High performance computing for large-scale genomic prediction

In the past decades genetics was studied intensively leading to the knowledge that DNA is the molecule behind genetic inheritance and starting from the new millennium next-generation sequencing methods made it possible to sample this DNA with an ever decreasing cost. Animal and plant breeders have always made use of genetic information to predict agronomic performance of new breeds. While this genetic information previously was gathered from the pedigree of the population under study, genomic information of the DNA makes it possible to also deduce correlations between individuals that do not share any known ancestors leading to so-called genomic prediction of agronomic performance. Nowadays, the number of informative samples that can be taken from a genome ranges from one thousand to one million. Using all this information in a breeding context where agronomic performance is predicted and optimized for different environmental conditions is not a straightforward task. Moreover, the number of individuals for which this information is available keeps on growing and thus sophisticated computational methods are required for analyzing these large scale genomic data sets. This thesis introduces some concepts of high performance computing in a genomic prediction context and shows that analyzing phenotypic records of large numbers of genotyped individuals leads to a better prediction accuracy of the agronomic performance in different environments. Finally, it is even shown that the parts of the DNA that influence the agronomic performance under certain environmental conditions can be pinpointed, and this knowledge can thus be used by breeders to select individuals that thrive better in the targeted environment