Bread-making quality is a complex score that integrates several physical measures of dough and loaf. It is expensive to phenotype and requires several kilograms of grains that are only available in last generations of a breeding cycle. Genomic predictions could help improving genetic gain per unit of time and cost through prediction of performance before or without phenotypic characterization of new candidates. Estimated genetic values are model-based molecular scores, with the hypothesis that all markers have a small effect.

This study aims at optimizing genomic prediction of bread-making scores used in France to register varieties and classify them according to baking quality.

To meet this objective, we analyzed a reference population of 900 lines from the INRA-Agri-Obtentions breeding program that was evaluated for almost 30 traits between 2000 and 2017and that have been genotyped for 35 000 SNPs developed within the Breedwheat project (www.breadwheat.fr). Some of the lines have also been genotyped for 12 molecular markers developed in two genesGlu-1-1and Glu-1-2that are present on each of the three genomes, i.e. six genes in total. These genes encode the high-molecular-weight glutenin subunits (HMW-GS) which are important determinants of technological quality of wheat.

We investigated several strategies to improve those predictions, including single-trait genomic prediction models(ST), single-trait genomic prediction models with the 12 markers associated with glutenins (ST-glu), multi-trait genomic prediction models (MT), trait-assisted genomic prediction models (TA). In MT and TA models, we used information from a correlated trait, the alveograph parameter W (dough strength), which is less expensive to measure. Only individuals of the training set have been phenotyped for the trait of interest in both MT and TA models. In MT model, only the individuals of the training set have been phenotyped for the correlated trait W. In TA model, the training set and a proportion of the validation set have been phenotyped for W. The costs of these approaches have been evaluated and compared.

For a fixed budget, TA methods allow to improve the accuracy of bread score predictions compared to ST and MT methods. By contrast, MT methods do not improve the accuracy of bread score predictions compared to ST models. Although ST-glu model improves the accuracy of prediction of W, it does not significantly improve the accuracy of prediction of bread-making scores. Implications of these results for optimizing breeding schemes of bread wheat are discussed.