For each gene, LASSO, the elastic net and the simple polygenic score were used to provide an estimate of GReX̂ (using equation 2, with the effect size estimates w^k,gLASSO, w^k,gEN and w^k,gPS, respectively). We included only local SNPs (within 1Mb of the gene start or end). In order to determine the optimal modeling method, we compared the 10-fold cross-validated prediction R2 (the square of the correlation between predicted and observed expression) for the simple polygenic score (GReX^PS) at several p-value thresholds (single top SNP, 1×10−4, 0.001, 0.01, 0.05, 0.5, 1) with that from LASSO (GReX^LASSO) and elastic net (GReX^EN).
