Predicting Combining Ability and the Stability of Genomic Selection

Identifying a set of crosses that produces desirable offspring is a major goal of breeding. To achieve this goal, natural and artificial selection recurrently select and recombine a population’s heritable variation. Each selection procedure may be evaluated by its ability to explore a population's potential crosses and exploit the set that will bear desirable offspring or increase the population's propensity to generate future crosses bearing desirable offspring. We propose a supervised pairwise kernel machine useful in predicting the combining abilities of potential crosses in a genotyped population. Using a similar modeling framework, we infer the stability of established genomic prediction models with respect to changes in a population’s allele frequencies. Algorithmic selection of optimized sets of crosses based upon predicted combining abilities was also performed. The utility of these modeling and selection approaches is revealed through simulation and case studies in densely genotyped populations of Drosophila melanogaster (Drosophila Genetic Reference Panel, DGRP) and Zea mays (Nested Association Mapping Panel, NAM). The methods show promise in optimizing a breeder’s allocation of evaluation resources and will greatly assist in decisions regarding the crosses to be generated for hybrid development or population improvement.