“…Several studies (Clayton, 2009;Cleynen et al, 2016;Hamshere et al, 2011;Pashayan et al, 2015;Pharoah, Antoniou, Easton, & Ponder, 2008;Sawcer, Ban, Wason, & Dudbridge, 2010) have shown that use of a limited number of top ranking SNPs can help discriminate diseased cases from unaffected controls, or between different disease sub-phenotypes, but that the utility for individual risk prediction generally falls far short of clinically useful levels. However, even when the combined set of SNPs explain a large proportion of variance, much larger sample sizes are required to achieve high prediction accuracy (Yang et al, 2017) because the individual SNP effects are substantially smaller than the total variance they explain. For example Dudbridge (2013) showed that the disappointing AUCs reported by Machiela et al (2011) were entirely consistent with the theoretical AUC values of 52-54% predicted on the basis of their discovery set sample size, but that these values could be increased to ≈ 80 − 90% if the samples were infinitely large.…”