Genomic prediction across two breeding cycles in maize silage using a factorial approach
Résumé
Genomic selection enables the prediction of all possible single-crosses between candidate lines by using a training set composed of genotyped and phenotyped individuals to calibrate an equation of prediction. The design used as training set, its size, composition and relationship with the set of validation can affect prediction accuracies. Previous simulation and experimental studies have shown the potential of using a sparse factorial design instead of tester designs as training set. This work aimed at: (i) evaluating the efficiency of a factorial training design in the context of prediction across breeding cycles and (ii) investigating optimization strategies to construct the training set. The study relies on two breeding cycles issued from a multiparental connected reciprocal population generated from the flint and dent complementary groups. In each group, about 800 inbred lines were generated and evaluated for their hybrid value in different factorial designs for their silage performance. The best 30 lines from each group were selected based on genomic predictions and intercrossed to produce the next generation evaluated in a factorial design. We showed that using the first cycle to predict the next one gave good predictive abilities, reaching 0.6 for dry matter yield. By adding hybrids from the new cycle to the training set, we increased predictive ability by 0.1 for dry matter yield. This indicates the benefit of recalibrating across breeding cycles. To optimize the training set and determine the interest of recalibrating model by phenotyping part of the G1 hybrids, a criterion maximizing the mean of the expected reliabilities (CDmean) was used. Samples chosen based on CDmean gave higher predictive abilities than random samples for various calibration set sizes. Our results confirm the potential of sparse factorial designs for revisiting genomic hybrid breeding schemes.