Abstract
Population breeding through recurrent selection has recently regained attention in the plant breeding community with the new possibility to reduce the breeding cycle length by selecting on genomic prediction (GP) rather than progeny testing. For several decades, the CIAT-Cirad rainfed rice (Oryza sativa L.) breeding program has been using a two-parts breeding scheme with population improvement based on recurrent selection and a cultivar development following pedigree breeding. More recently, effort have been made to implement GP to shorten the breeding cycle.The objective of this thesis was to evaluate the potential of GP in CIAT-Cirad program, test calibration strategies using the existing infrastructure to later implement early genomic selection of recombinant parents.A population was genotyped at generation S0 before being divided in two subpopulations: the PCT27A and the PCT27B. PCT27A was advance to generation S0:2 and S0:3 and phenotyped at those generations while PCT27B was advanced up to S0:4 and phenotyped. Four traits were measured in a target site and a surrogate evaluation site. The predictive ability of the GP models was estimated using several scenarios and models, according to the presence of one or two growing environments, one or several phenotyping generations, the presence of genetic by environment interaction and the size and composition of the training set.First, we assessed by cross-validation the GP in a single population. Then, we used the same data to predict more advanced material in an external validation population. To complete the field experiments, a simulation study was realized to assess the long-term effect of the integration of GP into the breeding program and the response of calibration scenario to dominance and genotype-environment interaction (GxE) variance.In this last study, the effect of three levels of GxE and two levels of dominance were assessed on two breeding schemes based on either two generations of phenotyping or a single generation.Two-sites calibrations never strongly outperformed single site calibrations. Early generation PCT27A phenotypes from target site could be used to predict generation S0:4 from PCT27B. However, when the calibration confounded generation and site effects, the precisions were lower. Hence it seems so far unappropriated to phenotype different generations in different sites. The simulations allowed to highlight that forward prediction, which is the base of recurrent selection, is possible with either breeding schemes, calibration with two generations being systematically better than single generation one. With the increase of GxE, the accuracies dropped for both schemes with similar intensity. Only the level of GxE had an impact on predictive ability and genetic gain.To conclude, GP can replace progeny testing in recurrent selection. The utility of the different sites and different generations in the calibration depend on the traits predicted and compromises will have to be done when the breeding scheme will be design to reach the best possible accuracy for each trait, while staying within the program financial and logistical limitations. Those results from the simulation and field experiments will be valuable for the improvement of the CIAT-Cirad breeding program toward a faster genetic progress and more sound use of resources.