In this study we failed to find an association between markers rs7775228 and rs10947262 and knee OA in European descent samples in spite of a sample size large enough to have sufficient statistical power to confirm the effect sizes previously reported and even much more modest genetic effects. We could not replicate in European descent samples the findings of Nakajima and co-workers 
who identified a genome-wide significant association between these two SNPs and knee OA in Japanese subjects. Furthermore, Shi et al 
were unable to find an association in Chinese OA patients for these same SNPs.
There have been several studies analyzing genomewide significant variants directly across ethnicities with varying results 
. Because of differences in underlying LD structure and variant frequency between ancestral populations the tagging relationship between a SNP and a pathogenic variant may be disrupted. Therefore a non-replication of a particular variant must be considered inconclusive and warrant further investigation as demonstrated in a recent study of breast cancer at the ESR
1 locus 
. Having analyzed the LD between the two SNPs ( rs7775228 and rs10947262) identified by GWAS observation in the Japanese study by Nakajima and co-workers 
and the classical loci in the HLA class II region, we can infer that the DRB1*1502 DQA1*0103 DQB1*0601 haplotype is protective for risk of OA in Japanese samples. Thus, association of these two SNPs with OA in the Japanese samples may be due to their property of tagging this haplotype. The frequency of this HLA haplotype varies between ethnic groups; it is present in 8% frequency in Japanese samples, 2.3% in Chinese samples and is only in 0.7–0.8% frequency in European samples. Moreover, the two SNPs are not in LD with this haplotype in European or in Chinese samples, possibly explaining the non-replication of these SNPs. A suitable tag SNP for this particular DRB1-DQB1 haplotype in European samples is rs2858880 
which has r2
~1.0 with DRB1*1502 DQA1*0103 DQB1*0601. The minor allele at this SNP is very rare and to achieve 80% power and p<0.05 even assuming a strong protective effect (OR
0.7, i.e. non carriage of this haplotype results in OR
1.4) it would require 10,586 knee OA cases and 10,586 controls, i.e., almost twice as many cases as were tested in our study. Ideally, however, the HLA loci should be genotyped directly (e.g.using PCR-based sequence specific oligonucleotide probe methods) in order to achieve conclusive validation results and to assess the role of any other DRB1-DQA1-DQB1 haplotypes in risk of OA.
We note that there are some potential study limitations, in particular the different definitions of knee OA used by the various studies, some of which focus on total knee replacement as an outcome and others that have focused on radiographic features and for OA phenotypic definitions which reflect a different subset of OA have been shown to influence the ability to detect genetic associations 
. The initial discovery was made in Japanese samples using a definition of radiographic OA (K/L≥2) and pain 
. The association with rs10947262 initially was replicated in a Greek study where the outcome as total knee replacement (TKR) but the association with rs7775228 was not replicated in that same cohort 
. In fact, rs7775228, the Icelandic, Spanish and Greek studies all used TKR as the outcome and show no evidence of association. For rs10947262 we cannot distinguish a clearly different pattern for studies where only a radiographic definition was used (e.g. Chingford, TwinsUK, HCS, Estonia) from those where symptoms or even TKR were part of the recruitment criterion (e.g. Nottingham, GOAL, deCODE). Further, for knee OA it has been shown in a case where the functional variant of a gene is truly involved in risk of disease (the promoter variant in the GDF-5
gene) that, these differences in phenotype definition do not deter from observing an association in Caucasians with genome-wide statistical significance even if it had been originally discovered in Asian samples 
Another study limitation to our conclusions is that the data used to derive inferences concerning LD with HLA haplotypes comes in part from imputed data. Therefore, although imputation for this SNP appears to be reliable, it is unlikely but possible that the conclusions might have been different using directly typed data.
Neither our results nor those from Shi et al 
preclude an important role for variation in the HLA region in genetic susceptibility to OA. It is indeed possible that class II antigen presenting molecules may be involved in the pathogenesis of OA. In order to determine if DRB1 -DQA1-DQB1 haplotypes (and antigen presenting molecules in general), are involved in risk of knee OA in Caucasian populations, further studies which take into account the complex LD structure and the extremely high genetic diversity of this region are needed.