Genomewide prediction accuracy within 969 maize biparental populations

In genomewide selection, the expected correlation between predicted and true genotypic values (r_MG) has been previously derived as a function of the training population size (N), heritability (h²), and effective number of chromosome segments (Me) affecting the trait. Our objectives were to determine: (i) the mean and variability of r_MG in 969 biparental maize (Zea mays L.) breeding populations for seven traits, (ii) if r_MG can be predicted in advance, and (iii) how N, h², and number of markers (N_M) affect r_MG. We modified a previous equation for expected r_MG to account for linkage disequilibrium (r²) between a marker and a quantitative trait locus (QTL). Across the 969 populations, the mean and range (in parentheses) of observed r_MG was 0.45 (-0.59, 1.03) for grain yield, 0.59 (-0.34, 0.96) for moisture, 0.55 (-0.24, 1.10) for test weight, 0.49 (-0.22, 1.04) for stalk lodging, 0.41 (-0.30, 0.93) for root lodging, 0.47 (-0.45, 0.97) for plant height, and 0.42 (-0.43, 0.94) for ear height. The observed r_MG values were centered around the expected r_MG when r² was accounted for, but the observed rMG had a large spread around the expected r_MG. The r²(Nh²)^1/2 had the strongest association with observed r_MG. When r²(Nh²)^1/2 exceeded 8, the proportion of r_MG equal to or larger than 0.50 reached 90% among all the population- trait combinations. We conclude it is difficult to predict r_MG in advance, but that rules of thumb based on r²(Nh²)^1/2 can help achieve a high r_MG.