Background

Type 2 diabetes is a cause of poor health and early death that is spreading worldwide and exerting a fearsome human and economic toll [1, 2]. Prevention and control of diabetes requires a better understanding of its basic molecular causes. Type 2 diabetes is a heterogeneous disease arising from physiological dysfunction in the pancreas, skeletal muscle, liver, adipose and vascular tissue. Much of the heterogeneity of type 2 diabetes has a genetic basis. A full picture of the complex genetic architecture of diabetes has been elusive [3–7].

Among type 2 diabetes susceptibility genes few, if any, individual loci are expected to carry alleles of major effect explaining a substantial proportion of cases, although a few genes could have a substantial population effect but not give a strong genetic signal if the causal alleles were common and the increase in risk were modest [6, 7]. Such genes have proven hard to detect using linkage-based approaches, although recent rapid advances in genetic association methodologies have led to some successes. The P12A polymorphism in the gene encoding the peroxisome proliferator-activated receptor-g (PPARG) [7], the E23K polymorphism in the gene encoding the islet ATP-dependent potassium channel Kir6.2 (ABCC8-KCNJ11) [8–10] and common variants in the gene encoding the transcription factor 7-like 2 gene (TCF7L2) [11, 12] were all found using well-powered association mapping, and all have been reproducibly associated with diabetes in diverse samples at highly significant p-values.

Current gene discovery strategies have focused on coding regions, but regulatory variants also influence disease [11, 13, 14]. A comprehensive picture of diabetes genetics will require a wide and adequately dense search across coding and conserved non-coding genomic regions using an association analysis approach, where power is superior to linkage analysis when seeking common variants of modest effect [6]. Resources are now becoming available to perform such genome-wide association (GWA) studies of type 2 diabetes [15–18].

In this report we describe the Framingham Heart Study (FHS) Affymetrix 100K SNP genome-wide association (GWA) study resource for type 2 diabetes. This resource complements the several other large extant type 2 diabetes GWA studies in three major respects: it is population-based (not diabetes proband-based), studies two generations, and has decades of longitudinal, standardized, detailed follow-up. We describe results of a simple low p-value-based SNP selection strategy and an alternate novel SNP selection strategy that takes advantage of the unique FHS diabetes-related quantitative traits data. We use FHS 100K SNPs in an in silico replication analysis that tests the hypothesis that SNPs in LD with published causal variants in PPARG, ABCC8, TCF7L2, CAPN10, and HNFa are associated with diabetes and related quantitative traits.

Methods

Study subjects

The study sample is described in the Overview Methods section [19]. With respect to diabetes-related traits, Offspring subjects provided genotypes and diabetes-related traits to the analyses, and Offspring parents from the Original FHS Cohort contributed genotypes for linkage analysis and FBAT statistics. Of 1,345 FHS subjects with 100K SNP data, 1,087 were Offspring and of these 560 were women, the mean age at exam 5 was 52 years, and the mean age at last follow-up was 59 years. Every study subject provided written informed consent at every examination, including consent for genetic analyses, and the study was approved by Boston University's Institutional Review Board.

Genotyping and annotation

Affymetrix 100K SNP and Marshfield STR genotyping are described in the Overview Methods section [19]. Genotype annotation sources are described in the Overview Methods section [19].

Diabetes phenotyping

Diabetes and related quantitative traits have been ascertained at every FHS exam for every generation. Diabetes-related quantitative traits available in the FHS 100K resource are displayed in Table 1. FPG data for the analyses came from all 7 Offspring exams, but the remainder of the data came from exam 5 (1991–94), when subjects without diagnosed diabetes underwent a 75 gram oral glucose tolerance test, or exam 7 (1998–2001), the most recent exam. We defined diabetes as chart-review-confirmed diabetes, new or ongoing hypoglycemic treatment for diabetes at any exam, or a FPG > 125 mg/dl at two or more of the seven exams. Diabetes age-of-onset was defined as the subject's age at the exam at which diabetes was first identified. Among Offspring with diabetes, >99% have type 2 diabetes [4]. Of the 1,083 Offspring with 100K genotypes and known diabetes status, 91 had diabetes. The mean age of onset of was 58 yr; through exam 7, 9.3% of diabetic subjects had developed diabetes by age 40 yr, 33.0% by age 50, 68.1% by age 60, and 99.7% by age 80.

Table 1 Type 2 diabetes-related quantitative traits in 1087 Framingham Offspring Study subjects with 100K genotype data
Full size table

In this presentation we focus on six (three glucose and three insulin) primary Offspring diabetes-related quantitative traits. Glucose traits are fasting plasma glucose (FPG) and hemoglobin A1c (HbA1c) measured at exam 5, and up to 28 yr time-averaged FPG (tFPG) level obtained from the mean of up to seven serial exams. Glucose traits included all subjects, including those with diabetes regardless of treatment, as these were the most informative subjects with respect to hyperglycemia. Subjects with diabetes had the highest glucose values when subjects were ranked with respect to any glucose trait; those on treatment had the highest values. The three insulin traits are fasting insulin, homeostasis model-assessed insulin resistance (HOMA-IR), and Gutt's 0–120 min insulin sensitivity index (ISI_0-120) measured at exam 5. Subjects with insulin-treated diabetes were removed from all insulin trait analyses, as we had no information on insulin dose and so measured insulin values were confounded by insulin treatment [20–22]. We also analyzed incident diabetes from first exam through last follow-up. We previously have described FHS laboratory methods for these diabetes-related quantitative traits [4, 23–25]. In addition to glucose and insulin traits, levels of adiponectin and resistin are available in the FHS dbGaP resource. Plasma adiponectin and resistin concentrations were measured using a commercial ELISA (R&D Systems, Minneapolis, MN); inter- and intra-assays CVs were 5.3%–9.6% for adiponectin and 7.6%–10.5% for resistin.

SNP prioritization

We used two approaches to prioritize SNPs potentially associated with diabetes or diabetes related traits. In the first, we simply ordered SNPs from lowest to highest p-value for association with one or more of the six primary glucose and insulin traits. We also ordered SNPs or Marshfield STRS by highest to lowest LOD score for linkage to one or more of the six primary traits, and present LOD scores > 2.0. In an alternative SNP prioritization strategy, we selected SNPs associated with multiple-related traits. In this approach, we selected SNPs with consistent nominal associations (p < 0.01 in GEE or FBAT) with all three glucose traits OR all three insulin-related traits OR (two glucose and two insulin traits). Among these we used extent of LD to select a non-redundant set of SNPs; when several were perfect proxies for each other (r2 ≥ 0.8) only one SNP was selected, based on the highest genotyping call rate.

Statistical analysis

The general statistical methods for linkage and GWA analyses are described in the Overview Methods [19]. For diabetes-related quantitative traits we used additive GEE and FBAT models, testing associations between SNP genotypes and age-age2-sex-adjusted residual trait values. We kept 70,987 SNPs in the analyses that were on autosomes, had genotypic call rates ≥ 80%, HWE p ≥ 0.001 and MAF ≥ 10%.

We tested association of 100K SNPs with incident type 2 diabetes in two additional models using the same adjustment strategy. First, Martingale residuals were created to measure the age-of-onset of type 2 diabetes; residuals were analyzed with FBAT [26]. Individuals with lower values of this 'martingale residual' trait developed diabetes at younger ages, and those with the highest values had been observed for the longest time without development of diabetes [27]. Second, we used a Cox proportional hazard survival analysis with robust covariance estimates in order to find SNPs associated with development of diabetes over all seven exams [28].

Results

Diabetes-related quantitative traits available in the FHS 100K SNP resource are listed in Table 1 and posted on the NCBI web site [29]. Each trait is available as an age-age2-adjusted or age-age2-BMI-adjusted residuals from sex-specific models. In this analysis we only consider the age-age2-adjusted traits. Among these, the following were the primary traits used in this analysis: exam 5 fasting plasma glucose (FPG; n with data = 1,027; mean, SD 99, 24.7 mg/dl); exam 5 HbA1c (n = 623; 5.28, 0.9%); 28-year time averaged FPG (tFPG; n = 1,087; 98, 16.2 mg/dl); exam 5 fasting insulin (n = 982; 30.1, 16.4 uU/ml); exam 5 HOMA-IR (n = 980; 7.8, 7.3 units); and the 0–120 min insulin sensitivity index (ISI_0-120; n = 935; 26.1, 7.6 mg·l2/mmol·mU·min). Among 1,087 Offspring with 100K SNP data there were 91 cases of type 2 diabetes. Additional diabetes-related quantitative traits not used in this analysis but that are available in the FHS 100K SNP dbGaP resource include, at exam 7: FPG (n = 987; 103, 26 mg/dl); fasting insulin (n = 999; 15.8, 12.8 uU/ml); HOMA-IR (n = 969; 4.2, 4.1 units); HbA1c (n = 893; 5.59, 0.97%); resistin (n = 831; 14.5, 7.4 ng/dl); adiponectin (n = 828; 9.9, 6.2 ng/dl).

The six primary quantitative traits had significant associations with 415 SNPs in GEE models and 242 SNPs in FBAT models, using p-value < 0.001, and only considering SNPs with call rate ≥ 0.80, HWE p-value ≥ 0.001, and MAF ≥ 10%. Additionally, there were 91 significant associations with incident diabetes in the survival analyses and 42 significant associations with age-of-onset in FBAT, representing 128 non-overlapping SNPs. The 25 SNPs with lowest p-values in GEE or FBAT models, and LOD scores > 2.0 in linkage analyses, are displayed in Table 2. After accounting for the overlap between sets of significant associations, 736 non-overlapping SNPs were identified by the p-value approach for SNP prioritization.

Table 2 Twenty five lowest p-values from GEE and FBAT models and LOD scores > 2 for 100K SNPs and FHS diabetes-related quantitative traits
Full size table

The FHS has multiple measures of diabetes-related quantitative traits. We used a multiple-related trait approach in a strategy different from prioritizing SNPs based solely on small p-values. This approach yielded 203 SNPs associated with multiple traits. Of these, 53 were also associated with incident diabetes (p < 0.01 by GEE or FBAT). We defined redundant SNPs as those in LD with r2 >= 0.80 to select 168 non-redundant SNPs associated with multiple traits; 42 of these non-redundant SNPs also were associated with incident diabetes (Table 3). Examination of the multiple trait-based approach revealed 1) consistent associations of traits with SNPs that were in LD (providing reassurance that the signal was due to an association of traits with a particular genomic region rather than to technical error); 2) several putative associations of traits with SNPs in the same gene but not in perfect LD (suggesting that the association signal may be due to a functional role of that gene rather than a statistical fluctuation); and 3) associations of traits with SNPs in a variety of novel but plausible biological candidate genes.

Table 3 Forty two (42) SNPs associated with (FPG, HbA1c, and tFPG) OR (fasting insulin, HOMA-IR, and ISI_0-120) OR (any two of either) AND incident DM
Full size table

We used the UCSC Genome Browser (http://genome.ucsc.edu/; accessed September 2006) to annotate SNP details [30, 31]. Of the 823 (736 + 203; 116 overlapped) SNPs identified by both prioritization methods without removing SNPs in LD (r2 >= 0.80), 304 (36.9%) were in genes, 173 (21%) were within 60 kb of a known gene and 5 (0.61%) were coding. For comparison, of the 70,987 SNPs included in this analysis, 25,916 (36.5%) were in genes, 14,333 (20.2%) were within 60 kb of a known gene and 421 (0.59%) were coding.

Some SNPs had p-values < 0.001 overlapping more than one analytical method. For instance, 18 SNPs were associated at p < 0.001 with at least one quantitative trait in both the GEE and the FBAT analyses. For incident diabetes, 5 SNPs were associated with diabetes survival in the Cox models and with age-of-onset in the FBAT analyses.

We used the FHS 100K array data to verify, in silico, replicated associations of reported diabetes candidate genes (Table 4). We found 7 SNPs in or near TCF7L2. One 100K SNP (rs7100927) was in moderate LD (r2 = 0.5) with TCF7L2-associated SNP rs7903146 and was nominally associated with a 56% increased relative risk of diabetes (p = 0.007) and with tFPG (GEE p = 0.03). We found 6 SNPs in or near ABCC8, but no SNPs in strong LD with ABCC8 A1369S (rs757110) or KCNJ11 E23K (rs5219), and thus could not replicate these associations. One 100K SNP (rs878208) ~25 kb upstream of ABCC8 showed nominal association with risk of diabetes, but it was not in LD with rs757110 in ABCC8 (r2 = 0.04). We found 15 SNPs in or near PPARG, but none were associated with diabetes. Four SNPs were associated (p < 0.05) with quantitative traits but were not in LD (r2 < 0.03) with PPARG P12A (rs1801282), the variant previously associated with type 2 diabetes [7]. We found no polymorphic (MAF > 1%) 100K SNPs in, near, or in LD with CAPN10 or HNFA.

Table 4 FHS 100K SNP Test of Association with SNPs in Established Candidate Genes for Type 2 Diabetes
Full size table

We also assessed our approach for confirmation of 4 SNPs associated with FPG reported on the Boston University Department of Genetics and Genomics public site http://gmed.bu.edu/about/index.html that displays selected associations with FHS 100K data. We found no association (all p-values > 0.6) of incident diabetes or levels of FPG with SNPs rs10495355, rs9302082, rs10483948, or rs1148509.

Discussion and conclusion

In this paper we describe the characteristics and initial GWA results for type 2 diabetes and related quantitative traits in the FHS 100K SNP resource. Over 1000 men and women from a community-based sample have detailed linkage and association of diabetes-related phenotypes and 100K dense array SNP results available on the web. About 0.3%–0.6% of SNPs in the 100K array with MAF > 10% are associated at p < 0.001 with six diabetes-related quantitative traits or with incident type 2 diabetes. A similar proportion of SNPs in the array (0.21%) are associated with multiple related diabetes traits. These several hundred SNPs likely contain more false positive than true positive associations with diabetes and related traits, however, they offer logical next targets for the follow-up replication studies in independent samples necessary to resolve true diabetes risk genes. The FHS 100K data replicate the otherwise widely-replicated TCF7L2 association with diabetes [11, 12, 32–40] in an in silico analysis.

The FHS 100K SNP data resource has potential value to detect and replicate novel type 2 diabetes susceptibility genes. The 100K SNP array is limited by relatively sparse coverage in some regions, accounting on average for just 30%–40% of the human genome in whites [17, 41]. Association with the risk SNP in TCF7L2 is detectable at p < 0.05, but there are no SNPs in adequate LD with ABCC8 or PPARG to assess replication of causal SNPs in these accepted diabetes susceptibility genes. Thin coverage will be remedied to a large degree by the incipient availability in FHS of Affymetrix 500 k SNP array data as part of the planned FHS SHARe Study. (http://www.nhlbi.nih.gov/meetings/nhlbac/sept06sum.htm; accessed September 2006) Our analysis also demonstrates that true positive diabetes susceptibility gene signals are likely to be associated with modest p-values and will remain challenging to detect at the stringent p-values required for GWA studies. The enormous datasets generated by GWA scans have the potential to greatly advance understanding, or conversely to overwhelm the field with false leads. SNP prioritization strategies that leverage the complexity of the diabetes phenotype may offer some advantages over strictly p-value driven approaches. Replication, fine mapping, and functional studies are required to determine which approaches are most efficient and which SNPs are true positive diabetes risk factors. Integration with other GWA scans in similar cohorts will allow in silico replication of significant findings, increase power and reveal generalizability.

This report details the FHS contribution to publicly available diabetes-related genetic data. An important key to efficiently and economically achieving adequate power to detect association will be to integrate information from several GWA scans. While several cohorts have been assembled to perform GWA scans in type 2 diabetes, few possess the wealth of longitudinal, multigenerational phenotypic data available in Framingham. The FHS complements extant type 2 diabetes GWA studies. This report guides the way to harness the power of the FHS 100K SNP GWA resource to identify type 2 diabetes susceptibility genes.