Prediction accuracy of single-step BLUP for growth and wood quality traits in the lodgepole pine breeding program in British Columbia

Abstract

Single-step best linear unbiased prediction (HBLUP) is a method used to predict breeding values by combining pairwise relatedness information derived from a pedigree with the realized relationships estimated from DNA markers. It is an ideal approach for the lodgepole pine (Pinus contorta Dougl. ex. Loud.) breeding program which has an extensive progeny testing program but a small proportion of trees that are genotyped. However, it is unclear what level of genotyping is required to affect prediction accuracy and genetic parameters, the performance across test sites and test cycles of different ages, and the ability to accurately rank trees within half-sib and full-sib families. Lodgepole pine trees were sampled from four progeny test sites in British Columbia, Canada. A SNP array was used to genotype 1569 trees which resulted in 19,584 high-quality SNPs. The prediction accuracy of HBLUP was compared to (1) an uncorrected relationship matrix (ABLUP) and (2) BLUP using a realized relationship matrix based on SNP markers (GBLUP) using various cross-validation scenarios for height growth at age 10 and 5 wood quality traits. Combining average and realized pairwise relationship information through the H-matrix resulted in heritability and Type B genetic correlation estimates that were generally a compromise between estimates for ABLUP (0% genotyping) and GBLUP (100% genotyping). The highest heritability was for average wood density (0.57 for ABLUP; 0.51 for HBLUP; 0.47 for GBLUP) and the lowest was for height (0.24 for ABLUP; 0.27 for HBLUP; 0.25 for GBLUP). GBLUP always had the lowest Type B genetic correlations (except for earlywood density) of the three models (0.46 to 1.0) assessed. The prediction accuracy for HBLUP increased slightly for genotyped trees (0.77 to 0.80), but not for non-genotyped trees as genotyping effort increased. Furthermore, prediction accuracy was high when predicting between environments (0.46 to 0.85) and test cycles (0.33 to 0.76) when connected through pedigree, and prediction was more accurate when using older first-cycle tests to predict breeding values for younger second-cycle tests for all traits, except microfibril angle. Rank correlations for trees within half-sib and full-sib families when predicting values across test cycles (the training population is phenotyped and the validation population is genotyped) were very low using HBLUP (0.08) compared to GBLUP (0.38) but increased to 0.25 when 40% of the trees in the training population were genotyped (HBLUP40). HBLUP should be regarded as an effective way to combine average and realized relationship information in a breeding program for more precise estimates of genetic parameters and breeding values and can be used for predicting and ranking trees within families without phenotypic data when genotyped trees from the same families are included in the training population (genotyped and phenotyped).