We analyzed SNP data for 33 GV candidate genes in a large non-Hispanic white case–control data set that was subjected to stringent data and population quality control and adjustment for population stratification in our previous GWAS (Jin et al., 2010a
). This resource provides 80% power to detect significant association with common alleles at ORs in the range 1.20–1.25 across a wide range of allele frequencies.
We found evidence of primary genetic association with GV for only three of the candidate genes tested, FOXP3, TSLP, and XBP1, all of which met the Bonferroni-corrected significance thresholds adjusted on the basis of both the number of genes (0.05/33) and the number of LD blocks (0.05/80) tested, although with marginal significance. Nevertheless, the meta-analysis provided strong support for the association of GV with XBP1 (meta-P= 9.5E-09, OR= 1.21), with the pattern of associated SNPs indicating that the same causal allele may exist in both the Chinese and the non-Hispanic white populations. In contrast, association of GV with SNPs in the TAP1-PSMB8 region of the MHC seems to derive from LD with primary association signals in the MHC class I and class II regions. Furthermore, as suggested by several previous studies, the apparent association of GV with CTLA4 seems to be secondary, driven by primary association of CTLA4 with other autoimmune diseases that are epidemiologically associated with GV.
We also subjected the SNP data to permutation analysis to assess whether the observed P-values were truly significant. The permutation analysis of SNP data for individual genes generally supported significance for FOXP3 (rs3761547, P= 3.3E-02; rs11798415, P= 5.6E-03), TSLP (rs764916, P= 5.9E-03), and XBP1 (rs2269577, P= 1.8E-02; rs6005863, P= 8.0E-03). However, combined permutation analysis of SNP data for all 33 genes studied failed to support the significance of any of the observed results. This may indicate that the actual number of total LD blocks represented is much greater than 80, and a more conservative Bonferroni-adjusted significance threshold would be more appropriate. Alternatively, the combined permutation analysis may itself be too conservative, as several of the genes tested are quite large (and thus are represented by a large number of SNPs) and have only questionable status as valid biological candidate genes in the first place.
Finally, we observed no evidence of association of GV with SNPs tagging 28 of the 33 candidate genes tested: ACE
, and VDR.
In the case of FOXD3
, this might reflect the involvement of the gene in only one unusual family with both atypical presentation and inheritance of GV. In the case of PDGFRA
, this might reflect the inability of association methods to detect rare causal gene variants. For many of the others, genetic association has previously only been tested by relatively small studies carried out in populations other than non-Hispanic whites, and it is possible that different populations have different causal variants. Nevertheless, it is widely recognized that small case–control candidate gene association studies are very often flawed by statistical fluctuation, inadequate correction for multiple testing, and population stratification, and that the great majority of such reported “associations” are therefore spurious (Hirschhorn et al., 2002
; Freedman et al., 2004
). Such studies must thus be interpreted with great caution until validated by repeated replication or by studies that use more robust methods. This analysis, by far the largest association study of GV candidate genes ever carried out, failed to support the association of most candidate genes reported for GV. Although our results do not completely exclude the possible involvement of these genes in disease pathogenesis, our findings nevertheless underscore the extreme unreliability of candidate gene studies in identifying true causal genes for disease susceptibility.