From mass selection to genomic selection: one century of breeding for quantitative yield components of oil palm (Elaeis guineensis Jacq.)

Abstract

More efficient methods are required to breed oil palm (Elaeis guineensis Jacq.) for yield maximization in order to meet the increased demand for palm oil while limiting environmental impacts. This review article analyzes the evolution of breeding schemes for oil palm yield and its quantative components and the changes expected to take place with genomic selection (GS). Genetic improvement of oil palm yield started in the 1920s through mass selection. Later, several disruptive improvements dramatically increased the rate of genetic progress: (1) understanding the heredity of fruit form and the adoption of tenera, with thicker mesocarp, in plantations; (2) the discovery of hybrid vigor and the adoption of modified reciprocal recurrent selection; and (3) clonal selection, exploiting intra-hybrid variability. In addition, the use of linear mixed models to estimate genetic values has made selection more efficient. Today, GS appears to be a new disruptive improvement that can speed up breeding schemes by avoiding field trials in some cycles and increase selection intensity by evaluating more candidates. The genetic potential for oil palm yield has increased considerably over one century of breeding. GS is expected to bring the rate of genetic progress to a previously unprecedented level. The future studies on oil palm GS will aim at making it efficient for all yield components. For this purpose, they should focus in particular on the optimization of training populations and on the improvement of prediction models. Minimizing environmental impacts will also require improvement in other aspects (resistance to diseases, cultural practices, etc.).

Appendix. Estimation of oil palm genetic values using the BLUP methodology and R software

Here, we present a practical example of the estimation of oil palm genetic values using BLUP with R software (R Core Team 2017) and the breedR package (Muñoz and Sanchez 2018). It was chosen because the authors are familiar with its use, but other packages can be used, including sommer (Covarrubias-Pazaran 2016), RR-BLUP (Endelman 2011) and ASReml-R (Butler et al. 2009). In this example, we will estimate the GCA of parents from group A and group B evaluated in hybrid progeny tests while taking the pedigree-based relationships into account. This example can be very easily adapted for genomic prediction (GBLUP) as it only requires replacing the genealogical relationship matrices by genomic matrices, which could include individuals that have been genotyped but not progeny tested. The data files and R script are available at https://github.com/david-cros/article2018.

The data concern eight crosses made according to an incomplete NCM2 mating design between four group A Ds and four group B Ps (Table S1). The crosses were planted according to a RCBD with three replicates. The pedigree is given in Fig. 4. The yield obtained per cross in the different replicates (y) is listed in Supplementary Table S1. A simple linear mixed model was used, with replicates as fixed effect (β) and the parental GCAs as random effects (u_A et u_B):

Table S1 Incomplete NCM2 mating design (the asterisks represent the number of crosses)

Fig. S1

Pedigree of the example population. Green: individuals from group A; blue: individuals from group B; turquoise: A × B hybrid crosses

$$ y=\boldsymbol{X}\beta +{\boldsymbol{Z}}_{\mathbf{1}}{u}_A+{\boldsymbol{Z}}_{\mathbf{2}}{u}_B+\varepsilon $$

In matrix form, the model is:

${u}_A\sim \mathrm{N}\left(0,0.5{\boldsymbol{A}}_{\boldsymbol{A}}{\sigma}_{a_A}^2\right)$_, ${u}_B\sim \mathrm{N}\left(0,0.5{\boldsymbol{A}}_{\boldsymbol{B}}{\sigma}_{a_B}^2\right)$ and for example, for the eight individuals in the pedigree of group A, the coancestry matrix

$$ 0.5{\boldsymbol{A}}_{\boldsymbol{A}}=\left(\begin{array}{cc}\begin{array}{cc}\begin{array}{cc}\kern0.75em 0.5& \kern1em 0\kern1.75em \\ {}\kern0.5em 0\kern0.75em & \kern1.25em 0.5\kern1.25em \end{array}& \begin{array}{cc}0.25\kern1.5em & 0\\ {}0.25\kern1.75em & 0\end{array}\\ {}\begin{array}{cc}\kern0.5em 0.25\kern0.75em & 0.25\kern1em \\ {}0\ & 0\kern1.75em \end{array}& \begin{array}{cc}0.5\kern1.5em &\ 0\\ {}0\kern1.75em & 0.5\end{array}\end{array}& \begin{array}{cc}\begin{array}{cc}0.125& 0\\ {}0.125& 0.25\end{array}& \begin{array}{cc}0.063& 0\\ {}0.188& 0\end{array}\\ {}\begin{array}{cc}\kern0.5em 0.25& \kern0.5em 0.125\\ {}0.25& 0.25\end{array}& \begin{array}{cc}0.188& 0\\ {}0.25& 0\end{array}\end{array}\\ {}\begin{array}{cc}\begin{array}{cc}\ 0.125& 0.125\kern0.5em \\ {}0& 0.25\end{array}& \begin{array}{cc}0.25& 0.25\\ {}0.125& 0.25\end{array}\\ {}\begin{array}{cc}0.063& 0.188\\ {}0& 0\end{array}& \begin{array}{cc}\ 0.188& 0.25\\ {}0& 0\end{array}\end{array}& \begin{array}{cc}\begin{array}{cc}0.5\ & 0.188\\ {}0.188& 0.5\end{array}& \begin{array}{cc}0.344&\ 0\\ {}0.344&\ 0\end{array}\\ {}\begin{array}{cc}0.344& 0.344\\ {}0& 0\end{array}& \begin{array}{cc}0.594&\ 0\\ {}0&\ 0.5\ \end{array}\end{array}\end{array}\right). $$

The estimates of the variance components were obtained from the syntax:

where remlf90 is the function that analyzes the linear mixed model using the REML, fixed is the argument representing the fixed effects (here, replicates), generic the argument representing the random genetic effects (GCAs) and indicating for each the associated incidence and variance-covariance matrices (parent_A and parent_B are the columns in the table yield_data). The objects Z.mat_A and Z.mat_B are the incidence matrices Z₁ and Z₂, respectively. The objects A.mat_A and A.mat_B are the matrices 0.5A_A and 0.5A_B generated by the function kinship (package kinship2) that computes the genealogical coancestry coefficients between the individuals in the pedigree.

The analysis gives the following variance estimates (± standard error): ${\sigma}_{a_A}^2$ = 192.15 ± 164.58, ${\sigma}_{a_B}^2$ = 195.36 ± 164.51, ${\sigma}_{\upvarepsilon}^2$= 7.32 ± 2.68, and the solutions for the block effects (BLUE): β₁=6.85±8.98, β₂=5.78±8.98, β₃=6.47±8.98. The solutions for the GCAs (BLUP) are given in Table S2. According to the parental GCAs, the best possible cross would have been A₇ × B₃, with an expected yield of 29.60 (β+${u}_{A_7}$+${u}_{B_3}$), while the best cross in the trial was A₇ × B₉, with an expected yield of 20.91 (and a mean observed yield of 21.98).

Table S2 Additive genetic values of progeny tested individuals