Low predictive performance on WES

Nora Verplaetse

Hi,

I'm performing predictive analysis for Inflammatory Bowel Disease on WES using different WES representations and models. Overall I can not get my test set performance over 55% ROC AUC, also with linear models on some associated set of SNPs. I'm selecting age and sex matched controls out of the cohort with ‘ICD10-diagnoses is NULL’. Any experiencing the same of any tips?

Thank you!

Comments

1 comment

Date Votes
  • Comment author
    David Curtis

    I'm not sure how you're treating the WES data but if you're using individual variants (not gene-wise burdens) and if you only have variants in exons then I certainly wouldn't expect the AUC to be anything higher than 55%. Frankly, you're lucky to get that. You have to consider the SNP heritability and the fact that the exons only contain 2% of SNPs. And rare variant heritability is very low unless you can specifically identify the (very small number of) variants which are pathogenic. (You can't.) You can check out my papers on rare variant analyses in WES data to get a better understanding.

    1

Please sign in to leave a comment.