Testing the accuracy of polygenic trait predictions.

A. Polygenic scores were computed based on the phenotypic measurements for two replicates, and correlated with the phenotype observed for the third replicate. Correlation was tested with a Spearman rank correlation test Rho..Nr_SNPs: number of SNPs associated with each trait at p<10–4. B. SNPs associated with the phenotype at sub-significant level improve significantly the phenotype prediction but random SNPs show that population structure plays an important role. Rho_associated shows the correlation between polygenic score and the genotypic values. Based on 1000 random samples of an equal number of SNPs, a distribution of random Zscores was computed and compared to the spearman correlation of the prediction of associated variants to the input phenotypes (Rho_associated). The distributions of spearman correlations of the 1000 random sets is described with the median (Rho_random_median), 95th quantile (Rho_random_95quantile) and the maximal Rho (Rho_random_max). The correlation obtained with random SNP set is also often significant at p<0.05 (Percentage_significant), but the maximum correlation coefficient (Rho_random_max) is always markedly lower than the one obtained with sub-significant SNPs (Rho_associated).
(XLSX)