Genomic Prediction in Pea: Effect of Marker Density and Training Population Size and Composition on Prediction Accuracy

Pea is an important food and feed crop and a valuable component of low-input farming systems. Improving resistance to biotic and abiotic stresses is a major breeding target to enhance yield potential and regularity. Genomic selection (GS) has lately emerged as a promising technique to increase the accuracy and gain of marker-based selection. It uses genome-wide molecular marker data to predict the breeding values of candidate lines to selection. A collection of 339 genetic resource accessions (CRB339) was subjected to high-density genotyping using the GenoPea 13.2K SNP Array. Genomic prediction accuracy was evaluated for thousand seed weight (TSW), the number of seeds per plant (NSeed), and the date of flowering (BegFlo). Mean cross-environment prediction accuracies reached 0.83 for TSW, 0.68 for NSeed, and 0.65 for BegFlo. For each trait, the statistical method, the marker density, and/or the training population size and composition used for prediction were varied to investigate their effects on prediction accuracy: the effect was large for the size and composition of the training population but limited for the statistical method and marker density. Maximizing the relatedness between individuals in the training and test sets, through the CDmean-based method, significantly improved prediction accuracies. A cross-population cross-validation experiment was further conducted using the CRB339 collection as a training population set and nine recombinant inbred lines populations as test set. Prediction quality was high with mean Q (2) of 0.44 for TSW and 0.59 for BegFlo. Results are discussed in the light of current efforts to develop GS strategies in pea.