Transethnic insight into the genetics of glycaemic traits: fine-mapping results from the Population Architecture using Genomics and Epidemiology (PAGE) consortium

This study was performed in accordance with the tenets of the Declaration of Helsinki and approved by the Institutional Review Boards of each participating study. All study participants provided written informed consent.

The PAGE consortium was funded by the National Human Genome Research Institute (NHGRI) to investigate the epidemiological architecture of well-replicated genetic variants associated with human diseases or traits [12]. This analysis includes self-reported H/L, AA, ASN and AI/AN individuals without diabetes, aged 18 years or over, from the Multiethnic Cohort Study (MEC), the Women’s Health Initiative (WHI), Atherosclerosis Risk in Communities (ARIC), Coronary Artery Risk Development in Young Adults (CARDIA), the Hispanic Community Health Study/Study of Latinos (HCHS/SOL) and the Mount Sinai School of Medicine’s (MSSM) DNA biobank (BioMe). Further details about each cohort can be found in the electronic supplementary materials (ESM) Methods (study population and trait measurement section).

Fasting glucose and fasting insulin concentrations were measured using standard assays, at laboratories specific to each PAGE site (ESM Table 1). Individuals self-reporting that they had ever been diagnosed with diabetes or taken diabetes medications or who had fasting blood glucose levels ≥ 6.99 mmol/l (≥ 126 mg/dl) were excluded from analyses. Individuals with BMI < 16.5 kg/m² or BMI > 70 kg/m² were also excluded on the assumption that these extremes could be attributable to data coding errors or underlying illness or could reflect a familial syndrome. Prior to analyses, each study removed race/ethnicity outliers using ancestry informative principal components.

After exclusions, fasting glucose analyses consisted of 14,953 H/L, 10,380 AA, 998 ASN and 429 AI/AN individuals. Fasting insulin analyses involved fewer individuals: 12,895 H/L, 8361 AA, 998 ASN and 420 AI/AN. Fasting insulin was not available for BioMe. Race/ethnicity was self-reported. Descriptive characteristics of PAGE study participants by cohort can be found in ESM Table 2. While ASN and AI/AN were included for transethnic meta-analysis, population-specific analyses were underpowered due to small sample sizes. As such, ASN and AI/AN population-specific analyses were used as a comparison for consistency in the direction of effect.

Genotyping was performed using the Metabochip, the design of which has been described elsewhere [10]. In brief, the 200K Metabochip is designed to cost effectively analyse putative association signals identified through GWAS of many glucose- and insulin-related metabolic and cardiovascular traits and to fine-map established loci [10]. More than 122,000 SNPs were included to fine-map 257 GWAS loci for 23 traits [10]. Fine-mapping loci were defined as the GWAS index SNP and all correlated SNPs (r ² ≥ 0.5) that were within 0.02 cM of the index and having a minor allele frequency (MAF) > 1% in at least one HapMap Phase I population. SNPs were excluded if the Illumina design score was < 0.5 or there were SNPs within 15 bp of the SNP of interest with MAF of > 2% among Europeans (CEU [HapMap Population Code for Utah residents (CEPH) with Northern and Western European ancestry]).

Metobochip genotyping was performed for MEC, ARIC, CARDIA, HCHS/SOL and WHI [13] individuals. Standard quality control filters were applied for samples and SNPs, including missing rate and Hardy–Weinberg equilibrium (p < 1 × 10⁻⁷). A portion of WHI individuals of AA had both Metabochip and the Affymetrix 6.0 genotype data available from the SNP Health Association Resource (SHARe); this was used to impute Metabochip SNPs in the remaining SHARe participants with only Affymetrix 6.0 GWAS [8] and only dosages with imputation R ² > 0.3 were included in the analyses. In BioMe, genotypes from the Illumina HumanOmniExpress array were imputed to 1000 Genome Phase I haplotype panels (March 2012) [14]. Metabochip SNPs with ‘proper info’ score ≥ 0.4 were included in the analysis. Principal components were determined within each study using the Eigensoft software [15]. We excluded SNPs with a minor allele count less than 5 within each study by racial/ethnic population. The sample success rate and concordance rate for duplicate pairs across all studies was ≥ 95% and ≥ 99%, respectively. Further genotyping and analytical characteristics of the participating studies are further summarised in ESM Methods (genotyping and quality control section) and ESM Table 1.

The overall study design for replication, fine-mapping and discovery of novel loci is summarised in Fig. 1. For replication of known loci, unconditional association analyses were performed for previously reported index SNPs listed in ESM Table 3. A nominal significance level (α = 0.05) was used to define replication of a locus. Next, unconditional association analyses were performed for all SNPs in a locus by race/ethnicity and by transethnic meta-analysis. A locus-specific p value threshold was defined as 0.05 divided by the number of SNPs passing quality control in each region (ranging from α = 1.4 × 10⁻⁵ to α = 4.1 × 10⁻⁴, Table 1). Locus-specific significance was used to conservatively adjust for multiple testing, while also acknowledging that genetic variation is known to influence glycaemic traits in these regions. Linkage disequilibrium was calculated for PAGE H/L, AA and Asian samples with 500 kb sliding windows using PLINK [16]. Metabochip linkage disequilibrium and frequency information in Europeans was provided by the 1000 Genomes Phase 3 population. These linkage disequilibrium patterns were used to evaluate locus refinement. Additionally, LocusZoom plots [17] were used to graphically display the fine-mapping results and linkage disequilibrium for these plots used 1000 Genomes Phase I Super Populations (European ancestry [EUR], admixed American ancestry [AMR], African ancestry [AFR]). After identifying the most significant lead SNP in each region, we searched for additional independent association signals by including the lead SNP in the conditional model and then testing each of the remaining SNPs in a region. These conditional analyses were repeated, adding in the lead SNP and conditional lead SNP(s), until no SNP in the model had a conditional p value less than the locus-specific significance. Sequential conditional analyses were performed for each race/ethnicity and transethnic meta-analysis. Further details on our approach to locus refinement are provided in ESM Methods (replication and fine-mapping of known glycaemic trait loci section).

Metabochip-wide analyses were performed to identify novel associations with fasting glucose and fasting insulin. Statistical significance for the Metabochip-wide analysis was set at 0.05 divided by the number of Metabochip SNPs passing quality control (α = 2.7 × 10⁻⁷). Results were examined through qq plots and Manhattan plots for each model, highlighting known regions defined in ESM Table 4. Further details are provided in ESM Methods (strategy for selecting novel associations section).

First, in each study with unrelated individuals we performed race/ethnic-specific analyses for fasting glucose and natural log-transformed fasting insulin, excluding ancestry outliers and first-degree relatives. In HCHS/SOL, a weighted version of generalised estimation equations was used to account for unequal inclusion probabilities and complex family-based sampling designs [18]. Models adjusted for age, sex (except WHI), study site (as applicable), smoking status (current vs former/never), continuous BMI and ancestry principal components. Like previous studies [11], primary analyses adjusted for BMI because it is a major risk factor for type 2 diabetes and is correlated with glycaemic traits. For comparison, all models were also run without adjustment for BMI. Next, fixed-effect models with inverse-variance weighting were used to pool the study-specific SNP effect estimates and their standard errors by race/ethnicity as implemented in METAL [19]. Finally, summary statistics from METAL for H/L, AA, NA/AI and ASN were combined using inverse-variance weighted fixed effects meta-analysis in METAL. Q statistics and I ² were used to evaluate heterogeneity across studies and race/ethnicity. Further details are provided in ESM Methods (statistical analysis section).

Detailed information on the functional annotation methods and various datasets used is provided in ESM Methods (functional annotation section). In brief, it is expected that the lead SNPs are more likely to be functional or to be in stronger linkage disequilibrium with underlying functional variant(s). Therefore, lead SNPs and all correlated SNPs (r ² > 0.2 in 1000 Genomes Phase 3 AFR/AMR populations) were annotated using publicly available functional datasets. Potential functional effects were assessed using PolyPhen2 [20] (http://​genetics.​bwh.​harvard.​edu/​pph2/​, accessed 24 August 2016) for non-synonymous variants, SPANR (http://​tools.​genes.​toronto.​edu/) [21] for variants near splice sites, TargetScan miRNA Regulatory Sites for 3′-UTR regions [22], ENCODE/NIH Roadmap data [23‐25] and GTEx (https://​www.​gtexportal.​org/​home/​) [26] to identify non-coding variants positioned in predicted regulatory elements.

We included a total of 26,760 participants (14,953 H/L, 10,380 AA, 998 ASN, and 429 AI/AN) in fasting glucose analyses. The sample sizes for fasting insulin analyses were slightly smaller, with a total of 22,674 participants (12,895 H/L, 8361 AA, 998 ASN and 420 AI/AN). The mean age across the five cohorts was 55 years for men and 59 years for women (range 18–93 years). Study-specific descriptive characteristics are shown in ESM Table 2. Particularly due to the inclusion of the WHI cohort, the proportion of women in the total study population was high, with the highest fraction observed in AA (82.6% for fasting glucose and 97.1% for fasting insulin). Glycaemic trait distributions were similar across studies and ethnicities, with average fasting glucose levels ranging from 4.7 ± 0.7 mmol/l to 5.5 ± 0.6 mmol/l and average fasting insulin levels ranging from 43.3 ± 23.6 pmol/l to 75.9 ± 38.8 pmol/l.

We found that 31/39 (79.5%) fasting glucose loci and 14/17 (82.3%) fasting insulin loci had a p value smaller than 0.05. Index SNP associations were directionally consistent in our transethnic PAGE meta-analysis and only four SNPs had heterogeneity p values less than 0.05 (Table 2). The effect estimates (βs) of index SNPs in the transethnic meta-analysis were very similar to those published in Metabochip analysis of individuals of European descent (Pearson’s r ² = 0.86, 95% CI 0.78, 0.91; p < 2.2 × 10⁻¹⁶; ESM Fig. 1). At three loci (WARS, GIPR and DPYSLS) we observed replication in only H/L and not the transethnic meta-analysis. Interestingly, while the sample sizes were much smaller for Asian individuals than for H/L and AA individuals, the transethnic meta-analysis of the PROX1 index (rs340874) was only nominally significant and directionally consistent in the Asian samples. In the remaining loci that did not replicate in transethnic meta-analysis or the race/ethnic-specific analyses, the effects were generally similar or at least in the same direction. Analyses without inclusion of BMI as a covariate were generally similar, with slightly lower significance at some loci. Full summary statistics for models with and without BMI covariate are reported in ESM Table 5 and ESM Table 6, respectively.

Among the 15 glycaemic trait loci for which fine-mapping was attempted on the Metabochip, ten fasting glucose loci and two fasting insulin loci had one or more SNPs that reached locus-specific significance (α = 0.05/number of SNPs in the locus) in the transethnic meta-analysis. The p values ranged from 1.0 × 10⁻²⁹ at G6PC2-rs560887 to 1.5 × 10⁻⁴ at PROX1-rs10494973 (Table 3). Although AI/AN ancestries were included in the transethnic meta-analysis, the AI/AN results are not shown because the small sample size was underpowered for population-specific analysis. At four fasting glucose loci, the most significant lead SNP in PAGE transethnic meta-analysis was the same as the European index SNP from prior Metabochip evaluation (G6PC2, ADCY5, MTNR1B and FADS2). For six fasting glucose loci (PROX1, GCKR, SLC2A2, DGKB, GCK and GLIS3) and the one fasting insulin locus (GCKR), the lead SNP in PAGE transethnic meta-analysis was in moderate or weak linkage disequilibrium with the index SNP in 1000 Genomes Population EUR (r ² > 0.2). At these six fasting glucose loci and one fasting insulin locus, the PAGE lead SNP and EUR index SNP were not independent of each other as only one of the two SNP associations maintained nominal significance in transethnic conditional meta-analysis where both lead and index variants were included in the model. This was further supported by investigation of potential fine-mapping through locus zoom plots.

For each of the 11 glycaemic trait loci with potential transethnic fine-mapping (fasting glucose loci–PROX1, G6PC2, ADCY5, MTNR1B, FADS2, GCKR, SLC2A2, DGKB, GCK and GLIS3; fasting insulin locus–GCKR), we found that the number of SNPs in linkage disequilibrium with the most significant marker in the transethnic results (r ² ≥ 0.2 in the 1KG super populations AFR and AMR) were less than the number of SNPs tagged by the EUR marker (r ² ≥ 0.2 in EUR). Visual inspection of locus zoom plots indicated that transethnic meta-analysis refined each of these loci by reducing the number of highly correlated SNPs reaching the same level of significance and/or narrowing the genomic region containing putative causal SNPs (ESM Fig. 2). On average, the number of variants in high linkage disequilibrium was reduced by 72.5% with the number of linkage disequilibrium SNPs ranging from one at MTNR1B to 162 at SLC2A2 in the PAGE transethnic meta-analysis results. Refinement was most evident at the SLC2A2 locus (Fig. 2). Bioinformatic functional follow-up was performed for each of the eleven glycaemic trait loci with one or more variants passing the region-specific significance threshold in our transethnic meta-analysis. We observed an overlap of promoter and enhancer sequences at each locus and identified potential target genes. These data not only provided further support for the fine-mapping results but also revealed additional insights into the aetiology of glycaemic traits. UCSC Genome Browser images of each locus are provided in ESM Fig. 3. The results of our in silico functional annotations are summarised in ESM Table 7.

To identify additional independent association signals at significant loci, conditional analyses were performed. Results of these analyses and population-specific associations are shown in Table 4. For transethnic conditional meta-analyses, ten fasting glucose loci and two fasting insulin loci were analysed. Independent secondary associations were identified at two fasting glucose loci (G6PC2-rs477224 and GCK-rs2908286). The second round of conditional analyses did not identify significant tertiary signals. Bioinformatic follow-up of rs477224 suggested that the variant is positioned within a pancreatic islet enhancer. The rs2908290 variant was in weak linkage disequilibrium (AMR r ² = 0.26, AFR r ² = 0.23) with a variant, rs2971677, predicted to alter splicing efficiency of GCK.

To identify population-specific loci, we conducted separate conditional analyses for significant loci in the primary H/L (GCKR-rs1260326, G6PC2-rs560887, SLC2A2-rs1280, DGKB-rs1005256, GCK-rs1799884, FADS3-rs12577276, MTNR1B-rs10830963, C2CD4A-rs7167881), AA (G6PC2-rs77719485, GCK-rs2908286, CRY2-rs117493014, MADD-rs77082299, ADCY5-rs11708067, MTNR1B-rs10830963) and Asian populations (GLIS3-rs4395942). A population-specific variant was detected in the AA analysis of the G6PC2 locus. The lead fasting glucose SNP, rs77719485, is less frequent in AA population (MAF 2.4%) and rare or monomorphic in the other populations (MAF 0.4% in H/L). Like the transethnic lead SNP, rs560887, bioinformatic follow-up suggested that rs77719485 may affect splicing efficiency for exon 4 for G6PC2.

In secondary analyses, we conducted a Metabochip-wide scan to identify potential novel or pleiotropic variants, given that the chip included variants with suggestive signals in established loci for many known metabolic traits. Models were run with and without BMI as a covariate (ESM Table 8, ESM Figs 4,5). Using the Bonferroni significance threshold (0.05/182,055 = 2.7 × 10⁻⁷), we identified one novel association for fasting insulin (rs75862513, p = 4.3 × 10⁻⁸, Fig. 3) at the SLC17A2 locus previously associated with height and blood pressure [27, 28]. After BMI adjustment (ESM Fig. 5), the association was attenuated suggesting that the effects may be mediated by BMI.