One strong biological candidate gene in this region is insulin growth factor-1 receptor (IGF1R; MIM: 147370). The IGF1R gene is composed of 21 exons spanning 308 kilobases (kb) on chromosome 15q26.3. IGF1R is a transmembrane tyrosine kinase receptor widely expressed in human tissues with 50% amino acid homology with the insulin receptor (IR). In beta cell-specific IGF1R null mice, beta cells develop normally, but GLUT2 and glucokinase expression are reduced, leading to diminished glucose-stimulated insulin secretion and impaired glucose tolerance [7]. In a combined IGF1R and IR beta cell-specific knockout, beta cells again develop normally, but undergo rapid apoptosis [8], suggesting that insulin and IGF1 signaling are critical in regulating beta cell mass, and defective signaling may contribute to T2DM development.

Of the 54 polymorphic SNPs, none were significantly associated (P<0.002) with variation in Insulin30 (Figure 2A). However, six SNPs with P<0.05 (P=0.011–0.047) were identified (Intron 1: rs874305, rs4966009, rs4966012, rs1319859; Intron 2: rs2670504, rs2715440) (Supplemental Table 2). As with Insulin30, none of the SNPs were significantly associated with ISI at P<0.002 (Figure 2B). Associations for ISI of P<0.05 were observed for three clusters of SNPs in intron 1 (rs874305, rs4966012, rs1319859), intron 2 (rs2670504, rs2684778, rs2715440, rs7165181, rs4966035), and intron 5 (rs3743259, rs4966038) (Supplemental Table 2).

While no 4-SNP haplotypes demonstrated association with Insulin30 (data not shown), the CTTG haplotype of SNPs rs17847195, rs2715439, rs8034284, and rs12440962 (haplotype 48) in intron 20 was associated with lower ISI levels (β = 0.18; 95% CI: 0.08, 0.29; genotypic means: CTTG/CTTG 1.37±0.09 vs. CTTG/CTCG 1.54±0.09 vs. CTCG/CTCG 1.73±0.11) (Figure 3), exceeding our multiple-testing-adjusted significance threshold of α=0.002 with P=0.001. Haplotypes 46 (rs8038056-rs41497346-rs17847195-rs2715439), 47 (rs41497346-rs17847195-rs2715439-rs8034284), 49 (rs2715439-rs8034284-rs12440962-rs2593053), and 50 (rs8034284-rs12440962-rs2593053-rs17847203), which flank haplotype 48, displayed associations with lower ISI consistent with haplotype 48 at P<0.05 (Supplemental Table 4).

We generated posterior estimates of power to detect association of Haplotype 48 with ISI levels in the sample of 468 AFDS members with complete haplotypic and phenotypic data using QUANTO [18] and found that we had 86% power to detect a statistically significant difference of 0.18 in the ISI levels between genotype at α=0.002 under an additive model for a haplotype with frequency of 0.16.

We tested association of the SNPs in haplotype 48 with T2DM in 139 AFDS T2DM cases and 468 AFDS controls. While none of the SNPs were individually associated with T2DM, the CTTG haplotype was modestly associated with elevated T2DM risk (P=0.02) in a model adjusted for sex, age, age², BMI, and Insulin0, though not at the gene-wide significance threshold of α=0.002.

While some associations of genes with diabetes may be spurious [19], studies in mice suggest that the association of IGF1R variants with insulin secretion may underlie a true relationship. As described previously, mice with beta cells lacking IGF1R demonstrate defective glucose-driven insulin secretion [7], suggesting IGF1R is critical for proper beta cell function. Similarly, mice with beta cell-specific knockouts of both IR and IGF1R are born with a normal complement of beta cells, but glucose-stimulated insulin secretion is blunted. As a result, by 3 weeks, the mice developed diabetes [8].

Previous studies of IGF1R sequence variation and T2DM risk are conflicting. Rasmussen et al. [18] analyzed IGF1R coding sequence in 82 Danish subjects with T2DM, and found no variants associated with T2DM or related endophenotypes including reduced birth weight and insulin sensitivity index. In contrast, in the Finnish Diabetes Prevention Study [19], a common synonymous coding variation in IGF1R (GAG1043GAA) (rs2229765) was significantly associated with differences in conversion rates from impaired glucose tolerance (IGT) to T2DM in subjects participating in a weight loss program, suggesting that this polymorphism increases T2DM risk among those with IGT. In our study, rs2229765 was not associated with differences in Insulin30 (P=0.9) or ISI (P=0.9) and was marginally associated with diabetes (P=0.008) using α=0.002 for statistical significance. Rs2229765, however, is present in a haplotype associated with ISI (haplotype 42), and is in strong LD with several of the SNPs in the associated haplotype in intron 20. One possible explanation for the lack of association at rs2229765 in our study may be differences in the LD patterns underlying IGF1R among different populations.

Several features of our study may have improved our power to detect these associations with IGF1R. First, our study was performed in subjects from a genetic isolate, thereby, minimizing confounding effects from population stratification. Second, our background LD is likely higher than in most Northern European samples [20], such that we may have better gene coverage than other outbred populations. Third, AFDS participants have similar occupational and environmental exposures [21] as they are predominantly farmers and homemakers from Lancaster, Pennsylvania, thus decreasing confounding effects from these factors. Fourth, our power to detect association with IGF1R may have been increased by analyzing T2DM subphenotypes (Insulin30, ISI) since these subphenotypes are likely to be less genetically complex. Finally, the existence of linkage to insulin levels and T2DM in more than one population near IGF1R [6,22,23] may have reduced the likelihood of a false-positive error due to multiple testing [24].

Despite the strengths of our study, the study was subject to several limitations. While we found our strongest association with a haplotype, type I error may be present in our analyses as a result of inferring haplotypes with only nuclear pedigrees. Another potential source of error may be the use of the most probable inferred haplotypes, though the average posterior probability for inferred haplotypes was 0.66, indicating that it is unlikely to be a major source of error. To minimize these limitations, an adjusted threshold for statistically significant association was used. Among the genotyped SNPs, we observed high replication rates; however, our call rates were lower than anticipated (average 89%). Similar replication and call rates have been reported by others using SNPStream UHT for genotyping [25]. To ensure that no systematic error was present, we examined genotyping plates and found the proportion of uncalled genotypes to be consistent across plates. Moreover, the fact that all SNPs were in HWE provided no strong evidence for under-calling or over-calling of heterozygosity. Furthermore, no statistically significant differences in trait distributions among those with and without genotype calls were found. Another potential limitation was that 45 of 99 SNPs were monomorphic, had a MAF <0.05 or were unsuccessfully genotyped. Despite this problem, the remaining 54 SNPs demonstrated a high degree of pairwise LD (average r²=0.94) and achieved a high degree of coverage of the gene. Finally, as the Amish are genetically isolated, it is possible that our association of IGF1R with ISI may not be generalizable to other populations; however previous genetic studies of T2DM and related traits in the Amish have found results concordant with other Caucasian populations [26,27], suggesting that our findings in the Amish will likely be relevant to more outbred populations.

While we did not observe statistically significant single-SNP associations, there are several reasons to believe that the observed haplotypic association is real. Haplotypic association may have increased power to detect genetic effects at ungenotyped causal SNPs through indirect association, as haplotypes may be better than individual SNPs at tagging causal SNPs via LD, as exemplified in Martin et al. [28] for APOE and Alzheimer’s disease and Veal et al. [29] for PSORS1 and psoriasis, among others [30,31]. Haplotype frequencies may approximate more closely the allele frequency of a causative SNP [32], which may be the case with haplotype 48. Additionally, haplotypes themselves may contribute to risk for disease, as several studies have identified important haplotypic risk variants for disease such as the APOE ε3/ε4 alleles for Alzheimer’s disease [33], β2-adrenergic receptor (β2AR) in bronchodilator response [34], and complement factor H (CFH) in age-related macular degeneration [35]. However, given that the haplotype is only modestly associated with ISI after adjustment for multiple hypothesis testing, this association needs to be replicated in an independent population.

The experience of our study mirrors that of many others insofar as it illustrates the challenges of going from a positive linkage result to identifying a causative SNP. One possibility (among several) is that the effect detected through linkage analysis may be attributable to one or more rare variants, while our strategy for identifying associated SNPs was predicated on a tagging SNP approach geared towards detecting common SNPs. Identifying the functional SNPs whose frequencies are rare remains a very challenging enterprise that may require both additional approaches for SNP discovery (e.g., sequencing) as well as functional assessment of associated SNPs.

Supp Tables