The Bayesian Variable Selection Approaches In Quantitative Trait Locus Mapping And Genomic Selection

In the past 20 years,quantitative trait loci(QTL)mapping and genomic selection(GS)became the one of the hot topics in animals and plants as well as in human ge-nomics.With the development of the high-density genetic maps,the genome is easily saturated with tens of thousands,even a few millions of molecular markers,as such,the traditional approaches in QTL mapping and GS are no longer suitable.Therefore,a large number of experimental lines have been developed to increase the resolution of QTL mapping and the prediction accuracy of genomic selection.Based on those ex-perimental lines,constructing suitable statistical analysis method is very necessary and crucial.In the article,for the Collaborative Cross(CC)mice developed in animals,we constructed a linear mixed model by introducing a random effects term to express the unbalanced relationship between individuals to jointly model the strain effects of the eight founders and the parent-of-origin(PoO)effects.For the Brassica napus doubled haploid(DH)lines developed in the breeding of the plants,homogeneous and hetero-geneous residual variance models across environments are built to jointly model the environmental effects,the genetic effects and the interaction effects between the geno-types and the environments(GxE).Considering the discrete genotypes with values 0,1 or 2 and the linkage disequilibrium between nearby markers,several Bayesian variable selection approaches by embedding a binary indicator variable acting like a 0-1 switch into the model are proposed to achieve the aim of variable selection and parameter es-timation.Embedding a binary indicator variable into the model to indicate whether a marker should be included into or excluded from the model can substantially reduce the computational time through the block Gibbs sampling method that updates the blocks of the selected and the unselected components of the parameters in turn.By doing that,we can overcome one of the drawback of the Bayesian method that the Markov chains often take a long time to converge.Extensive simulation studies and real data analysis show that our proposed the whole genomic models which incorporate multiple effects,such as PoO effects,the GxE effects and so on,perform better than the models which ignore those effects and the existing methods in model interpretation,QTL mapping and genomic predictions.