We built prediction models in the DGN whole blood cohort using LASSO, the elastic net (α=0.5), and polygenic score at several p-value thresholds (single top SNP, 1×10−4, 0.001, 0.01, 0.05, 0.5, 1). We assessed predictive performance using 10-fold cross-validation (R2 of estimated GReX vs. observed expression) as well as in an independent set. We found that LASSO performed similarly to the elastic net and that LASSO outperformed the polygenic score at all thresholds, although all methods are highly correlated (see Supplemental Figure 1). For subsequent analyses, we focused on the prediction models using the elastic net because we found it to perform well and to be more robust to slight changes in input SNPs (potentially due to variations in imputation quality between cohorts).
