The Efficiency Of Genomic Selection Methods For GR In Simulated Sheep Populations

Selection plans in plant and animal breeding are driven by genetic evaluation. Recent developments suggest using massive genetic marker information, known as "genomic selection." There is little evidence of its performance in sheep, though. The fat measurements at GR (tissue depth of 110 mm from midline over the 12th rib) are used in this study. We simulated 4 sheep populations almost equally in number, including 2076 individuals in total, and 1000 QTLs and 30,000 SNPs markers. Sampling of validation and training data sets was carried out across and within populations, which allows comparing across- and within-population information. We analyzed four predictional medthods, including GBLUP, Bayes A, Bayes CPi and Bayesion Lasso, and empirically compared tow strategies for selection:(1) use of genomewide markers information, and (2) the combination of genomewide markers and pedigree information. We used the QMSim software to simulated sheep popolations, and the GS3 program to implement genome-wide genetic evaluation (genomic selection). Cross-validation techniques were used to providing assumption-free estimates of predictive ability.We get better accurate results by using of the within populations. If the individual GEBV need to be predicted for several populations, the individuals of training population should come from each populations. In this study, prediction models builded by using within-population information was more accurate than using across-population information, and populational linkage disequilibrium are more tightly between training population and validation population.There are 30000 SNPs were used in this study. The model only use the SNPs information got better accurate than the model use the SNPs and polygenic information in spite of being a simpler one. The polygenic component be thought of as fitting the genes not accounted for by the marker-locus effects in the two effects (Polygenic and SNPs) model. One could expect the opposite, The most likely explanation is polygenic genetic values and "marker-explained" global genetic values are expected to be extremely collinear, which deteriorates performance of the estimation.Bayes Lasso and Bayesian CPi alwayes have the best performence of predition in genomic selection. The distribution of QTL effects be simulated in this study is a L-shaped gamma distribution, it have a large frequency close to zero. Bayes CPi and Bayesian Lasso give more appropriate prior distributions, but GBLUP and Bayes A make unreasonable assumptions about the distribution of QTL effects. These results suggest that different traits have varying genetic architecture and thus require appropriate prior distribution.In the same prediction models and methods, we get more accurate results when using bigger training population. The amount of the number of training population required is 1000 in general, the prediction accurates are incresing rapidly with the number of training population growing when the number of training population less than 1000, and the rate are reduced when the number of training population more than 1000 individuals.For the MCMC sampling in bayesian methods, the most important features are the number of iterations and residual correction interval (to avoid numerical problems). Although the performence of prediction in this study are little dependent on this two features, if SNPs information are included in the modles and need to estimate variances, the minimum requirement for MCMC is a number of iterations of 100000 with a burn-in of 20000 and a residual correction every 10000 iterations. The numerical problems in MCMC sampling are not significantly, so we can use a bigger residual correction interval, such as 2000 to short the computational time.In this study, the effects infuence the prediction accurate of GEBV were analyzed, these results provide scientific bases for the application of sheep genomic selection using low density SNP chip.