Validation of cross progeny variance genomic prediction using simulations and experimental data in winter elite bread wheat.
Abstract
Genomic prediction is used in many crop breeding programs to select parental lines to ensure high performance of their progeny. Taking into account progeny variance estimates on top of parental mean is expected to increase the probability to get outstanding progenies. Several Cross Selection Criteria have been proposed and the Usefulness Criterion (UC) that accounts for Parental Mean (PM) and progeny Standard Deviation (SD) has been shown as a good compromise to secure genetic gain as well as genetic diversity in the next generation using simulation studies. In this study, we predicted the three cross value components (PM, SD and UC) of 73 winter bread wheat crosses whom progenies have been evaluated in the field. The Training Population (TP) used to estimate marker effects was composed of 2,146 French varieties registered between 2000 and 2021 and INRAE-AO breeding lines. We first evaluated different factors influencing the prediction ability of the cross value components based on simulations, starting from the same crosses as the experimental design, simulating phenotypes with increasing heritability, number of QTLs and progeny size. As expected, increasing the number of QTL decreased the prediction ability for all cross value components, and increasing heritability or increasing progeny size improved prediction abili ties. The prediction of SD was the most impacted. We used as a reference a TRUE scenario, i.e. an optimal situation where TP is optimal and where marker effects are perfectly estimat ed. Once again, SD was strongly impacted by the quality of marker effect estimates. For poly genic traits (10 QTL), Bayesian models showed higher prediction ability. For quantitative traits (more than 300 QTL) with low heritability, using a progeny variance estimate that takes into account the error of marker effect estimates improved significantly SD prediction ability. We validated our findings using experimental data for four traits evaluated on the same crosses: yield, grain protein content, plant height and heading date. Prediction abilities were assessed for each cross value component, and overall, predictions aligned well with experi mental values. PM and UC were reasonably predicted for most traits, while SD was more challenging, especially for yield. To our knowledge, this study is the first to experimentally validate the genomic prediction of progeny cross variance and showed that prediction abili ties strongly depend on trait architecture. This study also revealed that it is essential to gener ate a very large number of progenies per cross to obtain reasonable prediction abilities of SD.