689
Aggregation of methods for genetic prediction
We evaluated the predictive performance (PP) of a forecaster combining 11 elementary DNA-based predictors of litter size in pigs. Data are litter size phenotypes belonging to lines A, B and the cross AB with respectively 2598, 1604 and 1879 individuals. Their predictions were obtained with a 60K SNP chip. The predictors used were Bayesian Ridge Regression, Bayesian LASSO, GBLUP, Reproducing Kernel Hilbert Space, and Neural Networks, using pedigree, marker matrix, principal score matrix or additive genomic relationship matrix. PP was measured as the correlation between predicted and realized phenotypes. Although the forecaster did not systematically yield PP higher than those yielded by the elementary predictor with the higher PP for both lines and their cross, the distribution of the ranking of predictors according to PP shows a consistent stability of the forecaster. This qualifies the forecaster as the most stable predictor.
Keywords:
genetic prediction, genomic selection, aggregation methods
