Genomic Selection for Growth and Wood Traits in Castanopsis hystrix

Abstract

Castanopsis hystrix, a precious tree species in Southeast Asia, has the advantages of rapid growth and high-quality wood materials. However, there are problems such as its long breeding cycle and low efficiency, and being time-consuming, which greatly restricts the industrial development of C. hystrix. Performing genome selection (GS) for growth and wood traits for the early selection of superior progeny has great significance for the rapid breeding of new superior varieties of C. hystrix. We used 226 clones in the main distribution and 479 progenies within 23 half-sib families as experimental materials in this study. Genotyping datasets were obtained by high-throughput re-sequencing technology, and GS studies were conducted on the growth (tree height (H), diameter at breast height (DBH)) and wood (wood density (WD), fiber length (FL), and fiber length–width ratio (LWR)) traits. The coefficient of variation (CV) of five phenotypic traits ranged from 10.1% to 22.73%, the average CV of growth traits was 19.93%, and the average CV of wood traits was 9.72%. The Pearson correlation coefficients between the five traits were almost all significantly positive. Based on the Genomic Best Linear Unbiased Prediction (GBLUP) model, the broad-sense heritabilities of growth traits were higher than those of wood quality traits, and the different number of SNPs had little effect on the heritability estimation. GS prediction accuracy first increased and then reached a plateau at around 3K SNPs for all five traits. The broad-sense heritability of these five traits was significantly positively correlated with their GS predictive ability (r = 0.564, p < 0.001). Bayes models had better GS prediction accuracy than the GBLUP model. The 15 excellent progeny individuals were selected, and their genetic gain ranged from 0.319% to 2.671%. These 15 superior offspring individuals were 4388, 4438, 4407, 4468, 4044, 4335, 4410, 4160, 4212, 4461, 4052, 4014, 4332, 4389, and 4007, mainly from three families F5, F6, and F11. Our research lays out the technical and material foundation for the rapid breeding of new superior varieties of C. hystrix in southern China.

1. Introduction

Genomic selection (GS) is a method of facilitating the rapid selection of superior genotypes using dense genomic markers, which has the advantages of there being no need to detect major genes, the efficient capture of genomic genetic variation, and a greatly shortened breeding cycle, especially for low-level heritability traits and difficult-to-measure complex traits [1]. Meuwissen et al. claim that, for the basic assumption of GS theory, at least one SNP among the dense genomic markers is in direct linkage with the quantitative trait locus, which affects the target trait [1]. GS mainly consists of two steps [2]: first, the GS prediction model is constructed based on the genotypic data and phenotypic data of the reference population; and second, the model is used to estimate the genomic-estimated breeding value (GEBV) of the candidate population, and early selection is performed.

GS was first applied to livestock breeding [3]. Subsequently, this tool was introduced into crops, such as maize [4] and rice [5], improving the predictive ability of complex phenotypic traits. Compared with livestock and crops, the GS research on forest trees is still in the initial stage due to factors such as long breeding cycles, the complex genetic structure of traits, and a weak research foundation. At present, GS research of forest trees has mostly been on commercial timber species, such as eucalyptus, pine, and spruce, namely, Eucalyptus grandis × E. urophylla [6], Eucalyptus grandis [7], Eucalyptus robusta [8], Eucalyptus nitens [9], Pinus taeda [10], Pinus pinaster [11], Picea glauca [12], Norway spruce [13], and Pseudotsuga menziesii [14], focusing on the growth and wood traits. There are relatively few GS studies on economic forest tree species, such as Elaeis guineensis [15], Hevea brasiliensis [16], and Macadamia integrifolia [17], mainly focusing on yield-related traits. Overall, the GS research on forest trees is only reported in a few traditional commercial timber tree species, and most of them are concentrated in theoretical research [18]. There are still a few cases of GS breeding practices on forest trees, especially precious tree species.

Castanopsis hystrix, an evergreen tree mainly distributed in southern China, is one of the important native broad-leaved tree species and precious timber afforestation tree species, as well as an important tree species for long-term large-diameter timber forests in the 14th Five-Year Plan [19]. Due to the advantages of a straight trunk, strong corrosion resistance, and excellent beautiful heartwood, it is widely used in wooden furniture, interior decoration, and other industrial productions [20]. So far, studies on C. hystrix have mainly focused on the conventional genetic variation analysis of growth and wood traits [20,21,22,23] and genetic diversity analysis [24,25]. Although these studies are helpful to the mining and utilization of the germplasm resources of C. hystrix, there are still some problems such as its long breeding cycle and low efficiency, and being time-consuming. Compared with traditional commercial tree species, GS studies on precious tree species, including C. hystrix, are still not reported. Therefore, it is of great significance to carry out genome selection for the rapid breeding of new varieties in C. hystrix.

In this study, 226 clones in the whole distribution area of C. hystrix were used as the GS reference population, and 479 offspring covering 23 half-sib families were used as the candidate population. Genotyping datasets were obtained by high-throughput re-sequencing technology, and GS studies were conducted on the growth (tree height (H), diameter at breast (DBH)), wood (wood density (WD), fiber length (FL), and fiber length–width ratio (LWR)) traits. The effects of 13 different numbers of SNPs and 5 different GS models on heritability estimation and GS prediction accuracy were assessed using 5-fold cross-validation. Then, the candidates’ genomic estimated breeding values (GEBVs) were predicted based on the GS model, and early selection of superior progeny individuals was implemented by Brekin’s multi-trait evaluation method.

2. Materials and Methods

2.1. Experimental Materials

A total of 226 clones were selected as the reference population in the C. hystrix gene bank [25] located at the Longyandong Forest Farm in Guangdong Province. The gene bank material was collected in 2001 from six provinces in southern China. The details of the gene bank material were shown in the reference [25] (Genes (Basel, Switzerland), 2022,13,2383). The branches of the excellent trees were planted in the gene bank in January 2003 after grafting, using a completely random design with 5 plots per clone. In January 2020, the seeds of the above gene bank were collected for seedling culture, and planted in the Yangjiang Forest Farm, Yunyong Forest Farm, and Yunfu Forest Farm in Guangdong Province in April 2021. A total of 479 2-year-old disease-free offspring from 23 families were selected as the candidate population from April to May 2022.

2.2. Phenotypic Trait Determination and Descriptive Statistics

At the end of 2018, at the age of 16 years, 226 clones of the reference population were measured for tree height (H, m) and diameter at breast height (DBH, cm), and each clone measured three sample trees. Wood traits were measured from the reference population using two increment cores extracted at breast height. One core was used for wood density determination and the other one was used for fiber trait determination. Wood density (WD, g·cm⁻³) was determined using the conventional water replacement method [26]. The other core wood sample was segregated with 1:1 hydrogen peroxide:glacial acetic acid, heated until the wood core turned white and soft, and washed with water more than 5 times until the samples had no obvious smell. After the specimen had been broken up with a fiber decomposition breaker, the average fiber length (FL, mm) and average fiber width (μm) were automatically measured using a Valmet fiber image analyzer (Valmet FS5). The fiber length–width ratio (LWR) was obtained using the ratio between fiber length and fiber width. Higher WD, longer FL, and higher LWR served as the criteria for preferable precious timber. All five phenotypic traits were analyzed by R package psych.

2.3. Genotyping and Quality Control

Leaf DNA samples from 226 clones and 479 progenies of the candidate population were extracted using a MagMax^TM Plant DNA Isolation Kit (Thermo Fisher, Waltham, MA, USA), and used for genotyping through resequencing at Novogene (Shanghai, China). A total of 86,031,588 SNPs were obtained for genotyping in each sample, and VCFTools software [27] was used for quality control. The quality control criteria were (1) SNP missing rate less than 20%; (2) minor allele frequency greater than 0.01; and (3) Hardy–Weinberg equilibrium p value less than 0.0001. The missing genotypes were then imputated using Beagle 5.0 software [28], and 790,877 SNPs per individual were finally obtained for subsequent GS analysis.

2.4. Statistical Models for Genomic Prediction

Statistical models are one of the key factors affecting the accuracy of genomic selection. Herein, five statistical models, including the GBLUP model and four Bayes models, were used to analyze the effect of the GS model on the accuracy of genomic prediction.

2.4.1. GBLUP Model

GBLUP method [3] was carried out by the following model:

$$y=Xb+Zu+e$$

(1)

where y is the vector of phenotypic values, b is the fixed effect vector, u is the random genetic effect vector, X and Z are the incidence matrices of b and u, respectively, and e is the random residual effect. GBLUP method directly estimated the individual GEBVs through the genomic relationship matrix G constructed from the SNP markers. The GBLUP model was analyzed using the R package AFEchidna 1.68 [29].

2.4.2. Bayesian Models

Four Bayesian methods [2] were analyzed with the following model:

$$y=Xb+W{a}_m+e$$

(2)

where y is the vector of phenotypic values, b is the fixed effect vector, X is the incidence matrix of b, W is the matrix of numeric genotypes for each SNP marker, $a_m$ is the random marker effect, and e is the residual effect.

Bayes A model assumes that each SNP marker (j) has an effect with a normal distribution $p\left(j|{\sigma }_j^2\right)=N\left(0,{\sigma }_j^2\right)$, and the variance of each effect ${\sigma }_j^2$ is different with a scaled inverted chi-square distribution $p\left({\sigma }_j^2\right)={\chi }^{-2}\left(v,S\right)$ with two hyperpriors: degrees of freedom (v) and scale (S). Based on the Bayes A model, Bayes B assumes that only a few SNPs are effective and that the proportion of effective SNPs ${\sigma }_j^2$ is 1 − π (π usually takes a value of 0.95). Bayes C model assumes that all the effective markers have a common variance, and solves the π as an unknown parameter through the model. The Bayes ridge regression (BRR) model assumes the same random effects across markers, similar to the GBLUP model. The above Bayes model analyses were performed using the R-package BGLR [30].

2.5. The Effect of Different SNP Numbers on Genomic Prediction Accuracy

In order to explore the effect of different numbers of SNP markers on genomic prediction accuracy, different SNPs were randomly selected according to the 50 Kb window on each chromosome, and a total of 13 marker numbers were set as follows: all available markers were 791K (100%), 633K (80%), 475K (60%), 316K (40%), 158K (20%), 79K (10%), 40K (5%), 16K (2%), 5K, 3K, 2K, 1K and 0.5K SNPs, K is 1000. The Bayes C model was implemented, and the 5-fold cross-validation method was used to evaluate the accuracy of genome prediction.

2.6. Genomic Prediction Accuracy Assessment

Genomic prediction accuracy (PA) was assessed using a 5-fold cross-validation method, in which the reference population was randomly divided into five subsets, four of which were used as the training population and the remaining one as the validation population, were used to evaluate the prediction accuracy [2]. Model prediction accuracy (PA) is defined as the Pearson correlation coefficient between the GEBV and its true phenotypic value (y) in a validation population divided by the square root of heritability [31], as follows:

$$\mathrm{P}\mathrm{A}={\displaystyle \frac{cor\left(\mathrm{G}\mathrm{E}\mathrm{B}\mathrm{V},y\right)}{\sqrt{H^2}}}$$

(3)

where PA is the prediction accuracy of GS, cor(GEBV,y) is the Pearson correlation coefficient between the GEBV and the phenotypic value (y) of the validation population, and $H^2$ is the trait heritability.

The 5-fold cross-validation was repeated 20 times, averaging a total of 100 values as prediction accuracy.

2.7. Heritability Estimate

The GBLUP model was used to estimate the broad-sense heritability as follows:

$$H^2={\displaystyle \frac{{\sigma }_g^2}{\left({\sigma }_g^2+{\sigma }_e^2\right)}}$$

(4)

where $H^2$ is the broad-sense heritability, ${\sigma }_g^2$ is the genetic variance, and ${\sigma }_e^2$ is the residual variance. The generalized heritability of traits was estimated by the GBLUP model using the R package AFEchidna 1.68 [29].

2.8. Early Selection of Superior Progeny Individuals in the Candidate Population

Based on the genomic prediction model of the reference population of 226 clones, the individual GEBVs of 479 progenies in the candidate population were predicted with genotypic data. Brekin’s multi-trait assessment method [32] was used to comprehensively evaluate the GEBV of the candidate population with the following formula:

$$Q_i=\sqrt{\sum _j^n{\displaystyle \frac{X_{ij}}{X_{jmax}}}}$$

(5)

where $Q_i$ is the comprehensive evaluation value of individual i, j is the trait j, n is the number of traits, $X_{ij}$ is the GEBV value of trait j of individual i, and $X_{jmax}$ is the maximum GEBV value of trait j.

Genetic gain (∆G) of selected superior progenies for each trait was calculated with the following formula:

$$∆\mathrm{G}={\displaystyle \frac{H^2\times ({\mu }_s-{\mu }_p)}{{\mu }_p}}\times 100\%$$

(6)

where $∆\mathrm{G}$ is the genetic gain, $H^2$ is the broad-sense heritability, ${\mu }_s$ is the mean GEBVs of the selected superior progeny individuals, and ${\mu }_p$ is the mean phenotypic values of the reference population.

3. Results

3.1. Descriptive Statistics of Growth and Wood Traits of C. hystrix

The descriptive statistical results of growth and wood traits of 226 clones in the reference population are shown in Figure 1, the diagonal histogram plots indicated that all five traits nearly obeyed a normal distribution. The coefficients of variation of growth traits were higher than those of wood traits, among them, the coefficient of variation of DBH was highest (26.59%), and the coefficient of variation of FL was lowest (10.10%). Moreover, the Pearson correlation coefficients between the five traits were almost significantly positive, among them, the highest one was between FL and LWR (r = 0.68, p < 0.001). Interestingly, there is a significant positive correlation between WD and growth traits, indicating that we may improve WD and growth traits simultaneously.

3.2. Statistics of Genotyping Data

After quality control, the distribution of SNPs on the C. hystrix genome was relatively uniform (Figure 2), and two regions in both Chr3 and Chr9 and one region in Chr6 had a large marker density, greater than 3500 SNP·Mb⁻¹. As shown in

Table 1. The distribution of SNPs in C. hystrix genome.

Chromosome NO	Chr1	Chr2	Chr3	Chr4	Chr5	Chr6	Chr7	Chr8	Chr9	Chr10	Chr11	Chr12
Chromosome size (Mb)	109.75	106.04	103.31	87.42	77.66	68.54	66.29	64.51	63.76	61.94	56.79	50.54
SNP number	83,426	89,409	97,191	72,637	65,273	52,119	59,774	56,819	60,835	53,556	54,151	45,687
Density(per Mb)	760.13	843.17	940.81	830.86	840.45	760.41	901.77	880.78	954.18	864.57	953.56	903.95

, Chr1 was the longest with a length of about 110 Mb and a marker density of 760 SNP·Mb⁻¹, while Chr12 was the shortest with a length of about 51 Mb and a marker density of 904 SNP·Mb⁻¹. The marker density of all chromosomes ranged from 760 to 954 SNP·Mb⁻¹, the highest was Chr9 (954.18 SNP·Mb⁻¹), the lowest was Chr1 (760.13 SNP·Mb⁻¹), and the average marker density was about 870 SNP·Mb⁻¹.

3.3. Effect of Varying the Number of SNPs on the Estimates of Heritability

To evaluate the impact of varying the number of SNPs on the estimates of heritability (H²), we used the GBLUP model, the results showed that the 13 different numbers of SNPs had little effect on broad-sense heritability (Figure 3). Heritabilities increased slightly and reached a plateau of around 5K SNPs for all traits. The highest heritability was H (around 0.5), followed by DBH (around 0.48), and the lowest one was WD (around 0.25). The broad-sense heritabilities of growth traits were higher than those of wood quality traits.

3.4. Effect of Varying Number of SNPs on the GS Prediction Accuracy

The effect of different SNP numbers on GS prediction accuracy was studied with the Bayes C model; the results showed that PA first increased and then reached a plateau of around 3K SNPs for all traits (Figure 4). Further comparison of prediction accuracy showed that the PA of H was higher than those of the other four traits and that the other four traits had similar GS prediction accuracy. The results of Pearson correlation analysis (Figure 5) showed that there was a significant positive correlation (r = 0.564, p < 0.001) between the broad-sense heritability of these five traits and their GS predictive ability.

3.5. Effect of Varying Statistics Models on the GS Prediction Accuracy

As shown in Figure 6, the prediction accuracy of four Bayesian models was better than that of GBLUP. However, there was little difference between the Bayesian models, and the Bayes C model was the highest and relatively robust. GS of H had the highest PA, followed by wood quality traits, and their PAs were close.

3.6. Early Selection of Superior Progeny Individuals in the Candidate Population

Based on the above comparison results of the different SNP numbers on the broad-sense heritability estimation and the prediction accuracy of GS, 5K SNPs were selected for the comprehensive selection of superior offspring. The top 15 superior progeny individuals were selected with a 3% selection rate. As shown in

Table 2. The GEBV and the comprehensive evaluation value of growth and wood traits of the selected 15 candidates.

ID	GEBVH	GEBVDBH	GEBVWD	GEBVFL	GEBVLWR	${\mathit{Q}}_{\mathit{i}}$	${\mathit{Q}}_{\mathit{i}}$ Rank	Family NO
4388	10.320	16.096	0.532	1141.030	59.876	2.213	1	F8
4438	9.863	17.014	0.530	1122.925	58.798	2.207	2	F5
4407	10.105	16.480	0.545	1099.743	58.360	2.205	3	F5
4468	10.055	16.618	0.528	1108.143	59.491	2.205	4	F5
4044	9.930	16.331	0.536	1139.370	58.493	2.204	5	F4
4335	9.625	16.599	0.530	1139.671	59.802	2.203	6	F11
4410	10.032	16.349	0.534	1102.054	58.874	2.200	7	F5
4160	9.567	16.727	0.520	1157.510	58.784	2.199	8	F11
4212	9.629	16.454	0.521	1134.130	60.471	2.199	9	F6
4461	10.129	16.352	0.530	1107.883	57.795	2.197	10	F29
4052	9.736	17.009	0.516	1119.941	58.657	2.197	11	F12
4014	9.653	16.191	0.525	1134.845	60.074	2.196	12	F6
4332	9.617	16.262	0.526	1118.131	60.458	2.195	13	F11
4389	9.946	16.592	0.526	1120.556	57.268	2.195	14	F8
4007	9.699	16.221	0.523	1134.877	59.434	2.195	15	F6
Mean.s	9.860	16.486	0.528	1125.387	59.109	-	-	-
Mean.p	9.375	15.633	0.520	1115.206	57.006	-	-	-
∆G(%)	2.671	2.625	0.375	0.319	1.217	-	-	-

GEBV is genomic estimated breeding value. H is height, DBH is diameter at breast, WD is wood density, FL is fiber length, and LWR is fiber length–width ratio, respectively. $Q_i$ is the comprehensive evaluation value of individual i. Mean.s is the mean GEBV of each trait of the selected individual. Mean.p is the mean phenotypic value of each trait in the reference population. $∆\mathrm{G}$ is genetic gain.

, the $Q_i$ value of these 15 selected progeny individuals ranged from 2.195 to 2.213, belonging to 7 families, of which F5 had the most, accounting for 26.7%, followed by F6 and F11, accounting for 20%, respectively. The mean GEBVs of the selected individuals almost exceeded the mean phenotypic values of the reference population, and their genetic gain ranged from 0.319% to 2.671%.

4. Discussion

4.1. Effect of Trait Heritability on GS Prediction Ability

Some studies have shown a positive correlation between trait heritability and GS predictive ability [10,15,33]. For example, Estopa et al. declared that the heritability of carbohydrate content was the highest and its GS predictive ability was also the highest in E. benthamii, while the heritability of volume was relatively tiny and its predictive ability of GS was also weak [33]. It is also reported that the GS model accuracy was positively related to trait heritability for growth and wood traits in Picea mariana [34]. In this study, there was a significant positive correlation between the generalized heritability of the five traits and their GS predictive ability (r = 0.564 p < 0.001). However, Cao et al. reported that there was no significant correlation between the heritability of traits and their GS predictive ability in peach [35]. Therefore, the effect of trait heritability on its GS predictive ability may be related to species and trait types. In addition, the number of SNP markers in the range of 3K–791K had little effect on the broad-sense heritability of the five traits, similar to the results of E. benthamii [33], indicating that when the number of SNPs reached a certain number, the number of SNPs had little effect on the heritability estimation.

4.2. Effect of Varying Number of SNPs on the GS Prediction Accuracy

Many studies have shown that the prediction accuracy of GS increases with the increase in the number of SNP markers, but when reaching a certain number, the prediction accuracy of GS tends to stabilize and reaches a plateau [13,33]. Nsibi et al. found that when the number of SNPs was less than 6103 (10% of the total markers), the prediction accuracy continued to increase, and then the increase in the number of SNPs had little effect on the prediction accuracy [36]. Estopa et al. reported that the GS predictive ability of all traits in E. benthamii reached a plateau after the number of SNPs reached 3K [33]. In P. mariana, it is sufficient to obtain the GS prediction accuracy with 0.5K to 1K SNP markers [34]. Herein, the GS prediction accuracy was lowest when the number of SNPs was 0.5K, and the reason may be due to the lower SNP density reducing the degree of linkage between SNP markers and QTLs, thus resulting in decreasing the GS prediction accuracy. However, it has also been reported that increased SNP labeling density can sometimes cause some noise that affects the accuracy of GEBV estimation [37]. The reasons may be as follows: (1) the increase in SNP marker density adds a large number of markers that are not related to the target trait, which interferes with the estimation of GEBV, and then affects the prediction accuracy of GEBV to some extent; or (2) the number of markers is much larger than the sample size, which may lead to model overfitting.

4.3. Effect of Varying Statistics Models on the GS Prediction Accuracy

Different traits have different genetic structures, and different statistical models should be used to evaluate the prediction efficiency [38]. Isik et al. observed similar GS prediction abilities of GBLUP, BRR, and Bayes LASSO (BL) in growth and trunk straightness traits in maritime pine, while when using the BL model, the prediction ability was higher, but the prediction bias was also larger [11]. In another study of maritime pine, the prediction accuracy of ABLUP was higher than that of GBLUP or BL [28]. In E. benthamii, the prediction ability of the six GS models for all target traits was similar, but all of them were better than the ABLUP model [33]. Our results show that the prediction accuracy of the Bayesian models for the same trait is very close, but it is higher than that of the GBLUP model, indicating that the assumptions of the Bayes models have little effect on GS prediction. Nevertheless, when more complex external environmental factors are considered, such as the statistics model accounting for external spatial effects and competitive effects, the GBLUP model may present its advantages for complex models [39].

4.4. The Efficiency of GS Early Selection

One significant advantage of GS is that it can greatly shorten the breeding cycle [18]. For example, in dairy cows [40], GS shortened the breeding cycle from 7 years to 1 year and increased genetic gain by 35%, while in rice [41], GS increased rice yield by 16% compared to traditional cross-breeding. In Eucalyptus, GS improved the efficiency of genetic gain per unit of time when the breeding cycle was reduced and selection intensity increased [42,43]. In this study, the genetic gain of the 5 traits of the 15 selected individuals ranged from 0.319% to 2.671%, compared with the reference population, regardless of the selection period. Assuming that the selection period is shortened from 16 years to 2 years, the genetic gain will increase by seven times.

4.5. The Limitations of Our Research

The prediction accuracy of GS in this study was still not high, and was lower than the average prediction accuracy (0.45) in the reported GS studies of forest trees [18]. The reasons may be as follows: (1) The reference population is tiny—Bartholome et al. suggested that the reference population should be greater than 2000 trees [28]. (2) The relationship between individuals in the reference population is weak—Isik reported that the GS prediction accuracy of the full-sib family was significantly higher than that of the half-sib family [18]. Our reference population in this study came from six provenances, and their genomic relationship matrix values were generally small. Therefore, in the future, half-sib and full-sib families should be added to the reference population and the GS prediction should be re-evaluated. Chen et al. presented that the inclusion of a large QTL SNP in the GS model as a fixed effect could improve GS prediction accuracy [44]. We can try including the large QTL SNP in the future GS prediction. Moreover, machine learning methods might be considered for use in the GS. Finally, in this paper, the GS method is mainly focused on a single site; genotype by environment (G×E) is known to have effects on GS prediction [45], so multiple environment trails should be considered in the second round of GS.

5. Conclusions

In this study, the GS prediction model for growth and wood traits in C. hystrix was established, and the effects of different SNP numbers and statistical models on the prediction accuracy of GS were also analyzed. The Bayes models had better prediction accuracy than the GBLUP model, and the number of SNPs had a limited effect on GS prediction accuracy, consistent with the results of GS studies in other forest trees. Meanwhile, the early selection of superior individuals was carried out according to the GEBVs of growth and wood traits of the candidate population, these selected individuals could be focused on in future experiment trials, which laid a technical and material foundation for the rapid breeding of excellent new varieties of C. hystrix. In addition, the prediction accuracy of GS in our research was not high; therefore, the reference population should be enlarged (ideally a minimum of 2000 trees), and the large QTL SNPs and G×E effects can also be included, for future more advanced GS prediction.