Introduction

The IGF1 gene encodes insulin-like growth factor 1 (IGF-I). This hormone has many biological functions involving cell growth, proliferation, and apoptosis.1 Circulating IGF-I concentrations have been associated with several human diseases, including cardiovascular mortality and cardiovascular risk factors such as age, body mass index, total cholesterol, the presence of diabetes, glomerular filtration rate, and alcohol consumption.2, 3 IGF-I levels are inversely correlated with insulin resistance3 that may be explained by the insulin-like effects of IGF-I on glucose-uptake. IGF-I is structurally comparable to insulin and they both crossreact with the other’s receptor.

Genome-wide association studies (GWASs) of fasting insulin (FI) levels revealed common noncoding single-nucleotide variants (SNVs) near the IGF1 gene.4, 5 SNV rs35767:A>G (hg18 chr12:g.101399699A>G), located 1.2 kb upstream of IGF1, was associated with a 0.010 pmol/l per G allele increase in FI level (P=3.3 × 10−8) in a large GWAS meta-analysis.4 Another large GWAS meta-analysis, in largely overlapping samples, revealed rs2114912:G>T (hg18 chr12:g.101453133G>T) as the variant most strongly associated with FI in the IGF1 region.5 This variant is located 54.7 kb upstream of the IGF1 gene and is associated with a 0.024 pmol/l increase in FI per copy of the T allele. These findings have inspired further assessment of the role that the IGF1 gene plays in insulin biology.

In this paper we hypothesize that the associations of SNVs near the IGF1 gene with FI (hence insulin resistance) are mediated by circulating IGF-I levels, and that the GWAS variants tag other common or rare functional variants in the IGF1 region associated with FI levels. To test the first hypothesis, we performed mediation analyses using imputed genotyping array data, and to test the second hypothesis we performed association analyses using deep, high-throughput next-generation targeted sequencing data around IGF1. We also examined ENCODE Consortium data sets6 of regulatory elements by viewing the IGF1 region in the UCSC Genome Browser7 in order to generate testable hypotheses about direct functional roles and mechanisms of the noncoding FI-associated GWAS variants.

Materials and methods

An overview of the study design is shown in Supplementary Figure 1.

Study populations

Individuals of European ancestry from four cohorts of the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium were included in this study: the Atherosclerosis Risk in Communities (ARIC) study, Cardiovascular Health Study (CHS), Framingham Heart Study (FHS), and the Rotterdam Study (RS).8

Mediation cohorts

A total of 5141 nondiabetic individuals of CHS (n=1717), FHS (n=3293), and RS (n=140) were available to contribute to mediation analyses. Genotypic data and both FI and circulating IGF-I levels were available on these participants.

Sequencing cohorts

A total of 3539 nondiabetic individuals (ARIC n=1761; CHS n=967; and FHS n=811) who were part of the CHARGE Targeted Sequencing Study were available for analyses of targeted sequence data with the outcome FI. Of these, 567 of the CHS and 78 of the FHS participants included in these analyses were also included in the mediation analyses. The design of the CHARGE Targeted Sequencing Study, including the cohort sampling design, has been described in detail in Lumley et al9 and Lin et al.10 In summary, to set up the analytic sample a case–cohort design was used in which both a cohort random sample and participants with extreme phenotypes for each of the 14 cardiometabolic traits (atrial fibrillation, blood pressure, BMI, bone mineral density, C-reactive protein, carotid intima–media thickness, echocardiography, electrocardiographic PR and QRS interval, FI, hematocrit, pulmonary function, retinal venule diameter, and stroke) were included. For FI (≥8 h fast), this included a sample of 200 participants (100 ARIC, 50 CHS, and 50 FHS) from the high tail of the distribution in individuals without diabetes, defined as either being diagnosed by a physician (ARIC), treated for diabetes, or having a fasting glucose (FG) >7 mmol/l (ARIC, FHS, and CHS). Three FHS participants with type I diabetes were excluded from selection.

Quantitative trait measurement

FI was measured from fasting plasma (FHS) or fasting serum (CHS and ARIC). In FHS, plasma was collected after a ≥8 h overnight fast and FI was measured on frozen specimen using the DPC Coat-A-Count RIA (total immunoreactive insulin) assay (assay sensitivity 1.2 μU/ml). In CHS (≥12 h fast), FI was measured using a competitive RIA (Diagnostic Products Corp., Malvern, PA, USA). In ARIC (≥8 h fast), FI was measured by radioimmunoassay (125Insulin kit; Cambridge Medical Diagnosis, Bilerica, MA, USA) (assay sensitivity 2 μU/ml). In CHS, circulating IGF-I levels were measured by ELISA (Immunodiagnostic Systems Ltd, Boldon Business Park, Boldon, Tyne & Wear, UK) and in RS by a radioimmunoassay (Medgenix Diagnostics, Brussels, Belgium). BMI was measured using standard methods as previously described.5

Genotyping in mediation cohorts

In CHS, genotyping was performed at the General Clinical Research Center’s Phenotyping/Genotyping Laboratory at Cedars-Sinai using the Illumina (San Diego, CA, USA) 370CNV BeadChip system. The following exclusions were applied: call rate <97%, Hardy–Weinberg equilibrium (HWE) P-value <10−5, >2 duplicate errors or Mendelian inconsistencies (for reference CEPH trios), heterozygote frequency=0, and SNV not found in HapMap. Samples were excluded from analysis for sex mismatch, discordance with prior genotyping, or call rate <95%. Imputation was performed using BIMBAM v0.9911 with reference to HapMap CEU using release 22. In the FHS, genotyping was conducted using the Affymetrix (Santa Clara, CA, USA) 500K SNP arrays supplemented with the MIPS 50K array. Samples with call rate <97%, excess Mendelian errors (≥1000), or average heterozygosity outside of 5 SD of mean (<5.758% or >29.958%) were excluded. SNPs with minor allele frequency (MAF) ≥1%, call rate ≥97%, differential missingness P-value ≥10−9, and <100 Mendelian errors were used for imputation based on the haplotypes of the HapMap CEU release 22 using the MaCH12 software. In the Rotterdam Study, genotyping was performed using 550 and 610K Illumina arrays. Exclusion criteria for individuals were excess autosomal heterozygosity, mismatches between called and phenotypic gender, and outliers identified by an IBS clustering analysis. SNVs were excluded for HWE P-value ≤10−6, or SNP call rate ≤98%. Genotypes with MAFs >1% were used for imputation using HapMap CEU release 22 as a reference panel. Imputation was performed using MaCH.12

Targeted next-generation deep sequencing

Target selection in the CHARGE Targeted Sequencing Study included regions that had been associated with one of 14 cardiometabolic traits by previous GWASs and regions that had been shown to exhibit pleiotropy, and included the IGF1 gene.10 Four regions in or near the IGF1 gene were sequenced at a mean depth of 50 ×, including 1 kb downstream, all five exons plus flanking regions, and five SNVs upstream that were associated with FI in GWAS:4, 5 rs35767:A>G, rs860598:G>A (hg18 chr12:g.101422576G>A), rs855213:A>G (hg18 chr12:g.101432427A>G), rs35747:G>A (hg18 chr12:g.101436688G>A), and rs2114912:G>T (Supplementary Figure 2). A total of 57.5 kb per copy of the IGF1 region was sequenced. Sequencing methods were described in detail in Lin et al.10 An extensive quality control (QC) pipeline was implemented, consisting of QC procedures in the sequencing laboratory followed by a series of variant-level filtering steps. These included the exclusion of variants mapping more than 100 base pairs from the requested target capture region, exclusion of variants with a Phred-scaled base quality score13 <30, with less than two reads of the alternate alleles, and variants with a depth of coverage of <10 total reads. Heterozygote genotypes were removed if their alternate to reference allele ratio was disproportionate (<0.2 or >0.8 for one allele). For strand bias, only variants with alternate allele reads obtained from both the positive and negative strands were kept. Finally, SNPs that had >20% missingness across all samples, more than two observed alleles, or were part of an overly dense SNP cluster (≥3 variants in a 10-bp window) were removed. Using only samples from the cohort random sample subjects, SNPs with HWE P-value <1 × 10−5 were filtered. This criterion was not applied in the samples selected based on extreme phenotypes, potentially enriched for rare variants, to prevent filtering out interesting rare variants with a possible role in disease etiology. To validate sequence-based genotypes, cross-validation was performed with data from the Affymetrix Gene Chip 500K Array Set and 50K Human Gene Focused Panel in 1096 FHS samples. A total of 558 SNPs were shared between the two platforms. After excluding missing genotypes, 98.0% of genotypes were concordant between the two platforms, suggesting high accuracy of the sequenced genotypes. The targeted sequencing data have been submitted to dbGaP (phs000651.v6.p10 (FHS), phs000667.v2.p1 (CHS), and phs000668.v1.p1 (ARIC)).

Variant classification and annotation

Variants identified by sequencing of the IGF1 locus were classified as common if the MAF was ≥1% and rare if the MAF was <1%. Novel variants were those not found in dbSNP, the 1000 Genomes Project, or ESP 6500 (Exome Sequencing Project).14, 15 Variants were annotated using several bioinformatics sources. ANNOVAR16 was used to determine whether a variant was synonymous, nonsynonymous, intergenic, upstream (within 1 kb upstream of a transcription start site), downstream (within 1 kb downstream of a transcription end site), intronic, in a 3′ untranslated region (3′UTR), or in a 5′UTR. Variants other than synonymous or nonsynonymous were defined as noncoding. Noncoding variants were predicted to be functional if they were predicted to be highly conserved across species using phastCons,17 predicted to lie in transcription factor binding sites extracted from the HMR Conserved Transcription Factor Binding Site track of the UCSC Genome Browser,7 in DNAse I hypersensitive sites or transcription factor binding sites identified by the ENCODE Project,6 or predicted to be functional using the ORegAnno database.18 In addition to this functional annotation of the variants present in the targeted sequencing data, we examined GTEx19 and the ENCODE Consortium regulatory element data sets (including DNAseI hypersensitive sites and histone modifications as well as TFBS Chip-seq) and public transcriptome data in the UCSC Genome Browser to determine whether the known common noncoding FI-associated GWAS variants might be directly functional.

Follow-up genotyping in FHS and lookup of selected rare variants

To verify the influence of variant rs151098426:C>T (hg18 chr12:g.101337467C>T) on FI levels, the variant was genotyped in 1745 FHS offspring and 3372 FHS generation 3 participants with FI levels available that did not overlap with the FHS participants included in the targeted sequencing analyses. Genotyping was performed using TaqMan (ABI PRISM 7700 HT Sequence Detection System, Applied Biosystems, Foster City, CA, USA) at the Joslin Diabetes Center Advanced Genomics and Genetics Core (Boston, MA, USA). We also did a lookup of the variant in FI exome chip meta-analysis results from the CHARGE diabetes-glycemia working group, including 38 528 samples.

Statistical analyses

All analyses were adjusted for age, sex, BMI, and study design variables (ie, clinic site for CHS and ARIC and recruitment cohort for FHS). FI, in pmol/l, was natural log transformed before analyses to improve normality.

Mediation analyses

To test whether association of FI with GWAS variants in the IGF1 region (rs35767:A>G, rs860598:G>A, rs855213:A>G, rs35747:G>A, and rs2114912:G>T, pairwise r2 0.272–1.00 in HapMap2 CEU (see Supplementary Table 1)) is mediated by IGF-I levels, in each cohort (CHS, FHS, and Rotterdam Study) two linear regression models per SNP were fitted, assuming an additive allelic effect. In both models, ln(FI) was the outcome variable. Results from the three cohorts were combined using inverse variance weighted fixed effects meta-analysis as implemented in the R package rmeta.20 In the first model, age, sex, and BMI were included as covariates, and in the second model IGF-I was added as a covariate. From the models, a ratio βSNP_model2/βSNP_model1 <1 would suggest that IGF-I levels explained part of the SNP–FI association.

Analyses of targeted sequence data

The analytic strategy of the targeted sequence data, described briefly below, followed the approach outlined in Lumley et al9 and Lin et al.10

Four subsets based on functional annotation of rare variants within the IGF1 locus were tested for association with ln(FI) using the Sequence Kernel Association Test (SKAT).21 The subsets included (1) nonsynonymous variants, (2) novel nonsynonymous variants, (3) noncoding variants that were predicted to be functional, and (4) novel noncoding variants that were predicted to be functional. FHS used a SKAT test that accounted for family structure.22 SKAT tests were conducted within the three cohorts (CHS, FHS, and ARIC) and meta-analyzed using a weighted sum of squares of z-statistics from single-variant score tests. These variant scores were squared, weighted based on combined allele frequencies across all studies, and summed to create a Q-statistic. The significance of the Q-statistics was determined using an asymptotic distribution, as described in Wu et al.21 The weighted squared z-score for each variant divided by the total Q-statistic can be used to identify variants contributing most to the signal. To control type 1 error for this part of the analysis, a P-value <0.05/4=0.0125 (corrected for four tests: 1 trait × 4 subsets of variants) was used to define statistical significance for the SKAT tests.

To test whether rare variant associations were independent of the known FI GWAS hits near the IGF1 gene, conditional analysis was performed by additionally adjusting for the two common variants rs35767 (FI top hit Dupuis et al4) and rs2114912 (FI top hit Manning et al5) (r2 between these variants=0.272 in HapMap2 CEU) in the rare variant analysis. As these two variants were not present in the targeted sequence data, rs2162679:C>T (hg18 chr12:g.101395389C>T) was used as a proxy for rs35767:A>G (r2=0.915 in HapMap2 CEU) and rs2607988:G>A (hg18 chr12:g.101454013G>A) was used as a proxy for rs2114912:G>T (r2=0.882 in HapMap2 CEU). Conditional SKAT analyses were performed in each cohort seperately and then meta-analyzed. Similar P-values in unconditional and conditional analyses suggest that rare variant associations are independent of the known common variant signals.

Although tests of rare variation were the primary aim of the targeted regional sequencing study, we also tested association of all variants with minor allele count (MAC) ≥50 identified by sequencing with ln(FI). In ARIC and CHS, standard additive genetic linear regression models were used, whereas in FHS mixed effects models were used to account for familial correlation. Results from each cohort were meta-analyzed using standard fixed effect inverse variance weighted meta-analysis.23 P-values were obtained from unweighted regression models. Analyses weighted by the inverse of the sampling probability were used to obtain unbiased estimates of effect size.9 The significance treshold for common variant analyses was set at P-value <1.0 × 10−3 (0.05/49 effective number of independent variants calculated using the Li and Ji approach24).

For analyses of follow-up genotyping data in FHS, we used linear mixed effect model to compare the average trait values by genotype category. As we performed two tests (offspring and generation 3 cohorts separately), we considered a P-value <0.025 (0.05/2) as significant.

Results

Descriptions of the CHARGE cohort characteristics are depicted in Table 1. Both in the individuals contributing GWAS data and in the targeted sequence samples, women were slightly overrepresented. The mean age ranged from 39 to 71 years in the GWAS samples and from 54 to 72 years in the targeted sequence samples. BMI was in the overweight range in all cohorts. As previously observed, FI values varied widely across studies.4 The same was observed for the IGF-I levels in the GWAS samples.

Table 1 Descriptions of the study populations
Full size table

Mediation analyses

Mediation analyses results are depicted in Table 2. Meta-analyses P-values were nominal to borderline significant for each SNV in both models (P=0.05–0.15). However, effect estimates were similar to the effect estimates in up to 51 750 samples in the discovery meta-analysis5 and in FHS, the largest contributing cohort, P-values were nominally significant for each SNV in both models (P=0.01–0.04) (Table 2). Both in the meta-analysis and in FHS alone, effect estimates were similar between model 1 (ln(FI)~SNP+age+sex+BMI) and model 2 (ln(FI)~SNP+age+sex+BMI+IGF-I). This is consistent with an effect of the variants near IGF1 on FI levels that is not mediated by circulating IGF-I levels.

Table 2 Association of known fasting insulin GWAS SNPs in the IGF1 region with fasting insulin levels without and with IGF1 levels as covariate in the model
Full size table

Analyses of targeted sequence data

Table 3 and Supplementary Table 2 show descriptions of known and novel variants identified by targeted sequencing of the IGF1 locus. Deep (mean read depth 50 ×) sequencing across the locus identified 1393 variants, 1143 (82.1%) of which were rare and novel. A total of 11 coding nonsynonymous variants were present, including 6 that were novel. Of the 1376 noncoding variants, 188 (14%) were predicted to be functional, including 156 that were novel. The large majority of the variants at the IGF1 locus had MAF <0.1% (Supplementary Figure 3). Of all variants present at the locus, 893 (64%) were only observed one time in our samples. Of the novel variants, 198 (17%) were present in at least two of the three cohorts.

Table 3 Descriptions of known and novel SNPs in the IGF1 region in the CHARGE Targeted Sequencing Study cohorts combined
Full size table

Meta-analyzed SKAT results (Table 4) showed that the subset of 11 rare coding nonsynonymous variants was significantly associated with ln(FI) (P=5.7 × 10−4). One rare variant (rs151098426:C>T, MAF=0.1%) accounted for 92.16% of the overall SKAT Q-statistic (Supplementary Table 3 and Supplementary Figure 4). This variant resulted in an alanine-to-threonine substitution and was predicted to be damaging by PolyPhen-2,25 LRT,26 and MutationTaster.27 In contrast to the positive effect estimate for the rare T allele of rs151098426:C>T in the SKAT targeted sequencing analysis (Supplementary Table 3), 3 of the 1745 FHS offspring participants and 11 of the 3372 FHS generation 3 participants with follow-up genotyping of rs151098426:C>T carrying the rare allele had lower FI levels than the noncarriers (offspring: β=−0.05; generation 3: β=−0.15). These differences between carriers and noncarriers were nonsignificant (offspring: P=0.734; generation 3: P=0.313). The geometric means and the corresponding confidence intervals in carriers and noncarriers are shown in Supplementary Figure 5. Lookup of the variant in CHARGE exome chip results revealed a positive, but also nonsignificant, effect of rs151098426:C>T on FI levels (MAF=0.14%, β=0.02, P=0.471).

Table 4 SKAT meta-analyses results for fasting insulin (BMI adjusted) from different subsets of rare (MAF <1%) SNPs in the IGF1 region
Full size table

Conditioning on proxies of the known FI GWAS variants rs2114912 and rs35767 attenuated the significant SKAT result to a nominal significant P-value (Pconditioned=0.019, Table 4), suggesting that the GWAS signal explains part of the rare variant signal and the presence of a residual independent rare variant effect. Examination of ENCODE Consortium regulatory element data sets and public transcriptome data in the UCSC Genome Browser suggested that GWAS variants in the vicinity of IGF1 might have a direct functional role. In particular, rs35767 is ∼1.2 kb upstream of the IGF1 promoter and merely a few bases away from a strong FOXA1 binding site that is observed in ENCODE ChIP-seq data across a variety of human cell lines. Similarly, rs2114912:G>T is ∼1.7 kb away from a strong ENCODE DNAseI hypersensitive site seen in multiple cell lines, including pancreatic islets, that overlaps an ENCODE transcription factor binding site ChIP-seq cluster for several transcription factors, including FOXA1. This combination of open chromatin as delineated by the DNAse I hypersensitive site with transcription factor binding in ChIP-seq data constitutes a regulatory element signature that warrants experimental validation. Rs2607988:G>A, a SNP in high LD with rs2114912:G>T (r2=0.882 in HapMap2 CEU), is located in a ChIP-seq site for FOXA1 and alters a motif for FOXA. Interrogating the GTEx database, we did not find evidence for the GWAS variants to influence gene expression in any of the available tissues.

Single-variant analyses did not reveal significant associations with FI for any of the common variants present in the targeted sequence data (Supplementary Figure 6), including the proxies of the known FI GWAS hits rs35767:A>G (Pmeta=0.69) and rs2114912:G>T (Pmeta=0.54) (Supplementary Table 4), most likely because of the much smaller sample size in these targeted sequence data compared with the original, very large discovery sample sizes.

Discussion

This study suggests that previously observed associations between SNVs near IGF1 with FI levels were not mediated by circulating IGF-I levels. Further investigation of the IGF1 gene, using deep sequencing data, revealed a large number of rare variants at the locus that had not been previously described, the large majority of which was very rare. A subset of rare coding nonsynonymous variants, including six novel variants and five variants that had been previously identified, was significantly associated with FI levels. Conditional analysis suggested that the common noncoding variants near IGF1 that were identified in GWAS4, 5 explain part of the rare variant signal and the presence of a residual independent rare variant effect. Examination of ENCODE Consortium regulatory element catalogs showed that the GWAS variants were located in the proximity of FOXA1 binding sites and DNAseI hypersensitive sites, suggesting that they might have a direct functional role. This finding is noteworthy because FOXA1 is a key transcriptional regulator implicated in glucose metabolism and insulin secretion.28, 29 Studies in human cell culture and animal models will be needed to interrogate and validate the function of these noncoding variants in insulin biology.

One variant, rs151098426:C>T, resulting in an alanine-to-threonine substitution and predicted to be damaging by several annotation tools, seemed to drive the rare variant association. However, follow-up genotyping of rs151098426:C>T in an independent set of samples and lookup of the variant in CHARGE exome chip results did not reveal significant differences in FI levels between carriers and noncarriers of the rare allele, suggesting the absence of a single-variant effect for rs151098426:C>T on FI levels. Several recently published studies have demonstrated the need for large sample sizes to robustly identify associations of low-frequency variants with complex traits.30, 31, 32, 33, 34, 35, 36 Because of the low MAF of rs151098426:C>T and thus the relatively small number of carriers, analyzing the variant in large numbers of additional samples will be required to definitively conclude whether this variant is associated with FI levels. Taking the FHS log FI distribution as an example and using a replication α of 0.05, if the effect was as large as we find in the SKAT results (1.32 SD), we would need 1657 samples to demonstrate the effect. However, this effect is likely to be an overestimate because of the winner’s curse. If the effect was modest as we found in the FHS offspring (0.17 SD), a sample size of 97 128 would be needed.

We did not find a mediation effect of circulating IGF-I levels on the association of SNVs near IGF1 with FI levels. However, measurement errors in IGF-I levels might be responsible for the absent observation of a mediation effect. Circulating IGF-I levels measured with an imperfect assay and at a single point in time may not sufficiently characterize the biologically relevant levels. However, although circulating levels of IGF-I decline with aging,37 the levels do not undergo large short-term fluctuations.38 Furthermore, in 3977 FHS participants, circulating IGF-I levels correlated negatively with insulin resistance, diabetes, and metabolic syndrome,3 suggesting that these measures do represent biologically relevant levels and thus making measurement errors a less likely cause for not observing a mediation effect of IGF-I in our study.

The identification of variants at the IGF1 locus that had not been previously described has increased our insight into the variation present at the locus. In line with previous sequencing studies,34, 39, 40 we identified a large number of very rare variants, the majority (64%) even observed only one time in our samples. The presence of large numbers of very rare variants in the human genome is likely explained by recent explosive human population growth.40, 41 It has been hypothesized that these variants might harbor larger effects than those observed for common variants, as selection can have influenced only the most deleterious variants.40 However, even for rare variants with larger effects, large sample sizes are needed to definitely conclude whether they influence complex traits because of the low MAF.

The strengths of this study in the CHARGE Targeted Sequencing framework include the high average sequence depth combined with stringent QC applied across the three cohorts, increasing confidence that even the rarest observed variation is real variation and not a technical artifact. Furthermore, we genotyped variant rs151098426:C>T in non-overlapping samples serving as replication cohort and as further evidence that the variant is real. A limitation of this study is type 2 error, both in mediation and targeted sequence analyses, where limited sample sizes have limited power to detect common and rare variant associations. The targeted sequence samples included only seven heterozygous carriers of the variant of interest rs151098426:C>T. With 3539 samples in this discovery set and a significance level of 0.001, for modest differences such as 0.1 SD in log FI, our power was 1% for MAF=1% and 22% for MAF=10%. Furthermore, because of the limited number of individuals with both targeted sequence data and IGF-I levels available in our study, it was not possible to test whether association of the subset of rare nonsynonymous variants with FI was mediated by IGF-I levels. Mean BMI was in the overweight range in all cohorts. However, evidence exists that effect sizes of known glycemic trait-associated variants do not differ between BMI strata.5 As previously observed, FI values varied widely across studies, likely because of limited standardization across assays. Previous gene discovery studies, however, despite the same observation were successful in identifying FI-associated variants.4, 5 Finally, our study only included individuals of European ancestry, and this might limit the generalizability to other ancestries of the observed IGF1 variants and variant associations in this study.

In conclusion, our analyses suggest that association of SNVs near the IGF1 gene with FI is not mediated by circulating IGF-I levels. Furthermore, our study increased insight into variation present at the IGF1 locus and thus into the specific local coding as well as noncoding genetic architecture underlying FI levels, showing a large number of novel rare variants present at the locus and suggesting association of both rare coding nonsynonymous variants and a potential direct functional effect of common noncoding GWAS SNVs in the region on FI levels.