Efficient Use of Historical Data for Genomic Selection: A Case Study in Wheat

Genomic selection (GS) is a breeding technology that can accelerate rates of genetic gain. With GS, a genotyped and phenotyped population is used to train a statistical model in order to predict a set of selection candidates. The training population (TP) size, heritability, and relationship to the selection candidates are key factors affecting accuracy. This study uses empirical data from a recurrent genomic selection program for stem rust (Puccinia graminis f.sp. tritici) adult plant resistance in wheat (Triticum aestivum) to 1) determine the relative accuracies of a historical and ‘population specific’ training set consisting of individuals from the same selection candidate population, 2) evaluate optimization methods based on the coefficient of determination (CD), prediction error variance (PEV) and Fst, 3) determine if historical data of high heritability could augment a low heritability population specific training set. We found that the historical training set was 42% less accurate than a population specific TP. Optimization of the historical TP based on CD and PEV improved accuracy by up to 28%. Adding individuals from a high heritability historical TP to a low heritability population specific training set did not significantly increase accuracy. These results indicate that using historical data for GS model training can be useful if a population specific TP is unavailable. Accuracy from a historical TP can be improved by subsampling from the population to improve the relevance to the selection candidates. Further work to develop TP subsampling methods is warranted.