounded to the nearest integer to represent the individual’s risk allele score (range 16–44). Along the x-axis, individuals in each risk allele category are shown (grouped ≤21 and ≥38 at the extremes), and the mean BMI (+/− SEM) is plotted (y axis on right), with the line representing the regression of the mean BMI values across the risk-allele scores. The histogram (y-axis on left) represents the number of individuals in each risk-score category. (b) The area under the ROC curve (AUC) of two different models predicting the risk of obesity (BMI = ≥30 kg/m²) in the n = 8,120 genotyped individuals of European descent in the ARIC Study. Model 1, represented by the solid line, includes age, age², and sex (AUC = 0.515, P = 0.023 for difference from AUC_null = 0.50). Model 2, represented by the dashed line, includes age, age², sex, and the n = 32 confirmed BMI SNPs (AUC = 0.0575, P < 10⁻⁵ for difference from AUC_null = 0.50). The difference between both AUCs is significant (P < 10⁻⁴).

All 32 confirmed BMI-increasing alleles showed directionally consistent effects on risk of being overweight (BMI ≥25 kg/m²) or obese (≥30 kg/m²) in stage 2 samples, with 30 of 32 variants achieving at least nominally significant associations. The BMI-increasing alleles increased the odds of overweight by 1.013 to 1.138-fold, and the odds for being obese by 1.016- to 1.203-fold (Supplementary Table 2). In addition, 30 of the 32 loci also showed directionally consistent effects on the risk of extreme and early-onset obesity in a meta-analysis of seven case-control studies of adults and children (binomial sign test P = 1.3×10⁻⁷) (Supplementary Table 3). The BMI-increasing allele observed in adults also increased the BMI in children and adolescents with directionally consistent effects observed for 23 of the 32 SNPs (binomial sign test P = 0.01). Furthermore, in family-based studies, the BMI-increasing allele was over-transmitted to the obese offspring for 24 of the 32 SNPs (binomial sign test P = 0.004) (Supplementary Table 3). As these studies in extreme obesity cases, children and families were relatively small (N_range = 354 − 15,251) compared to the overall meta-analyses, their power was likely insufficient to confirm association for all 32 loci. Nevertheless, these results show that the effects are unlikely to reflect population stratification and that they extend to BMI differences throughout the life course.

All BMI-increasing alleles were associated with increased body weight, as expected from the correlation between BMI and body weight (Supplementary Table 2). To confirm an effect of the loci on adiposity rather than general body size, we tested association with body fat percentage, which was available in a subset of the stage 2 replication samples (n = 5,359–28,425) (Supplementary Table 2). The BMI-increasing allele showed directionally consistent effects on body fat percentage at 31 of the 32 confirmed loci (binomial sign test P = 1.54×10⁻⁸) (Supplementary Table 2).

We also examined the association of the BMI loci with metabolic traits (type 2 diabetes¹⁸, fasting glucose, fasting insulin, indices of beta-cell function (HOMA-B) and insulin resistance (HOMA-IR)¹⁹, and blood lipid levels²⁰) and with height (Supplementary Tables 2 and 4). Although many nominal associations are expected because of known correlations between BMI and most of these traits and because of overlap in samples, several associations stand out as possible examples of pleiotropic effects of the BMI-associated variants. Particularly interesting is the variant in the GIPR locus where the BMI-increasing allele is also associated with increased fasting glucose levels and lower 2-hour glucose levels (Supplementary Table 4)^19,21. The direction of the effect is opposite to what would be expected due to the correlation between obesity and glucose intolerance, but is consistent with the suggested roles of GIPR in glucose and energy metabolism (see below)²². Three loci show strong associations (P < 10⁻⁴) with height (MC4R, RBJ/ADCY3/POMC and MTCH2/NDUFS3). Because BMI is weakly correlated with height (and indeed, the BMI-associated variants as a group show no consistent effect on height), these associations are also suggestive of pleiotropy. Interestingly, analogous to the effects of severe mutations in POMC and MC4R on height and weight^23,24, the BMI-increasing alleles of the variants near these genes were associated with decreased (POMC) and increased (MC4R) height, respectively (Supplementary Table 2).

Potential functional roles and pathways analyses

Although associated variants typically implicate genomic regions rather than individual genes, we note that some of the 32 loci include candidate genes with established connections to obesity. Several of the 10 previously identified loci are located in or near genes that encode neuronal regulators of appetite or energy balance, including MC4R^12,25, BDNF²⁶, and SH2B1^11,27. Each of these genes has been tied to obesity, not only in animal models, but also by rare human variants that disrupt each of these genes and lead to severe obesity^24,28,29. Using the automated literature search programme, Snipper (Online Methods), we identified various genes within the novel loci with potential biological links to obesity-susceptibility (Supplementary Note). Among the novel loci, the location of rs713586 near POMC provides further support for a role of neuroendocrine circuits that regulate energy balance in susceptibility to obesity. POMC encodes several polypeptides including α-MSH, a ligand of the MC4R gene product³⁰, and rare mutations in POMC also cause human obesity^23,29,31.

In contrast, the locus near GIPR, which encodes a receptor of gastric inhibitory polypeptide (GIP), suggests a role for peripheral biology in obesity. GIP, which is expressed in the K cell of the duodenum and intestine, is an incretin hormone that mediates incremental insulin secretion in response to oral intake of glucose. The variant associated with BMI is in strong LD (r² = 0.83) with a missense SNP in GIPR (rs1800437, Glu354Gln) that has recently been shown to influence the glucose and insulin response to an oral glucose challenge ²¹. Although no human phenotype is known to be caused by mutations in GIPR, mice with disruption of Gipr are resistant to diet-induced obesity³². The association of a variant in GIPR with BMI suggests that there may be a link between incretins/insulin secretion and body weight regulation in humans as well.

To systematically identify biological connections among the genes located near the 32 confirmed SNPs, and to potentially identify new pathways associated with BMI, we performed pathway-based analyses using MAGENTA³³. Specifically, we tested for enrichment of BMI genetic associations in biological processes or molecular functions that contain at least one gene from the 32 confirmed BMI loci (Online Methods). Using annotations from the KEGG, Ingenuity, PANTHER, and Gene Ontology databases, we found evidence of enrichment for pathways involved in the platelet-derived growth factor (PDGF) signaling (PANTHER, P = 0.0008, FDR = 0.0061), translation elongation (PANTHER, P = 0.0008, FDR = 0.0066), hormone or nuclear hormone receptor binding (Gene Ontology, P < 0.0005, FDR < 0.0085), homeobox transcription (PANTHER, P = 0.0001, FDR = 0.011), regulation of cellular metabolism (Gene Ontology, P = 0.0002, FDR = 0.031), neurogenesis and neuron differentiation (Gene Ontology, P < 0.0002, FDR < 0.034), protein phosphorylation (PANTHER, P = 0.0001, FDR = 0.045) and numerous other pathways related to growth, metabolism, immune and neuronal processes (Gene Ontology, P < 0.002, FDR < 0.046) (Supplementary Table 5).

Identifying possible functional variants

We used data from the 1000 Genomes Project and the HapMap Consortium to explore whether the 32 confirmed BMI SNPs were in LD (r² ≥ 0.75) with common missense SNPs or copy number variants (CNVs) (Online Methods). Non-synonymous variants in LD with our signals were present in the BDNF, SLC39A8, FLJ35779/HMGCR, QPCTL/GIPR, MTCH2, ADCY3, and LBXCOR1 genes. In addition, the rs7359397 signal was in LD with coding variants in several genes including SH2B1, ATNX2L, APOB48R, SULT1A2, and AC138894.2 (Table 1, Fig. 3, Supplementary Table 6 and Supplementary Fig. 2). Furthermore, two SNPs tagged common CNVs. The first CNV was previously identified and is a 45-kb deletion near NEGR1⁹. The second CNV is a 21-kb deletion that lies 50kb upstream of GPRC5B; the deletion allele is tagged by the T-allele of rs12444979 (r² = 1) (Fig. 3). Although the correlations with potentially functional variants does not prove that these variants are indeed causal, these provide first clues as to which genes and variants at these loci might be prioritized for fine-mapping and functional follow-up.

Figure 3. Regional plots of selected replicating BMI loci with missense and CNV variants.

SNPs are plotted by position on chromosome against association with BMI (−log₁₀ P-value). The SNP name shown on the plot was the most significant SNP after stage 1 meta-analysis. Estimated recombination rates (from HapMap) are plotted in cyan to reflect the local LD structure. The SNPs surrounding the most significant SNP are color-coded to reflect their LD with this SNP (taken from pairwise r² values from the HapMap CEU database, www.hapmap.org). Genes, position of exons, and direction of transcription from UCSC genome browser (http://genome.ucsc.edu) are noted. Hashmarks represent SNP positions available in the meta-analysis. (a, b, c) Missense variants noted with their amino acid change for the gene noted above the plot. (d) Structural haplotypes and BMI association signal in the GPRC5B region. A 21 kb deletion polymorphism is associated with 4 SNPs (r²=1.0) that comprise the best haplogroup associating with BMI. Plots were generated using LocusZoom (http://csg.sph.umich.edu/locuszoom).

As many of the 32 BMI loci harbor multiple genes, we examined whether gene expression (eQTL) analyses could also direct us to positional candidates. Gene expression data were available for human brain, lymphocytes, blood, subcutaneous and visceral adipose tissue, and liver^34–36 (Online Methods, Table 1 and Supplementary Table 7). Significant cis-associations, defined at the tissue-specific level, were observed between 14 BMI-associated alleles and expression levels (Table 1 and Supplementary Table 7). In several cases, the BMI-associated SNP was the most significant SNP or explained a substantial proportion of the association with the most significant SNP for the gene transcript in conditional analyses (P_adj>0.05). These significant associations included NEGR1, ZC3H4, TMEM160, MTCH2, NDUFS3, GTF3A, ADCY3, APOB48R, SH2B1, TUFM, GPRC5B, IQCK, SLC39A8, SULT1A1, and SULT1A2 (Table 1 and Supplementary Table 7), making these genes higher priority candidates within the associated loci. However, we note that some BMI-associated variants were correlated with the expression of multiple nearby genes, making it difficult to determine the most relevant gene.

Evidence for the existence of additional associated variants

Because the variants identified by this large study explain only 1.45% of the variance in BMI (2–4% of genetic variance based on an estimated heritability of 40–70%), we considered how much the explained phenotypic variance could be increased by including more SNPs at various degrees of significance in a polygene model using an independent validation set (Online Methods)³⁷. We found that including SNPs associated with BMI at lower significance levels (up to P > 0.05) increased the explained phenotypic variance in BMI to 2.5%, or 4% to 6% of genetic variance (Fig. 4a). In a separate analysis, we estimated the total number of independent BMI-associated variants that are likely to exist with similar effect sizes to the 32 confirmed here (Online Methods)³⁸. Based on the effect size and allele frequencies of the 32 replicated loci observed in stage 2 and the power to detect association in the combined stage 1 and stage 2, we estimated that there are 284 (95% CI: 132–510) loci with similar effect sizes as the currently observed ones, which together would account for 4.5% (95% CI: 3.1–6.8%) of the variation in BMI or 6–11% of the genetic variation (based on an estimated heritability of 40–70%) (Supplementary Table 8). In order to detect 95% of these loci, a sample size of approximately 730,000 subjects would be needed (Fig. 4b). This method does not account for the potential of loci of smaller effect than those identified here to explain even more of the variance and thus provides an estimated lower bound of explained variance. These two analyses strongly suggest that larger GWA studies will continue to identify additional novel associated loci, but also indicate that even extremely large studies focusing on variants with allele frequencies above 5% will not account for a large fraction of the genetic contribution to BMI.

Figure 4. Phenotypic variance explained by common variants.

(a) Variance explained is higher when SNPs not reaching genome-wide significance are included in the prediction model. The y-axis represents the proportion of variance explained at different P-value thresholds from stage 1 meta-analysis. Results are given for three studies (RSII, RSIII, QIMR), which were not included in the meta-analysis, after exclusion of all samples from The Netherlands (for RSII and RSIII) and the United Kingdom (for QIMR) from the discovery analysis for this sub-analysis. The dotted line represents the weighted average of the explained variance of three validation sets. (b) Cumulative number of susceptibility loci expected to be discovered, including those we have already identified and others that have yet to be detected, by the expected percentage of phenotypic variation explained and sample size required for a one-stage GWA study assuming a GC correction is utilized. The projections are based on loci that achieved a significance level of P < 5×10⁻⁸ in the joint analysis of stage 1 and stage 2 and the distribution of their effect sizes in stage 2. The dotted red line corresponds to the expected phenotypic variance explained by the 22 loci that are expected to be discovered in a one-stage GWAS with the sample size of stage 1 of this study.

We examined whether selecting only a single variant from each locus for follow-up led us to underestimate the fraction of phenotypic variation explained by the associated loci. To search for additional independent loci at each of the 32 associated BMI loci, we repeated our GWA meta-analysis, conditioning on the 32 confirmed SNPs. Using a significance threshold of 5 × 10⁻⁶ for SNPs at known loci, we identified one apparently independent signal at the MC4R locus; rs7227255 was associated with BMI (P = 6.56 × 10⁻⁷) even after conditioning for the most strongly associated variant near MC4R (rs571312) (Fig. 5). Interestingly, rs7227255 is in perfect LD (r² = 1) with a relatively rare MC4R missense variant (rs2229616, V103I, minor allele frequency = 1.7%) that has been associated with BMI in two independent meta-analyses^39,40. Furthermore, mutations at the MC4R locus are known to influence early-onset obesity^24,41, supporting the notion that allelic heterogeneity may be a frequent phenomenon in the genetic architecture of obesity.

Figure 5. Second signal at the MC4R locus contributing to BMI.

SNPs are plotted by position in a 1 Mb window of chromosome 18 against association with BMI ( log₁₀ P-value). Panel (a) highlights the most significant SNP in stage 1 meta-analysis, panel (b) the most significant SNP after conditional analysis where the model included the most strongly associated SNP from panel A as a covariate. Estimated recombination rates (from HapMap) are plotted in cyan to reflect the local LD structure. The SNPs surrounding the most significant SNP are color-coded to reflect their LD with this SNP (taken from pairwise r² values from the HapMap CEU database, www.hapmap.org). Genes, exons, and direction of transcription from UCSC genome browser (genome.ucsc.edu) are noted. Hashmarks at the top of the figure represent positions of SNPs in the meta-analysis. Regional plots were generated using LocusZoom (http://csg.sph.umich.edu/locuszoom).

Discussion

Using a two-stage genome-wide association meta-analysis of up to 249,796 individuals of European descent, we have identified 18 additional loci that are associated with BMI at genome-wide significance, bringing the total number of such loci to 32. We estimate that more than 250 (i.e. 284 predicted loci – 32 confirmed loci) common variant loci with effects on BMI similar to those described here remain to be discovered, and even larger numbers of loci with smaller effects. A substantial proportion of these loci should be identifiable through larger GWA studies and/or by targeted follow-up of top signals selected from our stage 1 analysis. The latter approach is already being implemented through large-scale genotyping of samples informative for BMI using a custom array (the Metabochip) designed to support follow-up of thousands of promising variants in hundreds of thousands of individuals.

The combined effect on BMI of the associated variants at the 32 loci is modest, and even when we try to account for as-yet-undiscovered variants with similar properties, we estimate that these common variant signals account for only 6–11% of the genetic variation in BMI. There is a strong expectation that additional variance and biology will be explained using complementary approaches that capture variants not examined in the current study, such as lower frequency variants and short insertion-deletion polymorphisms. There is good reason to believe (based on our findings at MC4R and other loci – POMC, BDNF, SH2B1 – which feature both common and rare variant associations) that a proportion of such low-frequency and rare causal variation will map to the loci already identified by GWA studies.

A primary goal of human genetic discovery is to improve understanding of the biology of conditions such as obesity⁴². One particularly interesting finding in this regard is the association between BMI and common variants near GIPR, which may indicate a causal contribution of variation in postprandial insulin secretion to the development of obesity. In most cases, the loci identified by the present study harbor few, if any, annotated genes with clear connections to the biology of weight regulation. This reflects our still limited understanding of the biology of BMI and obesity-related traits and is in striking contrast with the results from equivalent studies of certain other traits (such as autoimmune diseases or lipid levels). Thus, these results suggest that much novel biology remains to be uncovered, and that GWA studies may provide an important entry point. In particular, further examination of the associated loci through a combination of resequencing and fine-mapping to find causal variants, and genomic and experimental studies designed to assign function, could uncover novel insights into the biology of obesity.

In conclusion, we have performed GWA studies in large samples to identify numerous genetic loci associated with variation in BMI, a common measure of obesity. Because current lifestyle interventions are largely ineffective in addressing the challenges of growing obesity^43,44, new insights into biology are critically needed to guide the development and application of future therapies and interventions.

Supplementary Material

1

Supplementary Figure 1 Study design. Stage 1 - Meta analysis of genome-wide association data was performed in stage 1 across 46 studies of white European Ancestry. A total of 42 SNPs representing the best associating (P < 10⁻⁶) loci (shown) were taken forward for replication. Nineteen of these SNPs (loci in bold) reached already genome-wide significance at stage 1. Stage 2 – The 42 SNPs were genotyped in 16 de novo replication studies and extracted from 18 in silico replication studies, all adults of European ancestry and were tested for association with BMI. In a joint analyses of stage 1 and stage 2 data, 32 SNPs (loci in bold) reached genome-wide significance (P < 5×10⁻⁸). Follow-up analyses – The 32 confirmed loci were taken forward for additional analyses.

Supplementary Figure 2 Regional plots of the 32 confirmed BMI loci with missense and CNV variants. SNPs are plotted by position on chromosome against association with BMI (−log₁₀ P-value). The SNP name shown on the plot was the most significant SNP after stage 1 meta-analysis. Estimated recombination rates (from HapMap) are plotted in cyan to reflect the local LD structure. The SNPs surrounding the most significant SNP are color-coded to reflect their LD with this SNP (taken from pairwise r² values from the HapMap CEU database, www.hapmap.org). Genes, position of exons, and direction of transcription from UCSC genome browser (genome.ucsc.edu) are noted. Hash-marks represent SNP positions available in the meta-analysis. Plots were generated using LocusZoom (http://csg.sph.umich.edu/locuszoom).

Supplementary Figure 3 Quantile-quantile plot of SNPs at stage 1 GIANT meta-analysis (black) and after removing any SNPs within 1 Mb of the 10 previously reported genome-wide significant hits for BMI (blue), after additionally excluding the four loci for waist/weight (green) and after excluding all 32 confirmed loci (red).

Supplementary Figure 4 Relationship between the predicted BMI, based on the 32 confirmed BMI loci combined, and the actual BMI in the ARIC Study (N=8,120). Panel (a) shows the comparison of the predicted BMI (grey), with the actual BMI (green) and quartile ranges (orange) in sets of 500 individuals, suggesting good average predictions. Panel (b) shows the comparison of the individual predicted and observed BMI values, suggesting poor individual predictions.

Supplementary Table 1 The 42 SNPs, associated with BMI at P < 5.10⁻⁶ at stage 1, that were taken forward for replication in stage 2.

Supplementary Table 2 Association of 32 replicated SNPs with other anthropometric traits.

Supplementary Table 3 Association between 32 replicated SNPs with risk of extreme obesity in children and adults, and with BMI in population-based childhood studies.

Supplementary Table 4 Association of the 32 confirmed BMI SNPs with metabolic traits.

Supplementary Table 5 Gene set enrichment analysis (MAGENTA) of biological pathways with one or more genes from the 32 confirmed BMI loci, using the BMI meta-analysis.

Supplementary Table 6 Non-synonymous or splice-site variants in linkage disequilibrium (r² > 0.75) with lead SNPs.

Supplementary Table 7 Significant associations between BMI SNPs and cis gene expression (cis-eQTLs) in lymphocyte, blood, adipose and brain tissues.

Supplementary Table 8 Estimated number of BMI loci for each of the effect sizes observed in Stage 2 for the SNPs that reached a genome-wide significance of 5×10⁻⁸ in the joint analysis of stage 1 and stage 2, given the power to detect the association in the joint analysis of stage 1 and stage 2.