Predictions of Genetic Merit in Tree Breeding Using Factor Analytic Linear Mixed Models and Blended Genomic Relationship Matrices.

摘要

Increase in computer power and efficiency in DNA sequencing technologies is providing new opportunities to plant and animal breeders to fit more complex statistical models for predictions of genetic merit of individuals. Such models can be powerful to account for heterogeneity in the data and as a result can increase the accuracy of predictions and genetic gains from breeding programs.;In this study, I first evaluated the efficiency of factor analytic (FA) linear mixed models for a large, multi-environmental trial of loblolly pine (Pinus taeda L.). Height was assessed on 37,269 trees at age six years in a diallel experiment. Among models fitted, FA models produced the smallest AIC model fit statistics. FA models captured both the variance and covariance at the genetic level better than models with simpler covariance structures, and they provided more accurate predictions of genotypes. The mean narrow-sense heritability estimates for height was about 0.20 when more complex variance structures were used, compared to 0.13 when simpler variance structures were employed. FA models were parsimonious compared to US structures. The FA models provided a natural framework for modeling genotype by environment interaction. Genotype by environment interaction was non-significant as suggested by high genetic correlations both for additive (0.83) and dominance (0.91) effects.;Molecular marker data, especially single-nucleotide polymorphic (SNP) markers have been commonly used to predict genetic merit in animal breeding. However, marker data probably have missing genotypes and they need to be imputed. The effects of percent (level) and pattern (random or structured) of missing data, and mating designs on the accuracy imputation of genotypes were investigated. I used linkage based BLUP to impute missing genotypes for an empirical (unbalanced) data set for loblolly pine. For simulated (balanced) data sets, both BLUP and Hidden Markov Model (HMM) approaches were used. The actual data had 178 clones that were genotyped at 3,461 biallelic SNP markers. The simulated data consist of double-pair and half-diallel mating design with 2880 and 2940 individuals, respectively. For empirical data, accuracy of imputation was higher for the structured pattern of missing data at any level of missing percentages. Regardless of the pattern of the missing data, imputation accuracy was less than 0.70 when the data had greater than 40% missing values. For the simulated data, the imputation accuracy was not affected by mating design for the BLUP approach when the pattern of the missing data was structured. For HMM approach, when the pattern of the missing data was evenly-spaced, the mating design had no effect on the imputation accuracy.;Combining information from pedigree and DNA markers might improve prediction accuracies in tree breeding programs. In the third chapter, a cloned population of loblolly pine and simulated data sets were used to examine the efficiency of blended additive genetic relationships and realized genomic relationships were examined. Cloned 166 individuals were genotyped at 3,461 SNP markers out of a total 354 clones. For simulated data 300 or 600 trees genotyped at 1000 markers out of 1200. For the empirical data, the accuracy of predictions based on the ABLUP was 0.79. Predictions based on HBLUP had accuracy of 0.72 to 0.76, depending on the genomic relationships used. For the simulated data set, the accuracy values for HBLUP models were higher compared to ABLUP model. Also, as the genotyped population size increased the accuracies increased. HBLUP uses all the available phenotype, pedigree, and genotype data in a single-step and is easy to implement for genomic based selection. The major advantage of using genomic relationships matrices is that markers can capture the Mendelian sampling effect within full-sib families for selection.