Increasing accuracy and reducing costs of genomic prediction by marker selection

Genotyping costs can be reduced without decreasing the genomic selection accuracy through methodologies of markers subsets assortment. Thus, we compared two strategies to obtain markers subsets. The former uses the primary and the latter the re-estimated markers effects. Moreover, we analyzed each subset via prediction accuracy, bias, and relative efficiency by main genotypic effect model (MGE) fitted, using genomic best linear unbiased predictor linear kernel (GB), and Gaussian nonlinear kernel (GK). All scenarios (subset of markers × kernels models) were applied to a public dataset of rice diversity panel (RICE) and two hybrids maize datasets (HEL and USP). The highest prediction accuracies were obtained by MGE-GB and MGE-GK for grain yield and plant height when we decrease the number of markers. Overall, marker subsets via re-estimated effects method showed a higher relative efficiency of genomic selection. Based on a high-density panel, we can conclude that it is possible to select the most informative markers in order to improve accuracy and build a low-cost SNP chip to implement genomic selection in breeding programs. In addition, we recommend REE (re-estimated effect) strategies to find markers subsets in training population, increasing accuracy of genomic selection.