Training population optimization for genomic selection improves the predictive ability of a costly measure in bread wheat, the gliadin to glutenin ratio

International audienceEnd-use value of wheat four depends strongly on the concentration and composition of storage proteins, namely the gliadins and glutenins. As protein concentration in wheat grain is negatively correlated with grain yield, monitoring the gliadin to glutenin ratio is a mean to maintain end-use quality in modern varieties. However, the measurement of this ratio is expensive and time consuming. As genomic selection (GS) has proved very successful for traits controlled by many Quantitative Trait Loci and is already used for breeding, we decided to apply it to the gliadin to glutenin ratio. Therefore, we phenotyped for this trait and genotyped with a 420,000 SNP (Single Nucleotide Polymorphism) array a set of 88 modern varieties and 325 core-collection varieties. A GS model taking into account the genotypic, environmental and genotype x environment interaction efects was tested. Its predictive ability depends on the composition of the training population (TP). Adding signifcant SNPs as fxed efects did not improve the predictive ability. However, we observed improvements by optimizing the TP with fve methods based on relatedness between genotypes and obtained a maximum predictive ability of 0.62 and a minimum Root Mean Square Error of 0.056 for the gliadin to glutenin ratio. To conclude, our results are promising and strongly suggested that GS can be efciently applied to the gliadin to glutenin ratio. In addition, genotypes phenotyped and genotyped in previous breeding generations could be useful to train the model