To explore whether or not this is indeed the case, we simulated a series of disease models and conducted the same two-stage, genome-wide scoring delineated by MAF class (Fig. 3). Surprisingly, the strongest R2 signals in both populations are for simulated diseases arising entirely from rare and uncommon risk alleles, with modes overlapping the observed peak at the 0.3 ≤ MAF < 0.4. For AAs the observed R2 values fall slightly below those generated for the model based on 100 causal loci (with a maximum of 0.022 variance explained by any individual variant; goodness of fit R2 = 0.78, P = 0.046), whereas the best fitting model for EAs is for 1,000 causal loci (maximum variance explained of 0.0037; R2 = 0.49, P = 0.19). For disease models representing the other part of the frequency spectrum (i.e., common alleles), the fit to the observed results is poor for EAs (R2 = 0.07, P = 0.68 for 5,000 causal loci), with the genome-wide scores explaining substantially less of the variation in the disease phenotypes. For AAs the signals are more concordant;
