Multiple Trait Genomic Selection (MT-GS) in Plant Breeding via Multivariate Bayesian Modeling

Genetic correlation between traits is pervasive in plant breeding. These correlations mean that measurement of one trait carries information on another trait. Univariate analysis of traits does not take advantage of this information. Multivariate genomic selection models for plant breeding are not currently available. In this talk, I will demonstrate two multivariate Bayesian genomic selection models. The first model uses parameter shrinkage to estimate the genetic effects for all markers on the multiple traits (MT-BayesA). The developed MT-BayesA adopts a full Bayesian hierarchical approach to estimate optimal prior parameters from the data. The second model is different from MT-BayesA in that only a portion of markers is fit in the model via variable selection while all selected markers share a common variance (MT-BayesCpi). The selected marker proportion is estimated from the data. These new multivariate Bayesian models are compared with their corresponding univariate models and with uni- and multivariate ridge regression models. I will discuss how genetic architecture, heritability and genetic correlation affect multiple trait genomic prediction. Those models will be tested and applied to simulated and empirical plant breeding data. I will also show how MT-GS can be used to expand the training population by phenotype imputation. Finally, I will show computation performance for methods programmed in R versus in C.