Challenges in phenotypic data analysis and genomic selection application for maize hybrid prediction in tropical regions

Abstract

The process to obtain maize single-cross hybrids (SH) is very dynamic. In the initial stages, numerous genotypic combinations are produced. As the process progresses, many SH are discarded, and the number of repetitions and sites that these SH are evaluated increases. Thus, it is common that data from maize breeding programs are very unbalanced due to the low coincidence of SH that are evaluated over different sites and years. In the recent years, genomic selection has been proposed as a tool to accelerate the genetic gains and to reduce the cost with the selection and the recommendation of maize SH. However, a pending question is if the data obtained from these highly unbalanced experiments, can contribute to increase the accuracy of predictive models. Therefore, the purpose of the present study, in a first moment, was to critically analyze the grain yield data of 2770 maize hybrids, from experiments conducted in different years and sowing seasons; and to verify the impacts of unbalanced designs to the estimates of genetic and phenotypic parameters. Additionally, in a second step, using the same data set, the predictive capacity of the GBLUP (Genomic Best Linear Unbiased Prediction) model was compared considering additive and dominance effects. For this, the 447 parental lines were genotyped using 23,153 Darts markers. The results show the complexities of analyzing information from all crop seasons under conditions of high experimental unbalance and significant effect of genotype by environment interaction. These factors compromise the estimates of variance components, heritability, genetic values of individuals and, consequently, may affect the predictive accuracy of genomic selection models. Nonetheless, the analysis involving genomic information showed that it is possible to obtain genetic gains with the prediction of SH not evaluated and that the inclusion of dominance effects in the GBLUP model can improve its predictive ability.