Introducing sparsity into selection index methodology with applications to high-throughput phenotyping and genomic prediction

Research in plant and animal breeding has been largely focused on the development of methods for a more efficient selection by altering the factors that affect genetic progress: selection intensity, selection accuracy, genetic variance, and length of the breeding cycle. Most of the breeding efforts have been primarily towards increasing selection accuracy and reducing the breeding cycle.Genomic selection has been successfully adopted by many public and private breeding organizations. Over years, these institutions have developed and accumulated large volumes of genomic data linked to phenotypes from multiple populations and multiple generations. This data abundance offers the opportunity to revolutionize genetic research. However, these data sets are also increasingly heterogeneous, with many subpopulations and multiple generations represented in the data. This translates into potentially heterogeneous allele frequencies and different LD patterns, thus leading to SNP-effect heterogeneity.Genomic selection methods were developed with reference to homogeneous populations in which SNP-effects are assumed constant across the whole population. These methods are not necessarily optimal for the contemporary available data sets for model training. Therefore, a first focus of this dissertation is on developing novel methods that can leverage the large-scale of modern data sets while coping with the heterogeneity and complexity of this type of data.In recent years, there have also been important advances in high-throughput phenotyping (HTP) technologies that can generate large volumes of data at multiple time-points of a crop. Examples of this include hyper-spectral imaging technologies that can capture the reflectance of electromagnetic power by crops at potentially thousands of wavelengths. The integration of HTP in genetic evaluations represents a great opportunity to further advance plant breeding; however, the high-dimensional nature of HTP data poses important challenges. Therefore, a second focus of this dissertation is on the development of a novel approach to efficiently incorporate HTP data for breeding values prediction.Thus, this dissertation aims to contribute novel methods that can improve the accuracy of genomic prediction by optimizing the use of large, potentially heterogeneous, genomic data sets and by enabling the integration of HTP data. We present a novel statistical approach that combines the standard selection index methodology with variable-selection methods commonly used in machine learning and statistics, and developed software to implement the method. Our approach offers solutions to both genomic selection with potentially highly heterogeneous genomic data sets, and the integration of HTP in genetic evaluations.