Prediction models were compared across four different regression methods; elastic net (prediXcan), ridge regression (using the TWAS method16), Bayesian sparse linear mixed modelling (BSLMM; TWAS), and linear regression using the best eQTL for each gene (Supplementary Figure 1a). Mean Rcv2 values were significantly higher for elastic net regression (mean Rcv2 = 0.056) than for eQTL-based prediction (mean Rcv2 = 0.025), BSLMM (mean Rcv2 = 0.021) or Ridge Regression (mean Rcv2 = 0.020). The distribution of Rcv2 values was also significantly higher for elastic net regression than for any other method (Kolgorov-Smirnov test, p < 2.2 × 10−16).
