Validating genomewide predictions of genetic variance in a contemporary breeding program

Predicting the genetic variance among progeny from a cross-prior to making said cross-would be a valuable metric for plant breeders to discriminate among possible parent combinations. The use of genomewide markers and simulated populations is one proposed method for making such predictions. Our objective was to assess the predictive ability of this method for three relevant quantitative traits within a breeding program regularly using genomewide selection. Using a training population of two-row barley (Hordeum vulgare L.) lines, we predicted the mean (µ), genetic variance (V_G), and superior progeny mean (µ_SP, mean of the best 10% of lines) of 330,078 possible parent combinations for Fusarium head blight (FHB) severity, heading date, and plant height. Twenty-seven of these combinations were chosen to develop biparental populations, which were subsequently phenotyped for the same traits. We found that the predictive abilities (r_MP) for µ and µ_SP were moderate to high (r_MP = 0.46-0.69), whereas those for V_G were lower (r_MP = 0.01-0.48). Unsurprisingly, predictive ability was likely a function of trait heritability, as r_MP estimates for heading date (the most heritable trait) were highest, and r_MP estimates for FHB severity (the least heritable trait) were lowest. We observed strong negative bias when predicting V_G (on average -83 to -96%), but the relative consistency of this bias across validation families indicates that it may have little impact when selecting crosses. We concluded that accurate predictions of V_G and µ_SP are feasible, but as with any implementation of genomewide selection, reliable phenotypic data are critical.