We evaluated three prediction schemes: i. cis-eQTL, the single most significantly associated SNP in the training set was used as the predictor; ii. the best linear predictor (BLUP)30, estimates the causal effect-sizes of all SNPs in the locus jointly using a single variance-component; iii. The Bayesian linear mixed model (BSLMM)31, which estimates the underlying effect-size distribution and then fits all SNPs in the locus jointly. For the BLUP and BSLMM, prediction was done over all post-QC SNPs using GEMMA31. We note that the BLUP/BSLMM both perform shrinkage of the SNP weights but not variable selection, so all SNPs are included in the predictor. Recent work in parallel to ours also evaluated expression imputation using polygenic risk scores, LASSO, and elastic net16.
