Genomic prediction models, selection tools and association studies for genotype by environment data

Plant breeding is complicated by the fact that genotypes respond differently to different environments, a phenomenon known as genotype by environment interaction (GEI). Despite its importance, however, many plant breeding programmes still use inefficient methods for handling GEI. This thesis develops a wide-array of methods that leverage GEI for efficient prediction, selection and discovery in plant breeding. The methods are demonstrated using a collaborating cotton breeding dataset from Bayer CropScience as well as publicly available and simulated datasets. Chapter 1 presents a brief overview of plant breeding design and analysis, with a focus on genomic prediction models, selection tools and association studies for genotype by environment data. Chapter 2 develops genomic prediction models that predict the response of different genotypes across different growing environments. The models are referred to as integrated factor analytic (IFA) models. The IFA models integrate known genotypic covariates derived from marker data and known environmental covariates derived from weather and soil data along with latent environmental covariates estimated directly from the phenotypic data. These models have great potential to improve predictive plant breeding in the presence of GEI. Chapter 3 develops selection tools that provide breeders with information to select and deploy well-adapted genotypes to their target environments. The tools provide measures of overall performance and stability, which summarise average genotype performance across environments and the variability in performance. A new directional stability measure is also introduced that partitions genotype stability into components that reflect favourable and unfavourable adaptation. These tools are becoming increasingly important with the presence of rapidly changing environments amidst climate change. Chapters 4 and 5 develop fast exact methods for conducting genome-wide association studies (GWAS). The methods produce all required test statistics from the fit of a single linear mixed model, instead of a very large number of models for all markers of interest. Fast methods are also introduced for GWAS using complex models for GEI. These methods have great potential to improve discovery in a wide-array of genetic studies, particularly with the advent of large-scale datasets and complex genotype by environment interactions. Chapter 6 develops a general framework for simulating GEI using the class of multiplicative models. The framework can be used to simulate realistic multi-environment trial (MET) datasets and model breeding programmes that better reflect the complexity of real-world settings. This framework provides a general basis for plant breeders and researchers to evaluate different breeding methods in the presence of GEI. Chapter 7 presents a discussion and concluding remarks, with a focus on placing the thesis in the wider agricultural community.