Genomic Prediction Accuracy of Single-Family Versus Multiple-Family Training Populations in the Soybean Nested Association Mapping Population.

Genomic selection for plant breeding aims to increase genetic gain by increasing selection intensity and reducing the time of breeding cycles. Our objectives were to evaluate the effect of training population composition and marker imputation on predictive ability (PA). Genotypic and phenotypic data from 40 bi-parental crosses comprising the Soybean Nested Association Mapping (NAM) population were used to evaluate PA in two related studies.

In the first study, three types of training sets were used for genomic predictions of each of the 40 families: (i) within-family prediction, where the training set included data only from the family being predicted; (ii) across-family prediction, where the training set did not include data from the family being predicted; and (iii) combined-family prediction, where the training set consisted of data from all families including the family being predicted. For the second study, the families were divided in three groups as follows: g1, eight families with an exotic parent; g2, 15 families with a parent of diverse ancestry; and g3, 17 families with a high yielding, elite parent. Here, the aim was to perform predictions of each family of g3 using three combinations of training sets: (a) accession in g1, g2 and g3; (b) g2 and g3; and (c) only g3.  

Taking as baseline the results derived from (i) scheme (within-family prediction), for the first part, we have improvements in PA of 10-46% and 42-85% with (ii) and (iii) schemes for grain yield and days to maturity, respectively.  Regarding the analysis of different types of families (part 2), little difference was observed and direction of change in PA was highly variable between families.