Enviromics, nonlinear kernels and optimized training sets for a climate-smart genomic prediction of yield plasticity in maize

• Autor(a) principal: Costa Neto, Germano Martins Ferreira
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/11/11137/tde-11102021-134352/
• topic: Adaptabilidade; Adaptability; Ciência de dados; Data science; Envirotyping; Genomic selection; Seleção genômica; Tipagem de ambientes

Large-scale envirotyping (environmental + typing) or simply enviromics, is an emerging field of data science, applied both in agronomic research and plant breeding. This \"omics\" consists of gathering and processing reliable environmental information, respecting the crop-specific ecophysiology aspects, then for further integration of this data into quantitative genetics and prediction-based breeding. However, most of the current prediction-based platforms are based on genotype-phenotype relationships (i.e., the phenotype-genotype association enabled by whole-genome markers), in which the state-of-art of this approach in the context of predictive breeding is so-called genomic selection or prediction (GP). Despite the success of its use in preliminary breeding stages, mostly conducted under restricted environmental variations (e.g., few number of environments or a single environment), the occurrence of low accuracy values are still a reality under multiple environmental conditions, in which is detected the presence of the so-called \"genotype by environment interaction\" (G×E). On the other hand, knowledge of crop ecophysiology can be the alternative to boost the accuracy of GP under G×E. This environmental variation shapes genotype-specific phenotypic responses to a given gradient of soil, climate and management factors i.e., the reaction norm. In this thesis, we conducted three studies aimed to investigate the use of GP enviromics under G×E scenarios, using for this the grain yield of two datasets of tropical maize hybrids. The first study of this thesis involves the development of the first open-source software dedicated to envirotyping in genomic prediction. In this study, we elucidate the use of remote sensing to popularize the use of envirotyping, as well as aspects of ecophysiology useful to understand and define the concepts of \'environment\', \'enviromics\' and \'envirotyping\'. In the second chapter, we verify the accuracy gains acquired by the adoption of non-linear kernels (Gaussian Kernel, GK; Deep Kernel, DK) for modeling non-additive effects (e.g., dominance and envirotyping-enabled reaction-norms) using the traditional GBLUP (genomic best linear unbiased predictor) as a reference method. Our results suggest that non-linear kernels (GK and DK) are the best alternative to model non-additive and reaction norm effects. The adoption of GK or DK reduced the computational time in running the models, as well as increased the accuracy to predict complex G×E interactions (variations in the rank of genotypes across environments). Finally, we observe that the use of GK or DK for modeling non-additive effects is critical to expand GP\'s resolve to predict the interaction of a particular maize hybrid across multiple environments. Finally, in the third chapter we propose the concept of \'envirotype marker\', developed by reconciling classical concepts of ecophysiology (Shelford\'s Law) and characterization of the environmental typology (i.e., frequency of occurrence of qualitative classes of environmental factors over time and over time. space). The approach was exemplified with two case studies covering the hypothetical use of GP under evaluation trials in maize hybrids in different environments. The combined use of enviromics and genomics made it possible to design a prediction platform (called E-GP) that reconciles selective phenotyping (reduction of training populations for GP) and prediction of future scenarios (i.e., unknown G×E). We observed that the increase in phenotypic information in various environments does not always correspond to the increase in the accuracy of GP. Therefore, the representativeness of each hybrid under evaluation at the experimental network (most representative genotypes, evaluated in \"key\" environments) is more important than the number of genotypes and environments considered for training GP. Through E-GP together with genetic algorithms, we were able to select the most representative G×E combinations, which directly reflected in a drastic reduction in the size of the experimental network, reconciling increased accuracy. Finally, we found that GBLUP without any envirotyping information is inefficient in predicting the phenotypic plasticity of maize hybrids under multiple environments and unknown G×E. With E-GP it was possible to screen the best hybrids, in terms of phenotypic plasticity, using reduced phenotypic information and supplemented by the wide use of genomics and enviromics. Such results allow us to envision smart approaches to climate, involving the drastic reduction of field-testing efforts as the conscious use of enviromics (and envirotyping) combined with genomics increases.