Association of Common Variation in the HNF1α Gene Region With Risk of Type 2 Diabetes

The characteristics of our patient samples have been described elsewhere (2,9,27,28). They include 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. The type 2 diabetic patients met the 1998 World Health Organization 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.

The case-control samples from Genomics Collaborative, Inc. (GCI) are comprised of 2,452 individuals of U.S. Caucasian ancestry and 2,018 subjects from Poland. These samples were matched for sex, age, and ethnicity/geographic origin (for three generations). The phenotypic characteristics of all samples are described in Table 1. 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).

Previous association studies were found through review citations and by searching PubMed for the following queries: “polymorphism, tcf1, diabetes” and “polymorphism, hnf1, diabetes.” To be included in our analysis, patients must have been defined as late onset and non-MODY. Results for the subsamples were combined using Mantel-Haenszel meta-analysis of the odds ratios (ORs) (29).

Genotyping 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. (30,31). Most tag SNPs were genotyped twice, and the average genotype completeness for working assays was 96%. The genotyping consensus error was determined to be 0.6%, using both duplicate genotypes (69,611 comparisons) and errors in Mendelian inheritance.

To determine the association of each particular SNP with type 2 diabetes, we used simple χ² analyses in the case-control samples: the transmission disequilibrium test (32) for the trios and the discordant allele test (33) for the sibling pairs (using the oldest unaffected sib and a random affected sib). For multimarker analyses, the frequency of each combination was estimated in the individual sample using an expectation maximization algorithm (N. Patterson, unpublished software). Results for the subsamples were combined using Mantel-Haenszel meta-analysis of the ORs (29). Homogeneity among studies was tested using a Pearson χ² goodness-of-fit test, as previously described (29).

To evaluate the haplotype structure of the HNF1α region, we genotyped 158 SNPs from dbSNP (all available SNPs through build 118) and Celera in a multigenerational panel of 12 Centre d’Etude du Polymorphisme Humain (CEPH) pedigrees containing 96 chromosomes. We also included nine SNPs discovered by resequencing (34) 11 kb in 32 diabetic patients (targeted regions include the HNF1α promoter, AK096009 mRNA, and upstream mouse conserved regions). In total, these SNPs span 118 kb, from ∼49 kb upstream of the gene start site to ∼45 kb downstream of the end of the HNF1α 3′ untranslated region. SNPs were initially selected based on an evenly spaced grid across the region, with additional SNPs added based on the extent of LD. Thirty-nine of the SNPs attempted (25%) were technical failures (failing either Hardy-Weinberg equilibrium or to attain a 75% genotyping percentage), and 49 of the remaining 128 SNPs (38%) were either monomorphic in this population or had a minor allele frequency <5%, totaling a final set of 79 working, high-frequency SNPs. The average spacing between these 79 SNPs is 1.5 kb. Haplotype blocks were determined as described in 2.

This study was performed over a period of 3 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 (P.I.W.D., M.J.D., D.A., unpublished software). Tagger combines the simplicity of pairwise methods (35) with the potential for added efficiency of using multimarker predictors. Specific multimarker tests (combinations of alleles) that predict another site are explicitly recorded and included as hypothesis tests in the association analysis. We avoid overfitting by constraining markers of such specific haplotypes to be in strong LD with one other. Tagger is available as a web server at http://www.broad.mit.edu/mpg/tagger/.

To determine how well the final tag SNP set captured variation in the HNF1α region, SNPs were evaluated for their correlation to one another in the CEPH samples described above. Specifically, we recorded the maximal pairwise r² of each tag SNP to the complete set of other variants typed in the region. 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 (>5% frequency), which had an r² > 0.8, to all of the markers typed in the CEPH panel. We note that this is a nonconservative estimate of power, since we have not determined the LD patterns for all SNPs in the region but rather for the one common SNP per 1.5 kb found in dbSNP. Since the total number of SNPs with >5% frequency is ∼1 per 500 bp on average across the human genome, our tags likely capture about one-third of all such sites already in dbSNP. In addition, by choosing a set of tags from such a dense set of SNPs, most (but not all) of the remaining sites will likely also show a high r² value to one of the tags (P.I.W.D., M.J.D., D.A., unpublished observations).

We began our study of common variation in the HNF1α gene region by performing a meta-analysis of previously published literature on this gene in late-onset type 2 diabetes (21–26) (Table 2). None of these studies were individually significant, but there was some consistency in both magnitude and direction of ORs for two common missense polymorphisms, I27L and S487N. The most common missense SNP, I27L, had a suggestive association with type 2 diabetes when all samples were combined (OR 1.25, P = 0.005). The A98V missense polymorphism had a slightly higher estimated effect size, but given its lower frequency, the statistical significance was weaker (1.54, two-tailed P = 0.03).

To see whether we could replicate any of these hypotheses and to extend the analysis to noncoding regulatory regions, we started by evaluating LD patterns across a 118-kb region spanning the HNF1α gene, genotyping 167 SNPs in a panel composed of 30 CEPH trios. The promoter and mouse-conserved regions were also resequenced in 32 diabetic patients, and all discovered SNPs were genotyped in the CEPH panel. (Since the gene has been deeply studied in many labs for a role in MODY, including in one of our labs [L.G.], we did not believe it was necessary to sequence the exons any further.) In total, 79 SNPs were in Hardy-Weinberg equilibrium, with a minimum of 95% genotyping and 5% minor allele frequency (Fig. 1). The gene region shows extensive LD and limited haplotype diversity, with 94% of the sequence in regions of strong and consistent LD (haplotype “blocks”) (31). The combination of strong LD and high marker density suggests that most undiscovered SNPs are likely to be highly correlated to the SNPs already studied in the region.

We chose 21 SNPs from the 79 SNPs in the CEPH LD map to use in association studies, including the three common missense variants. (These 21 SNPs are not an efficient set, as they were chosen over time and contain partially redundant markers.) They provide an r² > 0.9 to all other (untyped) markers in the CEPH panel, indicating that the tag SNPs should provide strong power for both untyped reference panel SNPs and any undiscovered common SNPs when genotyped in the disease panel.

The 21 tag SNPs were genotyped in 2,042 type 2 diabetic patients from Scandinavia and Canada, plus their matched control subjects (family based or unrelated) (Table 1). Eleven tests showed a nominally significant association to type 2 diabetes in this initial panel (Table 3). This includes the I27L missense variant, for which we observed a similar-sized effect as in the meta-analysis (Table 1) (current study: OR 1.13, one-tailed P = 0.01). The most statistically significant result observed was for rs1920792, located 12 kb upstream of the HNF1α start site, which had an OR of 1.17 (two-tailed P = 0.002). The largest OR was observed for the rare missense variant A98V (OR 1.24), although due to its low frequency, this resulted in a one-tailed P value of 0.07.

In light of the prior functional and genetic data surrounding this gene, these initial results were quite encouraging: common variants in HNF1α might play a role in type 2 diabetes. To more conclusively address whether any of the positive common variants in our study are associated with type 2 diabetes, we performed two additional analyses. First, we genotyped the most strongly supported hypotheses from our study (rs1920792, I27L, and A98V) in an additional 4,470 Caucasian type 2 diabetic patients and matched control subjects of U.S. and Polish ancestry. Second, we collaborated with Weedon et al. (36), who were already studying the same gene locus, to align our tag SNPs so that we could directly compare the results of the two studies.

In the Polish and U.S. samples, none of the previously implicated SNPs show evidence for association with type 2 diabetes (Table 4). Similarly, neither the I27L nor the rs1920792 associations were seen by Weedon et al. (36). Of the SNPs evaluated jointly in our complete sample and by Weedon et al., the Val variant of A98V was most consistently associated with increased risk of diabetes in both studies and in the previous literature. However, the suggested OR was very modest, as was the frequency of the SNP (3–6% in our European populations), such that even if the effect turns out to be correct, it will explain very little individual or population risk and be difficult to prove even with collections of 5,000–10,000 samples.

HNF1α is the gene responsible for the most common form of MODY and is found in a region implicated by linkage results in multiple studies, and meta-analysis of previous studies suggests that missense SNPs might be associated with late-onset type 2 diabetes. Thus, before our study and that of Weedon et al. (36), the Bayesian prior probability was quite high that common variation in HNF1α might play a role in late-onset type 2 diabetes. Moreover, in our initial >4,000 patient/control samples, we saw encouraging evidence for association of several variants in the region. When studied in two large, independent samples, however, none of these putative associations were consistently observed. The most conservative conclusion of this study, therefore, is that common variation in the HNF1α region either has no role in late-onset type 2 diabetes, a very modest role (e.g., a modest effect of the quite rare SNP A98V), or a role that cannot be consistently observed without consideration of currently unmeasured genetic or environmental modifiers.

It is not unusual that an initial study will show modest signals for association (as in our initial screen with 21 SNPs in 4,100 people) that fails to be replicated in additional samples. Many explanations can be invoked to explain the lack of replication of our initial reports. First, the initial association may have been a statistical fluctuation. The P values in the initial study were modest after correction for the number of variants studied (on the order of 0.01), and HNF1α is one of many genes being studied by ours and other groups. Thus, encountering such a result by chance is not unexpected, and the similarity to the published literature could reflect past publication bias toward positive results.

A second potential explanation is that common variation at this locus does have an effect on diabetes risk, but that it is even more modest than seen in previous studies and our original panel of patients. That is, the apparent effect sizes for the most promising variants could be inflated by the so-called “winner’s curse” (29). However, given the large sizes of the two nonreplicating studies, any such true effect would have to be quite modest; for example, our analysis places the upper bound on the OR for the minor allele of I27L at 1.15 based on the upper 95% CI of a meta-analysis for all published studies.

A third possibility is that the initial association signal is real (at least in those samples), but that there is heterogeneity among populations, and the variant in question is not associated with risk in our GCI sample or the Weedon et al. (36) sample due to an unmeasured environmental or genetic modifier. We note that a formal test for heterogeneity across the three studies was negative and that such an explanation remains speculative unless such a genetic or environmental modifier is found and a gene-gene or gene-environment interaction is demonstrated. We also note that two widely replicated associations in type 2 diabetes, peroxisome proliferator–activated receptor γ P12A and Kir6.2 E23K, are observed as statistically significant in both the Scandinavian/Canadian and the GCI subsamples (2,9,37,38). However, it could be that the relationship between genotype and particular disease phenotypes varies for different SNPs such that some SNPs will be consistently observed, whereas others may be more sensitive to specifics of case ascertainment and/or phenotypic measurement in each study.

Excepting a consistent and reproducible association, the most straightforward interpretation is that although rare variants in HNF1α can cause an early-onset, autosomal-dominant form of type 2 diabetes, no common variants exist that contribute more modestly to the disease in its typical form. The case of MODY3 is a subset of a question of general importance: whether genes implicated in monogenic forms of disease also explain the heritability of the common form of the disease. Over the last decade, there have been many genes identified that cause Mendelian forms of common diseases, including MODY, maternally inherited diabetes and deafness (39), 20 inherited forms of blood pressure regulation (40), early-onset breast cancer (BRCA1, BRCA2, and ATM), Alzheimer’s (APP, PS1, and PS2), and others. To the extent that genes for common and rare forms of disease turn out to be nonoverlapping, it will suggest that the selective impact of these different forms of the disease, and/or the underlying pathogenic mechanisms, can be less similar than suggested by the shared clinical end point. Furthermore, it will mean that other methods (beyond positional cloning of rare, highly inherited subtypes) will be required to find those genes that explain the evident heritability of the disease.

J.N.H. is a recipient of a Burroughs Wellcome Career Award in Biomedical Sciences. D.A. is a Charles E. Culpeper Scholar of the Rockefeller Brothers Fund and a Burroughs Wellcome Fund Clinical Scholar in Translational Research, the latter of which supported this work. L.G. 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 European Community (BM4-CT95-0662, GIFT), the Swedish Medical Research Council, the JDF Wallenberg Foundation, and the Novo Nordisk Foundation.

We thank T. Frayling and colleagues 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.