Identification of Novel Candidate Genes for Type 2 Diabetes From a Genome-Wide Association Scan in the Old Order Amish

RESEARCH DESIGN AND METHODS

Study population and phenotype assessment.

Individuals with type 2 diabetes were identified from the AFDS. Details of the AFDS have been previously described (10). Phenotypic characterization of participants included medical and family history, anthropometry, and a 3-h 75-g oral glucose tolerance test (OGTT) with insulin levels. We based our primary analyses on 124 type 2 diabetic case and 295 normal glucose tolerant (NGT) control subjects. Type 2 diabetes was defined by fasting plasma glucose level (≥7 mmol/l), 2-h OGTT plasma glucose level (≥11.1 mmol/l), random plasma glucose level (≥11.1 mmol/l), the use of insulin or prescription oral glucose-lowering agents, or a diagnosis of diabetes documented by a physician. To minimize potentially misclassifying subjects with type 1 diabetes as having type 2 diabetes, case subjects with age of diagnosis <35 years were excluded. NGT control subjects were aged >38 years at the time of study and were selected based on fasting plasma glucose level (<6.1 mmol/l) and 2-h OGTT plasma glucose level (<7.8 mmol/l).

We performed secondary quantitative analyses of our mostly highly associated signals (P < 0.01) in a set of 427 nondiabetic Amish study participants, 132 of whom had impaired glucose tolerance and 295 of whom were part of the NGT control group used in our primary analysis. We estimated the mean levels of two OGTT-derived quantitative glucose traits (fasting glucose and glucose area under the curve [GAUC]) and three insulin traits (insulinogenic index [ISI], insulin area under the curve [IAUC], and homeostasis model assessment of insulin resistance [HOMA-IR]) according to the SNP genotypes in these individuals. Total GAUC and IAUC were calculated based on measurements at 0, 30, 60, 90, 120, 150, and 180 min using the trapezoidal method. The ISI was calculated as (insulin at 30 min − fasting insulin)/(glucose at 30 min − fasting glucose). HOMA-IR was calculated as fasting insulin (mU/l) × fasting glucose (mmol/l)/22.5. Table 1 describes the characteristics of this sample. The study protocol was approved by the institutional review board at the University of Maryland School of Medicine, and informed consent was obtained from each study participant.

Genotyping.

Genomic DNA from leukocytes were genotyped using the Affymetrix GeneChip Mapping 100K array set, which consists of two microarray chips (XbaI and HindIII) (Affymetrix, Santa Clara, CA). Total genomic DNA (250 ng) was digested with XbaI or HindIII restriction enzymes and processed according to the Affymetrix protocol. The GeneChip Genotyping Analysis Software (GTYPE 4.0) was used to generate dynamic modeling algorithm–derived genotypes that were reanalyzed with the BRLMM (Bayesian RLMM) genotype calling algorithm (confidence threshold of 0.33) to improve the proportion of heterozygote calls (19). As an initial quality-control measure, BRLMM-generated chip files with call rates <90% for both enzymes across all SNPs were excluded. The resulting median call rate across all of the remaining 419 case-control samples was 97.5% (97.6% for XbaI and 97.4% for HindIII). We further removed individual SNPs with genotype call rates <90%, monomorphic SNPs and SNPs with minor allele frequency < 5%, and those deviating from Hardy-Weinberg equilibrium in control subjects (P < 0.001). The number of monomorphic and low-frequency SNPs (n = 26,816) in the Amish was not appreciably different from that observed in more outbred Caucasians of the HapMap CEU sample. For this report, we focused our analysis on the 82,485 autosomal SNPs that passed our quality-control standards.

The concordance rate for 11 quality-control samples that were run twice on the Affymetrix GeneChip mapping panel was 97.5%. We also calculated a cross-platform concordance rate of 98% for 419 samples in which 61 SNPs were genotyped using the Affymetrix GeneChip Mapping 100K panel and an independent Illumina 1536-plex GoldenGate assay. Supplementary Table 3 (available in an online appendix at http://dx.doi.org/10.2337/db07-0457) summarizes the quality checks and informativeness of the data.

Association testing and SNP prioritization scheme.

Our GWAS analysis and SNP prioritization scheme is shown in Fig. 1. We selected the SNPs most highly associated with type 2 diabetes in our Amish case-control dataset based on P value rankings (P value cutoff <0.01) and then used two complementary approaches to further prioritize them. In one approach, we evaluated the most highly type 2 diabetes–associated SNPs for association with diabetes-related quantitative traits in an expanded set of 427 nondiabetic Amish subjects, 295 of whom were NGT control subjects from the primary type 2 diabetes association analysis (internal consistency). In a parallel approach, we assessed replication of the most highly associated type 2 diabetes–associated SNPs in the Amish in four independent GWASs from other populations (external replication).

Type 2 diabetes association analysis.

We performed case-control association analysis using a variance component approach as implemented in SOLAR software (20). Using a liability threshold model, we modeled the probability that the individual was a case or control subject as a function of the individual's age, sex, and genotype, conditional on the correlations in phenotype among relative pairs. Statistical testing was performed using a likelihood ratio test, in which we compared the likelihood of the data under a model in which the genotype effect was estimated with the likelihood of a nested model in which the genotype effect was constrained to be zero. Odds ratios (ORs) were computed from variance components models. We chose to report the additive model as our primary analysis. Supplementary analyses using a dominant or recessive model did not yield any SNP showing genome-wide significance. Of 82,485 SNPs, 611 had P < 0.01 under a dominant model and 569 had P < 0.01 under the recessive model. Our complete dataset with results from all models is available online (available at http://www.medschool.umaryland.edu/amishstudies/index.asp). Pairwise linkage disequilibrium (LD) correlation statistics (r²) were computed using the Helixtree software, version 5.0.2 (GoldenHelix, Bozeman, MT).

Quantitative trait analysis.

For quantitative trait analyses performed in nondiabetic Amish subjects, we used the measured genotype approach, in which we estimated the likelihood of an additive genetic model given the pedigree structure (21). Before analysis, all insulin traits (IAUC, ISI, and HOMA-IR) were transformed by their natural logarithm to reduce skewness. Parameter estimates were obtained by maximum likelihood methods, and the significance of association was tested by the likelihood ratio test. Within each model, we simultaneously estimated the effects of age and sex. These analyses were performed using the SOLAR program (20).

Power calculations.

Power calculations, based on the genetic power calculator of Purcell et al. (22), indicated that our sample would provide 80% power to detect a diabetes susceptibility allele having a genotype relative risk of 1.8 (for allele frequency of 30%, 124 case and 295 control subjects, 8% population prevalence of diabetes, assuming a multiplicative model) and 80% power to detect a quantitative trait loci accounting for 4% or higher of the trait variance for a continuously distributed phenotype (427 subjects).

In silico replication samples.

We considered whether our best type 2 diabetes association signals (P value cutoff <0.01) replicated in at least one of three distinct populations (Framingham Caucasians, Mexican Americans, and Pima Indians), each with different study designs but performed using the same Affymetrix 100K genotyping platform. Descriptions of each of the Type 2 Diabetes 100K GWAS Consortium study populations are provided in accompanying articles (23–25) and in supplementary Table 4. We directly checked whether any of the 1,093 SNPs with the best type 2 diabetes association signals (P < 0.01) in the Amish were also associated with type 2 diabetes based on generalized estimating equations and family-based association tests in the Framingham Heart Study, Fisher's exact allelic association test in the Mexican-American study, and case-control and sib-based association tests in the Pima Indian study. We also utilized publicly available prereleased data (March 2007) from a type 2 diabetes GWAS carried out in a Scandinavian cohort of 1,464 type 2 diabetic case and 1,467 matched control subjects and genotyped using the Affymetrix 500K platform by the Broad-Lund-Novartis Diabetes Genetics Initiative (DGI) (available at http://www.broad.mit.edu/diabetes/) (18). We specifically checked replication of 295 of 1,093 of our most highly type 2 diabetes–associated SNPs that were present on both 100K and 500K Affymetrix genotyping arrays. Since LD structure may differ across populations, and to limit multiple comparisons, we defined replication only if the same SNP was associated with type 2 diabetes at P < 0.05 with an OR in the same direction (i.e., reflective of the same allelic risk).

RESULTS

Following quality-control and Hardy-Weinberg equilibrium checks, 82,485 informative SNPs were included in our analyses. The median physical inter-SNP distance was 11.3 kb, and the average distance between SNPs was 29 kb. Under the additive model, a total of 1,093 SNPs, some of which were in LD, were associated (P < 0.01) with type 2 diabetes (Fig. 2). The 50 most strongly type 2 diabetes–associated SNPs (i.e., lowest P values) are shown in Table 2. The complete dataset is available online (available at http://www.medschool.umaryland.edu/amishstudies/afds.asp). No SNP was associated with type 2 diabetes at a conservative Bonferroni-corrected level. The strongest association (P = 1.07 × 10⁻⁵) was for rs2237457 on chromosome 7, which is located in intron 4 of growth factor receptor–bound protein 10 (Grb10), an adaptor protein known to regulate signaling of insulin and IGF receptors (26–28). In addition to Grb10, 15 SNPs were associated with type 2 diabetes at P < 1 × 10⁻⁴ (Fig. 2 and Table 2). These SNPs are located in or near MSH6 (chromosome 2), PRKG2 (chromosome 4), COL13A1 (chromosome 10), MTHFSD (chromosome 16), and SPECC1 (chromosome 17), none of which are obvious candidate genes for type 2 diabetes. Adjustment for BMI did not have a large impact on the strength of the associations of these SNPs with type 2 diabetes (Table 2).

As a measure of internal consistency, we tested whether the 1,093 SNPS associated with type 2 diabetes (P < 0.01) were also associated with OGTT-derived quantitative traits in nondiabetic individuals. In these analyses, we considered two OGTT glucose traits (fasting glucose and GAUC) and three OGTT insulin traits (IAUC, HOMA-IR, and ISI), with P < 0.01 as our threshold for significance. Thirty-eight nonredundant (r² < 0.80) type 2 diabetes–associated SNPs were also associated with at least one glucose trait and showed the same allelic association as that for diabetes (i.e., the diabetes risk allele was also associated with higher glucose levels), while 29 nonredundant type 2 diabetes–associated SNPs were also associated with at least one insulin-related trait (Fig. 1; Table 3). Of the top 16 SNPs associated with type 2 diabetes at P < 1 × 10⁻⁴, rs2237457 in Grb10 was the only one also associated with an OGTT trait (P = 0.001 for GAUC). Two perfectly correlated (r² = 1) type 2 diabetes–associated SNPs in ADAMTS1 (chromosome 5) (P = 0.004–0.005) were associated with one glucose trait (P = 0.006 for GAUC) and one insulin trait (P = 0.007 for IAUC).

We next sought to determine which of our 1,093 most highly type 2 diabetes–associated SNPs were also associated with type 2 diabetes in any of three independent populations for which the same 100K Affymetrix platform was used or in the DGI Scandinavian population for which the 500K Affymetrix platform was used. We identified 80 nonredundant SNPs for which the same risk allele was also associated with type 2 diabetes in one of the three studies from the Type 2 Diabetes 100K GWAS Consortium (P < 0.05) and 11 nonredundant SNPs that showed consistent association in the DGI sample (P < 0.05) (Fig. 1; supplementary Table 3). In total, three SNPs demonstrated associations in the Amish as well as in two independent populations. The T-allele for rs2540317 in MFSD9 on chromosome 2 was associated with decreased risk of type 2 diabetes in the Amish (OR 0.72, P = 0.007) and showed nominal association in the Pima Indian dataset (case-control OR 0.67, P = 0.042; sib-based OR 0.50, P = 0.043; and summary OR 0.63, P = 0.016) and also in the Mexican-American sample (case-control OR 0.75, P = 0.047). The G-allele in rs10515353 on chromosome 5 was associated with decreased risk of type 2 diabetes in the Amish (OR 0.61, P = 0.005) and also with decreased type 2 diabetes risk in Mexican-American (0.69, P = 0.035) and DGI (0.79, P = 0.007) samples. The T-allele in rs2242400 in BCAT1 on chromosome 10 was associated with decreased risk of type 2 diabetes in the Amish (0.71, P = 0.004) and also in the Pima Indian dataset (sib-based OR 0.66, P = 0.019; summary OR 0.78, P = 0.034) and the Mexican-American dataset (OR 0.67, P = 0.009); borderline association was also seen in the DGI sample (0.86, P = 0.051). The direction of effect was the same for all studies.

Table 4 highlights our most consistent overall findings. We present 21 type 2 diabetes–associated SNPs in the Amish (P < 0.005) that also demonstrated either 1) association with a diabetes-related quantitative trait (P < 0.005) in the Amish or 2) in silico replication of type 2 diabetes association in one independent population (P < 0.005). Of interest, the T-allele in rs3845971 in FHIT was associated with increased risk of type 2 diabetes in the Amish (OR 1.42, P = 0.004) and also in Mexican Americans (1.46, P = 0.004) and with increased GAUC (P = 4.0 × 10⁻⁴) in nondiabetic Amish subjects.

DISCUSSION

In this article, we described the results of a GWAS of type 2 diabetes of 82,485 SNPs in the Old Order Amish, a genetically closed founder population with a homogeneous lifestyle. We reasoned that this population is likely to carry a subset of the same common type 2 diabetes susceptibility variants as those found in the general population and that these variants might be easier to identify.

GWAS studies are prone to false-positives due to the very large number of statistical tests that must be performed. We were restricted by our relatively modest sample size and also computationally in our attempts to define a genome-wide significance level for which follow-up was justified (i.e., variance components tests were not feasible for the many replications needed for case-control permuted family datasets in the Amish). Thus, we relied heavily on a prioritization of SNPs worthy of follow-up by testing for 1) internal consistency of type 2 diabetes–associated SNPs with OGTT-derived quantitative traits in nondiabetic Amish individuals, 2) external replication of type 2 diabetes associations in three independent non-Amish 100K SNP GWAS studies, and 3) external replication in a 500K SNP GWAS of type 2 diabetes in a large population of Scandinavians.

We found that no single SNP replicated consistently and in the same direction across all GWAS studies, nor were all SNPs associated with type 2 diabetes also associated with quantitative traits in nondiabetic individuals (supplementary Table 4). This is not particularly surprising since we expect that an appreciable number of type 2 diabetes–associated SNPs will be false-positives. Furthermore, a true susceptibility gene in one population might not be readily discernible in other populations due to inadequate sample sizes as well as differences in genetic background, LD, and environmental exposures. Similarly, a true susceptibility gene for type 2 diabetes might not show association with diabetes-related quantitative traits in nondiabetic individuals, especially since our OGTT-derived traits are only surrogates for gold-standard measures of insulin sensitivity and insulin secretion. Nevertheless, we were able to identify a number of candidate genes and loci that showed evidence for association with type 2 diabetes in more than one population and/or were also associated with OGTT-derived quantitative traits. These results are intriguing but must be interpreted with caution. None of these loci fall within previously identified linkage regions for type 2 diabetes (chromosomes 1 and 14) in the Amish.

Our strongest type 2 diabetes association signal in the Amish was observed on chromosome 7 in a functionally relevant type 2 diabetes candidate gene, Grb10. Grb10 encodes growth factor–binding protein 10 and has been shown to bind to activated insulin receptor and act as a negative regulator of insulin action and glucose uptake (26–28). Overexpression of Grb10 in mice causes postnatal growth retardation and insulin resistance (29). Our 100K GWAS contained a total of 12 SNPs in Grb10, 6 of which were associated with type 2 diabetes (P < 0.05) and were in partial LD with each other (r² = 0.16–0.78). Rs2237457, located in intron 4, provided the lowest P value for association (OR 0.61 for the G- vs. A-allele, P = 1.07 × 10⁻⁵). This SNP was also strongly associated with OGTT GAUC in nondiabetic Amish individuals (P = 0.001). Rs2237457 was not associated with type 2 diabetes in the other three populations in which this SNP was genotyped or in the 500K SNP Scandinavian type 2 diabetes GWAS; however, three SNPs (rs2190496, rs2237478, and rs7805310) in Grb10 that were genotyped in the Scandinavian cohort were associated with type 2 diabetes (P = 0.029, P = 0.01, and P = 0.004, respectively) and are in partial LD with rs2237457 (r² = 0.12–0.49 in HapMap CEU). Lack of replication could suggest a false-positive or that variation in Grb10 is a true positive specific to the Amish due to a founder effect or context-dependent phenotypic expression of the variant due to genetic background or environmental influences. Alternatively, this variant could be in LD with a functional variant, and extended LD in the Amish enabled a type 2 diabetes association to be detected in this population and not the others.

In a recent report by Di Paola et al. (30), the A-allele of rs4947710, a synonymous coding SNP in Grb10, was associated with decreased risk of type 2 diabetes in a relatively homogeneous population of Italian Caucasians (P = 0.0001). This SNP was not part of the 100K SNP panel nor was our most highly type 2 diabetes–associated SNP (rs2237457) genotyped in the Italian sample. We found that rs2237457 and rs4947710 are not in LD (r² = 0) in HapMap CEU samples. However, rs10486757, another Grb10 SNP associated with type 2 diabetes in the Amish (P = 0.024), is in LD with rs4947710 (r² = 0.64 in HapMap CEU). Further investigation of Grb10 is currently underway.

Our GWAS and replication strategy have several limitations. First, the relatively small sample size limits our ability to detect gene variants of modest effect size. Second, we recognize that the definition of external replication of our top SNPs across three independent 100K studies of type 2 diabetes might represent a skewed distribution of the overall results since replication was limited to our ∼1,000 most highly type 2 diabetes–associated SNPs. This approach was used to facilitate comparisons across populations and also to limit the number of false-positive replications due to multiple comparisons. To the extent that we attempted to pursue signals that represent the “lowest hanging fruit,” we believe that the approach we have taken is reasonable. A formal meta-analysis of the entire set of data from all four 100K studies is currently underway. Third, our replication approach was focused at the level of the SNP in order to avoid additional multiple comparisons. However, it is possible that we did not identify significantly associated SNPs in other populations that were in LD with our top SNPs. This is particularly relevant for our comparisons with the Scandinavian 500K GWAS, for which only 27% of the SNPs identified in the Amish with P < 0.01 were identified in the 500K SHP panel.

The likelihood that we missed common variants important to type 2 diabetes is high due to the relatively sparse density of the 100K SNP panel (mean intermarker distance = 29 kb) compared with other denser GWAS SNP panels. For example, SNPs in well-replicated genes (SLC30A8, IGF2BP2, CKDAL1, CDKN2A/CKDN2B, and HHEX/IDE) found in four recently published type 2 diabetes GWAS studies (15–18), as well as previously known type 2 diabetes–associated variants in TCF7L2, KCNJ11, HNF4A, or CAPN10 (9), were not adequately covered on the 100K genotyping panel (i.e., r² < 0.8 between the SNP of interest and SNPs on the 100K panel). As a positive control, we previously demonstrated that TCF7L2 SNP rs7903146 and the HNF4A promoter SNP rs2425640, neither of which is present on the 100K panel, were associated with type 2 diabetes and impaired glucose tolerance in the Amish Family Diabetes Study (OR 1.57, P = 0.008; 1.60, P = 0.04, respectively) (12,25). Interestingly, rs10509645 in HHEX on the 100K panel (r² = 0.7 with rs7923837 found previously to be strongly associated with type 2 diabetes in other GWAS studies) was significantly associated with type 2 diabetes in the Amish (OR 1.30 for the G-allele; P = 0.02). Rs9300039 on chromosome 11, shown to be associated with type 2 diabetes in the other GWAS studies (17), was present on the 100K panel but was not significantly associated with type 2 diabetes in the Amish (OR 1.09 for the C-allele; P = 0.67).

In summary, we presented results from our initial examination of a GWAS of type 2 diabetes in the Amish. Although we did not identify any genes associated with type 2 diabetes that reached genome-wide significance, we report a number of genes and loci that are worthy of further study based on replication in other studies or on quantitative trait loci consistency. This report (and the three companion articles) provides a valuable resource for other investigators to utilize in the search for the pathogenic variants for type 2 diabetes.