查看文章

Background: In the field of common genetic variant analyses, polygenic risk scores (PRS) have become standard practice for estimating risk on a target test set using GWAS summary statistics of a training set, much like in typical machine learning (ML) applications. To date few groups have reported systematic comparisons of ML techniques applied to common disorders, especially with regard to PRS. Moreover, the use of genomic annotations that incorporate prior functional information can enhance this risk prediction. Here we extract pathway-specific polygenic scores and test their performance as features in a number of prediction and classification schemes. We thus investigate their potential as a flexible tool for individual prediction, that leverages both multi-dimensional information and the advantages of ML.

Methods: We used simulated genome-wide cohorts created with HAPGEN2, as well as real data from the …