Improving genomic prediction models

Animal breeding aims to increase the performance of livestock populations continuously. With the development of genotyping technology, genomic prediction (GP) has become a routine process in modern animal breeding schemes. Genomic prediction uses genotype information to prediction genomic breeding values and has largely shortened the generation interval and increased genomic prediction accuracy. Conventional genomic prediction methods like GBLUP and Bayesian alphabet methods assume linear relationships between genotypes and phenotypes. Therefore, these models cannot cope with non-linear relationships between genotypes and phenotypes, like epistasis and dominance, which are known to be important for complex traits. To overcome this limitation, non-linear machine learning (ML) models, like kernel methods and artificial neural networks, have been proposed to predict phenotypes. Although ML looks promising, it is still unknown which ML method has the highest prediction accuracy and how this depends on the heritability of the trait and training population size, how to include prior information about causal genes in ML models and whether this can improve prediction accuracy, and how well ML models would perform when used for selection across multiple generations. Therefore, the objectives of this research are 1) to explore which (class of) ML algorithm(s) has the highest genomic prediction accuracy of phenotypes with different combinations of training population sizes and heritabilities; 2) to investigate the benefit of using preselected genetic variants in ML models; and 3) to investigate the longterm effects of selection based on predicted phenotypes using ML. In order to investigate these objectives, analyses will be based on simulated data.