Model assisted dissection of the Genotype x Environment interaction

Conférence InvitéeEven when field or platform experiments are repeated in a large number of sites and years, they still represent a very small proportion of the most likely environmental scenarios in a given region. This considerably complicates the genetic analysis of tolerance to environmental cues, because the positive or negative value of a given trait or allele largely depend on these scenarios. Hence, we consider multi-year-multi-site network of experiments as the result of one single random sampling among many other possible, rather than a result applicable to a target population of environments. We argue that the role of phenotyping platforms in this context is to estimate reproducible characteristics of genotypes, which allow simulation of the performance of studied genotypes in a wide range of environmental scenarios. We are developing "dissection" approaches to identify heritable traits in large panels of genotypes, based on hidden variables such as sensitivity of growth to a given environmental condition, radiation use efficiency and its response to environmental conditions, or changes in intercepted light due to genetic differences in plant architecture. Models for 'dissection' are used to disentangle the complexity of phenotypes and identify the hidden variables that account for the genetic and environmental determinisms of raw traits. This in turn allows us to perform genetic analyses of hidden variables, and to compare the genomic regions that affect them. A surprising result is that commonality of QTLs between processes or organs may well be the rule rather than the exception. The second step of the approach is a simulation that predicts yield of studied genotypes, or the positive or negative effects of a given combination of alleles, in different climatic scenarios. Network of field experiments are still essential to test the approach and refine models and genetic analyses. It is noteworthy that the models for the first "dissection" step and for the second "simulation" step have not the same requirements, so a single model may not be suited for both steps