Bayesian genomic selection: the effect of haplotype length and priors

Breeding values for animals with marker data are estimated using a genomic selection approach where data is analyzed using Bayesian multi-marker association models. Fourteen model scenarios with varying haplotype lengths, hyper parameter and prior distributions were compared to find the scenario expected to give the most correct genomic estimated breeding values for animals with marker information only. Five-fold cross validation was performed to assess the ability of models to estimate breeding values for animals in generation 3. In each of the five subsets, 20% of phenotypic records in generation 3 were left out. Correlations between breeding values estimated on full data and on subsets for the "leave-out" animals varied between 0.77-0.99. Regression coefficients of breeding values from full data on breeding values from subsets ranged from 0.78-1.01. Single-SNP marker models didn't perform well. Correlations were 0.77-0.89 and predicted breeding values were biased. In addition the models seemed to over fit the genomic part of the variation. Highest correlations and most unbiased results were obtained when SNP markers were joined into haplotypes. Especially the scenarios with 5-SNP haplotypes gave promising results (distance between adjacent SNPs is 0.1 cM evenly over the genome). All correlations were 0.99 and regression coefficients were 0.99-1.01. Models with 5-SNP markers seemed robust to hyper parameter and prior changes. Haplotypes up to 40 SNPs also gave good results. However, longer haplotypes are expected to have less predictive ability over several generations and therefore the 5-SNP haplotypes are expected to give the best predictions for generations 4-6.