Journal list menu

Volume 60, Issue 3 p. 1499-1512
ORIGINAL RESEARCH ARTICLE
Open Access

Recurrent genomic selection for wheat grain fructans

Lynn D. Veenstra

Lynn D. Veenstra

Plant Breeding and Genetics Section, School of Integrative Plant Science, Cornell Univ., 240 Emerson Hall, Ithaca, NY, 14853 USA

Search for more papers by this author
Jesse Poland

Jesse Poland

Plant Pathology, Kansas State Univ., Throckmorton Hall, Manhattan, KS, 66506 USA

Search for more papers by this author
Jean-Luc Jannink

Jean-Luc Jannink

USDA-ARS, R.W. Holley Center for Agriculture and Health, Cornell Univ., Ithaca, NY, 14853 USA

Search for more papers by this author
Mark E. Sorrells

Corresponding Author

Mark E. Sorrells

Plant Breeding and Genetics Section, School of Integrative Plant Science, Cornell Univ., 240 Emerson Hall, Ithaca, NY, 14853 USA

Correspondence

Mark E. Sorrells, Plant Breeding and Genetics Section, School of Integrative Plant Science, Cornell Univ., 240 Emerson Hall, Ithaca, NY 14853, USA.

Email: [email protected]

Search for more papers by this author
First published: 16 February 2020
Citations: 14

Assigned to Associate Editor Jason Gillman.

Abstract

Fructans are carbohydrates found in many plants, including wheat (Triticum aestivum L.), and they serve physiological roles in both plants and humans. Genomic selection (GS) could facilitate the rapid development of climate-resilient, nutritionally improved wheat cultivars, such as high-fructan cultivars, while decreasing resource-intensive phenotyping requirements. However, few empirical studies have examined GS for nutritional quality breeding. Although GS can accelerate gain from selection, loss of genetic variation and inbreeding may limit the potential for long-term gain. The objectives of this study were (a) to determine realized gain from GS for wheat grain fructan content with simple truncated selection (TS) and optimized contribution selection (OCS) methods, (b) to determine if gains agree with theoretical expectations, and (c) to compare impacts of selection on inbreeding, genetic variance, and indirect selection on agronomic characteristics. Over 2 yr, two cycles of GS were performed with equal contribution TS and inbreeding-constrained OCS selection. Genomic selection with TS and OCS led to a 25 ± 12% and 34 ± 6.4% increase in wheat grain fructan content, respectively. Although positive gains from selection were observed for both populations, OCS populations exhibited these gains while simultaneously retaining greater genetic variance and lower inbreeding levels relative to TS populations. Selection for wheat grain fructan content did not change plant height but significantly decreased days to heading in OCS populations. In this study, GS effectively improved the nutritional quality of wheat, and OCS controlled the rate of inbreeding.

Abbreviations

  • BLUP
  • best linear unbiased prediction
  • C0
  • Cycle 0
  • C1
  • Cycle 1
  • C2
  • Cycle 2
  • GBLUP
  • genomic best linear unbiased prediction
  • GBS
  • genotyping-by-sequencing
  • GEBV
  • genomic estimated breeding value
  • GS
  • genomic selection
  • G × E
  • genotype × environment
  • LOOCV
  • leave-one-out cross-validation
  • OCS
  • optimum contribution selection
  • RGS
  • recurrent genomic selection
  • SNP
  • single nucleotide polymorphism
  • TP
  • training population
  • TS
  • simple truncated selection
  • 1 INTRODUCTION

    Fructans are naturally occurring plant polymers composed of fructose molecules found in ∼15% of flowering plant species, including wheat (Triticum aestivum L.). Fructans confer stress tolerance by acting as osmoregulators in times of cold, drought, and excess salinity (as reviewed in Hendry, 1993; Joudi et al., 2012; Kafi, Stewart, & Borland, 2003; Pilon-Smits et al., 1995). Fructan consumption, particularly inulin-type fructans, stimulates growth of healthy gut bacteria and supports overall gut health in hosts (as reviewed in Di Bartolomeo, Startek, & Van den Ende, 2013; Peshev & Van Den Ende, 2014; Roberfroid et al., 2010). The average daily fructan consumption in the United States is estimated to be 1.0–4.0 g, with the predominant source being wheat-based products (Van Loo, Coussement, de Leenheer, Hoebregs, & Smits, 1995). Fructan content within wheat grains averages between 1.28 and 1.40 g 100 g−1 (Andersson et al., 2013; Fretzdorff & Welge, 2003) with known variation between 0.7 and 2.9 g 100 g−1 (Huynh, Palmer, & Mather, 2008). The important physiological roles of fructans and known variation allow them to be a desirable breeding target for developing nutritionally improved, climate-resilient wheat cultivars (Veenstra, Jannink, & Sorrells, 2017).

    The development of nutritionally improved wheat cultivars often requires extensive resources. High-throughput phenotyping methods for measuring wheat grain fructan content have been developed (Li et al., 2017), but the cost and time requirements associated with phenotyping remain high. The use of genomic selection (GS) in breeding for wheat grain fructan content would shorten the selection cycle and reduce phenotyping requirements. Genomic selection (as reviewed in Heffner, Sorrells, & Jannink, 2009; Lorenz, Chao, & Asoro, 2011) uses phenotype and genotype data from a relevant population to train a statistical model and predict breeding values of genotyped selection candidates based on markers alone. Selecting candidates based on predicted breeding values allows for crossing of selected individuals before phenotyping. For crops with long growth cycles, such as winter wheat, young plants can be genotyped and selected for crossing with GS prior to flowering.

    Although GS can accelerate the breeding cycle, conservation of genetic variance and management of inbreeding levels are necessary for optimizing long-term gain from selection (Sonesson & Meuwissen, 2000). High rates of inbreeding per breeding cycle have been observed with GS in simulations (Lin et al., 2016) and empirical studies (Rutkoski et al., 2015). Several methods to control the rate of inbreeding have been proposed, including optimum contribution selection (OCS; Meuwissen, 1997). Rather than a simple truncation selection with equal contribution of all parents, which will include the individuals with highest genomic estimated breeding values (GEBVs), but also mating of very closely related individuals, OCS is an approach to balance the contribution and mating of breeding parents for the next cycle to minimize genetic relatedness and inbreeding among the parents and subsequent progeny. Optimum contribution selection methodologies were originally proposed as a way to maximize genetic gain while controlling the rates of inbreeding in animal breeding. Although the use of OCS methodologies may result in a small reduction of short-term genetic gain due to including lower merit individuals, simulations suggest that long-term gains are comparable with truncation selection (Henryon, Ostersen, Ask, Sørensen, & Berg, 2015).

    Significant genotype × environment interaction (G × E) for wheat grain fructan content has been observed (Haskå, Nyman, & Andersson, 2008; Veenstra, Santantonio, Jannink, & Sorrells, 2018), but not consistently (Huynh et al., 2008). Although significant G × E has been observed, the impacts of G × E on prediction accuracies for GS were found to be minimal, with observed heritability ranging from .52 to .59 (Veenstra et al., 2018).

    The empirical application of GS for nutritional traits in crops has not been widely examined. Furthermore, no empirical studies have used OCS within the GS framework in plants. Applying GS for nutritional quality traits has great potential in many crops; however, the rapid rates of inbreeding and loss of genetic variance may limit long-term gain from selection. A recurrent genomic selection (RGS) scheme for increasing grain fructan content in winter wheat was implemented to evaluate the feasibility of using GS for nutritional quality breeding. The objectives of this study were (a) to determine realized gain from GS for wheat grain fructan content with two selection methods, (b) to determine if gains agree with theoretical expectations, and (c) to compare impacts of selection on inbreeding, genetic variance, and correlated changes in agronomic characteristics.

    2 MATERIALS AND METHODS

    2.1 Germplasm

    All germplasm used in this study was obtained from the Cornell Small Grains Master Nursery collection. The Master collection is an association mapping population consisting of 1,290 elite, F5-derived advanced soft winter wheat breeding lines and cultivars from 2007–2012.

    2.2 Training population

    The population used for GS model training consisted of 284 wheat lines. Within the 284 lines, 14 lines were breeding population founders, and the remaining 270 lines were Master Nursery collection lines selected with the mean coefficient of determination methodology (Rincent et al., 2012) to serve as a representative sample of the genetic diversity in the larger collection.

    2.3 Validation population

    The Cycle 0 population, C0, was founded with 14 elite lines selected from the Master collection based on diversity and seed availability. To generate the C0 population, the founders were randomly intermated for two generations by hand pollination to produce F1 and double-cross F1 progenies. All F1 and double-cross F1 progenies were confirmed with simple sequence repeat genotyping. The double-cross F1 families were subsequently self-pollinated to increase seed, thereby resulting in double-cross F2 seed representing 64 families. Ten double-cross F2 individuals from each family were sampled to create an initial C0 population of 640 individuals. To replicate the selection schemes, the 640 individuals were randomly split into two C0 populations of 320 individuals each. The initial C0 population was split to facilitate independent replication of selection schemes.

    2.4 Breeding scheme

    Two GS methods were used for increasing wheat grain fructan content. The first selection method, simple truncated selection (TS), determined the GEBVs for all selection candidates with the top 12 individuals selected for intermating. The second method, constrained selection, optimized the contribution of selected individuals in the next generation by applying OCS to maximize genetic gain and control inbreeding in the following generation. Simple truncated selection assumes equal contributions of all selected individuals to future progeny, whereas OCS selection determines the optimal contribution of each selection candidate to the progeny of the next generation to control inbreeding.

    Each C0 population underwent two cycles of GS with TS or OCS, thereby creating four selection populations (Figure 1). For the first GS cycle, C0 individuals were genotyped, and their GEBVs were predicted with the training population (TP) of 284 lines from the Master population. The best 12 C0 individuals were selected using TS and intermated to produce progenies with approximately equal parental contribution. For OCS, the selected individuals were intermated to produce progeny containing optimal contributions of selected parental lines.

    Details are in the caption following the image
    Genomic selection (GS) schemes. C0 is the Cycle 0 population; C1 is the Cycle 1 population; C2 is the population of intermated S1 families obtained from crossing F2 seed from the Cycle 1 population; GH is greenhouse; OCS is GS with previously imposed constraints on inbreeding; TS is GS with no inbreeding constraints

    After C0, the genotypes and phenotypes of parental material for each population were added to the TP. For Cycle 2 (C2), candidates in the C1 populations were selected with the same methodology as the C0 selections, and S1 seed of individuals from C1 were intermated to create C2 populations consisting of S1 families.

    Selection intensities across cycles ranged between 0.943 and 2.002 for the four programs with 12–22 individuals selected (Table 1). Selection intensities were calculated per Equation 1, with GEBVs originating from estimates based on the initial TPs at the time of selection. All genotyped material was selected directly, and crossing of selected individuals was performed in the greenhouse.
    i = μ s μ σ where μ s = Σ c j e GEB V j (1)
    where i is selection intensity, μs is the mean GEBV of selected individuals, μ is the mean GEBV of selection population, σ is the standard deviation of fructan content for selection population, c j e is the realized contribution of individual j, GEBVj is the best linear unbiased prediction (BLUP) estimated breeding values of individual j, and j is the selected individual (1, …, n).
    TABLE 1. Selection intensities and number of individuals planted and selected
    Populationa Replicate No. of candidates No. selected ib
    C0 TS 1 320 12 2.022
      2 320 12 1.899
    C0 OCS 1 320 17 1.851
      2 320 15 1.781
    C1 TS 1 46 12 1.425
      2 46 12 1.169
    C1 OCS 1 45 21 1.344
      2 47 22 0.943
    • a C0, Cycle 0; C1, Cycle 1; C2, Cycle 2; TS, simple truncated selection; OCS, optimized contribution selection.
    • b i, selection intensity.

    2.5 Genotypic data

    Genotypic data for all populations, with the exception of C2 populations, were generated with genotyping-by-sequencing (GBS; Elshire et al., 2011) according to the protocol described by Poland et al. (2012). The GBS data for the TP and breeding populations (C0, C1) were processed through different versions of single nucleotide polymorphism (SNP) calling pipelines with UNEAK (Lu et al., 2013). To assimilate the GBS markers from separate pipelines, markers with matching SNP tags and variable positions within tags were retained from both genotypic datasets. The populations were filtered separately (minor allele frequency > 5%, marker presence > 50%, individual presence > 20%, heterozygosity < 20% for TP) and expectation maximization imputation was performed in the respective populations with the A.mat function in the R package rrBLUP (Endelman, 2011). After merging the processed SNP calls, the resulting 8,903 markers were used for selections.

    For the GBS SNPs aligned to the IWGSC RefSeq v1.0 reference genome, the resulting SNPs for all RGS materials were filtered for marker presence and individual coverage by markers (96-plex). After applying filters to retain markers with presence over 30%, a minor allele frequency above 1%, and heterozygosity under 10%, a total of 12,599 SNPs were obtained. Individuals missing over 70% of the 12,599 filtered SNPs were subsequently removed. Imputation with the expectation maximization imputation algorithm was performed for remaining missing data points with rrBLUP.

    In order to infer inbreeding to its across selection cycles, raw sequence reads from previous GBS genotyping were used in combination with TASSEL-GBS (Glaubitz et al., 2014) and the IWGSC RefSeq v1.0 reference genome (Appels et al., 2018) to recall SNPs that align to the reference genome for the TP, C0, and C1 populations prior to final analysis of inbreeding levels.

    2.6 Phenotypic data

    2.6.1 Fructan quantification

    Phenotypic data for total grain fructan content were measured using the enzymatic–spectrophotometric method (AOAC 999.03; McCleary, Murphy, & Mugford, 2000). Briefly, 3 g of seed was ground in a Geno/Grinder 2000, and 1.0 g of resulting flour was dissolved in 80 °C water for 15 min. The volume of the dissolved solution was adjusted to 100 ml, and supernatant produced from centrifugation of solution aliquots was stored for later analysis. All ground flour and dissolved samples were stored at −20 °C to minimize endogenous enzymatic degradation of fructans.

    All dissolved samples were analyzed with an enzymatic assay kit (K-FRUC, MegaZyme International, https://secure.megazyme.com/Fructan-Assay-Kit). The enzymatic assay protocol was adapted (downscaled to one-fifth) for 96-well plates (Newell et al., 2014) to increase throughput and reduce phenotyping costs. Sample absorbance was measured with a SpectraMax Plus 384 microplate reader, and total fructan content was calculated per Equation 2. To control for variation between plates, fructan content for each plate was adjusted by the MegaZyme fructan controls on each plate (Equation 3), thereby aligning all plates to a known concentration of the fructan control.
    fructan % w / w , as is = Δ A F V / 2 . 48 W (2)
    where ΔA = sample absorbance − sample blank absorbance (both read against the reagent blank); F = factor to convert absorbance values to micrograms of D-fructose; V = volume (ml) of extractant used (i.e., 50 or 100 ml); W = weight (mg) of sample extracted; and 2.48 = unit conversion factor.
    plate adjustment = 1 observed control / expected control (3)
    where observed control is the observed fructan content within the control well, and expected control is the expected fructan content value for control supplied in the K-FRUC enzymatic assay kit.

    2.7 Genomic selection model training

    All lines for the initial TP were grown at the Caldwell Farm, Cornell University, Ithaca, NY, in summer 2013 to maximize the ability to estimate genetic variance under a fixed field resource allocation. Phenotypes for updating TPs in each diverged breeding population were measured on bulked S1 greenhouse seed from selected individuals (Figure 1).

    2.8 Realized genetic gain evaluation

    The C0, C1, and C2 populations were evaluated in 1-m head rows of S1 progeny in two locations (Caldwell [42.449813° lat., −76.460692° long.] and Ketola [42.471043° lat., −76.438273° long.]) near Ithaca during the 2016 growing season to estimate cycle means, genetic variances, and genetic correlations. Due to the large size of the C0 populations, selected individuals and a random sample of 36 unselected individuals from each replicate population were planted in the evaluation experiment. A random incomplete block design with 24 blocks containing cycle populations, TP, and checks (Ava, Caledonia, E0028, OH751, Pioneer 25R47, Pioneer25W36, Richland, SW50, SW80, and 7730R) was used. Plant height and heading date were collected on each line evaluated in the trial. Phenotypic data for checks were reviewed to ensure uniformity across the blocks in the 2016 field trial, and no data adjustments were required. Fructan content was measured on bulked field seed in fall 2016. Monthly weather data for the Ithaca area were obtained from annual summaries published on the Northeast Regional Climate Center website (http://www.nrcc.cornell.edu/wxstation/ithaca/ithaca.html).

    2.9 Statistical models for selection

    2.9.1 Genomic selection with truncated selection

    Genomic selection for the TS populations was performed with the genomic BLUP model (GBLUP; Equation 4) in rrBLUP (Endelman, 2011). The additive genetic relationship was calculated with the A.mat function in rrBLUP (Endelman, 2011):
    y = μ + X β + Zg + ε (4)
    where y is a vector of fructan phenotypes, μ is the mean, X is the incidence matrix for environments, β is a vector of fixed environmental effects, Z is the incidence matrix for genotypes, g is a vector of random marker effects, ε is a vector of residuals, g ∼ N(0, σa2K), ε ∼ N(0, σe2), σa2 is the additive genetic variance, K is the additive genetic covariance matrix, and σe2 is the error variance.

    2.9.2 Genomic selection with constrained selection

    Genomic selection with the OCS method was performed in two stages. First, breeding values for individuals were estimated with GBLUP (Equation 4). The resulting GEBVs, additive genetic relationship matrix, and a list of top 12 selection candidates were subsequently used in the OCS method proposed by Meuwissen (1997) to maximize genetic gain and control inbreeding in the following generation.

    In brief, expected gain from selection of the top 12 individuals with TS is computed from selection candidate GEBVs assuming equal contribution of the selection candidates to progeny in the next generation (Equation 5). The baseline level of estimated inbreeding for TS is set equal to the average coancestry of selection candidates (Equation 6). Lagrangian multipliers are then used to obtain the optimal contribution of selected candidates for which expected genetic gain is equal to TS with estimated inbreeding levels equivalent to 50% of the baseline level:
    G t + 1 = c t EB V t (5)
    where Gt+1 is the genetic level of the next generation, ct is a vector of genetic contributions of the selected candidates to generation t + 1, and EBVt is a vector of BLUP estimated breeding values of the candidates of selection in generation t.
    C ¯ t + 1 = c t A t c t 2 (6)
    where C ¯ t + 1 is the average coancestry between selected candidates, and At is a matrix of additive genetic relationships among selection candidates in generation t.

    2.10 Realized genetic gain calculations

    2.10.1 Wheat grain fructan content

    Adjusted population means for each replicate of C0, C1, and C2 were calculated as the sum of population effects and mean (p + μ, Equation 7). Realized gains were calculated by subtracting from population means their respective C0 population means. Percentage gain for each population was calculated as realized gain divided by the corresponding C0 population mean. Paired two-tailed t tests with pairings of gains across replicate populations by selection methodology were used to test differences in realized gain per cycle between TS and OCS methods. Paired two-tailed t tests were also used to test between realized and expected gains for each GS method within each cycle. One-sided, one-sample t tests were used to test the significance of realized genetic gains between cycles for each method and overall genetic gain relative to C0:
    y i j k = μ + β i + r j + p k + ε i j k (7)
    where yijk is the fructan phenotype, μ is the mean, βi is the fixed environment effect (i = 1, 2), rj is the random effect for blocks (j = 1, …, 24), pk is the fixed population effect (= 1, …, 18), εijk is the residual, rk ∼ N(0, σr2) where σr2 is the block variance, and ε ∼ N(0, σe2) where σe2 is the residual error variance.
    Genetic values for breeding lines across environments were calculated with the mixed model in Equation 7. Genetic values of breeding lines within each environment were obtained by fitting the mixed model in Equation 7 with environment effects (βi) removed. Narrow-sense heritability was calculated for genotyped lines with variance estimates obtained from the mixed model in Equation 8:
    y j l = μ + r j + g l + ε j l (8)
    where yijl is the fructan phenotype, rj is the random effect for blocks (j = 1, …, 24), gl is the random genotype effect (l = 1,…, 970), εil is the residual, gl ∼ N(0, σa2K) where σa2 is the additive genetic variance and K is the marker-based relationship matrix, and ε ∼ N(0, σε 2).

    Genetic correlations between 2013 and 2016 field environments were calculated based on genetic values for each environment. Additionally, the genetic correlations between greenhouse and each field environment were computed.

    2.11 Expected genetic gain calculations

    Expected genetic gain in wheat grain fructan content from GS for each population was calculated with Equation 9. The selection intensity for each population equaled the previously calculated intensities observed at the time of empirical selection. To estimate genetic variances for grain fructan content, nongenetic effects were removed from Equation 10 with population-specific models fit (Equation 11) on the residuals (yl′) from the nongenetic model. Estimates of genetic variance from the validation trial, σg2, were used to approximate the additive genetic variance, σa2, for C2 populations.
    R = i r σ A (9)
    where R is the response to selection, i is the selection intensity, r is the selection accuracy, and σA is the square root of additive genetic variance of observed phenotypes.
    y i j = μ + β i + r j + ε i j (10)
    where yij is the fructan phenotype, μ is the mean, βi is the fixed environment effect (i = 1, 2), rj is the random effect for blocks (j = 1, …, 24), εij is the residual, rj ∼ N(0, σr2), and ε ∼ N(0, σe2).
    y l = μ + g l + ε l (11)
    where yl′ is the fructan phenotype, μ is mean, gl is the random genotype effect (l = 1,…, n), εl is the residual, gl ∼ N(0, σa2K) where σa2 is the additive genetic variance and K is marker-based relationship matrix (Cycles 0, 1, 1S, and 2S), gl ∼ N(0, σg2) where σg2 is the genetic variance estimated from markers (Cycle 2), and εl ∼ N(0, σe2).

    2.12 Genomic selection accuracies

    The realized accuracy of the GS models used in this study was computed as the Pearson correlation between the GEBV and observed phenotypes (Falconer & Mackay, 1996). For realized accuracy calculations, the GEBVs were obtained from 2013 field and greenhouse materials and observed phenotypes were obtained from 2016 field trials. To further confirm realized accuracies, the correlations between ranks of individual GEBV and observed phenotype values were examined for each population. Additionally, leave-one-out cross-validation (LOOCV) was performed per Equation 11 for each population on 2016 phenotypes.

    2.13 Correlated agronomic responses

    To calculate correlated response to selection on grain fructan content for plant height and relative maturity measured as heading date, adjusted population means were estimated with Equation 7 (p + μ) where y was the corresponding phenotype of plant height or plant heading date. Realized gain and percentage gain were calculated from population mean estimates. Realized gains were calculated by subtracting all population means by their respective C0 population means. Percentage gain for each population was calculated as realized gain divided by the corresponding C0 population mean. Paired two-tailed t tests were used to test differences in gain per cycle between the GS methods. One-sided, one-sample t tests were used to test if genetic gains were significant. One-sided t tests were used to test if heading dates for TS populations were significantly greater than OCS population heading dates within each cycle.

    Genetic values for individuals across environments for plant height and heading date were calculated with the mixed model in Equation 8. Pearson correlations between genetic values for fructan content and agronomic traits were computed.

    2.14 Inbreeding assessment

    To assess the impacts of GS on inbreeding, the level of observed inbreeding was compared with the expected level of inbreeding based on markers and pedigrees. The expected level of inbreeding for each population was set equal to the average coancestry of selected individuals, which served as parental material for the population (per Equation 6). Observed levels of inbreeding (f) based on markers for each population were estimated by computing the mean of the off-diagonal elements of the marker-based relationship matrix. The mean level of observed inbreeding for individuals within each population based on pedigrees was computed with the R package pedigreemm (Bates & Vazquez, 2014).

    Paired two-tailed t tests with pairs across selection population replicates were used to determine if mean observed inbreeding levels differed significantly from expected levels for populations in each cycle. Paired one-tailed t tests were used to determine if mean observed inbreeding levels for OCS populations were significantly lower than TS populations within each cycle. The inbreeding values used in the t tests were mean inbreeding values for populations inferred from markers or pedigrees as described above.

    2.15 Genetic variance assessment

    To test the significance of difference of genetic variances between selection populations, populations were grouped by selection method and cycle and assigned indicator variables. An analysis was performed on the genetic effects of the groups (yi′ per Equation 11), allowing heterogenous genetic variance for the groups of entries: C0, C1 TS, C1 OCS, C2 TS, C2 OCS, and TPs or checks. Subsequent analyses were performed where groups within the same cycle (e.g., C1 TS and C1 OCS) were placed in the same group, thereby forcing them to have the same estimate of genetic variance. To determine the significance of difference of variance, a likelihood ratio test was performed. Under the likelihood ratio test, the difference between the −2 restricted maximum likelihood (REML) log values for the full and reduced models were computed, and p values were calculated based on the χ2 distribution (Saxton, 2004) with one degree of freedom because there were 10 parameters in the first test and nine in the second test. All analyses for assessing genetic variance were performed in SAS PROC MIXED (SAS Institute).

    3 RESULTS

    3.1 Realized genetic gain trial

    Observed wheat grain fructan content in the 2016 field trial differed between selection populations (Figure 2). The observed fructan content for Caldwell and Ketola environments averaged 1.36 (± 0.28) and 1.29 (±0.26) g 100 g−1, respectively. Within the 2016 trial, observed fructan levels across lines trialed ranged from 0.62–2.38 g 100 g−1. Narrow-sense heritability across environments was .72. Estimated narrow-sense heritabilities for Caldwell and Ketola were .63 and .70, respectively. Estimates of genetic values were consistent across environments with a correlation of .83. Correlation of genetic values between greenhouse and field locations ranged from .29 at Caldwell to .28 at Ketola. Additionally, correlation of genetic values between 2013 and 2016 field environments was .51. Adjusted population means are summarized in Table 2.

    Details are in the caption following the image
    Observed fructan content of recurrent genomic selection (RGS) populations. GS, genomic selection; C0, Cycle 0; T, simple truncated selection; O, optimized constrained selection; C1, Cycle 1; C2, Cycle 2; R1, Replicate 1; R2, Replicate 2
    TABLE 2. Means for grain fructan content, plant height, heading date and level of inbreeding for each population and each selection method
    Populationa Replicate No. of individuals evaluated Mean fructan content Mean height Mean heading date f (markers) f (pedigree)
    g 100 g−1 cm Julian d
    C0 1 58 1.320 92.2 151.1 0.052 0.000
      2 54 1.327 88.1 151.5 0.038 0.000
      Mean 56 1.324 ± 0.004 90.2 ± 2.9 151.3 ± 0.264 0.045 ± 0.01 0.000 ± 0.000
    C1 TS 1 46 1.605 95.8 150.9 0.208 0.145
      2 43 1.491 85.9 151.8 0.225 0.061
      Mean 45 1.548 ± 0.08 90.8 ± 7.0 151.3 ± 0.598 0.217 ± 0.012 0.103 ± 0.059
    C1 OCS 1 43 1.586 93.0 150.7 0.080 0.086
      2 46 1.533 85.6 151.7 0.104 0.046
      Mean 44 1.559 ± 0.037 89.3 ± 5.2 151.2 ± 0.733 0.092 ± 0.017 0.066 ± 0.028
    C2 TS 1 12b 1.774 88.6 151.8 0.305
      2 13b 1.541 78.1 153.0 0.160
      Mean 13 1.657 ± 0.165 83.4 ± 7.4 152.4 ± 0.862 0.200 ± 0.102
    C2 OCS 1 11b 1.829 92.4 151.2 0.202
      2 9b 1.709 86.3 151.8 0.099
      Mean 10 1.769 ± 0.085 89.4 ± 4.3 151.5 ± 0.367 0.200 ± 0.073
    • a C0, Cycle 0; TS, simple truncated selection; OCS, optimized contribution selection; C1, Cycle 1; C2, Cycle 2.
    • b Families.

    3.2 Gain from selection for wheat grain fructan content

    Realized genetic gains for wheat grain fructan content between GS cycles ranged between 0.110 and 0.236 g 100 g−1 (Table 3) with total gains of 0.334–0.445 g 100 g−1 (Figure 3). Percentage total genetic gains from two generations of GS with TS and OCS were 25.5 ± 12.5 and 34 ± 6.5, respectively (Table 3). Total genetic gains were significant for both selection methods (p < .001). The differences between total expected and realized genetic gains were significant for C2 (p = .005, Figure 4). No significant differences in genetic gain per cycle were detected between the GS methods (p = .30–.76).

    TABLE 3. Realized and expected genetic gains from Cycle 0 (C0) and previous cycles
    Populationa Realized gain from C0 Gain from C0 Expected gain from C0 Realized gains from previous cycle Gain from previous cycle Expected gain from previous cycle
    % %
    C1 TS
    Replicate 1 0.284 21.5 0.178 0.284 21.5 0.178
    Replicate 2 0.164 12.4 0.058 0.164 12.4 0.058
    Mean 0.224 ± 0.085 16.9 ± 6.5 0.118 ± 0.085 0.224 ± 0.085 16.9 ± 6.5 0.118 ± 0.085
    C1 OCS
    Replicate 1 0.265 20.1 0.167 0.265 20.1 0.167
    Replicate 2 0.207 15.6 0.055 0.207 15.6 0.055
    Mean 0.236 ± 0.041 17.8 ± 3.2 0.111 ± 0.079 0.236 ± 0.041 17.8 ± 3.2 0.111 ± 0.079
    C2 TS
    Replicate 1 0.453 34.3 0.293 0.169 12.8 0.116
    Replicate 2 0.214 16.7 −0.030 0.050 4.3 −0.087
    Mean 0.334 ± 0.169 25.5 ± 12.5 0.132 ± 0.228 0.110 ± 0.084 8.6 ± 6 0.014 ± 0.144
    C2 OCS
    Replicate 1 0.508 38.5 0.181 0.243 18.4 0.014
    Replicate 2 0.382 29.4 0.144 0.176 13.9 0.089
    Mean 0.445 ± 0.089 34 ± 6.4 0.162 ± 0.026 0.209 ± 0.048 16.1 ± 3.2 0.051 ± 0.053
    • a C1, Cycle 1; TS, simple truncated selection; OCS, optimized contribution selection; C2, Cycle 2.
    Details are in the caption following the image
    Realized genetic gain of wheat grain fructan content over cycles for four populations. C0, Cycle 0; C1, Cycle 1; C2, Cycle 2; TS, simple truncated selection; OCS, optimized contribution selection
    Details are in the caption following the image
    Expected genetic gain of wheat grain fructan content over cycles for four populations. C0, Cycle 0; C1, Cycle 1; C2, Cycle 2; TS, simple truncated selection; OCS, optimized contribution selection

    3.3 Genomic selection accuracy

    Mean realized GS accuracies among cycles ranged from 0.146 to 0.335 (Table 4). The mean correlations between ranks of individual GEBV and observed phenotype values ranged between 0.168 and 0.305. The mean accuracies for the LOOCV with 2016 phenotypes within populations ranged between 0.851 and 0.937.

    TABLE 4. Accuracies for genomic selection (GS) and correlations of ranks between breeding values
    Populationa Realized accuracy Rank correlation between Obs Pheno, GEBVsb 2016 LOOCVc accuracy
    C1 TS
    Replicate 1 0.496 0.486 0.945
    Replicate 2 0.173 0.124 0.926
    Mean 0.335 ± 0.229 0.305 ± 0.256 0.936 ± 0.014
    C1 OCS
    Replicate 1 0.496 0.486 0.945
    Replicate 2 0.173 0.124 0.926
    Mean 0.335 ± 0.229 0.305 ± 0.256 0.936 ± 0.014
    C2 TS
    Replicate 1 0.738 0.711 0.796
    Replicate 2 −0.445 −0.375 0.905
    Mean 0.146 ± 0.837 0.168 ± 0.768 0.851 ± 0.077
    C2 OCS
    Replicate 1 0.071 −0.076 0.931
    Replicate 2 0.596 0.478 0.943
    Mean 0.333 ± 0.371 0.201 ± 0.392 0.937 ± 0.008
    • a C1, Cycle 1; TS, simple truncated selection; OCS, optimized contribution selection; C2, Cycle 2.
    • b Obs Pheno, observed phenotype; GEBV, genomic estimated breeding values.
    • c LOOCV, leave-one-out cross-validation.

    3.4 Correlated agronomic responses

    Selection for wheat grain fructan content had no effect on height over multiple cycles of GS (p = .22, Table 5). Additionally, no correlation was observed (r = −.08) between grain fructan content and plant height.

    TABLE 5. Realized genetic gain for plant height and heading date for each population and each selection method
    Populationa Realized gain from Cycle 0
    Height Heading
    %
    C1 TS
    Replicate 1 3.870 −0.125
    Replicate 2 −2.540 0.187
    Mean 0.666 ± 4.53 0.031 ± 0.22
    C1 OCS
    Replicate 1 0.820 −0.287
    Replicate 2 −2.840 0.151
    Mean −1.010 ± 2.59 −0.068 ± 0.31
    C2 TS
    Replicate 1 −3.920 0.477
    Replicate 2 −11.320 1.034
    Mean −7.620 ± 5.24 0.755 ± 0.39
    C2 OCS
    Replicate 1 0.190 0.090
    Replicate 2 −2.020 0.185
    Mean −0.920 ± 1.56 0.138 ± 0.07
    • a C1, Cycle 1; TS, simple truncated selection; OCS, optimized contribution selection; C2, Cycle 2.

    Wheat grain fructan content and heading date were negatively correlated (r = −.30), but selection for wheat grain fructan content did not significantly change plant heading date over multiple cycles of GS (p = .23, Table 4). Realized genetic gains and mean heading date indicate that OCS populations had significantly earlier heading dates than TS populations in C2 (p = .004) but not C1 (p = .098) (Tables 2 and 5).

    3.5 Inbreeding

    Observed inbreeding levels (f, Table 6) in OCS populations were significantly lower (p = .01) than inbreeding levels in TS populations over all cycles based on marker data. Estimates of inbreeding levels inferred from pedigrees were also significantly lower (p = .02) in OCS populations relative to TS populations for all cycles with the exception of the TS Replicate 2 population in C2. Observed inbreeding levels were significantly lower than estimated inbreeding values based on marker and pedigree data (p = .02 and p = .003, respectively).

    TABLE 6. Expected and observed inbreeding for simple truncated selection (TS) and optimized contribution selection (OCS) populations
    Populationa Expected inbreeding (markers) f (markers) Expected inbreeding (pedigree) f (pedigree)
    C1 TS
    Replicate 1 0.236 0.208 0.422 0.145
    Replicate 2 0.240 0.225 0.281 0.061
    Mean 0.238 ± 0.002 0.217 ± 0.012 0.351 ± 0.100 0.103 ± 0.059
    C1 OCS
    Replicate 1 0.123 0.080 0.231 0.086
    Replicate 2 0.133 0.104 0.178 0.046
    Mean 0.128 ± 0.007 0.092 ± 0.017 0.204 ± 0.038 0.066 ± 0.028
    C2 TS
    Replicate 1 0.527 0.667 0.305
    Replicate 2 0.448 0.427 0.160
    Mean 0.488 ± 0.056 0.547 ± 0.170 0.200 ± 0.102
    C2 OCS
    Replicate 1 0.220 0.379 0.202
    Replicate 2 0.220 0.253 0.099
    Mean 0.220 ± 0.000 0.316 ± 0.089 0.200 ± 0.073
    • a C1, Cycle 1; TS, simple truncated selection; OCS, optimized contribution selection; C2, Cycle 2.

    3.6 Genetic variance

    Estimated genetic variances between C0 and C1, as well as between C1 and C2, within each selection method indicated significant reductions (p < .001, Table 7). Although the average genetic variance of TS populations after multiple cycles of selection was 50% lower than genetic variance in OCS populations, genetic variances within cycles were not significantly different between selection methods (p = .31–.57, Table 7).

    TABLE 7. Contrast of estimated genetic variances of genomic selection methods based on likelihood ratio test
    Selection methoda Difference in variance P value
     Cycle 0 vs. Cycle 1
    TS −0.1021 <.0001
    OCS −0.1148 <.0001
    Cycle 1 vs. Cycle 2
    TS −0.0696 .0805
    OCS −0.1074 .0053
    TS vs. OCS
    Cycle 1 −0.0127 .5708
    Cycle 2 −0.0505 .3132
    • a TS, simple truncated selection; OCS, optimized contribution selection.

    4 DISCUSSION

    4.1 Effectiveness of selection and gains from selection

    For both GS methods, percentage total genetic gain was high (25–34%, Table 3). The success of GS in breeding programs has also been observed for end-use quality traits in spring wheat (Battenfield et al., 2016; Heffner, Jannink, Iwata, Souza, & Sorrells, 2011) and drought-tolerant corn (Zea mays L.) hybrids (Cooper et al., 2014),

    For GS cycles, percent genetic gain per cycle was higher in C1 (16.9–17.8%) than in C2 (8.6–16.1%). The large, observed deviations in realized and percentage gains in TS populations in C2 were due to the smaller genetic gains observed in Replicate 2 (Table 3). The significant gains in wheat grain fructan content (p < .001) for both selection methods suggest that RGS is an effective breeding strategy for increasing wheat grain fructan content.

    Realized total genetic gains were significant for both selection methods (p < .001) with total increase in grain fructan content ranging from 0.33 to 0.45 g 100 g−1 over two cycles of selection. This observed gain is in accordance with numerous studies that predict high levels of genetic gain with the implementation of GS (Gaynor et al., 2017; Lin et al., 2016; Voss-Fels et al., 2018). Estimates of total realized genetic gains exceeded total expected genetic gains for C2 (p = .005) though expected genetic gains from selection were negatively influenced by low selection accuracies (Table 4) and small selection differentials (Table 1). Although the use of OCS has been noted to result in a small reduction in short-term genetic gain relative to TS in animal mating simulations (Henryon et al., 2015), this phenomenon was not observed in this empirical study with plants. Unlike animals, wheat plants tiller, thereby producing multiple flowers that allow high-merit individuals to contribute to multiple mating pairings in the OCS framework as both females and males. This key difference in reproductive biology between animals and plants resulted in the elimination of penalties from inclusion of lower merit individuals and subsequently resulted in no reduction in short-term genetic gains in OCS relative to TS.

    4.2 Realized selection accuracies

    Realized accuracies of GS across populations averaged 0.287. Although there was a wide range of realized GS accuracies observed (0.146–0.335), the realized accuracies were low compared with the significant positive realized genetic gains observed for wheat grain fructan content. This highlights that predictions were closer to the true breeding values than observed phenotypes. Further examination of the correlations of ranks between GEBVs of material grown in the greenhouse and observed phenotypes of material grown in the 2016 field season found correlations were similar to the realized GS accuracies (0.146–0.305), with an average of 0.24.

    The 2016 growing season from which observed phenotypes for GS accuracies were estimated was an abnormally dry growing season. Observed rainfall between planting (October 2015) and harvest (July 2016) averaged 75% of normal (523 mm observed vs. 762 mm average). During the peak of the growing season (May–July), observed rainfall was 45.3% of normal (4,123 mm observed vs. 279 mm average). Given that fructans act as osmoregulators and are involved in physiological functions under drought conditions, it is possible some G × E interactions occurred in the field trial which reduced GS accuracies. The low correlation of greenhouse genetic values and 2016 field environment observed phenotypes (r = .28–.29) support the hypothesis that G × E may have been one factor that affected prediction accuracies in this study.

    Genomic selection accuracies were assessed using a LOOCV for each population with phenotypic data collected in the 2016 realized gain trial. Accuracies of the LOOCV were high with an average accuracy of 0.91. The similarity between the correlation of performance ranks and realized GS accuracies, along with high LOOCV accuracies, also supports the hypothesis that G × E contributed to the prediction accuracies.

    4.3 Impact of selection on inbreeding and genetic variance

    Inbreeding levels for all selected populations increased; however, inbreeding levels in OCS populations were significantly lower than inbreeding levels in TS populations based on markers (p = .01) and pedigree records (p = .02).

    Observed inbreeding levels were significantly lower (p ≤ .02) than predicted inbreeding values for all cycles. Expected inbreeding levels were predicted based on empirical contributions of selected individuals to the subsequent generation under the assumption of random mating. Given that individual plants were selected and intermated, it is possible that some intermating was nonrandom, leading to lower than expected levels of observed inbreeding. Use of more plants representing selected individuals would have facilitated intermating and led to observed inbreeding levels that were closer to predicted values.

    Genetic variance significantly decreased (p < .001) with the first cycle of GS (C0–C1) for both methods and with the second cycle of GS (p = .005, C1–C2) for the OCS method. This loss of genetic variance is in accordance with previous studies that found the empirical application of GS results in significant decreases in genetic variance (Asoro et al., 2013; Rutkoski et al., 2015). The populations resulting from the OCS approach contained 50% more genetic diversity than TS populations. Although OCS populations contained more genetic diversity than TS populations after two cycles of selection, the difference between populations was not significant for C2 (p = .31).

    4.4 Correlated agronomic responses to selection

    No previous studies have indicated a correlation between wheat grain fructan content and other traits. This study confirmed there was no correlation as TS for wheat grain fructan content was not associated with a change in plant height (p = .22) or heading date (p = .23). Days to heading for OCS populations were significantly earlier than TS populations for all selection cycles (p = .004) with the exception of C1 (p = .098). However, the short-term practical differences in heading date between the populations were minimal as the observed mean differences were <1 d.

    4.5 Improvements in wheat nutritional quality

    Given the prevalence of wheat within the human diet, improving nutritional quality of wheat has become a point of interest for many breeding programs. Although diets in whole grains have been proven to have positive physiological impacts for consumers (Shewry & Hey, 2015), there are areas of wheat nutrition that remain open for exploration and improvement. Important areas of research have included increasing bioavailability of key nutrients such as Fe and Zn in wheat (Kenzhebayeva et al., 2019). The preexisting genetic variability in fructan content within wheat and the key physiological roles fructans serve within plants and humans allow fructans to be an ideal target for developing nutritionally improved wheat cultivars using traditional breeding methods with the aid of tools such as GS.

    5 CONCLUSION

    This study showed that GS for wheat grain fructan content with two selection methods was successful over two cycles of selection. Over two selection cycles, OCS genetic gain was equal to TS and exhibited lower levels of inbreeding and greater levels of genetic variance.

    Realized genetic gains exceeded expected gains estimated with field computed selection differentials. Realized GS accuracies calculated from expected genetic gains were lower than expected based on observed realized genetic gain. Individual performance between greenhouse and 2016 environments found poor correspondence of ranks for selected individuals and low correlations of genetic values. Genotype × environment interaction, as well as other factors, may explain the reduced genetic gain in C2. Selection for wheat grain fructan content did not affect plant height but was associated with fewer days to heading in the OCS populations relative to TS populations.

    The GS scheme used in this experiment was well adapted for the generation time of winter wheat; however, the implementation of this GS scheme on a larger scale requires further considerations. The genotyping, selection, and crossing of individual plants can limit the ability to cross plants with differing heading dates. Further research is needed to determine if OCS methods provide reduced inbreeding rates relative to TS with similar levels of genetic gain from selection in long-term empirical GS studies.

    ACKNOWLEDGMENTS

    The authors thank Jessica Rutkoski for her assistance in designing the breeding scheme, James Tanaka for his valuable mentorship in wheat breeding and technical assistance in the greenhouse and laboratory, Nicholas Santantonio for assistance with recalling SNPs, Jean Koski for tending to research materials in the greenhouse, David Benscher for his assistance with field trials, and the International Wheat Genome Sequencing Consortium (IWGSC) for prepublication access to the reference sequence of wheat, IWGSC RefSeq v1.0. This research was supported in part by USDA–National Institute of Food and Agriculture (NIFA)–Agriculture and Food Research Initiative (AFRI) grants, Awards no. 2009-65300-05661, 2011-68002-30029, and 2005-05130, and by Hatch Project 149-447.

      AUTHOR CONTRIBUTIONS

      L.D. Veenstra performed this research with the assistance of J. Poland for genotyping support. J.L. Jannink and M.E. Sorrells provided advisory and editorial support.

        CONFLICT OF INTEREST

        All authors (L.D. Veenstra, J. Poland, J.L. Jannink, and M.E. Sorrells) certify they have no affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers’ bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or nonfinancial interest (such as personal or professional relationships, affiliations, knowledge, or beliefs) in the subject matter or materials discussed in this manuscript.