Reader Comments

Post a new comment on this article

Non-MHC SNPs fail to improve performance of the MHC-only model.

Posted by antoinelizee on 13 Mar 2014 at 23:26 GMT

I enjoyed reading your article and discussed it at our journal club.
We were surprised that the model with MHC-only SNPs had the same performance as the model with all SNPs (see figures 3b and S2, AUC = 0.87). Since the non-MHC SNP model had respectable performance (AUC = 0.72), we thought that adding those SNPs to the model would be helpful, assuming that non-MHC SNPs have independent effects from MHC SNPs.

Best,
Antoine

No competing interests declared.

RE: Non-MHC SNPs fail to improve performance of the MHC-only model.

gabraham replied to antoinelizee on 20 Mar 2014 at 23:12 GMT

Dear Antoine,

Thank you for your comment. In terms of predictive power, the MHC-only models were indeed very similar to the genome-wide models, even though they are based on somewhat different sets of SNPs (the top ones are in HLA in both cases). Our results suggest that there is a high degree of redundancy between the two signals, which means that non-MHC SNPs are somehow correlated with ("tagging") some aspects of the MHC SNPs. Also, the AUC is a rank-based statistic and is quite insensitive to small model differences which may still be biologically interesting in their own right.

Regards,
Gad

Competing interests declared: I am one of the paper's authors.