Title: Multiple environment genomic prediction on cotton yield and fibre quality traits using statistical predictive methods

Abstract: Genomic prediction (GP) is an emerging data driven technology for plant breeding, which provides predictive phenotypes or breeding values of crops based on both genomic and environmental information. Capturing genotype x environment interactions (G x E) is essential to reflect the complexity of breeding target environment and ensure the relevance, robustness and impact of predictive tools, from the statistical computation perspective, however, adding the interactions between environmental covariates and genome-wide markers may significantly increase the model dimension, and create a computational burden. We developed a dimensional reduction method based on the collinearity among the DNA markers and among the environmental covariates, which can represent the large amount of G x E using a small number of summary statistics. Then we proposed to use Bayesian slab and spike regression, as well as residual neural network as the genomic predictive models. The predictive performance of these methods was evaluated on a large-scale cotton breeding trial data sets collected from the CSIRO cotton breeding program, consisting of more than 3 000 genotypes, 50 environmental covariates and over 10 000 phenotype records. We showed promising prediction accuracies of the models for cotton breeding target economical traits ranging from lint yield and fibre qualities.