Evaluation of Common Variants in the Six Known Maturity-Onset Diabetes of the Young (MODY) Genes for Association With Type 2 Diabetes

Informed consent was obtained from all participants. The phenotypic characteristics of stage 1 and 2 samples are described in Table 1. The characteristics of our samples have been described elsewhere (19,20,22,39,40,41). Plasma glucose (fasting and during an oral glucose tolerance test) was measured by a glucose oxidase method with a Beckman Glucose analyzer (Beckman Instruments, Fullerton, CA). All subjects included in this study are of Caucasian ancestry; specific matching schemes are described below.

Stage 1 clinical samples included: 321 type 2 diabetic trios; 1,189 siblings discordant for type 2 diabetes; two Scandinavian case-control samples containing 942 and 1,028 subjects, respectively; and 254 subjects from the Saguenay Lac–St. Jean region in Quebec. These case-control samples were individually matched for age, BMI, and geographic region (by study center). The type 2 diabetic patients met World Health Organization 1998 criteria for type 2 diabetes. In the trios and discordant sibling collections, severe impaired glucose tolerance was defined as >10.0 mmol/l at 120 min, with blood glucose ≥8.5 mmol/l. Age of onset was available for 641 individuals in this study, with a mean of 53.0 ± 11.7 years. The λ was 1.03, based on >500,000 genomewide markers, indicating essentially no population stratification in these samples.

The case-control samples from Genomics Collaborative were comprised of 2,452 individuals of U.S. Caucasian ancestry and 2,018 subjects from Poland. Each sample was individually matched for sex, age, and ethnicity/geographic origin (for three generations). Two studies indicate there is no substantial population stratification in these samples (42,43).

All type 2 diabetic subjects were unrelated U.K. Caucasian individuals who had diabetes defined either by World Health Organization criteria (44) or by being treated with medication for diabetes, and they were recruited from four sources: 1) a young-onset collection (defined as ≥18 and ≤45 years at age of diagnosis), 2) probands from type 2 diabetes sibships from the Diabetes U.K. Warren 2 repository described previously (45,46), 3) a collection from the Warren 2 repository with type 2 diabetes diagnosis between 35 and 65 years, and 4) a collection of subjects from the Oxagen resource, selected in the same way as the Warren 2 case subjects. Patients with high GAD autoantibody levels (>3 units [normal <0.5]) were excluded from the study. Control subjects were recruited from three sources: 1) parents from a consecutive birth cohort (Exeter Family Study) with normal (<6.0 mmol/l) fasting glucose and/or normal A1C levels (<6%; Diabetes Control and Complications trial corrected) (45), 2) U.K. Caucasians (HRC [Human Random Control]) obtained from the ECACC (European Cell Culture Collection), and 3) a follow-up study of all people born in the U.K. during 1 week in 1958 (http://www.cls.ioe.ac.uk/Cohort/Ncds/mainncds.htm). All subjects from our family-based study were independent of those from the case-control study. Families fitting the following criteria were included: an affected proband with both parents, or one parent and at least one unaffected sibling.

Two different matching schemes were used with regard to BMI (BMI matched or not matched). This could potentially lead to false negatives; however, the following evidence indicates this is not the case in this study: 1) no variant in this study had a significant association in all subsamples with one matching scheme and not the other, and 2) tests of association between BMI and rs757210 and G(−30)A in all stage 1 control subjects demonstrated that these variants have no significant association with BMI (data not shown).

Genotyping for stage 1 and 2 samples was performed as previously described by primer extension of multiplex products with detection by matrix-assisted laser desorption ionization–time-of-flight mass spectroscopy using a Sequenom platform (47,48). For the U.K. samples, genotyping was outsourced to KBioscience (Herts, U.K.). rs757210 has a rare third allele; all results are given for A allele versus C+G.

To determine the association of each particular SNP with type 2 diabetes, we used simple χ² analysis in the case-control samples, the transmission disequilibrium test (49) for the trios, and the discordant allele test (50) for the stage 1 sibling pairs (using oldest unaffected sibling and a random affected sibling). The U.K. family study used SibTDT for the sibling pairs. Multimarker tests were evaluated using Haploview (51). Because this program cannot accommodate discordant sibling pairs, they were not included in the reported multimarker results. Results for the subsamples were combined using Mantel-Haenszel meta-analysis of the odds ratios (ORs) (52). Homogeneity among studies was tested using a Pearson r² goodness-of-fit test as previously described (52). Genetic models for rs757210 were evaluated in the five unrelated case-control samples in stage 1 and 2 using the program MODEL (http://pngu.mgh.harvard.edu/∼purcell/model/model.html). Reported results are not corrected for multiple hypothesis testing.

False-positive report probability was calculated using Wacholder et al.’s (53) Excel spreadsheet showing input and output for false-positive report probability calculations. The prior probabilities used are listed in Table 4; the step 3 likely OR values entered were 1.05, 1.15, and 1.25 (only 1.15 is reported in Table 4). The OR and 95% CI values in Table 3 were used for the calculations in step 4.

To evaluate the haplotype structure of the GCK, IPF1, TCF2, and NEUROD1 regions, we selected SNPs from dbSNP and Celera based on an evenly spaced grid across the regions. These were genotyped in a multigenerational panel of 12 Centre d’Etude du Polymorphisme Humain (CEPH) pedigrees containing 96 chromosomes, with additional SNPs added based on the extent of LD. To be included in the LD maps, SNPs were required to attain a 75% genotyping percentage, be in Hardy-Weinberg equilibrium, and have a minor allele frequency >5%. Additionally, for the family-based studies, multiple errors in Mendelian inheritance were cause for exclusion. GCK had 62 passing SNPs, spanning 83.2 kb with an average marker spacing of one per 1.3 kb. IPF1 had 23 working SNPs, covering 44.2 kb and averaging one marker per 1.9 kb. There were 80 passing SNPs in the TCF2 region, spanning 76.3 kb with an average SNP spacing of one per 0.95 kb. There were 10 passing SNPs in NEUROD1, spanning 44.1 kb with an average marker spacing of one per 4.4 kb. Haplotype structures for the GCK, IPF1, TCF2, and NEUROD1 regions are shown in supplemental Appendix 1, which can be found in an online appendix (available at http://dx.doi.org/10.2337/db06-0202).

This study was performed over several years, and as dbSNP coverage improved and methods for tag SNP selection evolved, additional tag SNPs were added. The final tag SNP set was selected using the program Tagger (54). To determine how well the final tag SNP set captured variation in these regions, SNPs were evaluated for their correlation to one another in the CEPH samples described above. On the hypothesis that any of these variants could be a putative causal variant or proxy thereof, we selected a final set of tag SNPs and multimarker combinations (>5% frequency) that had an r² >0.8 to all of the markers typed in the CEPH panel. For all tests, tag SNPs, and correlation values, see supplemental Appendix 2.

To assess whether any of the MODY variants created a genetic background in which the effect of a different MODY variant on phenotype could be observed, we performed two analyses. First, we used the set of 23 variants that showed a nominal single-marker association with type 2 diabetes. For each pair of variants A and B from this set, we first stratified the cohort based on genotype at variant A and then assessed the association of variant B with type 2 diabetes phenotype (χ² test) in each of the subcohorts defined by genotype at A (1,518 hypotheses tested in total). The second method used the fact that epistasis implies stronger LD between two interacting loci in case subjects compared with control subjects. The assumption of no LD in control subjects for unlinked genes allows for a case-only test (simply a test for LD in case subjects). To test for interactions, canonical correlation analysis (55) was used to measure the association between the sets of markers in two genes, testing the OR for the association between markers in each gene. Permutation was used to obtain empirical significance values for both methods.

A multistage study design offers a powerful approach to detect modest associations while minimizing the cost of genotyping large numbers of markers (56). Our study design is shown in Fig. 1. We started by evaluating LD patterns spanning each gene, including at least 20 kb upstream and downstream of the coding regions. A dense, evenly spaced grid of SNPs from each region was genotyped in a panel composed of 30 CEPH trios (for details see research design and methods and supplemental Appendix 1). SNP maps for HNF4α and TCF1 were reported previously (19,22). The IPF1 and NEUROD1 gene regions both showed extensive LD and limited haplotype diversity, indicating that most undiscovered SNPs are likely to be highly correlated to the SNPs already studied in these regions (48). The GCK and TCF2 regions showed more evidence of historical recombination, indicating that a higher number of tag SNPs are required to adequately capture the common variation within these genes.

We used the program Tagger (54) to analyze the CEPH LD maps, selecting a total of 135 tests comprised of 63 SNPs (from the four genes) to take forward into association studies. These tests included the common missense and previously implicated promoter variants. They provide an r² >0.8 to all other markers typed in the CEPH panels, indicating that they should provide strong power for both untyped reference panel SNPs and most undiscovered common SNPs when genotyped in the disease panel (supplemental Appendix 2).

The tag SNPs were genotyped in 4,206 type 2 diabetic patients and matched control subjects from Scandinavia and Canada (Table 1). For test details and complete association results, see supplemental Appendix 3. Including reanalysis of our previous data on HNF4α and TCF1 (19,22) (performed for consistency across the genes), we found a total of 30 tests that achieved a P value <0.1 in our stage 1 sample (Table 2). (Note that SNPs within genes are often correlated—for example, many of the nominally associated SNPs in TCF1 and TCF2 have high r² with one another.) We then genotyped the SNPs with P < 0.1 in stage 1 in a second independent sample of 4,470 subjects from Poland and North America. Eight SNPs attained nominally significant associations in the combined 8,676 individuals: rs730497—a perfect proxy for GCK G(−30)A—and seven SNPs in TCF2 (Table 3). The strongest statistical evidence was for rs3110641 (OR 1.16 [1.08–1.26], P = 0.0002) and rs757210 (OR 1.13 [1.06–1.21], P = 0.0003). Six of the TCF2 SNPs are located within ∼1 kb of rs3110641 and are in LD with it (supplemental Appendix 1c). Therefore, it seems likely that the more modest associations observed in TCF2 may be attributable to those SNPs’ correlation with the stronger signal at rs3110641.

To further evaluate the observed associations for rs3110641, rs757210, and GCK G(−30)A, we genotyped them in an additional 5,891 unrelated case and control subjects and 500 families from the U.K. The results from these samples did not support our hypothesis for rs3110641 (C allele OR of 1.04, P = 0.33), making it more likely that the initial observation may have been a false-positive. However, both samples did replicate the original effect seen for rs757210. The A allele had an OR of 1.10 (P = 0.02) in 3,640 control subjects and 2,251 case subjects and, similarly, an OR of 1.13 (P = 0.03) in the family study (n = 500 trios and duos families). To determine the genetic model by which rs757210 influences diabetes risk, we analyzed rs757210 using MODEL. The results indicate that while the additive, multiplicative, and dominant models all have P < 0.001, the best-fitting model was the dominant model. There was no significant effect under a recessive model.

The GCK G(−30)A variant had no effect on type 2 diabetes risk in the U.K. samples (case/control OR of 0.96, family OR of 1.00). However, the GCK G(−30)A association is also supported by a number of other previous reports of diabetes/diabetes-related traits for this allele (23–33) and thus may be considered less likely to be a false positive than an SNP without such evidence.

We performed a meta-analysis of the current results and previously reported association results for rs757210 and G(−30)A with type 2 diabetes (Fig. 2). Both achieved overall statistical significance in the combined analysis; however, their overall effect on disease risk was very modest [rs757210: OR 1.12 [1.07–1.17], P = 10^–6; G(−30)A: OR 1.08, P = 0.004], although the effect was of similar size to other well-established type 2 diabetes genes such as P12A in PPARγ and E23K in KIR6.2.

To avoid false negatives caused by testing SNPs one at a time, we evaluated the possibility that combinations of MODY genes interact to influence type 2 diabetes. We tested pairwise combinations of genes in two ways. First, we took the SNPs with P < 0.1 in the initial sample and looked for significant single-marker effects of each marker by stratifying on the genotype at each of the 22 other markers. Second, we looked for genetic interactions at the level of the gene (as opposed to the allele), testing for correlation between unlinked loci in the case subjects. Both methods evaluated significance using permutation testing. We also included data from our group for the well-established type 2 diabetes susceptibility loci P12A in PPARγ, E23K in KIR6.2, and rs7903146 in TCF7L2 in our interaction tests (39,40,57). No significant interactions (threshold of P < 0.05 after permutation testing) between these nine candidate genes were observed using either approach.

Mendelian mutations in the MODY genes cause the vast majority of cases of monogenic type 2 diabetes, and a number of previous association studies have suggested that missense SNPs within these genes might be associated with common type 2 diabetes. Thus, before beginning this study, we estimated that the prior probability of at least one of these genes being involved in type 2 diabetes was quite high. Despite this, we did not find any major effects on type 2 diabetes in the MODY genes, and only a single variant in TCF2 (with a very modest OR) was reproduced across multiple large studies and achieved a high level of overall statistical significance.

We found it useful to use Wacholder et al.’s (53) Bayesian approach to communicating the posterior probability that these associations are false positives. Table 4 shows the false-positive report probability over a range of possible prior probabilities for the OR of 1.15 in our combined 8,676-subject study. The table indicates that if the reader assumes a relatively high a priori likelihood that common variants in MODY genes play a role in type 2 diabetes (1–10%), our best result (rs757210) has a low probability of being a false positive (<7%). GCK G(−30)A has a 6–42% false-positive probability, although one could argue that because G(−30)A is implicated in >10 diabetes-related studies, an even higher prior probability may be indicated for this particular SNP. Given the known role of MODY genes in type 2 diabetes, these relatively high prior probabilities do not seem at all unreasonable, and thus one or both of these associations may be real.

A concurrent study on the MODY genes in type 2 diabetes was performed by Bonnycastle et al. (58). They also concluded that there are no common variants in these genes that have large effects on type 2 diabetes risk, but, based on nominal associations, they generated hypotheses that several variants may confer modest effects. None of their leading hypotheses were associated in our samples; conversely, our best result (rs757210) showed a consistent but nonsignificant OR (1.10 [0.93–1.30]) in their 1,408 case and control subjects from Finland.

One possible reason for the varying association results is that our groups used different statistical criteria to declare association and differently powered samples. The modest effects seen in our study would provide little power in a study of the size performed by Bonnycastle et al. (58). Another possible explanation is cryptic population heterogeneity or an unknown genetic or environmental modifier that differs between the two samples (although we note that all positive associations genotyped by both reports fail a formal test for heterogeneity; also, both studies examined large numbers of Scandinavian samples). We note that all populations studied are of European ancestry, and therefore one cannot extrapolate this result to other populations of more diverse ancestry.

Over the past few years, inconsistent replication has obscured the role of several MODY genes in common type 2 diabetes. One possibility is that a failure to replicate a positive result may be caused by some unknown genetic interaction with an allele that varies in frequency between study populations, limiting the power to detect the association in the second sample. It is plausible that combinations of variants in risk genes could be required to consistently demonstrate an effect on disease. To test this hypothesis, we evaluated pairwise combinations of the MODY genes and three well-established type 2 diabetes genes, PPARγ, KIR6.2, and TCF7L2. Using two methods, we found no evidence that any of these nine genes interact, suggesting that the observed inconsistency in association results is likely due to false positives. However, it remains possible that there are critical interactions with untested loci or environmental factors, with the criteria for defining case or control status, or interaction(s) among gene variants that will only be apparent on consideration of three or more loci.

In our study of the six MODY genes, we identified several common variants that may be associated with the common form of type 2 diabetes, particularly rs757210 in TCF2, which has been replicated in >15,000 samples and achieved strong overall statistical significance. Two other widely replicated common type 2 diabetes genes, KIR6.2 and PPARγ, also cause rare Mendelian forms of diabetes (2,3), yet each of these three monogenic diabetes genes carries only a modest effect on risk of common type 2 diabetes (allelic OR <1.25). This finding that in aggregate the genes responsible for most cases of monogenic type 2 diabetes contribute little to the common form of the disease has important implications for many complex traits. If the common forms of disease are most often influenced by inherited variants in genes distinct from those that cause Mendelian forms, it will suggest that the underlying pathogenic mechanisms of common forms may not be as similar as initially hoped, and that study of Mendelian subtypes alone would not necessarily identify all of the pathways and genes required to understand the common forms of disease. Combined with the fact that many type 2 diabetes candidate genes with strong biological support have been demonstrated to have no effect on the common subtype of type 2 diabetes, these results indicate that a large portion of the genetic component of type 2 diabetes risk may lie in genes and/or pathways not yet identified, thus requiring unbiased genomewide approaches. Whole-genome association studies (2) offer one such approach and should reveal other genes that harbor common variants that contribute significantly to risk of type 2 diabetes.

L.G., T.T., and the Botnia Study are principally supported by the Sigrid Juselius Foundation, the Academy of Finland, the Finnish Diabetes Research Foundation, the Folkhalsan Research Foundation, the Swedish Medical Research Council, the JDF Wallenberg Foundation, and the Novo Nordisk Foundation. We acknowledge use of DNA from the British 1958 Birth Cohort collection, funded by Medical Research Council Grant G0000934 and Wellcome Trust Grant 068545/Z/02. The U.K. work was funded by Diabetes U.K., the Medical Research Council, and Wellcome Trust. M.N.W. is a Vandervell Foundation fellow, and A.T.H. is a Wellcome Trust research leave fellow. J.N.H. is a recipient of a Burroughs Wellcome career award in biomedical sciences. D.A. is a Charles E. Culpepper scholar of the Rockefeller Brothers Fund and a Burroughs Wellcome Fund clinical scholar in translational research, the latter of which supported this work.

We thank the FUSION study for sharing their unpublished data, the Botnia research team for clinical contributions, and the members of the Altshuler, Hirschhorn, Daly, and Groop labs for helpful discussions.