P0508 Benchmarking genomic prediction in forestry – what works and what doesn't for some growth, developmental, and adaptative traits

Patricio R. Munoz , University of Florida, Gainesville, FL
We recently demonstrated that accurate prediction RRBLUP models can be developed for genome-wide selection (GWS) of growth traits in loblolly pine and that gains of over 150% in selection efficiency can be obtained (Resende et al 2011). However, various prediction methods have been found to perform differently depending on the species, and genetic architecture of the trait. Here we report the comparative results from GWS predictive models using four different methods (RRBLUP, BayesA, BayesCπ and Bayesian LASSO) for 26 different traits with different heritabilities and genetic architecture, measured in a loblolly pine breeding population. Two contrasting, 10-fold and leave-one-out, cross validation schemes were tested in RRBLUP. Because no difference was found between the two approaches, the 10-fold was chosen because requires significantly less computational time. We observed that for growth and developmental traits there was no significant difference in the accuracy of GWS methods. However, BayesCπ showed at least 15% better predictive ability than alternative methods applied to rust resistant traits. BayesCπ differs from RRBLUP in that it assumes some markers are not associated with the trait while RRBLUP assumes equal contribution to the observed variation. These results are consistent with the genetic architecture of rust resistant traits, where has been showed that is controlled by few genes with large effect. Not surprisingly, a selected subset of 100 markers based on the above results increased the accuracy (from 0.24 to 0.37) of RRBLUP outperforming other methods