Genomewide prediction of genotypic values and genetic variances within 969 maize biparental populations

Abstract

In plant breeding, selecting within biparental crosses and selecting parents to make new crosses are both important. My first study investigated the accuracy of genomewide selection (<italic>r</italic>_MG) within 969 biparental maize populations (<italic>Zea mays</italic> L.). My objectives were to determine: (i) the mean and variability of <italic>r</italic>_MG, (ii) if <italic>r</italic>_MG can be predicted, and (iii) how training population size (<italic>N</italic>), heritability (<italic>h</italic><super>2</super>), and number of markers (<italic>N</italic>_M) affect <italic>r</italic>_MG. I modified an equation for expected <italic>r</italic>_MG [E(<italic>r</italic>_MG)] to account for linkage disequilibrium (<italic>r</italic><super>2</super>) between markers and quantitative trait loci. Across the 969 populations, the mean and range (in parentheses) of observed <italic>r</italic>_MG was 0.45 (&minus0.59, 1.03) for yield, 0.59 (&minus0.34, 0.96) for moisture, and 0.55 (&minus0.24, 1.10) for test weight. The observed <italic>r</italic>_MGvalues were centered around E(<italic>r</italic>_MG) when <italic>r</italic><super>2</super> was accounted for, but had a large spread around E(<italic>r</italic>_MG). The <italic>r<super>2</super>(Nh<super>2</super>)<super>&half</super></italic> had the strongest association with the observed <italic>r</italic>_MG. In the second study, my objective was to determine whether related populations could be used to predict the genetic variance (V_G) of a segregating population from two parents (A and B). For each of 85 A/B populations, 2&ndash23 A/* and B/* populations were used as training populations, where * denotes a random parent. In the genomewide selection model, the testcross V_G in A/B was predicted as the variance among the predicted genotypic values of progeny from the simulated A/B population. In the mean variance model, V_G was estimated as the mean of V_G in A/* and B/* populations. The correlations between observed and predicted V_G were not significant (P = 0.05) for the genomewide selection model but were significant for the mean variance model (0.26 for yield, 0.46 for moisture, and 0.50 for test weight). The V_G of A/B population could therefore be predicted as the mean of V_G in A/* and B/* populations. Overall, the results indicated that genomewide selection can identify the best individuals within a cross, but it cannot reliably predict which parents would lead to the largest genetic variance.