Genomic relationships between training and testing sets affect genomic prediction accuracy of nodavirus resistance in Gilthead sea bream

Gilthead sea bream (Sparus aurata) has been reported to be susceptible to a reassortant betanodavirus strain (RGNNV/SJNNV), posing a new threat for sea bream industry (Volpe et al. 2020) and raising the attention to selective breeding as a plausible disease prevention action. Genomic selection might be beneficial for traits, such as disease resistance, characterized by difficult, expensive and time-consuming routine individual phenotyping. Genomic models are trained firstly using a reference population of full- and half-sibs of the future breeding candidates, but, in a long-term view, the prediction of the genetic merit of future breeding candidates should be satisfactory even when the reference population consists of distant relatives of the animals to be predicted. In this sense, the genomic predictive accuracy provided by random k-fold cross-validations might be unrealistic. In this study, we assessed the accuracy of a genomic prediction model for VNN symptomatology pseudo-phenotypes (estimated breeding values, EBV) in gilthead sea bream in three different validation settings: 1) a random cross-validation; 2) a cross-validation based on genomic clustering; 3) a leave-one-family-out (LOFO) validation focused on the parents of the fish subjected to the VNN challenge test.