Accuracy of genome enabled prediction in a dairy cattle population using different cross-validation layouts

The impact of extent of genetic relatedness on accuracy of genome-enabled predictions was assessed using a dairy cattle population and alternative cross-validation (CV) strategies were compared. The CV layouts consisted of training and testing sets obtained from either random allocation of individuals (RAN) or from a kernel-based clustering of individuals using the additive relationship matrix, to obtain two subsets that were as unrelated as possible (UNREL), as well as a layout based on stratification by generation (GEN). The UNREL layout decreased the average genetic relationships between training and testing animals but produced similar accuracies to the RAN design, which were about 15% higher than in the GEN setting. Results indicate that the CV structure can have an important effect on the accuracy of whole-genome predictions. However, the connection between average genetic relationships across training and testing sets and the estimated predictive ability is not straightforward, and may depend also on the kind of relatedness that exists between the two subsets and on the heritability of the trait. For high heritability traits, close relatives such as parents and full sibs make the greatest contributions to accuracy, which can be compensated by half-sibs or grandsires in the case of lack of close relatives. However, for the low heritability traits the inclusion of close relatives is crucial and including more relatives of various types in the training set tends to lead to greater accuracy. In practice, cross-validation designs should resemble the intended use of the predictive models, e.g. within or between family predictions, or within or across generation predictions, such that estimation of predictive ability is consistent with the actual application to be considered.