Some methodological developments in genomic selection

Genomic selection involves predicting the value of candidate breeding lines using their genotype on the basis of a model that has been trained on a population of lines with both genotype and phenotype. Since it's proposal in 2001, a number of statistical and machine learning approaches to genomic selection have been put forward. A fundamental divide exists between methods that use genotype data to construct a similarity matrix between lines and make predictions as a function of the similarity between training and candidate lines versus methods that require prior analysis of the genotypes and phenotypes of training lines. Method similarities can be assessed on the basis of the similarities of their predictions. From a purely predictive point of view, it may be useful to combine predictions from more than one method to obtain an "ensemble" prediction that has improved stability over that of any single method. Finally, predictions will commonly be required for more than one trait and joint analysis of multiple traits can capture information in the covariance between them.