Genomic Prediction in 3-Line Layer Data Using Different Linear Methods

In breeding programs with multiple lines, an alternative to increase the reference population for each line is to use multi-line reference populations. We evaluated prediction accuracy of several linear models in multi-line genomic prediction for three different lines of layer chickens. The following models based on genotypic information are compared: GBLUP, RRBLUP, principal component regression, Bayesian stochastic search variable selection and BayesC. The data contained two closely related lines of brown layers, and a third unrelated line of white layers. The lines had 1,006-1,026 animals in the training data and 238-240 validation animals. Seven different combinations of training animals were evaluated: three single line datasets, three combinations of two lines and one dataset with all three lines combined. Use of multiple lines marginally improved prediction accuracy of line B2 and did not harm prediction accuracy for lines B1 and W1. Genomic breeding values were highly related between methods for line W1 (correlations > 0.95), but less so for lines B1 and B2 (correlations > 0.87). Differences between models were more pronounced for related lines B1 and B2 than for line W1, probably partly because the achieved (GBLUP) accuracy for line W1 was higher (0.75) compared to lines B1 (0.48) and B2 (0.34). Differences in realized prediction accuracy for GBLUP across lines were largely in line with predicted accuracies based on observed heritabilities and estimated numbers of effective chromosome segments.