Thursday, October 27, 2016
Will high throughput phenotyping and genotyping techniques help us to better predict GxE interactions? Some considerations from statistics and crop growth modelling
Fred A. van Eeuwijk & Daniela V. Bustos-Korts
Biometris, Wageningen University and Research Centre, the Netherlands
Predictions of phenotypic traits for diverse sets of genotypes across broad ranges of environmental conditions are at the basis of attempts to maximize selection responses in plant breeding programmes. For the last decade, multi-environment trials (METs) provided the information for building such prediction models. These prediction models were mostly from the class of linear mixed models (LMMs). LMMs are a flexible class of models with facilities for modelling genetic and environmental correlations between traits and environments and allow for heterogeneity of genetic and environmental variances. In addition, at the genotypic side of the prediction models, LMMs provide the possibility to include marker and sequence information as genotypic covariates to improve phenotypic prediction, which is then called genomic prediction. For the environmental side, LMMs can be extended by including environmental characterizations as environmental covariates to improve prediction. Recently, the use of phenotyping platforms has led to an additional source of information that may be useful to improve prediction. In the context of LMMs, this information enters the prediction model as additional genotypic covariates. Models for phenotypic prediction need to address genotype by environment interactions (GxE). The statistical objective is to find functions of genotypic and environmental covariates that can predict GxE. Statistical criteria for choosing appropriate genotype-to-phenotype functions (linear, non-linear, parametric, non-parametric, univariate, multivariate, networks, graphical models) and selecting genotypic (markers, sequences, platform characterizations, resistances, tolerances) and environmental (environmental characterizations, sensors, crop growth models, stress indices) covariables may be insufficiently clear to guide the model building process and may not lead to accurate phenotypic predictions. As a complement to purely statistical approaches to phenotypic prediction, crop growth models have been proposed. On the positive side, crop growth models provide a causal ordering and identification of physiological parameters, component traits and target traits (yield, resistance, tolerance, quality) that can help streamline the process of phenotypic prediction, whereas statistical models may at best select and estimate approximate trait configurations and orderings from the data itself. Crop growth models also make explicit the environmental information that is required for phenotypic prediction. On the negative side, crop growth models often contain parameters and input variables that are hard to obtain and measure in practice on diverse sets of genotypes. We have been studying various ways of hybridizing statistical prediction models with crop growth models. A straightforward hybrid method of prediction inserts genomic predictions for component traits in crop growth models for yield and other target traits. More sophisticated hybridizations are possible and will be presented. Points of consideration are the design and analysis of individual phenotyping trials, both in field trials and on platforms, the choice of genotype-to-phenotype models, and the selection of genotypic and environmental covariates. Special attention will be given to the added value of high throughput genotyping, phenotyping and envirotyping information for prediction.
Contact : François Tardieu