Abstract only

Genomic selection (GS) is breeding tool used to accelerate rates of genetic gain for complex traits. With GS, selections are based on genomic predictions which are generated when a prediction model is applied to a set of genotyped selection candidates. This enables selections to occur before phenotypic evaluation. We are currently applying GS to breed for complex, slow-rusting adult plant resistance (APR) to stem rust Puccinia graminis f.sp. tritici using historical data generated 2005-2011 as the initial model training population. Using data from this selection program we have evaluated 1) the genomic selection accuracies using training sets consisting of either historical data or with close (approximately first and second degree) relative data 2) the utility of training population optimization, and 3) the relative utility of model training with historical or close relatives assuming various relative heritabilities. Genomic selection accuracies were 42% and 85% that of the phenotypic selection accuracies when using historical and close relative data respectively. Accuracies from training with historical data significantly improved when an optimum subset was used for model training. Lastly, when heritibilies were low, 0.2, for the close relative set, and medium to high , 0.4+, for the historical set the resulting GS accuracies were not significantly different. These results indicate that in the case of stem rust APR, although model training with close relatives is generally the most accurate approach, use of an optimum subset of historical data can also be effective and can be just as accurate when heritability in the historical dataset is much higher than that of the close relative dataset.

Global Wheat Program

English

Lucia Segura

INT0610

CIMMYT Staff Publications Collection