We use heritability estimates as our benchmark for the prediction R2 since this constitutes the upper limit of our prediction performance. For genes for which an elastic net model was available (n=10,427), the average heritability in DGN was 0.153. In comparison, the average 10-fold cross-validated prediction R2 for elastic net was quite close at 0.137; for the polygenic score (P<1×10−4) and top-SNP models, average prediction R2 values were sizably lower at 0.099 and 0.114, respectively. We show the performance R2 for each model in Figure 3, with the corresponding heritability estimate and confidence interval in the background for comparison. We also note that elastic net predictive performance reached or exceeded the lower bound of the heritability estimate for 94% of genes, while polygenic score (P<1×10−4) did so for just 76% of the genes and the top SNP for 80% of the genes (Figure 3), consistent with the performance ranking given by the average (across genes) R2.
