Re: "underlying genetic models of inheritance in established type 2 diabetes associations".

Genome-wide association studies have identified a large number of disease susceptibility loci of unknown function, characterized by high allele frequencies and low relative risks. This is also the case for type 2 diabetes for which Salanti et al. (1) identified 17 replicated associations from 19 studies. The odds ratios (ORs) for risk homozygotes ranged from 1.10 to 1.61. The authors compared genetic models for genotype effects and found that for 13 loci the data fitted best to the additive model, while for the remaining loci the discrimination between the additive and other models was not clear. The authors concluded that their results may be useful for predictive modeling and for designing biologic and functional experiments. In type 2 diabetes, as in some other common diseases, the identified variants explain a much larger proportion (approaching 100%) of the disease etiology, measured by the population attributable fraction, than of the familial risk (only about 2.5%) (1). Probably the most important reason for this apparent discrepancy is in the genotyping platforms used to identify susceptibility variants. Current platforms include common variants with allele frequencies of >10%, whereby the findings are constrained to high population attributable fraction and low attributable familial risks (2). The proven power of this approach is that marker signals point to putative functional variants that are in linkage disequilibrium with them. However, the use of markers instead of functional variants has 2 important consequences to the genetic parameters, as we have shown elsewhere: The unknown rare functional variants convey a higher familial risk than predicted by the marker, and the genetic inheritance model of the functional variant cannot be predicted from the marker (3). These restrictions also apply to loci identified for type 2 diabetes, for which functions are largely unknown. Even for gene regions with known functional variants, such as PPARG and KCNJ11 (4), it is unclear if they capture most of the genetic variation in these genes. In the simulation shown in Table 1, based on previously published code (3), we assume that the marker alleleM tags the causative variant C; M is more frequent than C, but C is always found together with M; that is, D# 1⁄4 1.0. There are thus 3 haplotypes, c-m, c-M, and C-M. The association signal for M is entirely due to the functional effect of C. The frequency of C is 1/10 of that of M (pM 1⁄4 0.5, pC 1⁄4 0.05). We further assume a dominant (ORC_Het 1⁄4 ORC_Hom), additive (ORC_Het 1⁄4 (ORC_Hom þ 1)/2), or recessive (ORC_Het 1⁄4 1) penetrance for C. When the odds ratio of the dominant causative allele C is 2.0, the odds ratio of M is 1.10 for heterozygotes (Het) and 1.19 for homozygotes (Hom). Notably, the population attributable fraction is 8.88%, and it is equal for C and M. When the odds ratio for homozygote carriers of the additive causative allele C is 2.0, the odds ratio of M is 1.05 for heterozygotes and 1.10 for homozygotes. Recessive C yields very low odds ratios for M, whereby these would be difficult to find in genome-wide association studies. These data show that the observed odds ratios for M are essentially lower than the underlying odds ratios for the causative allele. Moreover, a dominant effect for the causative allele results in nondominant associations between the marker and the disease. The penetrance mode ofM in Table 1 is close to additive under both dominant and additive penetrance for C. The present data caution against using marker genotype data to infer genetic models for