Dealing with the genotype × environment interaction in genomic prediction

SANTIAGO ALVAREZ PRADO; EMILIE J. MILLET; CLAUDE WELCKER; WILLEM KRUIJER; LAURENCE MOREAU; ALAIN CHARCOSSET; FRED VAN EEUWIJK; FRANÇOIS TARDIEU

Paris

Seminario; Séminaire SelGen 2017: La selection génomique ? bilan et perspectives.; 2017

INRA

Yield displaysa large genotype by environment interaction (G×E), especially under stressfulconditions. Even in environments classified as ?dry?, the ranking of genotypesand the allelic effects of QTLs largely differ between specific scenariosinvolving soil water status, evaporative demand and temperature (Millet et al2016). Hence, selecting genotypes for specific environmental scenarios representsa viable strategy to improve crop production in the context of climate changeand increasing occurrence of water deficits. Genomic prediction methods do notexplicitly take the G×E into account, therefore limiting the success of predictionsto highly heritable traits. Models (either statistic or crop models) may helpin solving this problem. They involve genotype-specific parameters (with noGxE) that allow calculating yield in a range of conditions based on the vectorof genotypic parameters and on local environmental conditions. A potentialapproach could therefore consist in carrying out the genomic prediction on thegenotypic parameters rather than on yield. This requires deriving the modelparameters from a training set of field experiments and genotypes, and/ormeasuring them in a phenotyping platform on the training set of genotypes. We are performinga proof of concept of this approach, based on two models, first the APSIM crop model,second a regression model in which yield components depend on environmentalvariables (Millet et al 2016). (i) A series of experiments was performed in thePhenoArch platform with contrasting soil water status and evaporative demandand in 15 field experiments across Europe. We used a training population of 200maize hybrids and 46 extra hybrids, which were genotyped with 50K polymorphicSNPs. Parameters used in the APSIM model have been measured in the platform,for example phyllochron and the response of stomatal conductance to evaporativedemand and soil water deficit. The phyllochron could be predicted based on 50Kmarkers (r2 = 0.46 in the test set of genotypes). Stomatalconductance and its response to environmental conditions was predicted based ona set of QTLs (genomic prediction under way), with a good prediction ofgenotypic + environmental variabilities (r2 = 0.86). The next stepwill be to run the APSIM model in the experiments of the test set, so we cancompare predicted and observed yield. The model is now set for carrying these calculationsin all experimental sites. (ii) The second proof of concept consists of predictingthe parameters of the regression model via genomic prediction based on thetraining set, and testing the yield prediction on the test set. Parameters havebeen calculated from data originating from both the platform and the trainingset of fields. We expect better prediction of yield via the regression modelfor the range of environmental conditions in which the regression model wastrained (Europe at mid-latitudes), but a wider applicability by using the APSIMmodel.