Genomic prediction using high-density and whole-genome sequence genotypes

Abstract

The main objective of this thesis was to investigate genomic prediction methods for high-density and whole-genome sequence genotypes, with emphasis on traits that may have difficulties achieving a high prediction accuracy with pedigree-based predictions, such as disease resistance and maternal traits. A Bayesian variable selection method that combines a polygenic term through a G-matrix and a BayesC term (BayesGC) was compared with Genomic Best Linear Unbiased Prediction (GBLUP), and for Paper I and II, it was also compared to BayesC.

Paper I aimed to investigate genomic prediction accuracy for the trait host resistance to salmon lice in Atlantic salmon (Salmo salar). Three genomic prediction methods (GBLUP, BayesC and BayesGC) were compared using 215K and 750K SNP genotypes through both within-family and across-family prediction scenarios. The data consisted of 1385 fish with both phenotype- and genotype, and the prediction accuracy was determined through five-fold cross-validation. The results showed an accuracy of ~0.6 and ~0.61 for across-family prediction with 215K and 750K genotypes and ~0.67 for within-family prediction for both genotypes. BayesGC showed a slightly higher prediction accuracy than GBLUP and BayesC, especially for the across-family predictions, but the differences were insignificant.

Paper II investigated the prediction accuracy of GBLUP, BayesC and BayesGC for six maternal traits in Landrace sows. The data consisted of between 10,000 and 15,000 sows, all genotyped and imputed to a genotype density of 660K SNPs. The effects of different priors for the Bayesian variable selection methods were also investigated. The ~1,000 youngest sows were used as validation animals to validate the prediction accuracy. Results showed a variation in genomic prediction accuracy between 0.31 to 0.61 for the different traits. The accuracy did not vary much between the different methods and priors within traits. BayesGC had a 9.8 and 3% higher accuracy than GBLUP for traits M3W and BCS. However, for the other traits, there were minor differences.

For within-breed prediction marker density and sizes of reference populations are often sufficient. However, when predicting across breeds, one might need a higher density, such as Whole Genome Sequence (WGS), or one could benefit from functional markers derived from WGS. Paper III investigates prediction accuracy for four maternal traits in two pig populations, a pure-bred Landrace (L) and a Synthetic (S) Yorkshire/Large White line. Prediction accuracy was tested with three different marker data sets: High-Density (HD), Whole Genome Sequence (WGS) and markers derived from WGS based on their pig Combined Annotation Dependent Depletion (pCADD) score. Two genomic prediction methods (GBLUP and BayesGC) were investigated for across- within- and multi-line predictions. For across- and within-line prediction, reference population sizes between 1K and 30K animals were analysed for prediction accuracy. In addition, multi-line reference population consisting of 1K, 3K or 6K animals for each line in different ratios were tested. The results showed that a reference population of 3K-6K animals for within-line prediction was usually sufficient to achieve a high prediction accuracy. However, increasing to 30K animals in the reference population further increased prediction accuracy for two of the traits. A reference population of 30K across-line animals achieved a similar accuracy to 1K within-line animals. For multi-line prediction, the accuracy was most dependent on the number of within-line animals in the reference data. The S-line provided a generally higher prediction accuracy than the L-line. Using pCADD scores to reduce the number of markers from WGS data in combination with the GBLUP method generally reduced prediction accuracies relative to GBLUP_HD analyses. When using BayesGC, prediction accuracies were generally similar when using HD, pCADD, or WGS marker data, suggesting that the Bayesian method selects a suitable set of markers irrespective of the markers provided (HD, pCADD, or WGS).

Overall, these three studies showed that BayesGC seemed to have a slight advantage over GBLUP, especially with large datasets, high-density genotypes, and when relationships between the reference and validation animals were lower. They also showed that the relationship between the animals in the reference and validation population, and the size of the reference population, had a more significant impact on the prediction accuracy than the prediction method.