Exome array analysis identifies novel loci and low-frequency variants for insulin processing and secretion

Abstract

Insulin secretion plays a critical role in glucose homeostasis, and failure to secrete sufficient insulin is a hallmark of type 2 diabetes. Genome-wide association studies (GWAS) have identified loci contributing to insulin processing and secretion^1,2; however, a substantial fraction of the genetic contribution remains undefined. To examine low-frequency (minor allele frequency (MAF) 0.5% to 5%) and rare (MAF<0.5%) nonsynonymous variants, we analyzed exome array data in 8,229 non-diabetic Finnish males. We identified low-frequency coding variants associated with fasting proinsulin levels at the SGSM2 and MADD GWAS loci and three novel genes with low-frequency variants associated with fasting proinsulin or insulinogenic index: TBC1D30, KANK1, and PAM. We also demonstrate that the interpretation of single-variant and gene-based tests needs to consider the effects of noncoding SNPs nearby and megabases (Mb) away. This study demonstrates that exome array genotyping is a valuable approach to identify low-frequency variants that contribute to complex traits.

Exome sequencing studies have discovered many low-frequency and rare coding variants³ that have yet to be examined systematically for association with complex traits. To determine the role of low-frequency coding variants in traits reflecting pancreatic beta-cell function, insulin sensitivity, and glycemia, we evaluated putative functional coding variants selected from exome sequences of >12,000 individuals (see Online Methods for a description of exome array design and content). We successfully genotyped 9,660 Finnish participants in the population-based Metabolic Syndrome in Men (METSIM) study⁴ for 247,870 variants on the Illumina HumanExome Beadchip. Clinical characteristics of 8,229 analyzed non-diabetic study participants are summarized in Supplementary Table 1. Among 242,071 variants passing quality control, 89,864 (38.1%) were variable in the studied individuals; of these, 71,077 were nonsynonymous, nonsense, or located in splice sites (Supplementary Table 2). We tested 59,029 variants with MAF>0.05% for association with insulin processing, secretion, and glycemic traits assuming additive allelic effects and using a linear mixed model to account for relatedness among study participants⁵.

We first evaluated rare and low-frequency coding variants at the nine signals previously identified by GWAS for fasting proinsulin level adjusted for fasting insulin (hereafter referred to as fasting proinsulin)¹. To recognize independent association signals, we carried out conditional analysis adjusting for the known GWAS variants, all of which were represented on the exome array and replicated in METSIM (P<.01; Figure 1, Supplementary Table 3). Coding low-frequency variants at the known SGSM2 and MADD loci showed strong evidence of association (P<5×10⁻⁸; Table 1, Supplementary Figures 1 and 2). Previous studies highlighted several possible candidate genes at these loci^1,6,7.

Figure 1.

Manhattan plot for the fasting proinsulin analysis. Association results of the single-variant analysis (−log₁₀ P values) are plotted against genomic position (NCBI Build 37). Previously identified loci are denoted in blue and loci identified by the current study in red. Fasting proinsulin levels were log-transformed and adjusted for fasting insulin, body mass index, age, and age².

Table 1.

Novel low frequency variants at fasting proinsulin loci previously identified by genome-wide association studies

SNP	Gene	Variant	Chr	Position	Minor/major allele	MAF	β̂ ± SE	Effect size in SD units ± SE	Proportion of trait variance explained	P value	Conditional P value
rs61741902	SGSM2	V996I	17	2,282,779	A/G	.014	.126 ± .021	.41 ± .07	.0047	8.7 × 10−10	4.8 × 10−10
rs35233100	MADD	R766X	11	47,306,630	T/C	.037	−.100 ± .013	−.32 ± .04	.0075	7.6 × 10−15	.0001

Chr, chromosome; MAF, minor allele frequency; SE, standard error; SD, standard deviation. Positions are from NCBI Build 37 with allele labels from the forward strand. Fasting proinsulin was log-transformed and adjusted for fasting insulin, BMI, age, and age². Effects are reported for the minor allele.β̂ coefficient units are ln(pmol/l). Conditional P values are reported after adjusting for the lead SNPs from GWAS signals (rs4790333 at SGSM2; rs7944584 and rs1051006 at MADD). Full results of conditional analysis are provided in Supplementary Table 4. For rs35233100, effect size in SD units (± SE) and proportion of trait variance explained after adjusting for rs7944584 and rs1051006, are −0.17 (± .05) and 0.0007, respectively. Based on analysis of 8,224 non-diabetic males.

At SGSM2, rs61741902 (MAF=1.4%, P=8.9×10⁻¹⁰) encodes Val996Ile and is independent of GWAS variant rs4790333 (P_cond=4.8×10⁻¹⁰, r²=.001; Table 1, Figure 1, Supplementary Table 4). SGSM2 (small G protein signaling modulator 2) is a GTPase activating protein (GAP) that interacts with members of the Rab and Rap small G protein pathways and may act in a cascade of Rab-mediated steps in insulin secretory vesicle transport^8–10. At rs61741902, the reference valine is well-conserved across vertebrates, and the isoleucine substitution is predicted to be damaging (Supplementary Table 5) Each additional copy of the minor allele was associated with an average increase of 0.41 standard deviations (SD) in fasting proinsulin (Table 1, Supplementary Figure 2). Still, the proportion of the trait variability explained is modest (0.47%; 95% CI = 0.22–0.82%) due to the low minor allele frequency. Identification of an independent and plausibly functional variant suggests that SGSM2 is the causal gene underlying the common fasting proinsulin GWAS signal.

At MADD, rs35233100 (MAF=3.7%, P=7.6×10⁻¹⁵) creates stop codon Arg766Ter and is in modest linkage disequilibrium (LD) with the lead GWAS variant rs7944584 (P_cond=.0001, r²=.17), and independent of the second GWAS variant rs1051006 (P_cond=5.0 × 10⁻¹⁶, r²=.02). The nonsense allele of rs35233100, associated with decreased proinsulin, is observed only on haplotypes containing the proinsulin-decreasing allele of rs7944584. Adjusting for one variant in a conditional analysis decreased, but did not eliminate, association for the other (P=4.9 × 10⁻²⁵ and P_cond=5.7 × 10⁻¹⁵ for rs7944584; Table 1, Supplementary Table 4), suggesting biological contributions from the nonsense variant and an additional causal variant tagged by rs7944584. Of note, the trait-decreasing alleles of the two common GWAS-identified variants rs7944584 and rs1051006 tend to occur on different haplotypes, causing the evidence of association for either SNP to become dramatically more significant when adjusting for the other (rs1051006, P=0.033 and P_cond=2.7 × 10⁻⁸; rs7944584, P=4.9 × 10⁻²⁵ and P_cond=8.3 × 10⁻³¹, Supplementary Table 4). Although the conditional association for the nonsense variant only achieves suggestive significance (P=.0001), it provides an especially plausible functional effect. The MADD nonsense variant is located in exon 13 of 36, suggesting that the mRNA would be targeted for nonsense-mediated decay¹¹. MADD can act as a guanine nucleotide exchange factor for RAB3 proteins including RAB3A and RAB3B¹², which are critical for insulin exocytosis^13,14. Identification of a nonsense variant that contributes to the evidence of association suggests that MADD is a causal gene underlying the common GWAS signals.

LD at chromosome 11 from 46–57 megabases (Mb) and encompassing MADD has been reported to extend long distances¹⁵. Consistent with this, we noted significant or suggestive (P<1×10⁻⁵) fasting proinsulin association with nonsynonymous variants up to ~9 Mb away from the lead (noncoding) GWAS variant. Proinsulin-associated variants included rs628524, located ~9 Mb away and encoding Ser171Asn in the olfactory receptor OR5M11 (P=3.7×10⁻⁶ for fasting proinsulin; P as low as 5.0×10⁻¹⁰ for related traits) and rs7941404, located 376 kb away and encoding Arg349His in AGBL2 (MAF=11.8%; P=4.7×10⁻²¹). After adjusting for the three MADD variants (rs7944584, rs1051006, rs35233100), association significance for the distant variants was reduced by orders of magnitude (Supplementary Table 6, Figure 2). The fact that these associations were not eliminated suggests that additional variant(s) in this region await identification or that we may be adjusting for imperfect proxies of causal variants. These results also demonstrate that LD should be considered when interpreting GWAS results in this region. For example, the recently reported¹⁶ novel fasting glucose locus at OR4S1, represented by rs1483121, is in LD (r²=.19) with the lead and nonsense MADD SNPs ~1 Mb away (Supplementary Table 6).

Figure 2.

The MADD gene is located in a region of unusually high linkage disequilibrium on chromosome 11 from 46 to 57 Mb. Regional association results of the single-variant analysis (−log₁₀ P values) are plotted against genomic position (NCBI Build 37) for fasting proinsulin before (a) and after (b) adjustment of the lead SNPs for the common GWAS signals (rs7944584 and rs1051006) and the nonsense variant rs35233100 (MAF 3.7%) at MADD. Fasting proinsulin levels were log-transformed and adjusted for fasting insulin, body mass index, age, and age². The conditioning SNPs are indicated in blue. For clarity, only a portion of the 11 Mb region and a subset of the genes are shown.

Next, we tested coding variants across the genome for association with 19 traits measuring pancreatic beta-cell function, insulin sensitivity, and glucose levels. We identified two genes harboring low-frequency nonsynonymous variants with novel associations for fasting proinsulin levels: rs150781447 encoding TBC1D30 Arg279Cys (MAF=2.0%, P=5.5×10⁻¹¹) and rs3824420 encoding KANK1 Arg667His (MAF=3.0%, P=1.3×10⁻⁸). The TBC1D30 variant was most strongly associated with late-phase proinsulin-to-insulin conversion (proinsulin AUC_30–120; P=1.3×10⁻¹⁶) and the KANK1 variant with early-phase proinsulin-to-insulin conversion (proinsulin AUC_0–30; P=1.6×10⁻⁹) (Table 2, Supplementary Figure 1). The TBC1D30 variant effect is large, with each additional copy of the minor allele resulting in an average increase of 0.50 SD in proinsulin AUC_30–120 (Table 2, Supplementary Figure 2). This variant explained 0.94% of the trait variability (95% CI = 0.55–1.44%). We also observed a novel locus for insulin secretion as measured by the insulinogenic index, represented by nonsynonymous SNPs in PAM (smallest P=1.9×10⁻⁸) and PPIP5K2, located 200 kb apart, both with MAF=5.3%, and in near-perfect LD (r²=0.997) (Table 2, Supplementary Figures 1,2, and 3).

Table 2.

Novel loci for insulin processing and secretion

SNP	Gene	Variant	Chr	Position	Minor/major allele	MAF	Lead trait	β̂ ±SE	Effect size in SD units ± SE	Proportion of trait variance explained	P value
Identified by low-frequency variants
rs150781447	TBC1D30	R279C	12	65,224,220	T/C	.020	Proinsulin AUC30–120	.204 ± .025	.50 ± .06	.0094	1.3 × 10−16
rs3824420	KANK1	R667H	9	712,766	A/G	.029	Proinsulin AUC0–30	.107 ± .018	.28 ± .05	.0045	1.6 × 10−9
rs35658696	PAM	D563G	5	102,338,811	G/A	.053	Insulinogenic index	−.152 ± .027	−.21 ± .04	.0044	1.9 × 10−8
rs36046591	PPIP5K2	S1228G	5	102,537,285	G/A	.053	Insulinogenic index	−.152 ± .027	−.21 ± .04	.0043	2.3 × 10−8
Identified by common variants
rs2650000	HNF1A	intergenic	12	121,388,962	A/C	.455	Insulinogenic index	−.076 ± .012	−.10 ± .02	.0054	5.0 × 10−10
rs505922	ABO	intronic	9	136,149,229	C/T	.471	Disposition index	−.038 ± .006	−.09 ± 02	.0043	3.8 × 10−9
rs60980157	GPSM1	S391L	9	139,235,415	T/C	.300	Insulinogenic index	.072 ± .013	.10 ± .02	.0041	1.4 × 10−8

Chr, chromosome; MAF, minor allele frequency; SE, standard error; SD, standard deviation; CI, confidence interval. Positions are from NCBI Build 37 with allele labels from the forward strand. The lead trait is the trait with smallest P value. Traits were log-transformed and adjusted for BMI, age, and age². Effects are reported for the minor allele. The SNPs at PAM and PPIP5K2 are tightly linked (D′=0.999, r²=0.997). Based on analysis of 8,103-8,191 non-diabetic males.

Common SNPs at GPSM1, HNF1A, and ABO, previously associated with other traits, are here associated with insulin secretion or beta-cell function in non-diabetic individuals (Table 2, Supplementary Figure 3). GPSM1 Ser391Leu is in LD with noncoding rs3829109 (r²=.69), previously associated with fasting glucose². At ABO, the T allele of rs505922 is a proxy for the O blood group and has been associated with diverse phenotypes, including decreased pancreatic cancer risk¹⁷ and increased risk for duodenal ulcer¹⁸. Near HNF1A, rs2650000 was previously associated with LDL-cholesterol¹⁹ and C-reactive protein²⁰; other HNF1A variants are associated with MODY3 (MIM #600496) and type 2 diabetes risk²¹.

TBC1D30 and KANK1 both function in G protein signaling and are strong biological candidates. TBC1D30 (TBC 1 domain family, member 30) encodes a GAP protein that likely regulates activity of specific Rab GTPases including RAB3A²² and RAB8A²³. Rab3A knockout mice show a severe decrease in glucose-induced first phase insulin release and a 75% decrease in plasma insulin levels, without insulin resistance²⁴. The reference arginine at rs150781447 is well-conserved across vertebrate species and the cysteine substitution is predicted to be damaging²⁵ (Supplementary Table 5). The variant is located within a Rab-GAP domain and within the Kozak sequence of one TBC1D30 isoform and may alter translation initiation.

KANK1 (KN motif and ankyrin repeat domain-containing protein 1) plays a role in cytoskeleton formation by regulating actin polymerization²⁶ and negatively regulates Rac1 and RhoA G protein signaling, pathways that have been implicated in insulin secretion^27,28. At rs3824420, the reference arginine is not well conserved across species and the protein structure is predicted to tolerate the histidine substitution without an effect on function; this variant may still affect KANK1 or may tag another nearby variant. While rs3824420 has low frequency in Europeans (MAF=2.9% in Finns), it is common in East Asians (MAF=16%; Supplementary Table 7).

PAM encodes peptidylglycine alpha-amidating monooxygenase, an essential secretory granule membrane enzyme that catalyzes α-amidation of peptide hormones such as proinsulin²⁹. Older mice heterozygous for Pam deficiency exhibit glucose intolerance³⁰. At rs35658696, the reference aspartic acid is well conserved across vertebrates and located in one of the catalytic domains, and the glycine substitution is predicted to be damaging (Supplementary Table 5). The nearby gene PPIP5K2 is involved in cell signaling but has no known connection to insulin pathways. At rs36046591 in PPIP5K2, in near-perfect LD (r²=0.997) with rs35658696, the glycine substitution is predicted to be tolerated and the reference serine is not well conserved across species. This difference suggests that the PAM variant is causal at that locus, rather than the PPIP5K2 variant, but it is impossible to dissect them genetically.

Next, we carried out gene-based tests to investigate further the role of rare and low-frequency variants in insulin secretion and processing. Gene-based tests offer an alternative to single-variant tests, which are often underpowered to detect association with rare variants. Tests were performed on trait residuals adjusted for relatedness and covariates (see Online Methods). To address the impact of less common and rare variants, we considered only SNPs with MAF<3% or MAF<1%. In total, we tested 10,515 genes having at least two such variants using the SKAT-O test³¹.

We found significant associations between fasting proinsulin and TBC1D30, SGSM2, and ATG13 when using a MAF upper bound of 3% (Table 3, Supplementary Figure 4); by conditioning on the low-frequency variants detected by single-variant analysis, we demonstrated that these signals are driven by low-frequency variants. After adjusting for the common and nonsense variant signals at MADD, significance for ATG13, ~609 kb away, decreased by orders of magnitude (Table 3), showing that this signal is partially driven by the MADD variants and suggesting that other variants in this region await identification or that we may be adjusting for imperfect proxies of the causal variant. No additional associations were detected with other traits, including type 2 diabetes (data not shown).

Table 3.

Genes associated with fasting proinsulin identified by gene-based tests of aggregated low-frequency nonsynonymous variants with MAF < 3%

Gene	Number of Variants	Variants (minor allele counts)	P value	Conditional P value
TBC1D30	2	R279C(324), P746L(427)	3.3 × 10−9	.75a
SGSM2	3	Y416C(78), T789P(3), V996I(236)	2.0 × 10−9	.68b
ATG13	7	L5V(20), I131V(1), Q249P(3), R392W(1),	1.8 × 10−8	.0055d
		L427Q(3), G434R(488), X406G(200)c		37e

Fasting proinsulin was log-transformed and adjusted for fasting insulin, BMI, age, and age². Residuals were adjusted for relatedness and gene-based testing was carried out using the SKAT-O test (see Online Methods). This analysis was based on 8,224 participants. Reported associations were significant after Bonferroni correction for testing 19 traits for 10,515 genes (significance threshold: 2.5 × 10⁻⁷).

^aAfter adjusting for the low-frequency nonsynonymous variant R279C (rs150781447) at TBC1D30 (MAF 2.0%).

^bAfter adjusting for the low-frequency nonsynonymous variant V996I (rs61741902) at SGSM2 (MAF 1.4%).

^cAnnotation relative to a non-canonical (longer) isoform.

^dAfter adjusting for lead SNPs of common GWAS signals (rs7944584 and rs1051006) and nonsense variant rs35233100 at MADD (MAF 3.7%).

^eAfter adjusting for the low-frequency nonsynonymous variant G434R (rs35619591) at ATG13 (MAF 3.0%).

In summary, we identified two low-frequency coding variants in genes at known loci and three novel genes with low-frequency variants associated with insulin processing or secretion. At least four of these genes play roles in G-protein signaling (Supplementary Figure 5). We show that the interpretation of both single-variant and gene-based tests needs to consider the effects of distant common SNPs, an especially important consideration when exome sequence data are analyzed without data on the surrounding noncoding regions. Although regions of long-range LD are unusual, at least 24 have been reported¹⁵ to extend >1 Mb in Europeans, a distance frequently used to claim independence of association signals in GWAS meta-analyses. Several of the identified exome array variants are plausibly functional, although ~25% and ~28% of low-frequency nonsynonymous variants on the exome array were annotated as conserved and plausibly damaging, respectively (Supplementary Table 2), and the exome array does not provide complete coverage of all functional variants at each locus. By the content of the exome array, this study was also limited in its ability to look at very rare variants. While sequencing will still be required to completely assess variants associated with insulin processing, secretion, and glycemic traits, this study provides proof-of-principle that exome array genotyping is a powerful approach to identify low-frequency functional variants and fine-map GWAS-identified loci in complex traits.

ONLINE METHODS

Study participants

We attempted exome array genotyping of 9,717 participants in the Metabolic Syndrome in Men (METSIM) study⁴. Male study participants were randomly selected from the population register of Kuopio, Eastern Finland (population 95,000). Participants undertook a 1-day outpatient visit to the Clinical Research Unit at the University of Kuopio. Participants with diagnosed type 1 diabetes or type 2 diabetes (previously diagnosed, on diabetes medication, with fasting glucose ≥ 7 mmol/l, or with 2-hour glucose ≥ 11.1 mmol/l) were excluded from quantitative trait analysis. Clinical characteristics of non-diabetic study participants are provided in Supplementary Table 1. The study was approved by the ethics committee of the University of Kuopio and Kuopio University Hospital; informed consent was obtained from all study participants.

OGTT and laboratory measurements

Clinical testing was performed following a 12-h overnight fast. A 2-h oral 75 g glucose tolerance test (OGTT) was performed with blood samples drawn at 0, 30, and 120 min, for measurement of plasma proinsulin, insulin, and glucose levels. Plasma specific proinsulin (Human Proinsulin RIA kit; Linco Research, St. Charles, MO; no cross-reaction with insulin or C-peptide) and insulin (ADVIA Centaur Insulin IRI, No. 02230141; Siemens Medical Solutions Diagnostics, Tarrytown, NY; minimal cross-reaction with proinsulin or C-peptide) were measured by immunoassay, and plasma glucose by enzymatic hexokinase photometric assay (Konelab System Reagents, Thermo Fisher Scientific, Vantaa, Finland).

Phenotypes

Association results are reported for five traits: fasting proinsulin (adjusted for fasting insulin), early-phase (ProinsAUC_0–30) and late-phase (ProinsAUC_30–120) glucose-stimulated proinsulin-to-insulin conversion measured as proinsulin area under the curve (AUC) during the first 30 min and remaining 90 min of an OGTT, insulin secretion assessed by the insulinogenic index³², and a disposition index measure of β-cell compensation for insulin resistance defined as InsAUC_0–30/GluAUC_0–30 × Matsuda index of insulin sensitivity (Matsuda ISI)^4,33. The reported associations were discovered by analyzing a total of 19 traits. Other measures of β-cell function included: oral glucose-stimulated proinsulin-to-insulin conversion during the first 30 min (ProinsAUC_0–30/InsAUC_0–30), and 30–120 min (ProinsAUC_30–120/InsAUC_30–120) of the OGTT, unadjusted fasting proinsulin, fasting proinsulin/insulin ratio, HOMA-β³⁴, fasting insulin, insulin at 120 min, insulin AUC during the first 30 min (InsAUC_0–30) and during 30–120 min (InsAUC_30–120), and early-phase glucose-stimulated insulin release (InsAUC_0–30/GluAUC_0–30) adjusted for Matsuda ISI³⁵. Indices of insulin sensitivity included HOMA-IR³⁴ and the Matsuda ISI³⁶. Associations with fasting and 120 min glucose were also tested. Supplementary Figure 6 shows correlations among traits. We calculated AUC measures using the trapezoid rule.

Exome array

The Illumina HumanExome-12v1_A Beadchip includes 247,870 markers focused on protein-altering variants selected from >12,000 exome and genome sequences representing multiple ethnicities and complex traits. Nonsynonymous variants had to be observed three or more times in at least two studies, splicing and stop-altering variants two or more times in at least two studies. Additional array content includes variants associated with complex traits in previous GWAS, HLA tags, ancestry informative markers, markers for identity-by-descent estimation, and random synonymous SNPs. Details about SNP content and selection strategies can be found at the exome array design webpage (see URLs).

Genotyping and quality control

9,717 study samples, 104 blind duplicate samples, and 116 HapMap samples of different ethnicities were genotyped at the Genetic Resources Core Facility (GRCF) at the Johns Hopkins Institute of Genetic Medicine. Genotype calling was carried out using Illumina's GenTrain version 1.0 clustering algorithm in GenomeStudio version 2011.1. Cluster boundaries were determined using study samples. After clustering, 5,574 non-autosomal and 3,379 autosomal variants identified through filtering strategies developed at GRCF were manually reviewed and clusters edited as necessary. After technical failure and marker-level quality control, 242,458 of 247,870 (97.8%) attempted markers were successfully genotyped and had call rate >95% (average call rate 99.95%).

We evaluated genotyping quality using concordance rates for HapMap samples genotyped in our study and (a) sequenced by Complete Genomics or the 1000 Genomes Project (on-target regions of integrated phase 1 release; see URLs) or (b) genotyped on the Illumina HumanOmni2.5 Beadchip by the 1000 Genomes Project. These comparisons were based on 60,574, 117,063, and 39,056 overlapping variants and 17, 49, and 86 individuals, respectively. Overall concordance rates were 99.933%, 99.972%, and 99.956% for Complete Genomics and 1000 Genomes sequence data, and HumanOmni2.5 Beadchip data, respectively. Considering the external data as truth, concordance rates for homozygous genotypes were 99.982%, 99.987%, and 99.974%, and for heterozygous genotypes were 99.678%, 99.529% and 99.886%, respectively.

In total, 9,660 of 9,717 (99.4%) individuals were successfully genotyped (call rate > 98%). For the 242,458 SNPs that passed quality control, genotype concordance among the 104 blind duplicate sample pairs was 99.998%. Three sex-mismatched individuals were identified and excluded from subsequent analyses. One individual per pair of 6 known twin pairs and 6 unexplained apparent duplicates were excluded.

We carried out principal components analysis (PCA) twice, once excluding HapMap samples to identify population outliers, and then including HapMap samples to help interpret outliers. To avoid artifactual results due to family relatedness³⁷, we computed principal components using SNP loadings estimated from a subset of 7,304 not-close-relatives. We defined close relatives as ones for whom estimated genome-wide identical-by-descent (IBD) proportion of alleles shared was >0.10. We estimated IBD sharing using PLINK's "--genome" option³⁸ and carried out PCA using SMARTPCA³⁷ on a linkage-disequilibrium-pruned set of 22,464 autosomal SNPs obtained by removing large scale high-LD regions^15,39, SNPs with a MAF < 0.01, or SNPs with HWE P value < 10⁻⁶, and carrying out LD pruning using the PLINK option: "--indep-pairwise 50 5 0.2". Inspecting the first 10 PCs, we identified 12 population outliers, 9 of whom had self-reported non-Finnish ancestry; we excluded these 12 individuals from subsequent analysis. After further removal of 25 individuals with diagnosed type 1 diabetes, 1,376 with type 2 diabetes, and 3 with missing phenotypes, 8,229 individuals remained for quantitative trait analysis.

Statistical analysis

Single-variant analysis

We tested for trait-SNP association assuming an additive genetic model using a linear mixed model to correct for relatedness using EMMAX⁵. We excluded SNPs with MAF < 0.05% or Hardy-Weinberg equilibrium (HWE) P value < 10⁻⁶. To reduce the impact of outliers, we log-transformed traits with skewed distributions and then Winsorized all traits at 5 standard deviations from the mean. All traits were adjusted for BMI, age, and age² prior to association testing. We analyzed both untransformed residuals and rank-based inverse-normal transformed residuals to assess robustness of association results to distributional assumptions. Since no appreciable differences were observed between the two analyses, we report results for untransformed residuals. Finally, we visually inspected genotype cluster plots and checked HWE P values for all described variants. The lowest HWE P value for a reported novel associated variant was 0.09.

Population stratification

To correct for population stratification, we modeled population structure as part of the random effects, indistinguishable from the relatedness effect⁵. To investigate residual population stratification, we calculated genomic control inflation factors⁴⁰ and inspected quantile-quantile (QQ) plots for test statistics both before and after removal of established and newly discovered loci (2 Mb segments centered on lead SNPs) (Supplementary Figure 7).

Conditional analysis

To identify additional association signals after accounting for the effects of known and newly discovered trait loci, we carried out conditional analyses where we included the allele count at the lead SNP(s) at the conditioning loci as covariate(s). To allow discovery of more than two association signals per locus, we used a stepwise procedure where additional SNPs were added to the model according to their conditional P value, as programmed in EMMAX⁵. We estimated LD metrics r² and D' using 9,633 METSIM individuals who passed genotyping quality control. LD with SNPs not included on the exome array was determined based on whole-genome sequence data for 1,479 Northern European individuals.

Gene-based analysis

For gene-based testing we used the SKAT-O³¹ test which encompasses burden tests and SKAT⁴¹ as special cases. SKAT-O has been shown to perform well under a range of scenarios, including scenarios in which protective, deleterious, and null variants are present, and in which a large number of variants are causal and associated in the same direction³¹. To account for relatedness, we adopted an approach similar to GRAMMAR⁴² by first obtaining trait residuals adjusted for relatedness using GenABEL⁴³ and then carrying out gene-based testing. We performed analyses using default weights³¹ and MAF upper bounds of 1% and 3% for the combination of nonsynonymous, stop-altering, and splice-site variants. In total, 10,515 genes with at least two variants were tested. The results of the naive SKAT-O analysis and the analysis adjusted for relatedness were highly correlated (Supplementary Figure 8). To evaluate whether common or low-frequency SNPs associated with the trait in the single-variant analysis can account for a gene-based test signal, we also carried out conditional analyses by including the allele count at such SNP(s) as covariate(s).

Statistical significance

We declared a single variant-trait association significant if the nominal P value was < 4.46 × 10⁻⁸, corresponding to a Bonferroni correction for 1,121,551 tests (19 phenotypes × 59,029 variants). We declared a gene-based test association significant if the nominal P value was < 2.50 × 10⁻⁷, corresponding to a Bonferroni correction for 199,785 tests (19 phenotypes × 10,515 genes).

Annotation

We annotated variants relative to GENCODE version 7 coding transcripts⁴⁴ using in-house developed software (unpublished). Amino acid substitution positions are relative to the canonical UniProt protein sequence⁴⁵.

URLs

Exome array design, http://genome.sph.umich.edu/wiki/Exome_Chip_Design;

Complete Genomics 69 Genomes Data, http://www.completegenomics.com/public-data/69-Genomes;

The 1000 Genomes Project, www.1000genomes.org;

PLINK, http://pngu.mgh.harvard.edu/~purcell/plink;

SMARTPCA, http://genetics.med.harvard.edu/reich/Reich_Lab/Software.html;

EMMAX, http://genetics.cs.ucla.edu/emmax;

GenABEL, http://www.genabel.org