Evaluating the accuracy of genomic prediction of growth and wood traits in two Eucalyptus species and their F1 hybrids

Abstract

Background: Genomic prediction is a genomics assisted breeding methodology that can increase genetic gains by accelerating the breeding cycle and potentially improving the accuracy of breeding values. In this study, we use 41,304 informative SNPs genotyped in a Eucalyptus breeding population involving 90 E.grandis and 78 E.urophylla parents and their 949 F₁ hybrids to develop genomic prediction models for eight phenotypic traits - basic density and pulp yield, circumference at breast height and height and tree volume scored at age three and six years. We assessed the impact of different genomic prediction methods, the composition and size of the training and validation set and the number and genomic location of SNPs on the predictive ability (PA).

Results: Heritabilities estimated using the realized genomic relationship matrix (GRM) were considerably higher than estimates based on the expected pedigree, mainly due to inconsistencies in the expected pedigree that were readily corrected by the GRM. Moreover, the GRM more precisely capture Mendelian sampling among related individuals, such that the genetic covariance was based on the true proportion of the genome shared between individuals. PA improved considerably when increasing the size of the training set and by enhancing relatedness to the validation set. Prediction models trained on pure species parents could not predict well in F₁ hybrids, indicating that model training has to be carried out in hybrid populations if one is to predict in hybrid selection candidates. The different genomic prediction methods provided similar results for all traits, therefore either GBLUP or rrBLUP represents better compromises between computational time and prediction efficiency. Only slight improvement was observed in PA when more than 5000 SNPs were used for all traits. Using SNPs in intergenic regions provided slightly better PA than using SNPs sampled exclusively in genic regions.

Conclusions: The size and composition of the training set and number of SNPs used are the two most important factors for model prediction, compared to the statistical methods and the genomic location of SNPs. Furthermore, training the prediction model based on pure parental species only provide limited ability to predict traits in interspecific hybrids. Our results provide additional promising perspectives for the implementation of genomic prediction in Eucalyptus breeding programs by the selection of interspecific hybrids.

Keywords: Bayesian LASSO; GBLUP; Genome annotation; Genomic heritability; Genomic relationship; High-density SNP-chip; Two-generation; rrBLUP.

Fig. 1

Correlation and distribution of phenotypes. Scatter plots (lower off-diagonal) and correlations with probability values (upper off-diagonal; H₀: r = 0) for adjusted phenotypes between pairs of traits. Color key on the right indicates the strength of the correlations. Diagonal: histograms of the distribution of adjusted phenotypes values

Fig. 2

Genetic structure and relatedness in the breeding population. (a) First two principal components of a PCA revealing population structure. Dots represent E.grandis (blue), E.urophylla (red) and their F₁ (green) individuals. (b) Heatmaps of the pairwise pedigree-expected relationships (blue, upper off-diagonal) and genomic-realized relationship (red, lower off-diagonal) of the 1117 individuals assigned to E.grandis (G), E.urophylla (U) and their hybrid progenies (H)

Fig. 3

Predictive abilities with different methods and increasing sizes of training sets. Predictive ability (y axis) estimated using five methods across five training set/validation set sizes in numbers of individuals (x axis) 558/559, 743/374, 836/281, 892/225 and 1003/114. Red and blue dashed lines indicate the pedigree-based (${\mathit{h}}_{\mathit{a}}^2$) and genomic-realized (${\mathit{h}}_{\mathit{g}}^2$) narrow-sense heritability respectively

Fig. 4

Predictive abilities with variable levels of relatedness between training and validation sets. CV₁: random assignment of individuals to either training set (TS) or validation set (VS); CV₂: all the G0 pure species parents assigned to the TS; CV₃: minimum relatedness between TS and VS individuals; CV₄: maximum relatedness between TS and VS individuals. Estimates were obtained using GBLUP and RKHS across five TS/VS sizes in numbers of individuals (x axis): 558/559, 743/374, 836/281, 892/225 and 1003/114

Fig. 5

Predictive abilities with increasing numbers of SNPs. Predictive ability estimated with GBLUP and RKHS with increasingly larger sets of SNP sampled at random from the total of 41,304 SNPs. Outliers are indicated by black dots. Letters indicate significant difference between the different models after Bonferroni adjustment (P < 0.05)

Fig. 6

Predictive abilities using SNPs located in different genomic regions. Predictive ability estimated with GBLUP and RKHS using 11,786 SNPs in coding DNA, 30,405 SNPs in genic regions (CDS, UTR, intron, and within 2 kb upstream and downstream of genes), 10,899 SNPs in intergenic regions and all 41,304 SNPs. Letters indicate significant difference between the different models after Bonferroni adjustment (P < 0.05)