Performance of genomic prediction within and across generations in maritime pine

Abstract

Background: Genomic selection (GS) is a promising approach for decreasing breeding cycle length in forest trees. Assessment of progeny performance and of the prediction accuracy of GS models over generations is therefore a key issue.

Results: A reference population of maritime pine (Pinus pinaster) with an estimated effective inbreeding population size (status number) of 25 was first selected with simulated data. This reference population (n = 818) covered three generations (G0, G1 and G2) and was genotyped with 4436 single-nucleotide polymorphism (SNP) markers. We evaluated the effects on prediction accuracy of both the relatedness between the calibration and validation sets and validation on the basis of progeny performance. Pedigree-based (best linear unbiased prediction, ABLUP) and marker-based (genomic BLUP and Bayesian LASSO) models were used to predict breeding values for three different traits: circumference, height and stem straightness. On average, the ABLUP model outperformed genomic prediction models, with a maximum difference in prediction accuracies of 0.12, depending on the trait and the validation method. A mean difference in prediction accuracy of 0.17 was found between validation methods differing in terms of relatedness. Including the progenitors in the calibration set reduced this difference in prediction accuracy to 0.03. When only genotypes from the G0 and G1 generations were used in the calibration set and genotypes from G2 were used in the validation set (progeny validation), prediction accuracies ranged from 0.70 to 0.85.

Conclusions: This study suggests that the training of prediction models on parental populations can predict the genetic merit of the progeny with high accuracy: an encouraging result for the implementation of GS in the maritime pine breeding program.

Keywords: Genomic selection; Growth; Multiple generations; Pinus pinaster; Progeny validation; Relatedness; Stem straightness.

Fig. 1

Strategy for selecting the reference population and validation methods for model evaluation. The reference population was designed in two steps. The first was based on breeding value and pedigree information and the second was based on the use of simulated data to optimize the population to be genotyped. The reference population was then used to evaluate the performance of prediction models with different validation methods

Fig. 2

Prediction accuracy (a) and status number (b) based on simulated data. Results are given for four methods for selecting G2 individuals (Random, HS: half-sib family, FS: full-sib family and CD: coefficient of determination). The prediction accuracy was calculated as Pearson’s correlation coefficient for the relationship between GEBV and true breeding values for the validation set assessed by the cross-validation method. The results obtained with A_PBLUP are in orange, those obtained with A_FBLUP are in green, and those obtained with GBLUP are shown in purple. A Tukey boxplot is used to represent the data

Fig. 3

Comparison between expected and realized genetic relationship coefficients. Expected additive genetic relationships from the pedigree (top panel) and realized genetic relationships from SNP markers (bottom panel), for the reference population

Fig. 4

Relationship between predicted breeding values (x-axis) and empirical breeding values (y-axis) for the progeny validation method. The three traits (circumference, height and stem straightness) and three different models (ABLUP, GBLUP and B-LASSO) are represented. The prediction accuracy (r) of genomic prediction models evaluated on the validation set (G2 genotypes are shown as open green circles) is indicated. Closed circles represent the calibration set with G0 genotypes (n = 46) in blue and G1 genotypes (n = 62) in orange