•Genomic selection (GS) is expected to cause a paradigm shift in tree breeding by improving its speed and efficiency. By fitting all the genome-wide markers concurrently, GS can capture most of the ‘missing heritability’ of complex traits that quantitative trait locus (QTL) and association mapping classically fail to explain. Experimental support of GS is now required.
•The effectiveness of GS was assessed in two unrelated Eucalyptus breeding populations with contrasting effective population sizes (Ne = 11 and 51) genotyped with > 3000 DArT markers. Prediction models were developed for tree circumference and height growth, wood specific gravity and pulp yield using random regression best linear unbiased predictor (BLUP).
•Accuracies of GS varied between 0.55 and 0.88, matching the accuracies achieved by conventional phenotypic selection. Substantial proportions (74–97%) of trait heritability were captured by fitting all genome-wide markers simultaneously. Genomic regions explaining trait variation largely coincided between populations, although GS models predicted poorly across populations, likely as a result of variable patterns of linkage disequilibrium, inconsistent allelic effects and genotype × environment interaction.
•GS brings a new perspective to the understanding of quantitative trait variation in forest trees and provides a revolutionary tool for applied tree improvement. Nevertheless population-specific predictive models will likely drive the initial applications of GS in forest tree breeding.
Genomic selection (GS) or genome-wide selection (GWS) was a landmark proposal made 11 yr ago providing a genome-wide paradigm for marker-assisted selection (MAS) in plants and animals (Meuwissen et al., 2001; Goddard & Hayes, 2009). This idea, based on the forecast of rapid technological advances and dropping costs of high-throughput genotyping, has now become reality, revolutionizing animal breeding (Hayes et al., 2009a,b; Hayes & Goddard, 2010) and raising awareness in plant breeding (Bernardo & Yu, 2007; Resende et al., 2008; Grattapaglia et al., 2009; Heffner et al., 2009; Jannink et al., 2009). Three major insights characterize this ground-breaking proposition. First, instead of the standard two-step MAS approach where marker–trait associations are first discovered and validated in one or a few representative families or populations and then used for selection, GS estimates all marker effects simultaneously, retaining all of them as predictors of performance, thus precluding the prior search for significant marker–trait associations. Secondly, GS assumes that linkage disequilibrium (LD) provided by dense genotyping is sufficient to inform about QTL effects which are expected to be in LD with at least some of the queried markers. Thirdly, it avoids prior marker selection for the development of a prediction model so that estimated marker effects are sufficiently precise, unbiased and accurate to mitigate the quandary of how to capture the ‘missing heritability’ of complex traits (Manolio et al., 2009; Yang et al., 2010; Makowsky et al., 2011) likely explained by large numbers of small-effect QTLs that association genetics studies do not capture.
No report exists to date on the actual use of QTL or association genetics (AG) information for operational tree breeding by MAS, in spite of the large volume of QTL mapping information and an increased number of AG reports in forest trees (Grattapaglia et al., 2009; Neale & Kremer, 2011). As proposed early on, reasons for this derive mainly from the rapid decay of LD in undomesticated tree genomes, such that marker–trait associations detected in specific mapping families do not hold in unrelated pedigrees (Strauss et al., 1992). Initial predictions that QTLs detected in biparental tree pedigrees would only be useful within the same or related families (Grattapaglia et al., 1995) have been confirmed by several reports showing that many more QTLs with small and variable effects across backgrounds and environments typically underlie complex traits in forest trees (Dillen et al., 2009; Rae et al., 2008; Ukrainetz et al., 2008; Mamani et al., 2010). Although AG was presented as a way to develop more efficient methods of MAS in forest trees (Neale & Savolainen, 2004), experimental results based on candidate genes (Gonzalez-Martinez et al., 2007; Wegrzyn et al., 2010) have not captured satisfactory fractions of the genetic variance to be valuable to breeding. These results challenge the significance of such mapping efforts for breeding practice that deals with a broad diversity and must capture large proportions of valuable allelic combinations for multiple traits.
Genomic selection in forest trees and perennial crops is particularly attractive because of the potential of increasing gain per unit time and enhancing selection for low heritability traits (Wong & Bernardo, 2008; Jannink et al., 2010; Grattapaglia & Resende, 2011). We originally proposed GS in the context of forest tree genetics and breeding as a promising way to capture larger proportions of variation for growth traits (Resende et al., 2008; Grattapaglia et al., 2009). Using deterministic simulations we subsequently showed that the extent of marker–QTL LD, modeled by varying the effective population size (Ne) and genotyping density, had the largest impact on the accuracy of GS. By shortening the breeding cycle we also predicted that GS could radically improve the efficiency of forest tree breeding practice (Grattapaglia & Resende, 2011), a conclusion also reached by a subsequent study using stochastic simulations for a conifer breeding program (Iwata et al., 2011). While the limited extent of LD in natural populations of forest trees (Neale & Savolainen, 2004) would make GS economically unfeasible, marker–QTL LD is readily increased by limiting Ne, a standard practice in tree breeding. The generation of new LD as a result of the reduction in Ne is therefore a key element for the prospects of GS in tree breeding at currently economically viable genotyping densities of two to five markers per cM (Eckert et al., 2010; Sansaloni et al., 2010), although higher marker densities are becoming practical at competitive costs (Elshire et al., 2011).
In contrast to animal breeding, where little control over the effective population size is possible, tree breeders have ample autonomy to establish elite populations. Tree breeding populations with Ne between 20 and 50 support selection with appreciable genetic gains for several generations (Namkoong et al., 1988). Such elite breeding populations have increasingly been used in advanced programs worldwide (McKeand & Bridgwater, 1998; Resende & de Assis, 2008). In Eucalyptus, major breakthroughs in productivity and wood quality have been achieved by advanced hybrid breeding involving populations of 10–50 elite parents exploiting the wide interspecific variation for cold and drought tolerance, growth and wood quality coupled to clonal propagation (Grattapaglia & Kirst, 2008; Resende & de Assis, 2008). Such strategies are particularly suited to the application of GS. It is in the framework of such specialized Eucalyptus breeding programs that the operational application of GS is envisaged.
Experimental validations of GS in forest trees and perennial plants in general are critically needed to provide the necessary support for its recommendation in applied breeding. From the fundamental standpoint, it is also appealing to assess the power of this approach in capturing the genetic variance controlling quantitative traits while learning about the genome positions of the markers that optimize predictive accuracy of GS. In this work we extend our initial assessments of GS and describe a comprehensive proof-of-concept study with two large genetically unrelated Eucalyptus breeding populations with contrasting Ne. We demonstrate that GS not only captures large fractions of trait heritability but achieves selection accuracies as good as, or better than, those attainable by conventional phenotypic selection for all growth and wood quality traits evaluated. Finally, we show that in spite of a largely coincident genomic distribution of the loci controlling the same trait in the two populations, GS models have low predictive accuracies across populations, highlighting the fact that population- and environment-specific genome-enabled predictions will likely drive the application of GS in tree breeding.
Materials and Methods
Breeding populations and phenotypes
The study was carried out with two standard, genetically unrelated Eucalyptus elite breeding populations belonging to two Brazilian pulp and paper companies, CENIBRA (CEN) and FIBRIA (FIB) (Table 1). Height growth (HG) was measured by means of a Suunto PM-5 clinometer in CEN and a Haglof Vertex Hypsometer in FIB. Circumference at breast height (CBH; c. 130 cm from the ground) over bark was measured with a diameter tape. In CEN wood specific gravity (WSG) and pulp yield (PY) were measured on felled trees. WSG was measured by the water displacement method using a 3- to 5-cm-thick wood disk sampled at breast height while PY was measured by batch kraft digestion of 150 g of wood chips at 15–18% effective alkali (NaOH) to obtain kappa of c. 17 per sample. In FIB, WSG was measured on the standing tree using the pin penetration of a 6 J Pilodyn (Proceq, Schwerzenbach, Switzerland) at breast height. A 12-mm-diameter core was taken from bark to bark through the centre of each tree at breast height and in a north–south direction. Each core was air-dried, ground to wood meal and used to indirectly measure pulp yield by taking a near infrared reflectance spectroscopy (NIRS) reading using a Foss NirSystem 5000 and applying calibration curves developed by Fibria using methods described previously (Raymond & Schimleck, 2002).
Table 1. General attributes of the two breeding populations studied
CENIBRA population (CEN)
FIBRIA population (FIB)
Total no. of trees in progeny trial
Total no. of families in progeny trial
Average no. of trees per family
No. of parents crossed
Eucalyptus species of parents
E. grandis × E. urophylla F1 hybrids
E. grandis, E. urophylla and E. globulus and F1 hybrids of these species
No. of experiments
Sabinópolis 18°34′S/42°58′W Virginópolis 18°34′S/42°31′W Antonio Dias 19°22′S/42°47′W
Temperature range (°C)
Randomized complete block with 36 reps; single-tree plots; spacing 3 × 2 m
Alpha-lattice with 40 reps; single-tree plots; spacing 3 × 2 m
Age of measurements (yr)
3 for all traits
3 for growth traits 3.7 for wood traits
No. of families sampled for genomic selection (GS)
Effective population size (Ne)
No. of individuals/family sampled for GS
No. of individuals sampled for GS
DArT marker genotyping
DArT genotyping was carried out at DArT Pty (Yarralumla, Australia) using a 7680 DArT-probe microarray as described previously (Sansaloni et al., 2010). DArT genotyping and marker evaluation were accomplished based on: reproducibility ≥ 95% as measured by the concordance of the genotype call between the two DArT clone replicates on the array; marker quality Q-value ≥ 70% which measures between-cluster variance as a percentage of total variance in signal distribution among the genotyped samples; marker call rate ≥ 80%, that is, the percentage of effective scores.
Genetic value predictions from field experiments and adjusted phenotypes
where y is the vector of the trait being analyzed; β is a vector of fixed effects (i.e. general mean and experiments effects); a is a vector of random additive genetic effects of individuals; b is a vector of random incomplete block effects; ε is the random residual effect; and X, Z and W are incidence matrices for β, a and b, respectively. The variance structure of the model was as follows:
A, matrix of additive genetic relationships among individuals; I, identity matrix.
The mixed model equations for the best linear unbiased predictor (BLUP) were as follows:
: individual narrow sense heritability across incomplete blocks.
: coefficient of determination of incomplete block effects.
Variance components were estimated by restricted maximum likelihood (REML) (Lynch & Walsh, 1998). Predicted additive genetic values were deregressed and corrected for parents’ effects to obtain the adjusted phenotypic values to be used for genomic predictions. The deregressing approach employed (Garrick et al., 2009) corrects the raw phenotypic data for environmental and parent average effects removing familial structure. Other nongenetic sources of variation were corrected by fitting environmental fixed and random effects in the mixed model.
Estimation of molecular marker effects and genome-enabled capture of heritability
The markers had their effects estimated adjusting all the allelic effects simultaneously using the random regression best linear unbiased predictor (RR-BLUP) (Meuwissen et al., 2001) as described previously (Resende et al., 2008, 2012). Differently from that previous study, however, dominant markers were used so that the values of Zij in the incidence matrix Z are 0 or 1 for marker genotypes mm and MM or Mm, respectively. This prediction equation assumes a priori that all loci explain an equal amount of the genetic variation, as previously described by (Meuwissen et al., 2001) and applied by others (Bernardo & Yu, 2007; Muir, 2007). Therefore, the genetic variation of each locus is given by where η is related to the number nm of markers used in the prediction model and is given by (Resende et al., 2008; Gianola et al., 2009). The total additive genetic variance was estimated by restricted maximum likelihood (REML) based on phenotypes. The recovered additive genetic variance and the corresponding trait heritability captured by the markers were estimated based on the proportion of the genetic variation explained by the markers using an approach described previously (Gianola et al., 2009; Resende et al., 2010). To assess the increase in captured heritability with increasing number of markers, this analysis was carried out for progressively larger sets of markers ranked by their absolute effect. To fit RR-BLUP, the genomics module of the software SELEGEN was employed (Resende & Oliveira, 1997; Resende et al., 2008) and also a script written in R (M.F.R. Resende Jr. unpublished).
Cross-validation and accuracy of genomic selection
The estimated markers effects were validated using a ‘leave-one-out’ cross-validation scheme. Briefly, a single individual from the population was used as the validation set, and the remaining individuals as the estimation or training set. This was repeated such that each individual in the sample was used once as the validation data. This process was repeated N times, using each time a different set of individuals for estimation and one different individual for validation until all individuals had their phenotypes predicted and validated. This method maximized the estimation and validation population sizes so that eventually the training population had (N − 1) individuals in each set and the validation population consisted of the whole set of N individuals genotyped and phenotyped.
Each individual tree had its GEBVs predicted by multiplying the incidence matrix Z for the marker by the vector of estimated marker effects and summing the estimated general mean according to the expression:
The accuracy of GS to predict breeding values was calculated by the correlation of the GEBV with the adjusted phenotypes (deregressed predicted additive genetic values corrected for parents effects) (y) and then dividing it by the square root of trait heritability. Theoretically the accuracy of GS depends on , the proportion of the genetic variation explained by the markers (degree of LD), which is a function of Ne and the chromosome segment length (cM) between markers as given generally by Sved’s equation (Sved, 1971); and , the accuracy of the prediction of the marker effects in LD with the QTL. The analyses were carried out using the information of all the markers that provided the maximum accuracy. The expected gain from genomic selection was compared with conventional phenotypic selection considering different reductions in the breeding cycle duration as a result of early GS as described previously (Grattapaglia & Resende, 2011).
Genome predictions across populations and genome positioning of GS fitted markers
Two comparative analyses of the GS results were carried out between the breeding populations. First, prediction accuracies of GS across populations were obtained for the four traits. In essence, we simply applied a predictive model developed for a particular trait in CEN to predict the same trait in individuals belonging to FIB and vice versa. Secondly, we compared the genome positions of the markers that were fitted into the genomic prediction models in the two populations for each trait separately. The corresponding DArT marker sequences (GenBank accession numbers HR865291–HR872186) were mapped on to the assembled Eucalyptus grandis genome (version 6.1 available in Phytozome) after partitioning it in 500 kb virtual bins corresponding to c. 1 cM based on a simple correspondence between the size of the assembled Eucalyptus genome (609 Mbp) and the best estimates of the size of the recombining Eucalyptus genome, estimated between 1200 and 1300 cM (Brondani et al., 2006).
DArT markers provide suitable genotyping density and coverage for GS
Given the large samples sizes of the populations, DArT markers with a minimum allele frequency (MAF) ≥ 0.01 corresponding to a gene diversity (GD) ≥ 0.02 were used in all subsequent analyses to allow the capture of rarer alleles at relevant loci. The total number of high quality and polymorphic markers scored was slightly higher for FIB (3564) than for CEN (3129), reflecting the most diverse genetic composition of the former. The distributions of percentage markers across the filtering criteria were similar between the two populations, with ≥ 80% of the markers with reproducibility ≥ 99%, call rate ≥ 90% and GD ≥ 0.25 in both populations (Supporting Information, Fig. S1). Over 87% of the genotyped markers (2720 markers for CEN and 3145 for FIB) could be physically mapped with high confidence on to the 11 main scaffolds of the current Eucalyptus genome assembly (608.5 Mbp) while the remaining markers (c. 13%) mapped to the additional unassembled scaffolds. A realized average density of one marker every 193–223 kb was achieved. At the physical level, using a conservative analysis based on 500 kbp bins that correspond, on average, to 1 cM recombination, 912 of the 1217 bins, that is, 75% of the genome, was tagged by the DArT markers (Fig. S2). These results indicate that in spite of the fact that the 7680 DArT probes in the array were selected based exclusively on polymorphism detection and high-quality signal-to-noise ratio (Sansaloni et al., 2010), they sample the genome in a relatively well distributed fashion, thus providing a good tool to carry out an experiment that requires a whole-genome analysis. From the recombination standpoint, assuming an average recombining genome size for Eucalyptus between 1200 and 1300 cM (Brondani et al., 2006), an expected average genotyping density ≥ two markers per cM was achieved in both populations. Such a density corresponds to an average intermarker distance equal to 0.5 cM, a value used to estimate the proportion of the genetic variation explained by the markers (see the Materials and Methods section).
Capture of trait heritability by genome-wide markers
For each trait, the heritability estimated from the phenotypes and pedigree data was regarded as the upper limit that could be explained by the GS model. Markers were ranked by the largest to the smallest effect in absolute value and used to estimate the captured fractions of the trait heritability (h2). In both populations for most traits, ≥ 80% of h2 could be captured with 200 markers of largest effect (Fig. 1). The increase in h2 captured was similar for CBH, HG and WSG, while for PY a plateau was reached earlier. For CBH, HG and WSG, the increase in the h2 captured was slower after 50 markers, reaching a plateau at 95–97% of captured h2 in CEN and 70–82% in FIB with 300–500 markers of largest effect, consistent with the smaller effective population size of CEN. For PY no increase in captured h2 was observed after considering 25 markers in CEN and 50 markers in FIB, and a slight drop was seen in FIB possibly as a result of overfitting for this trait for which a smaller training set was used to estimate marker effects.
Predictive abilities and accuracies of GS
Moderate to high prediction accuracies of GS were estimated from cross-validation for all four traits in both populations (Table 2). Higher estimates were obtained in CEN as expected, as a result of its smaller Ne that results in a more extensive LD. Accuracies of prediction models ranged between 0.74 and 0.88 in CEN, and 0.55 and 0.73 in FIB, depending on the trait. More markers were needed to maximize accuracy for CBH and HG in both populations consistent with the expectation of more loci controlling these traits when compared with WSG and PY. The number of markers that maximized prediction accuracy ranged from 564 for PY to 1543 for CBH in FIB. Genomic predictions were generally unbiased as revealed by regression coefficients close to one, indicating that the variability between predicted and true genetic values was similar. In both populations, PY was measured in a smaller set of individuals because of the higher cost of measuring this trait. The smaller training and validation set for this trait possibly contributed to a smaller number of markers in the model when compared with the other traits. However, while in FIB this fact might explain the smaller than expected accuracy, in CEN a higher accuracy was observed. This discrepancy could arise as a result of the fact that PY in CEN was measured by batch kraft digestion, a higher-precision method than NIRS.
Table 2. Experimental results of genomic selection (GS) for growth and wood quality traits in two Eucalyptus breeding populations (CENIBRA and FIBRIA)
CENIBRA (Ne = 11)
FIBRIA (Ne = 51)
Ne, effective population size; CBH, circumference at breast height; HG, height growth; WSG, wood specific gravity; PY, pulp yield.
1Cross-validation was carried out by a ‘leave-one-out’ jackknife procedure so that the size of training population was N − 1 while the size of the validation population was N.
2Correlation between the observed and predicted breeding values obtained by cross-validation (Fig. S3).
3Calculated as the ratio between the predictive ability and the square root of trait heritability (Dekkers, 2007).
Trait heritability (estimated from field experiments)
Number of individuals N in the training (N − 1) and validation (N) populations1
Number of markers that maximized accuracy of the GS model
Accuracy of GS (experimental)3
Accuracy of GS (simulations)4
Regression coefficient of observed on predicted phenotypes
Accuracy of conventional phenotypic BLUP selection5
Proportion of the accuracy of phenotypic BLUP selection recovered by GS
Estimated number of QTL controlling trait variation6
Accuracies of GS estimated from this proof-of-concept experiment were consistent with those expected from simulations, although slightly larger for most traits. The recent hybrid origin of both populations is probably contributing to a longer-range LD than the one expected from theory, resulting in an improved ability of capturing genetic variation given a fixed genotyping density. Accuracies of GS matched or surpassed those provided by phenotypic BLUP selection in CEN and recovered between 75 and 90% of such accuracies in FIB. The estimated number of QTLs controlling the trait variation ranged between 142 and 263 for CEN and between 97 and 276 for FIB. These numbers were consistent with the theoretical expectations given the effective population size and size of the recombining genome in Morgans (Hayes & Goddard, 2001), although in the case of PY it is probably underestimated as a result of the smaller training population.
Increase of selection efficiency by genomic-enabled predictions
The value of GS over phenotypic selection materializes when the breeding cycle is reduced by performing early selection for yet-to-be observed phenotypes. The expected increase of selection efficiency of GS in the two experimental populations was modeled for increasing reductions in the time necessary to complete a breeding cycle (Fig. 2). These increased efficiencies from experimental data matched those predicted from earlier simulations (Grattapaglia & Resende, 2011). The intersection at zero values on the X and Y axes correspond to the efficiency of traditional BLUP-based phenotypic selection. Positive values of Y correspond to gains in efficiency of selection by adopting GS. If a modest reduction of only 50% in the length of a breeding cycle is achieved, efficiency gains between 50% for FIB and 100% for CEN are expected for all traits. If reductions of 75% are biologically feasible, a remarkable increase of 300% for CEN and 200% for FIB for all traits could be achieved.
Suitability of GS prediction models across populations
A pertinent question is whether a GS model fitted to one population is suitable to predict phenotypes in an unrelated population. When a model developed for CEN was used to predict phenotypes in FIB or vice versa, the accuracies declined drastically (Table 3). The percentage of markers that could be used to carry out these predictions varied between 60 and 79%. For example, out of the 1177 DArT markers that maximized the accuracy of the GS model for HG in CEN, only 899 (76%) were also genotyped in FIB, the remaining were either not polymorphic or did not pass the quality filtering criteria. Therefore the prediction in FIB was carried out using these 899 markers only. However, irrespective of the proportion of markers shared, accuracies were close to zero, suggesting that this was not an important element to explain the reduction in accuracies. Different size effects associated with the markers, variable patterns of LD and genotype × environment interaction are the likely factors explaining the reduction (see the Discussion section).
Table 3. Prediction accuracies of genomic selection (GS) across the two breeding populations: CENIBRA (CEN) and FIBRIA (FIB)
GS model estimated in CEN predicting in FIB
GS model estimated in FIB predicting in CEN
CBH, circumference at breast height; HG, height growth; WSG, wood specific gravity; PY, pulp yield.
No. of DArT markers used in the GS model developed in the estimation population
No. of DArT markers of the GS model of the estimation population also genotyped in the prediction population
% DArT markers of the GS model that could be used for prediction
Accuracy of GS
Comparative genomic distribution of markers fitted into the GS models
A comparative analysis of the physical distribution of the markers fitted in the GS models in the two populations was carried out based on the presence or absence of at least one significant marker–trait effect in the bin (Fig. 3). The reasons for using this more conservative approach instead of the total number of significant markers per bin is the relatively long-range LD in these populations, so that when several trait-associated markers are found in the same 500 kb bin they are likely linked to the same effect. Furthermore, the number of markers per bin varies with the sequence polymorphism of that particular genome segment so that a test based on the total number of significant markers per bin would be biased. All tests at the whole-genome level (Pearson chi-squared) were highly significant (P <0.001); at the individual chromosome level, all 11 Fisher exact tests were significant for CBH, 10 for HG, eight for WSG and eight for PY (P <0.01). These results indicate that loci underlying trait variation in CEN and FIB overlap significantly more frequently than expected as a result of chance alone for all four traits. In other words, the distribution of the genomic segment maximizing selection accuracy was highly coincident between the two populations in spite of the low suitability of the prediction models between populations.
We have presented the first experimental results of genome-enabled prediction accuracies for growth and wood quality traits in Eucalyptus and among the first ones in trees in general. More importantly, however, these results were generated for two genuine tree breeding program settings with regard to the effective population size, genetic composition, sizes of training and validation populations and phenotyping procedures, providing real-life data on the prospects of GS for accelerating selection for complex traits in forest trees. Accuracies of GS varied between 0.55 and 0.88 depending on the trait and effective population size, closely matching the accuracies achieved by conventional phenotypic selection and corroborating the encouraging results of our earlier simulations. Substantial proportions (74–97%) of the trait heritability could be captured by fitting all of the genome-wide markers simultaneously, confirming that the GS approach brings a new perspective to the understanding of quantitative trait variation in forest trees and a revolutionary tool for applied tree breeding.
Genome-wide approaches capture large fractions of trait heritabilities
Trait heritabilities estimated by the resemblance between relatives in the two populations fell within the ranges typically estimated for these traits in Eucalyptus (Apiolaza et al., 2005; Rosado et al., 2009). Large proportions (74–97%) of these heritabilities were captured in both populations for all four traits, leaving a relatively small fraction of the estimated heritability unexplained or ‘missing’ (Fig. 1). These results are consistent with the expectations of modeling all genome-wide markers concurrently instead of the one-marker-at-a-time hypothesis testing of association genetics (Meuwissen et al., 2001). Such a whole-genome prediction approach has been recently shown to successfully capture similarly large fractions of the heritability for human height (Yang et al., 2010; Makowsky et al., 2011). Although our estimates of captured heritability in tree growth closely match those obtained for human height growth, a key difference exists between the experiments. The human study dealt with a much larger effective population size which evidently demanded a significantly higher genotyping density with over 40 000 single nucleotide polymorphisms (SNPs) needed to capture 90% of the heritability. In our populations the effective population sizes were much smaller, consistent with the sizes employed in advanced breeding of forest trees. With a much more extensive LD, only a few hundred markers were necessary to capture similar fractions of the heritability for height growth. Notwithstanding all other differences in the biology of the organisms involved, these results confirm that a genome-wide approach with genotyping density calibrated to the degree of LD in the population effectively explains a large fraction of the genetic variation of complex traits.
Recently, Hamblin et al. (2011) discussed the differences in population genetics properties, trait architecture and mode of reproduction that led to more successful AG results in crops than in humans. In trees, however, the recent domestication and large effective population size of natural populations used in AG experiments more closely resemble humans, making the ‘missing heritability’ dilemma fully relevant to forest tree populations in natural conditions. Experimental results have confirmed this expectation, with very small proportions of genetic variance explained by the associations found in forest trees (Gonzalez-Martinez et al., 2007; Thumma et al., 2009; Wegrzyn et al., 2010). Our results demonstrate, however, that the genome-wide approach of GS, applied to tree breeding populations that have undergone some amount of domestication bottleneck and directional selection, can capture large fractions of the genetic variation underlying quantitative traits. Interestingly, our results also show that, given the extensive LD of these populations, c. 200 markers of largest effect already capture over 80% of the heritability, although the individual effects of these markers hardly surpass 1%, and thus are comparable to the size effects found for SNPs in AG experiments of forest trees. Nevertheless, while in AG markers have been selected a priori in candidate genes aiming at allele mining in low-LD populations, in GS markers are selected a posteriori from a large number of genome-wide markers in high-LD breeding populations, therefore allowing a better opportunity to capture relevant alleles at multiple loci in disequilibria along the genome. Although GS makes no progress in discovering genes or functional polymorphisms underlying complex traits, larger effect markers fitted in the GS predictive models may provide unbiased genomic positions wherefrom allele mining efforts could be proposed.
Accuracy of GS matches that of conventional BLUP phenotypic selection
Confirming earlier simulation-based predictions for forest tree breeding scenarios (Grattapaglia & Resende, 2011; Iwata et al., 2011), the experimental accuracies estimated in this study confirm that genome-enabled predictions have true potential to radically improve the efficiency and speed of tree breeding (Table 2). The accuracies of our GS models matched or even exceeded those calculated for conventional pedigree-based phenotypic selection for all traits in both populations, with a slight advantage when the effective population size was smaller. The reduction of effective population size together with directional selection not only extended LD but possibly contributed by shifting average allele frequencies toward intermediate values, eliminating rare alleles so that QTLs tended to segregate at frequencies more similar to the frequency of marker alleles used for prediction (Hamblin et al., 2011). Given the clear, although modest, DArT marker divergence between Eucalyptus species (Steane et al., 2011), the recent hybridization that occurred in both breeding populations might also have contributed to an increased extent of LD with a positive impact on the estimated accuracies of GS.
Three elements should be highlighted regarding these estimates that make them useful benchmarks for assessing the prospects of GS in Eucalyptus and tree breeding in general. First, the use of deregressed phenotypes makes these GS models free of familial relatedness (Garrick et al., 2009), capturing only the relevant marker–trait LD in a way that accuracies should hold in independent samples and future generations with adequate model updating. Secondly, the leave-one-out validation maximized the training and validation population sizes with no cross-contamination of estimates, providing the best possible sampling of allelic effects (given the afforded training/validation set) and thus optimizing the power to estimate genomic breeding values. Thirdly, our experimental assessment of GS was carried out for several traits with variable heritability in two different populations and the accuracies were always close to those anticipated by simulations. Furthermore, our results are in line with those obtained for growth traits in loblolly pine (Resende et al., 2012) and the growing body of experimental reports of GS in annual crops (Crossa et al., 2010; Albrecht et al., 2011). Finally, our GS results, similar to the few others in plants, have been better than those reported for domestic animals (Hayes et al., 2009a,b; Moser et al., 2010). Reasons for this include not only those that have made AG results in plants more rewarding than in humans (Hamblin et al., 2011), but also the opportunity of controlling the Ne of the target population, together with the possibility of assembling large training populations and phenotyping them at high precision.
GS will likely require population-specific predictive models
The use of GS models across populations is an important issue for the operational use of GS by a program that sustains forest plantation with distinct breeding populations tailored to different environments. In plants, no study to date has looked into this issue across genetically unrelated populations. In domestic animals, the few studies to date showed that a GS model fitted to one population cannot predict phenotypes in an unrelated population unless genotyping density is increased (Hayes et al., 2009a,b). In our study we show that the predictive models had no appreciable accuracy for any trait when going from CEN to FIB or vice versa (Table 3). Our results indicate that GS prediction models will likely be population-specific. Multipopulation GS models might be feasible with increased genotyping density so that marker–QTL linkage phase would persist across populations. However, the genotype × environment interaction might supersede the persistence of LD relationships and eventually cause equally unacceptable accuracies. In fact, a recent GS study of a cloned loblolly pine population planted in different sites showed that genotype × environment interaction may severely affect the transferability of GS models across breeding zones (Resende et al., 2012). A safer and more accurate GS applicable across a range of environments is likely to come from fitting models taking into account the genotype × environment interaction effects.
Genomic regions underlying trait variation significantly overlap between populations
Genomic selection presupposes that phases of LD between markers and QTLs are the same in the selection candidates and the training population. Poor performance of GS across populations is therefore consistent with the genetics of outbred forest trees in linkage equilibrium. The availability of a Eucalyptus genome sequence allowed us to compare these results with the genome-wide positioning of the GS fitted markers in each population. This contrast revealed an interesting result: while GS accuracies across populations were poor, a highly significant level of coincidence across populations was observed regarding the location of the relevant genomic segments underlying complex traits (Fig. 3). These apparently incompatible results are probably expected, suggesting that although there is a significant coincidence of the genome position of the loci that explain trait variation, the allelic effects vary across populations, making predictions inaccurate. This variation has two major components to it that make phenotype predictions ineffective: when markers capture the same loci across populations they probably tag different alleles as a result of inconsistent coupling–repulsion relationships between marker and QTL alleles; when the same alleles are tagged by the same markers, that is, the linkage phase is consistent across populations, the relative allelic effect varies as a result of the surrounding genetic background or of genotype × environment interaction. Our experimental results at the genome-wide level essentially corroborate the theoretical predictions of poor transferability of marker–QTL associations across populations of forest trees (Strauss et al., 1992).
Breeding by GS in tropical Eucalyptus
We had previously outlined the opportunities and challenges of GS in forest trees based on deterministic simulations (Grattapaglia & Resende, 2011). The experimental results presented here substantiate those views and confirm that the assumptions underlying those simulations, that is, a strict infinitesimal model and build-up of LD following Sved’s equation (Sved, 1971), were generally valid. Gains expected from GS in these Eucalyptus populations will largely derive from the ability to shorten the breeding cycle by ultra-early selection for yet-to-be observed phenotypes at the seedling stage. Gains in selection efficiency should reach 50% if the breeding cycle time is halved and can exceed 100–200% if more aggressive tactics are taken (Fig. 2). GS in current tropical Eucalyptus breeding would reduce breeding cycle by at least 50% by eliminating the progeny trial and the primary clonal trial where a large number of trees selected only for growth traits are typically tested as clones. The net effect is increasing selection intensity not only for growth but for all wood properties simultaneously at the progeny level and anticipating the deployment of elite clonal material by 9 yr (Fig. 4). Using conservative calculations, for a progeny trial of 20 000 seedlings genotyped at $50–100 each, assuming that the GS selected clones would provide an increase of 1% in pulp yield 9 yr earlier than expected by conventional breeding, an economic return of at least 20 times on the investment made in GS is expected for a 500 000 ton yr−1 pulp mill operation.
Breeding for multiple traits simultaneously by GS should be no more complicated than the process of multiple-trait selection in conventional breeding. GEBV for all traits would be aggregated into a single selection index and different multiple-trait selection methods applied on the genomic data. The number of fitted markers per trait also indicate that the use of low-density marker panels could be a possible alternative to reduce genotyping costs (Habier et al., 2009). Current trends in genotyping-by-sequencing technologies (Elshire et al., 2011), however, point to the possibility of increasing marker densities at the same cost per sample so that considerations about single data-point costs might no longer be an issue. Higher marker densities will also mitigate the decay of marker–trait LD as a result of recombination, providing satisfactory accuracies even when the training and selection populations are several generations apart (Meuwissen & Goddard, 2010) or when multipopulation predictive models are pursued (Hayes et al., 2009a,b).
In conclusion, our experimental genomic-enabled predictions in two Eucalyptus breeding populations, together with those recently reported for loblolly pine (Resende et al., 2012), are very promising and should encourage additional assessments of GS in forest trees. We reiterate, however, our cautiously optimistic outlook about GS, given that some issues remain to be examined such as the accuracy of GS on selection candidates several generations ahead of the training set and the possibility of reduced long-term genetic gains as a result of a potentially faster build-up of inbreeding with GS when compared with conventional phenotypic selection (Grattapaglia & Resende, 2011). We predict, nevertheless, that reports on realized accuracies of operational GS in Eucalyptus breeding will likely follow in the next few years. While most technical aspects of GS in tree improvement will soon be covered and genotyping costs will become increasingly affordable, its adoption by breeders will depend to a large extent on a detailed financial evaluation of the achievable gains of deploying improved clones or seeds several years earlier, against the costs of genotyping and data handling of this new breeding technology. As with all innovations, this will only happen if one is willing to consider a new attitude, take some degree of risk and recognize that genomics may finally make its way into applied tree breeding.
This work was supported by CNPq grant 577047/2008-6, PRONEX-FAP-DF grant ‘NEXTREE’ 2009/00106-8 and EMBRAPA Macroprogram 2 grant 02.07.01.004. C. P. S., C. D. P. and M. F. R. R. Jr were sponsored by graduate fellowships from CAPES, and D. A. F. was sponsored by a postdoctoral fellowship from CNPq; M. D. V. R., G. J. P. Jr and D. G. have been awarded research fellowships from CNPq. We acknowledge the team at DArT Pty for their outstanding support to the genotyping work developed by C. P. S. and C. P. D. while there as part of their graduate training.