We also tested the prediction models trained in the DGN whole blood cohort on several independent test cohorts with available whole-genome genotype and transcriptome data. We used weights derived from the DGN whole blood data (“training set”) to predict gene expression levels (treated as quantitative traits) in GEUVADIS LCLs (lymphoblastoid cell lines) and nine GTEx pilot tissues (“test sets”). Figure 4 provides a Q-Q plot showing the expected (under the null, correlation between two independent vectors with the same sample size) and observed R2 (between observed and predicted) from the elastic net prediction in GEUVADIS LCLs. We find a substantial departure from the null distribution indicating that the elastic net model trained in DGN (equation 2 of Materials and Methods, with effect size estimates w^k,gEN) captures a significant proportion of the transcriptome variability. The average prediction R2 is 0.0197 for GEUVADIS LCLs. For GTEx tissues, the prediction R2 values are 0.0367 (adipose), 0.0358 (tibial artery), 0.0356 (left ventricular heart), 0.0359 (lung), 0.0269 (muscle), 0.0422 (tibial nerve), 0.0374 (sun exposed skin), 0.0398 (thyroid), and 0.0458 (whole blood). Interestingly, we also find
