A Candidate-Gene Approach Identifies Novel Associations Between Common Variants in/Near Syndromic Obesity Genes and BMI in Pediatric and Adult European Populations Free

Obesity has become a global epidemic, with the most recent data from the World Health Organization estimating a worldwide prevalence of 650 million cases in adults. It is a well-established risk factor for numerous comorbidities as well as increased mortality (1). With 65–80% of new cases being accounted for by overweight or obesity, excessive body weight is the main risk factor for incident type 2 diabetes (2). To date, most prevention programs have emphasized an “eat less, move more” global approach, but this has had limited success (3). Numerous behavioral (e.g., diet, exercise) and pharmacological treatments have been developed and implemented clinically, but have had only modest effects (4). Currently, bariatric surgery is the most effective treatment for severe obesity, but it is an invasive procedure with numerous complications and only a small fraction of eligible patients are referred (5). A better understanding of the underlying causes of obesity may be useful in developing more effective prediction, prevention, and treatment (6).

Genetics accounts for 40–75% of BMI interindividual variability (7). Recent estimates suggest that single nucleotide polymorphisms (SNPs) alone may explain up to 30–40% of BMI variance (8). Genome-wide association studies (GWAS) have identified a substantial number of SNPs associated with BMI (9). These results have helped determine causal genes and significantly expanded our understanding of obesity pathophysiology (6).

The Genetic Investigation of ANthropometric Traits (GIANT) Consortium and the UK Biobank (UKB) recently released results from the largest BMI GWAS meta-analysis to date, which identified 941 near-independent SNPs, accounting for 6% of BMI variance (9). Although gene prioritization analysis was conducted, the individual genes were not identified. Furthermore, only single-marker analysis was conducted, causing a substantial reduction of statistical power because of the multiple testing burden (10). Power can be increased by 1) targeting SNPs only at genes with strong pre-existing evidence for a role in obesity, 2) conducting tagSNP analysis, and 3) conducting gene-based analysis (10).

Although obesity in the general population is polygenic, it can also present with monogenic nonsyndromic mutations or monogenic syndromic mutations. Monogenic syndromic forms of obesity present with striking obesity along with features such as intellectual disability, dysmorphic features, and organ-specific abnormalities (11). These forms of obesity are often considered separate entities (6). However, accumulating evidence suggests that these forms exist on a continuum. For example, many loci at monogenic nonsyndromic obesity genes have also been associated with polygenic obesity (e.g., POMC, PCSK1, MC4R, and ADCY3) (6). Furthermore, several SNPs located in and near syndromic obesity genes have been reported by previous GWAS for BMI: BDNF, NTRK2, SIM1, BBS2, BBS4, SH2B1, and SDCCAG8 (9). We hypothesized that additional syndromic obesity genes were associated with polygenic variation of BMI.

The aim of this study was to identify loci associated with BMI variation at highly relevant candidate genes missed by hypothesis-free GWAS. We recently published a systematic review describing 79 genetic syndromes with obesity, of which 30 have been fully or partially elucidated genetically (11). We conducted single-marker, tagSNP, and gene-based analysis of these identified genes to investigate the role of syndromic obesity genes in the most recent BMI GWAS meta-analysis conducted by the GIANT Consortium and UKB, as well as meta-analysis conducted by the Early Growth Genetics (EGG) Consortium (9,12).

We used publicly available data from the GIANT Consortium and UKB European meta-analysis, which consisted of European cohorts only (N = 681,275; participants on average per SNP) (9). We also used publicly available data from the EGG Consortium meta-analysis of European children in a post hoc meta-analysis (N = 35,668) (12).

Sixty-three genes had been previously identified in a systematic review of obesity syndromes (11). Nine genes were excluded since they were located on the X chromosome, which was not included in the GIANT Consortium and UKB meta-analysis (9). See Supplementary Table 1 for a summarized list of the syndromes and the associated genes.

The summary statistics of all SNPs within 200 kb upstream and downstream of the remaining 54 genes were extracted from files released online by the GIANT Consortium using the R package biomaRt (13). GWAS-significant SNPs have been shown to have the greatest number of significant pathway-phenotype associations when using a gene definition of ±200 kb compared with others between 0 and 1,500 kb (14). Sensitivity analyses using different gene boundaries were conducted using 50, 20, 10, and 0 kb.

Yengo et al. (9) selected and analyzed 711,933 SNPs overlapping the GIANT Consortium and UKB SNP sets after imputation and quality control. Briefly, Yengo et al. (9) analyzed a subset of SNPs with consistent alleles between the GIANT Consortium and UKB and after further elimination using a linkage disequilibrium (LD) threshold of r² > 0.9. To increase power, we further pruned SNPs by selecting tagSNPs, variants that can infer nearby SNPs because of limited historical recombination in these regions (15). We selected tagSNPs genotyped in the largest sample from SNPs with r² > 0.8 calculated from 1000 Genomes Project Consortium phase 1 European data based on greatest sample size, then greatest minor allele frequency, then greatest number of SNPs tagged, then selecting the SNP closest to the gene, and finally by random selection (16,17).

The GIANT Consortium transformed BMI residuals using an inverse standard normal function in linear regression models after adjustment for age, age squared, and/or other study-specific covariates (18). In the UKB, linear mixed model association testing was conducted under an infinitesimal model and adjusted for age, sex, recruitment center, genotyping batches, and the top 10 principal components.

All gene-based tests in this study were applied using Versatile Gene-based test for Genome-wide Association Studies 2 (VEGAS2) (19). VEGAS2 uses a simulation approach to calculate gene-based empirical association P values by pooling P values of individual SNP associations to χ² statistics after correcting for LD (19). Since a 200-kb gene definition was not an available option, the largest alternative was used: 50 kb.

Meta-analysis was performed using a random-effects model in PLINK version 1.90 software. Study weights were generated using the inverse variance method. Heterogeneity between studies was assessed using the χ² test for homogeneity, the I² statistic (20).

The Bonferroni correction was applied to all tests (α = 0.05). Applying an experiment-wide Bonferroni correction reduces the chance of making type I errors but increases the chance of making type II errors. Therefore, we followed the strategy reported previously by Feise (21) and considered independent Bonferroni corrections for each question asked. We considered each type of analysis (single-marker, tagSNP, or gene-based analysis) using a specific gene definition as separate questions (e.g., single-marker analysis with gene boundaries of ±200 kb). A total of 15 questions was tested, of which 3 are the primary hypotheses of the study, 11 are sensitivity analyses, and 1 is the meta-analysis with the EGG Consortium cohort.

LocusZoom was used to create regional association plots (22). LDlink was used to determine LD between SNPs (using r² values), where 1000 Genomes Project Consortium phase 3 (version 5) European data were used (23).

To further delineate the potential causative SNPs, functional annotation of SNPs was conducted using the Ensembl Variant Effect Predictor (VEP) tool (https://useast.ensembl.org/Tools/VEP). VEP predicts potential functional consequences of variants using the position of each variant. We tested all fine-mapping SNPs, which were defined as lead signals from single-marker, tagSNP, and gene-based analysis as well as SNPs in moderate LD (r² > 0.5) (24), which had also been tested by Yengo et al. (9).

Using ±200 kb gene boundaries, 18,140 SNPs were extracted from 54 genes (Supplementary Tables 1 and 2). A total of 997 of 18,140 SNPs (5.5%) located at 27 loci in/near 29 genes were significantly associated with BMI (P < 2.80 × 10⁻⁶). Of these SNPs, 38 of 997 (3.8%) were in/near eight novel loci: MRAP2, BBS1, GNAS, IFT27, LZTFL1, NIPBL, MAGEL2/MKRN3/NDN, and BBS9 (Table 1 and Supplementary Table 2).

To test for independence from previously identified loci, the r² value was determined between top SNPs at novel loci and top SNPs at loci previously identified in the study by Yengo et al. (9) as well as genome-wide significant SNPs identified for BMI in the GWAS Catalog (25). The top SNP at BBS1 (rs12805133) was in LD with previously identified SNPs: rs682842, which was previously identified at C11orf80 (r² = 0.564) (9). There was no LD (r² > 0.1) between any of the other top SNPs at novel loci and previously identified SNPs.

Of the remaining 959 of 997 significant SNPs, 520 of 997 (52.2%) were GWAS significant (P < 5 × 10⁻⁸) and were located at the following 19 genes: BDNF, SH2B1, CREBBP, BBS4, KCTD13, SDCCAG8, BBS2, SIM1, NTRK2, IFT172, EHMT1, KMT2D, SOX10, WDPCP, MKS1, EP300, FGFR1, RAB23, and RAI1 (Supplementary Table 2). The remaining 439 of 997 (44.0%) SNPs were mapped to the 19 genes with genome-wide significant loci.

Of the SNPs identified with single-marker analysis, 4,942 of 18,140 (27.2%) were selected as tagSNPs (Supplementary Table 3). A total of 224 of 4,942 (4.5%) tagSNPs were significantly associated with BMI (P < 1.01 × 10⁻⁵) of which 3 of 224 (1.3%) were novel association signals in/near BBS12, TUB, and RAD21 (Table 2). None of these SNPs were in LD (r² > 0.1) with previously identified SNPs. The remaining 221 of 224 tagSNPs were mapped to the same 27 loci identified in single-marker analysis (Supplementary Table 3).

All 54 selected genes were analyzed with VEGAS2, and 22 of 54 (40.7%) were significantly associated with BMI (P < 9.26 × 10⁻⁴) (Supplementary Table 4). Three novel genes were identified: GHR, PROKR2, and CHD2 (Table 3). Top SNPs at these loci were not in LD (r² > 0.1) with previously identified SNPs. Of the remaining genes, 15 of 54 (27.8%) were gene candidates for GWAS-significant SNPs identified in single-marker and tagSNP analysis, whereas 4 of 54 (7.4%) were candidates for novel SNPs identified in single-marker analysis: BBS1, GNAS, IFT27, and LZTFL1 (Supplementary Table 4).

Single-marker, tagSNP, and gene-based analyses were repeated using different gene boundaries: ±0, ±10, and ±20 kb, as well as ±50 kb for single-marker and tagSNP analyses. SNPs at novel genes were not identified, although tagSNP analysis using ±10- and ±20-kb gene definitions identified SNPs at loci that had been identified previously by gene-based analysis (Supplementary Table 5).

We conducted a meta-analysis of the top SNPs from single-marker, tagSNP, and gene-based analyses from the GWAS meta-analysis by the EGG Consortium of BMI in children, investigating the pooled results of the GIANT Consortium and the UKB as well as the meta-analyses of the EGG Consortium (Table 4). Of the SNPs investigated, 17 of 33 (51.5%) were significantly associated (P < 1.52 × 10⁻³): BDNF, SH2B1, KCTD13, CREBBP, BBS4, SIM1, BBS2, EP300, SOX10, WDPCP, and RAB23. Five of those 17 SNPs were in/near novel loci: MRAP2, BBS9, BBS12, GHR, and TUB. Heterogeneity between the studies for these SNPs was not significant (I² < 5).

Fine-mapping SNPs were identified and VEP analysis was conducted for 715 variants (Supplementary Table 6). Most SNPs had multiple predicted functions, including nonsense-mediated decay transcript variants and/or noncoding transcript variants. The majority were intronic (501 of 715; 70.1%) and intergenic variants (113 of 715; 15.8%). Others were related to regulatory function, with 73 of 715 (10.2%) at regulatory regions, 22 of 715 (3.1%) at 3′ untranslated regions, and 5 of 15 (0.7%) at 5′ untranslated regions. Several SNPs were identified as coding variants, with 10 of 715 (1.4%) identified as missense variants and 1 of 715 (0.1%) identified as a stop-gained variant. Of the 10 missense variants, 1 was identified for BDNF (rs6265), 2 for SH2B1 (rs7498665, rs2904880), 1 for BBS4 (rs2277598), 1 for SIM1 (rs240780), 1 for BBS2 (rs11373), 3 for BBS1 (rs618838, rs540874, rs1671064), and 1 for RAB23 (rs3800023). The stop-gained variant was also identified for BBS1 (rs1815739).

We analyzed highly relevant candidate genes linked to syndromic obesity in publicly available data from the latest meta-analysis of BMI by the GIANT Consortium and UKB. Single-marker, tagSNP, and gene-based analyses revealed 14 novel associations at 16 of 54 (29.6%) candidate genes, and genome-wide significant SNPs were mapped to 19 of 54 (35.2%) syndromic obesity genes. A significant association for 17 of 33 (51.5%) loci was also observed with BMI in children.

Our data confirm the importance of conducting candidate gene approaches to identify SNPs with modest genetic effects missed by conventional hypothesis-free GWAS. As an illustration, the average effect size of the 11 novel SNPs identified in this study on BMI was two times lower in comparison with the 941 nearly independent SNPs identified in the most recent GWAS for BMI (β = 0.0105 [95% CI 0.0089, 0.0122] vs. β = 0.0205 [95% CI 0.0195, 0.0215] (9).

Most of the identified genes are related to leptin signaling and primary cilia function in neurons. Leptin signaling produces an anorexigenic effect by activating proopiomelanocortin (POMC)-expressing neurons, whereas dysfunction in primary cilia causes ciliopathies, which are well-established causes of complex syndromes that often present with obesity (6,11).

Eleven of the identified genes are related to the BBSome: BBS4, BBS2, MKS1, WDPCP, SDCCAG8, IFT172, BBS1, BBS9, BBS12, LZTFL1, and IFT27. The BBSome is a protein complex in primary cilia with a key role in leptin receptor (LepR) trafficking (26). Mutations in these genes cause Bardet-Biedl syndrome, which is characterized by truncal obesity as well as intellectual disability, retinal dystrophy, polydactyly, microorchidism in men, and renal malformations (11). Postnatal disruption of the primary cilia function in POMC neurons of mice results in hyperphagia and obesity (27). Previous evidence has demonstrated a potential role for the BBSome in common obesity, as higher BMI in heterozygous carriers of rare mutations in BBS genes without additional clinical features has been observed (28).

TUB and RAB23 may also be related to neuronal cilia. Mutations in TUB, which encodes tubby, produce a phenotype similar to that caused by mutations in the BBSome, including obesity and retinopathy (11). Tubby has demonstrated a critical role in trafficking G-protein–coupled receptors in neuronal cilia (29). The protein product of RAB23 also has a role in trafficking in primary cilia (30).

MRAP2, MAGEL2, NDN, and GNAS are also related to the leptin signaling pathway. Mutations in MRAP2 alter melanocortin receptor 4 (MC4R) signaling, a receptor downstream of POMC (31). Although large deletions involving MRAP2 have been reported in cases of syndromic obesity, pathogenic variants also have been identified in several cases of early onset obesity without additional features (11,31). Moreover, Magel2 (encoded by MAGEL2) promotes the cell surface presence of LepR (32). LepR subsequently interacts with necdin, encoded by NDN (32). POMC neurons in Magel2-deficient mice have impaired leptin response, which may be caused by reduced LepR trafficking to the cell surface (33). Finally, GNAS encodes G-protein α subunits, including the G-protein α subunits integral in MC4R (34). Mutations in SOX10, PROKR2, and FGFR1 may also affect leptin signaling via their effect on the migration of gonadotropin-releasing hormone neurons. Gonadotropin-releasing hormone is modulated by leptin and LepR, and inactivating mutations in leptin and LepR causes an absence of pubertal development as well as hypogonadotropic hypogonadism (35).

The KCTD13 locus lies within the well-described 16p11.2 interval, along with SH2B1. Deletion of the region is associated with macrocephaly, hyperphagic obesity, psychosis, epilepsy, schizophrenia, and autism spectrum disorder (ASD) (11). Although SH2B1 has been identified as the primary driver for the changes in feeding behaviors and body weight, KCTD13 has been identified as the primary candidate for the brain-related phenotypes (36,37). However, KCTD13 interacts with ciliopathy-associated genes, and KCTD13 knockdown-induced macrocephaly can be rescued by BBS7 overexpression (38).

SIM1 has been linked to obesity through its critical role in the development of key components of the hypothalamic-pituitary axis (39). Some evidence suggests that SIM1 may be downstream of MC4R signaling. For example, the effect of a melanocortin receptor agonist that reduces food intake in wild-type mice is blunted in Sim1 heterozygous mice (40).

Another pathway heavily implicated in obesity is the brain-derived neurotrophic factor (BDNF)-tyrosine-related kinase B (TrkB) axis. BDNF encodes BDNF, which binds to the TrkB receptor encoded by NTRK2 to produce an anorexigenic effect (41). Strong evidence has implicated the BDNF-TrkB axis in common obesity (42). There is also an interaction between the BDNF-TrkB pathway and the leptin signaling pathway, as the coadministration of a TrkB antagonist attenuates the anorexigenic effects of leptin (43).

Kleefstra, Kabuki, and Smith-Magenis syndromes are linked to mutations in EHMT1, KMT2D, and RAI1, respectively (11). Individuals with any of these three syndromes can present with ASD as well as obesity. Pathway analysis has demonstrated potential functional convergence, and these syndromes may be involved in common and/or overlapping pathways (44). One of those pathways may involve the BDNF-TrkB axis. EHMT1, a histone methyltransferase, has been shown to mediate BDNF expression during synaptic scaling (45). However, the observed association with BMI in our study may result from an interaction with ASD, which has been associated with a higher BMI (46). Mutations in CHD2 have also been associated with ASD as well as intellectual disability and photosensitive epilepsy. However, only one case report (11) has described obesity with a CHD2 mutation. Two patients harbored a 15q26.1 microdeletion of CHD2 and RGMA, and exhibited truncal obesity (11).

The remaining genes identified in this study relate to various syndromes whose underlying mechanisms are not well understood. Mutations in cohesin-related proteins rad21 and nipbl cause Cornelia de Lange syndrome (CdLS), which presents with characteristic facial features and variations of weight (11,47). However, none of the 23 reported cases of CdLS with Rad21 mutations are obese (47). Among the genes tested for in a suspected diagnosis of CdLS are CREBBP and EP300, which are causative genes for a similar disorder: Rubinstein-Taybi syndrome (RTS). RTS is characterized by features including growth restriction, microcephaly, and broad thumbs and toes (11). Case series have demonstrated that obesity presents in 30–40% of patients with RTS caused by mutations in CREBBP or EP300 (48). Finally, the deficiency of GHR produces Laron syndrome, which is characterized by abnormal growth with child-like proportions in adulthood (e.g., greater deviation of stature than head size). Obesity is observed in early childhood and progresses into adulthood as total body fat reaches 30–60% of weight (11). This is not attributed to increased caloric intake or reduced energy expenditure (49).

Our results confirm a large continuum between monogenic syndromic, monogenic nonsyndromic, oligogenic, and polygenic obesity (6). Substantial evidence, including candidate gene studies as well as independent GWAS, has previously suggested a role for many of the genes identified in this study in polygenic obesity (6).

In addition to the candidate-gene single-marker analysis, tagSNP and gene-based approaches were also conducted to explore the sensitivity of our analyses after further reduction of the Bonferroni correction. Although the original analysis conducted by the GIANT Consortium and UKB had pruned SNPs in high LD (r² > 0.9), additional pruning (r² > 0.8) in our tagSNP analysis identified all 27 loci from single-marker analysis as well as three additional SNPs at novel loci. Future GWAS may identify additional SNPs by using a lower threshold for LD. Furthermore, gene-based analysis rescued three loci where no individual SNP was significantly associated with BMI, but rather several variants had a moderate effect on BMI. We recommend that future GWAS systematically conduct gene-based analysis in tandem with classic single-marker analysis (50).

This study defined gene boundaries as ±200 kb. We found that this boundary captured the greatest number of significant SNPs after Bonferroni correction compared with boundaries of ±50, ±20, ±10, and ±0 kb. Although larger gene boundaries were not tested, this finding is consistent with a separate study that found the greatest number of significant pathway-phenotype associations for GWAS-significant SNPs when using a gene definition of ±200 kb compared with others between 0 and 1,500 kb (14).

Some of the included syndromic obesity genes in this study have not been replicated or confirmed by a second independent study (see Supplementary Table 1) yet have been previously confirmed as GWAS significant (P < 5 × 10⁻⁸): NTRK2, SH2B1, and SDCCAG8. Therefore, there is no direct correlation between the strength of known evidence for the role of a gene in a genetic obesity syndrome and its effects on common BMI variation. However, this study provides some additional support for a role for these genes in genetic obesity syndromes. This is particularly impactful for genes that have not been replicated or confirmed: SH2B1, SDCCAG8, and NTRK2, as well as LZTFL1, TUB, and CHD2.

Meta-analysis with the EGG Consortium cohort demonstrated significant signals in more than half of the top SNPs from single-marker, tagSNP, and gene-based analysis (17 of 33; 51.5%) after Bonferroni correction (P < 1.52 × 10⁻³). Since the EGG Consortium cohort was composed of children rather than adults, significant genetic associations reflect those that play a role in anthropometry throughout life. In contrast, SNPs that did not demonstrate a significant association all showed significant heterogeneity between the studies (I² > 20). This may suggest age-dependent effects for these SNPs. However, some SNPs showing significant heterogeneity are located in/near genes that have been previously associated with childhood obesity syndromes (e.g., EHMT1 in Kleefstra syndrome, GNAS in Allbright Hereditary Osteodystrophy) or have shown previous evidence of associations with obesity in children and adolescents (e.g., SDCCAG8).

This study demonstrates that candidate gene studies can be highly successful, with 64.8% of the candidate genes (35 of 54) demonstrating significant associations and 51.5% of significant loci (17 of 33) showing association after meta-analysis with a separate pediatric cohort. Two main components of this study differ significantly from most candidate-gene studies: 1) the selected candidate genes have strong evidence for a causal role in obesity (i.e., clinically evident cases of morbid obesity with evidence of drastic genetic mutations); and 2) we used a very well-powered BMI data set (up to 681,275 participants). Other candidate gene studies often use indirect evidence such as transcriptome and methylome analysis, which do not strongly distinguish obesity as a cause or consequence (10). We recommend that future well-powered GWAS analyses supplement hypothesis-free analysis with candidate-gene approaches using highly causal evidence, including novel syndromic obesity genes. Other potential sources of highly causal evidence include mouse models of obesity (6).

The genes analyzed in this study are likely a limited selection of potential candidates since only a proportion of the known genetic syndromes with obesity have been fully genetically elucidated (19 of 79; 24%) (11). Some have only been partially genetically elucidated (11 of 79; 14%), and there may also be syndromes that have yet to be identified (11). Also, neither rare variants nor potential long-range associations were included in these analyses. These genes were identified in a population of European descent only. Therefore, transferability studies in other ethnic groups are needed to investigate the generalizability of these results. Finally, as with all GWAS, the genes identified in this study are candidates, and the SNPs identified may only be proxies for the causative SNP. Other potential causative genes may lie within the proximity of the identified SNPs. The causative SNPs may lie within the LD blocks of the identified SNPs.

In summary, this study identified novel common variants associated with BMI in a large sample of children and adults of European ancestry using a combination of candidate-gene, single-marker, tagSNP, and gene-based approaches. This provides additional support to growing evidence of a continuum between monogenic syndromic obesity and common polygenic obesity.

Acknowledgments. The authors thank the members of the GIANT Consortium, the UKB, and the EGG Consortium for their data and participation.

Funding. D.M. is supported by a Canada Research Chair in Genetics of Obesity.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. D.X.W. designed the study, analyzed the data, wrote the paper and had responsibility for final content, and read and approved the final manuscript. Y.K. wrote the paper and had responsibility for final content, and read and approved the final manuscript. A.A. analyzed the data, and read and approved the final manuscript. D.M. designed the study, wrote the paper and had responsibility for final content, and read and approved the final manuscript. D.M. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Data Availability. The GIANT Consortium and the UKB have freely provided their data at http://portals.broadinstitute.org/collaboration/giant/index.php/GIANT_consortium_data_files. The EGG Consortium also provided their data freely at https://egg-consortium.org/.