Our PrediXcan method tests the mediating effects of gene expression levels by quantifying the association between the genetically regulated levels of expression and the phenotype of interest. To implement this, we developed prediction models of gene expression using large-scale transcriptome study datasets (DGN, GEUVADIS, GTEx). (Summary statistics on samples per tissue for every data release are available from the GTEx portal.) After extensive testing, we chose to use the elastic net model, which performed similarly to LASSO, but substantially outperformed simple polygenic approaches. Manor and Segal36 have published results on robust prediction of expression levels using K nearest neighbor (KNN) and elastic net approaches. Based on their conclusion that a combination of elastic net and KNN along with the use of genomic annotation such as GC content can improve prediction performance, it is reasonable to hypothesize that the incorporation of a more comprehensive functional annotation approach into the PrediXcan framework can yield additional performance gain.
