Common Variants at 10 Genomic Loci Influence Hemoglobin A_1C Levels via Glycemic and Nonglycemic Pathways

Cohort description, study design, and genotyping.

The cohorts included in this study were part of MAGIC (19). The characteristics of the population samples used in this analysis are shown in Table 1. All participants were adults of European ancestry from Europe or the U.S., and free of diabetes as assessed by either clinical diagnosis, self-reported diabetes, diabetes treatment, or undiagnosed diabetes defined by FG ≥7.0 mmol/l. HbA_1c (in percentages) was measured in all studies from fasting or nonfasting whole blood using NGSP-certified methods. We found remarkably consistent means and SD across studies, increasing confidence that laboratory variability had a minimal effect on the study results. A local research ethics committees approved all studies and all participants gave informed consent.

We carried out a meta-analysis including 35,920 participants from 23 cohorts with available HbA_1c measurements and genotype data including ∼2.5M genotyped and imputed autosomal SNPs. This sample size ensures 80% power to detect SNPs, explaining 0.12% of the trait variance at α = 5 × 10⁻⁸. For 5 SNPs (rs1046896, rs16926246, rs1799884, rs1800562, and rs552976) that had been previously selected from an interim analysis of the first 10 participating cohorts (n = 14,898), we obtained further data by genotyping up to 10,448 participants from 8 additional cohorts. The sample size for each SNP is thus related to the number of cohorts that were genotyped (up to 31) and to the specific call rate. Details on genotyping methodology, quality control metrics, and statistical analyses for each cohort are shown in supplementary Table S1 in the online appendix available at http://diabetes.diabetesjournals.org/cgi/content/full/db10-0502/DC1. Additional details on imputation and quality control applied by each study are given in the online supplementary methods.

Primary genome-wide association studies and meta-analysis.

In each cohort a linear regression model was fit using untransformed (percentage) HbA_1c as the dependent variable to evaluate the additive effect of genotyped and imputed SNPs. HbA_1c showed a mild deviation from normality in the majority of cohorts. Log-transformation did not significantly improve normality; nevertheless, such mild deviation did not result in an inflation of the test statistics suggestive of an excess of false positives, as indicated by a genomic correction λ very close to the expected value of 1.0; thus, we report untransformed (percentage) HbA_1c results. The model was adjusted for age, sex, and other cohort-specific variables as applicable. Further details are given in the supplementary methods and supplementary Table S1. Regression estimates for each SNP were combined across studies in a meta-analysis using a fixed effect inverse-variance approach (justified by nonsignificant heterogeneity of effect sizes at all validated loci), as implemented in the METAL software. The individual cohort analysis results were corrected prior to performing the meta-analysis for residual inflation of the test statistic using the genomic control method if the λ coefficient was >1.0 (20). Cohort-specific results for each of the 10 loci are given in supplementary Table S2. Heterogeneity across study-specific effect sizes was assessed using the standard χ² test implemented in METAL, Cochran's Q statistic and the I² statistics (21).

Association with related traits and diseases.

Secondary analyses were carried out on 10 SNPs (rs2779116, rs552976, rs1800562, rs1799884, rs4737009, rs16926246, rs1387153, rs7998202, rs1046896, and rs855791) reaching genome-wide significance and including only the stronger of the 2 significant ANK1 SNPs (see supplementary methods for additional information). A first goal was to detect “pleiotropic” effects on potentially related traits for the 10 loci. To this end we tested them for association with correlated intermediate traits (BMI, and glycemic and hematologic parameters, supplementary Table S3).

Further, we carried out association analyses of HbA_1c levels conditional on FG levels (Table 3) and hematologic parameters (supplementary Table S4) to formally test mediation by glycemia or erythrocyte traits. Mediation is used here to distinguish it from confounding. A confounder is a characteristic associated with both exposure and outcome but is not on the causal pathway linking the two together. By contrast, a mediator is also associated with both exposure and outcome, but is on the causal pathway that may explain the association between them. Our mediation analyses decompose the association between a SNP and HbA_1c into two paths. The first path links the SNP directly to HbA_1c, and the second path links the SNP to HbA_1c through a mediator, e.g., FG or hematologic parameters. A marked attenuation of the size of effect on HbA_1c of the SNP in the conditional “mediation” model implies that the SNP (e.g., rs552976) acts on the mediator (e.g., FG), which in turn acts on HbA_1c levels. Further details on these analyses are provided in the on-line supplementary methods.

Finally, we tested associations of the 10 loci with risk of type 2 diabetes or coronary artery disease (CAD) using adequately powered case-controlled meta-analyses. Association statistics with type 2 diabetes were obtained from a previous analysis of the MAGIC datasets or from the DIAGRAM+ meta-analysis (22). CAD associations were tested in this study using cohorts described in supplementary Table S5. The CAD analytic sample size assembled for this study had 80% power to detect associations at an α level of 5 × 10⁻⁸ for a genotype relative risk of 1.14, and a risk allele frequency of 0.2.

Estimates of genetic effect size.

We used several methods to evaluate the size of the genetic effect of HbA_1c-associated SNPs: 1) we used regression to estimate in percentages the total variance in HbA_1c explained by the 10 loci; 2) we calculated an additive genotype score based on the number of risk alleles at 7 (nonglycemic) or 10 (all) loci and then calculated the difference in HbA_1c (%) between individuals in the top 10% of the genotype score distribution and those in the bottom 10% (supplementary methods); and 3) we used net reclassification analysis to gauge the effect of individual genotype on HbA_1c distributions at the population level.

Net reclassification analysis.

Variation in the measured level of HbA_1c associated with nonglycemic genetic effects may affect the classification of individuals as diabetic or nondiabetic when screening general population samples using HbA_1c. We used this relationship as a way to understand the clinical influence of the HbA_1c loci when applied at the population level. We estimated the change in classification that occurred when accounting for effects of the seven loci presumed not to affect HbA_1c via primarily glycemic mechanisms (SPTA1, HFE, ANK1, HK1, ATP11A/TUBGCP3, FN3K, and TMPRSS6) using the method of Pencina et al. (23). For this analysis we combined the Framingham Heart Study (FHS), and Atherosclerosis Risk In Communities (ARIC) European ancestry cohorts (N = 10,110). ARIC and FHS have several characteristics suitable for this analysis: 1) they are population-based samples, thus allowing a test of diabetes screening in a truly unselected sample; 2) they are of large sample size, thus maximizing the number of diabetic subjects that can readily be folded back for reclassification analysis; 3) they have both fasting glucose and HbA_1c measured. We excluded as in previous analyses all individuals on diabetes treatment (diagnosed diabetes), but retained individuals with FG ≥7.0 mmol/l not on treatment (who we classified as having undiagnosed diabetes, N = 593) as well as all nondiabetic individuals (N = 9,517). We then sought to differentiate these individuals on the basis of their HbA_1c levels, using ≥6.5% as the cutoff indicating diabetes. We counted the cumulative frequency distribution of measured HbA_1c levels by diabetes status, then re-estimated the frequency distribution after regression analysis adjusting for the seven SNPs at the nonglycemic loci, recalibrating the distribution to have the same mean HbA_1c as in each original cohort. We counted the proportion of undiagnosed diabetic individuals with unadjusted HbA_1c ≥6.5% who had an adjusted HbA_1c <6.5%, and the proportion of nondiabetic individuals with unadjusted HbA_1c <6.5% who had an adjusted HbA_1c ≥6.5%. The difference between these proportions is called “net reclassification” and in this instance indicates the overall proportion of a population whose diagnostic status would change based on the influence of these seven common, nonglycemic genetic variants.

New common variants associated with HbA_1c.

We carried out a meta-analysis of SNP associations with HbA_1c levels in up to 46,368 participants of European ancestry from 31 cohorts. We identified 10 genomic regions associated with HbA_1c levels (Table 2, Figs. 1 and 2). Six associated regions were new, including FN3K (rs1046896, P = 1.57 × 10⁻²⁶), HFE (rs1800562, P = 2.59 × 10⁻²⁰), TMPRSS6 (rs855791, P = 2.74 × 10⁻⁴), ANK1 (rs4737009, P = 6.11 × 10⁻¹²), SPTA1 (rs2779116, P = 2.75 × 10⁻⁹), and ATP11A/TUBGCP3 (rs7998202, P = 5.24 × 10⁻⁹). A second, independent SNP near ANK1 was also associated with HbA_1c (rs6474359, P = 1.18 × 10⁻⁸; r² with rs4737009 = 0.0001; see also supplementary methods). In addition, SNPs in or near HK1 (rs16926246, P = 3.11 × 10⁻⁵⁴), MTNR1B (rs1387153, P = 3.96 × 10⁻¹¹), GCK (rs1799884, P = 1.45 × 10⁻²⁰), and G6PC2/ABCB11 (rs552976, P = 8.16 × 10⁻¹⁸) were associated with HbA_1c levels. These loci had previously been associated with HbA_1c (15,16), FG (9–12,14,15) and/or type 2 diabetes risk (9–12,15,16,19). Associations were generally similar across cohorts, showing no significant heterogeneity (Table 2). This lack of heterogeneity suggests that there is good consistency in trait measurement across different cohorts.

FIG. 1.

Manhattan plot and quantile-quantile (QQ) plot of association findings. The figure summarizes the genome-wide association scan results combined across all studies by inverse variance weighting. The blue dotted line marks the threshold for genome-wide significance (5 × 10⁻⁸). SNPs in loci exceeding this threshold are highlighted in green. A QQ plot is shown in the inset panel, where the red line corresponds to all test statistics, and the blue line to results after excluding statistics at all associated loci (highlighted in green in the Manhattan plot). The gray area corresponds to the 90% confidence region from a null distribution of P values (generated from 100 simulations). (A high-quality color representation of this figure is available in the online issue.)

FIG. 2.

Regional association plots at the HbA1c loci. Each panel spans ± 250 kb around the most significant associated SNP in the region, which is highlighted with a blue square (panel C spans ± 300 kb). At the top of each panel, comb diagrams indicate the location of SNPs in the Illumina HumanHap 550K and Affymetrix 500K chips, and of SNPs imputed. The SNPs are colored according to their linkage disequilibrium with the top variant based on the CEU HapMap population (http://www.hapmap.org). Gene transcripts are annotated in the lower box, with the most likely biologic candidate highlighted in blue; ± indicates the direction of transcription. In panel C, a few gene names were omitted for clarity. Here, genes are, from left to right, SCGN, HIST1H2AA, HIST1H2BA, SLC17A4, SLC17A1, SLC17A3, SLC17A2, TRIM38, HIST1H1A, HIST1H3A, HIST1H4A, HIST1H4B, HIST1H3B, HIST1H2AB, HIST1H2BB, HIST1H3C, HIST1H1C, HFE, HIST1H4C, HIST1H1T, HIST1H2BC, HIST1H2AC, HIST1H1E, HIST1H2BD, HIST1H2BD, HIST1H2BE, HIST1H4D, HIST1H3D, HIST1H2AD, HIST1H2BF, HIST1H4E, HIST1H2BG, HIST1H2AE, HIST1H3E, HIST1H1D, HIST1H4F, HIST1H4G, HIST1H3F, HIST1H2BH, HIST1H3G, HIST1H2BI, and HIST1H4H. In panel D, the names of the first two genes, UBE2D4 and WBSCR19, were also omitted for clarity. (A high-quality color representation of this figure is available in the online issue.)

Pleiotropy and mediation of SNP-HbA_1c associations.

HbA_1c levels are influenced by average ambient glycemia over the preceding 3 months, and possibly by erythrocyte turnover. We therefore investigated the novel HbA_1c loci for associations with several diabetes-related and hematologic quantitative parameters in the MAGIC cohorts (19,24) (supplementary Table S4). As previously shown (19), 3 of 10 loci, GCK, MTNR1B, and G6PC2, were associated with FG and HOMA-B (an index of β-cell function, Table 3 and supplementary Table S3), and GCK was additionally associated with 2-h glucose. In all cases, the allele associated with increased HbA_1c was also associated with increased FG and 2-h glucose. No HbA_1c-associated SNP was significantly associated with measures of insulin (supplementary Table S3). We further used conditional models to investigate whether FG levels mediated associations of SNPs with HbA_1c. In these analyses a marked attenuation of the effect size of the SNP in a model adjusted for FG compared with the original main effects model would be consistent with the hypothesis that glycemic pathways primarily account for, or mediate, the HbA_1c association. For the three loci associated with FG (GCK, MTNR1B, and G6PC2/ABCB11), effect sizes were substantially decreased in FG-conditioned models, whereas at the other seven loci, effect sizes remained essentially unchanged (Table 3), indicating that associations with HbA_1c at these loci are unlikely to be mediated by glycemic factors.

We also investigated associations of HbA_1c loci with several hematologic parameters in a subset of four populations with available data (KORA F3, KORA F4, SardiNIA, and NHANES III, supplementary Table S3). Two HbA_1c loci (encoding for functional alleles at HFE and TMPRSS6) showed genome-wide significant association with erythrocyte indexes, consistent with an influence of erythrocyte physiology on HbA_1c variability. The HbA_1c-raising alleles had diverse effects, including associations with lower hemoglobin, mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH) and iron, and higher transferrin (HFE and TMPRSS6). In addition, three loci (SPTA1, ANK1, and HK1) showed suggestive associations (P < 5 × 10⁻³) with erythrocyte indexes, with HbA_1c-raising alleles associated with increased MCV (SPTA1, ANK1), or lower hemoglobin (HK1).

We used these same four cohorts where those parameters were available to carry out a meta-analysis on HbA_1c levels, this time conditioning for the hematologic traits. We did not observe any difference at the three “glycemic” loci, although attenuation of β estimates was observed at HFE, TMPRSS6, and HK1 (supplementary Table S4). However, the sample size used for this analysis was relatively underpowered, resulting in nonsignificant differences (P value > 0.1) and we lacked power for other loci, indicating the need for future analysis in larger sample collections.

Associations with disease: type 2 diabetes and CAD risk.

HbA_1c has been shown to have strong epidemiologic associations with type 2 diabetes risk and with CAD risk in persons without diabetes. To ascertain if the novel loci affected type 2 diabetes risk, we tested associations in well-powered datasets. In a previous meta-analysis of 40,655 type 2 diabetes cases and 87,022 controls in MAGIC (19), MTNRB1, and GCK showed significant evidence of association (rs1387153 OR = 1.09, 95% CI 1.06–1.12, P = 8.0 × 10⁻¹³; rs1799884 OR = 1.07, 95% CI 1.05–1.10, P = 5.0 × 10⁻⁸), whereas G6PC2/ABCB11 did not (rs552976 OR = 0.97, 95% CI 0.95–0.99, P = 0.012). We tested the other novel loci reported here for associations with type 2 diabetes in a partly overlapping study of 8,130 cases and 38,987 controls from the DIAGRAM+ consortium (22) (supplementary Table S3). No other locus associated with HbA_1c was associated with type 2 diabetes risk.

We also tested for associations with CAD using data from nine case/control studies of European descent (13,925 cases and 14,590 controls, supplementary Table S5). None of the SNPs associated with HbA_1c were associated with CAD in the combined sample of 28,515 participants (supplementary Table S6).

Effect size estimates for HbA_1c-associated loci.

In a regression model, the 10 loci combined explained ∼2.4% of the total variance in HbA_1c levels, or about 5% of estimated HbA_1c heritability. We calculated a genotype score using four of the largest population-based studies (ARIC, SardiNIA, KORA F4, and FHS). Using the 10 HbA_1c loci, we estimated cohort-specific differences between the top and bottom 10% of the genotype score distribution (mean [SE] % HbA_1c) to be: 5.25% (0.01) and 5.50% (0.004), respectively (P = 3.61 × 10⁻³³) for ARIC; 5.37% (0.027) and 5.49% (0.027) (P = 1.36 × 10⁻³) for SardiNIA; 5.32% (0.024) and 5.58% (0.027) (P = 4.64 × 10⁻¹²) for KORA F4; and 5.07% (0.046) and 5.38% (0.046) (P = 1.45 × 10⁻⁶) for FHS. The corresponding weighted average difference between the top and bottom 10% of the HbA_1c distributions was 0.21%. For a genotype score using only the seven nonglycemic loci (FN3K, HFE, TMPRSS6, ANK1, SPTA1, ATP11A/TUBGCP3, and HK1), the weighted average difference between the top and bottom 10% of the HbA_1c distributions was 0.19%.

Net reclassification in diabetes screening with HbA_1c.

We used net reclassification analysis to estimate the population-level impact of the seven nonglycemic loci when HbA_1c ≥6.5 (%) is used as the reference cutoff for diabetes diagnosis, as recently proposed (18). We calculated the net reclassification around this threshold attributable to effects of the seven nonglycemic HbA_1c loci that might be expected when screening a general European ancestry population for undiagnosed diabetes using HbA_1c. We studied the FHS and ARIC cohorts combined (N = 10,110), and included individuals with undiagnosed diabetes for detection by screening. We compared the measured distribution of HbA_1c to the distribution adjusted for the seven nonglycemic SNPs (Fig. 3). The net reclassification was −1.86% (P = 0.002), indicating that the population-level effect size of the 7 nonglycemic HbA_1c-associated SNPs is equivalent to reclassification of about 2% of an European ancestry population sample according to HbA_1c-determined diabetes status.

FIG. 3.

Net reclassification when screening for undiagnosed diabetes, using HbA1c as a population-level measure of genetic effect size. The figure shows the distribution of HbA1c in the FHS and ARIC cohorts combined (N = 10,110), stratified by individuals with undiagnosed type 2 diabetes (UnDx DM, N = 593, black lines) or without diabetes (Non DM, N = 9,517, gray lines), and by HbA1c without adjustment (solid lines) or after adjustment for seven nonglycemic SNPs (dashed lines). The vertical dashed line is the diabetes diagnostic threshold at HbA1c ≥6.5(%). Net reclassification is the overall proportion of the population appropriately moved above or below this line by considering the genetic information. For instance, among individuals with undiagnosed diabetes, 39.5% had an unadjusted HbA1c level ≥6.5 (%) and 37.4% had a seven SNP-adjusted HbA1c level ≥6.5 (%), and among those with undiagnosed diabetes, 2.02% of those with undiagnosed diabetes were misclassified by the influence of the seven SNPs. The net reclassification is calculated as the difference −2.02% − (−0.17%) = −1.86%.

URLs. METAL, http://www.sph.umich.edu/csg/abecasis/Metal/index.html; HapMap, http://www.hapmap.org; R- project, http://www.r-project.org; 1,000 Genomes Project, http://www.1000genomes.org.