Date of this Version
THE CROP JOURNAL 8 (2020) 843–854
The characterization of genomes with great detail offered by the modern genotyping platforms have opened a venue for accurately predicting the genotype-by-environment interaction (GE) effects of untested genotypes in different environmental conditions. Already developed statistical models have shown the advantages of including the GE interaction component in the prediction context using molecular markers, pedigree, or both. In order to leverage the family information of highly structured populations when pedigree data is not available, we developed a model that uses the family membership instead. The proposed model extends the reaction norm model by including the interaction between families and environments (FE). A representative fraction of a soybean Nested Association Mapping population (16,187 grain yield records) comprising 38 bi-parental families (1358 genotypes) observed in 18 environments (2011, 2012, and 2013) was used to contrast the proposed model with three conventional prediction models. Two cross- validation scenarios (prediction of tested [CV2] and untested [CV1] genotypes) with a twofold design (50% for training and testing sets) were used for mimicking prediction situations that breeders face in fields. Results showed that the family factor in interaction with environments explains a sizable amount of the phenotypic variability. This helped to improve the predictive ability with respect to the main effects model (GBLUP) around 41% (CV2) and 49% (CV1), and about 17% with respect to the conventional reaction norm model. The inclusion of the FE term not only improved the global results but also significantly increased the prediction accuracy of those environments where the conventional models showed a very poor performance. These results show the importance of taking into consideration the family structure existing in breeding programs for improving the selection strategies in multi-parental populations.