A0742
Title: Statistical and ML options for prediction of GxE in plant breeding
Authors: Fred van Eeuwijk - Wageningen University (Netherlands) [presenting]
Abstract: In plant breeding and genetics, a central research objective is the prediction of phenotypic performance from genetic and environmental inputs. In its simplest form, the question is how to predict single or multiple phenotypes (responses) from genetic and environmental factors and covariates. The default model class for such predictions was that of linear mixed models. Over the last decades, genetic and environmental inputs have become high-dimensional, and additional classes of inputs, like phenomics, are accessible via new types of sensors and measurement devices. Environmental and phenomic inputs are often longitudinal. The mixed model framework for prediction of phenotypes has been extended to incorporate multiple sets of high-dimensional inputs via multiple kernels. The longitudinal aspect of environmental and phenomic data can be addressed by the insertion of various types of base functions or special types of variance-covariance matrices. Alternative methods for analysis and prediction of phenotypes try to exploit the high resolution of some of the environmental and phenomic inputs. Hierarchical Bayesian approaches that model longitudinal phenotypes by systems of differential equations are found. Machine learning and deep learning approaches have been proposed, too. The prediction of phenotypes is considered via different modeling classes with special attention, for the way in which genotype by environment interactions are addressed.
