Univariate and multivariate genomic prediction for agronomic traits in durum wheat under two field conditions

Genomic prediction (GP) has been evaluated in durum wheat breeding programs for several years, but prediction accuracy (PA) remains insufficient for some traits. Recently, multivariate (MV) analysis has gained much attention due to its potential to significantly improve PA. In this study, PA was evaluated for several agronomic traits using a univariate (UV) model in durum wheat, subsequently, different multivariate genomic prediction models were performed to attempt to increase PA. The panel was phenotyped for 10 agronomic traits over two consecutive crop seasons and under two different field conditions: high nitrogen and well-watered (HNW), and low nitrogen and rainfed (LNR). Multivariate GP was implemented using two cross-validation (CV) schemes: MV-CV1, testing the model for each target trait using only the markers, and MV-CV2, testing the model for each target trait using additional phenotypic information. These two MV-CVs were applied in two different analyses: modelling the same trait under both HNW and LNR conditions, and modelling grain yield together with the five most genetically correlated traits. PA for all traits in HNW was higher than LNR for the same trait, except for the trait yellow index. Among all traits, PA ranged from 0.34 (NDVI in LNR) to 0.74 (test weight in HNW). In modelling the same traits in both HNW and LNR, MV-CV1 produced improvements in PA up to 12.45% (NDVI in LNR) compared to the univariate model. By contrast, MV-CV2 increased PA up to 56.72% (thousand kernel weight in LNR). The MV-CV1 scheme did not improve PA for grain yield when it was modelled with the five most genetically correlated traits, whereas MV-CV2 significantly improved PA by up to similar to 18%. This study demonstrated that increases in prediction accuracy for agronomic traits can be achieved by modelling the same traits in two different field conditions using MV-CV2. In addition, the effectiveness of MV-CV2 was established when grain yield was modelled with additional correlated traits.