Combining GWAS and Genomic Prediction using Sequence Information on 5000 Holstein Bulls, effect on Prediction Accuracy

Our earlier works (van Binsbergen et al, 2015) showed that using whole genome sequence information did not improve the accuracy for genomic prediction within the Holstein population. In this study we therefore compared genomic prediction using selected SNP from single SNP GWAS with imputed whole genome information (from run4 of the 1000 bull genomes project) to genomic prediction obtained using either the full sequence, 50k or BovineHD SNP chips. The GWAS used highly accurate deregressed proofs from 3416 training bulls, all progeny tested bulls for protein yield, interval first to last insemination and SCC . The analyses were performed in the GCTA package. The genomic relationship matrix was included to account for population structure and was based on the SNP of the bovine HD chip. After the GWAS, the relevance of different selected SNP sets were tested by estimating the h2 and the prediction accuracy in 2287 validation animals, using the GRM calculated with different SNP subsets. From the 30 million SNP in the sequence information, only 13,789,029 where segregating in the Holstein population. For protein yield 2,194 SNPs were significant (-log10(p) >5), and 28 (160) of those were present on the 50k (HD) SNP chips. Within the Holstein population prediction accuracy as well as the h2 were lower when SNPs were selected based on the GWAS, and there was no advantage in comparison with the 50K or BovineHD SNPchips. But more selective selection procedures and training population might be required to benefit from the precision of full sequence genotypes.