Improving the baking quality of bread wheat by genomic selection in early generations
Genomic selection shows great promise for pre-selecting lines with superior bread baking quality in early generations, 3 years ahead of labour-intensive, time-consuming, and costly quality analysis.
The genetic improvement of baking quality is one of the grand challenges in wheat breeding as the assessment of the associated traits often involves time-consuming, labour-intensive, and costly testing forcing breeders to postpone sophisticated quality tests to the very last phases of variety development. The prospect of genomic selection for complex traits like grain yield has been shown in numerous studies, and might thus be also an interesting method to select for baking quality traits. Hence, we focused in this study on the accuracy of genomic selection for laborious and expensive to phenotype quality traits as well as its selection response in comparison with phenotypic selection. More than 400 genotyped wheat lines were, therefore, phenotyped for protein content, dough viscoelastic and mixing properties related to baking quality in multi-environment trials 2009–2016. The average prediction accuracy across three independent validation populations was r = 0.39 and could be increased to r = 0.47 by modelling major QTL as fixed effects as well as employing multi-trait prediction models, which resulted in an acceptable prediction accuracy for all dough rheological traits (r = 0.38–0.63). Genomic selection can furthermore be applied 2–3 years earlier than direct phenotypic selection, and the estimated selection response was nearly twice as high in comparison with indirect selection by protein content for baking quality related traits. This considerable advantage of genomic selection could accordingly support breeders in their selection decisions and aid in efficiently combining superior baking quality with grain yield in newly developed wheat varieties.
The genetic improvement of baking quality is one of the grand challenges in winter wheat breeding due to its complex inheritance pattern, which is governed mainly by wheat storage proteins, foremost the prolamins gliadin and glutenin (Payne 1987; Shewry et al. 1995, 2003) as well as their interaction with other fractions like the puroindolins that confer grain hardness (Bekes 2012a; Quayson et al. 2016; Würschum et al. 2016). The quality of these wheat storage proteins can be determined by various measurements, amongst others by dough rheological tests that assess the viscoelastic and mixing properties during dough preparation that are important for food-processing and baking of various products like bread, cookies, and pastry. Baking tests and dough rheology are furthermore part of official variety test and registration in various countries to offer varieties with fitting combinations of quality characteristics. A major obstacle is thereby that the assessment of the associated traits often involve time-consuming, labour-intensive, and costly testing as well as the too low amount of grains that is available per genotype in early breeding generations, thus sophisticated quality tests have usually been postponed to the very last phases of variety development.
Hence, selection decisions for baking quality improvement are beforehand guided by indirect phenotypic selection based on correlated traits like protein content. Protein quantity assessed in this way explains however merely a limited part of the genetic variation observed for traits related to baking quality, for which loci associated with the composition of the wheat storage proteins gliadin and glutenin play a major role (Payne et al. 1987; Lukow et al. 1989; Rogers et al. 1989). A pre-selection of lines in early generations by markers linked to the known Glu-1 and Glu-3 glutenin loci is accordingly an interesting option (Eagles et al. 2002; Zheng et al. 2009; Krystkowiak et al. 2016), but there are few successful reports of such an approach (Kuchel et al. 2007), and the respective markers have to be combined with additional small-scale tests to achieve a reasonable prediction accuracy for selection (Oury et al. 2010). Such a marker-assisted selection focuses mainly on major quantitative trait loci (QTL) that explain a substantial but limited amount of the underlying genetic variance, while most traits of interest in plant breeding show a polygenic inheritance and are, therefore, controlled mostly by many minor QTL.
Genomic selection has been implemented in recent years into many national and international wheat breeding programs (Guzmán et al. 2016a; He et al. 2016; Michel et al. 2016) to additionally target these small effect loci influencing quantitative traits with genome-wide distributed markers in genetically fingerprinted training and selection populations. Training populations in applied wheat breeding programs are normally comprised of advanced breeding material that has been thoroughly phenotyped for grain yield, disease resistances and numerous milling and baking quality traits (Guzmán et al. 2016a; He et al. 2016; Michel et al. 2016). The phenotypic information of the selection population is on the other hand very limited, and genomic estimated breeding values for numerous traits of interest can be derived for these genotypes via their genetic relationship with the training population in order to support breeders in their selection decisions (Heffner et al. 2009; Heslot et al. 2015). This promising selection method could be especially valuable for baking quality related parameters such as dough rheological traits whose assessment requires costly, labour-intensive and time-consuming tests. The aims of this study were thus (1) to investigate the prospect of genomic selection for these laborious to phenotype quality traits, (2) enhancing this approach by integrating prior knowledge about trait correlations and genetic architecture, and (3) compare the selection response of direct and indirect phenotypic with genomic selection.
Materials and methods
Plant material and phenotypic data
We analyzed a population of 840 genotyped winter bread wheat lines (Triticum aestivum L.), which was derived from multiple families and selected by the pedigree method until the F4:6 and F5:7 generation or generated by the double haploid method. Different subpopulations of these lines were phenotyped in multi-environment trials at locations in Austria, Hungary, Serbia, Croatia, Romania and Turkey from 2009 to 2016. Grain samples were collected and milled from 401 lines with a Quadrumat Junior milling system according to the method AACC26-50 approved by the American Association of Cereal Chemists (AACC 2000). The resulting flour samples were employed to create a dough rheological profile of each line, starting with the dough-mixing properties that were assessed by a Farinograph (Brabender GmbH and Co KG) equipped with a 300 g mixing bowl. The optimal water uptake of each flour sample was estimated in a preliminary test on a subsample of 100 g flour until it reached a dough consistency of 500 farinogram units (FU) according to the standard procedure AACCI 54e21 (AACC 2000). The dough development time was measured as the time in minutes from the first water uptake until the dough began to soften due to intensive mixing in the main test. Dough stability was assessed as the timeframe between which the kneading curve first intersected and left the 500 FU borderline, and the farino quality number was calculated as the time point when the dough consistency fell 30 FU after reaching its peak. Thereafter, the Extensograph (Brabender GmbH and Co KG) was used to determine the viscoelastic properties of the flour samples according to AACCI 54-10.01 (AACC 2000) of which the extensibility (mm), resistance to extension at 50 mm in extensogram units (EU), and the area under the curve, i.e., the dough energy (cm2) after a 135-min resting time were of prior interest in this study. The protein content (%) was determined by near infrared spectroscopy (NIRS; FOSS GmbH) for all 840 lines directly at harvest.
The measurement of baking quality by mixing and viscoelastic tests is typically costly, labour-intensive and time-consuming, thus the obtained phenotypic records from the 401 lines that were subject to dough rheological analysis were highly unbalanced between trials, and the data from different trials were mainly connected by several common check varieties replicated in each of the completely randomized trial designs. The other 439 lines were on the other hand thoroughly tested in multi-environment trials, and orthogonally phenotyped for their protein content across various locations in the above-described target population of environments from 2009 to 2013. Hence, they provided an additional source of information for a rapidly to assess quality parameter that is routinely generated in many wheat breeding programs.
Statistical analysis of phenotypic data
The additional 439 lines tested in multi-environment trials from 2009 to 2013, were analysed for their performance with regard to the protein content following a two stage analysis strategy. Each individual yield trial was first analysed with various models correcting for row and/or column effects as well as with an autoregressive variance–covariance structure (Burgueño et al. 2000). The best model was chosen by Akaike’s Information Criterion (AIC) to calculate best linear unbiased estimates (BLUE), while trials with a heritability larger than 0.3 were used for an across trial analysis for each individual year 2009–2013 following formula (2). All phenotypic analyses were conducted using the statistical package ASReml 3 for the R programming environment (R Development Core Team 2016).
DNA was extracted following the protocol by Saghai-Maroof et al. (1984) using leaf samples that were collected from F4:5 or doubled haploid lines by sampling minimum ten plants per line during early summer. All 840 lines were genotyped using the DarT genotyping-by-sequencing (GBS) approach. Quality control was applied by filtering out markers with a call rate lower than 90%, a minor allele frequency smaller than 0.05, and more than 10% of missing data. An MVN-EM algorithm (Poland et al. 2012) was employed to impute missing data of the remaining 7687 markers, and their pair-wise correlation was used as an ad hoc measure of linkage disequilibrium. One marker from each marker pair that exceeded the r2 = 0.8 threshold was dropped at random in order to remove strongly correlated predictor variables that would not contribute further to the prediction accuracy but elongate computation time, which resulted in a final set of 4598 markers.
A map position was available for 2637 of these markers with an average coverage of one marker every 1.8 cM. The usage of phenotypic data from the genotyped lines in a higher generation than the F4:5 was expected to introduce a small error due to a minor change in average heterozygosity, which was nevertheless seen to be acceptable considering the cost–benefit ratio of re-genotyping all lines in the advanced generations. Additionally, a subset of 444 lines was screened for their allelic state at the high-molecular weight glutenin subunit loci Glu-A1, Glu-B1, and Glu-D1 by sodium dodecyl sulphate polyacrylamide gel electrophoresis (SDS-PAGE). The missForest algorithm (Stekhoven and Bühlmann 2012) was used to impute the missing values of the other lines in a chromosome-wise manner, employing both the GBS and SDS-PAGE markers.
Single-trait genomic and marker-assisted selection
Prediction using the markers linked with the Glu-1 loci on chromosomes 1A, 1B, and 1D.
Performing a genome-wide associations study (GWAS) with the K model and previously determined population parameters (Yu et al. 2006; Zhang et al. 2010) in the estimation set with markers of known map position and subsequent prediction with the three most significant markers.
Combining the known Glu-1 loci markers with the de novo identified markers by GWAS.
Fitting the RR-BLUP model with all 4598 markers as random effects for predicting the various dough rheological traits.
Including the three marker sets used for marker-assisted selection as fixed effects into the RR-BLUP model, thereby excluding them from the random effects matrix.
Prediction by fourfold cross-validation within the estimation set used for the GWAS.
Prediction of the validation set, i.e., the left-out 20% of lines.
Forward prediction of lines tested in 2014–2016.
The forward prediction was thereby done for three independent validation populations that consisted of 70, 79, and 125 lines for which a dough rheological profile was generated in 2014, 2015, and 2016, respectively. It should be stressed out at this point that in the first scenario of cross-validation within the estimation set used for the GWAS, predictors are selected before leaving observations out. The left-out observations are however supposed to be unobserved for a correct application of cross-validation, and marker selection has to be conducted after leaving such validation data out (Hastie et al. 2009). Nevertheless, this wrong statistical method has been used in several recent publications about genomic selection in plant breeding. Hence, we like to highlight the outcome of such a procedure in comparison with correctly applied cross-validation and independent validation in this study, especially as the potential consequences could be dramatically when prediction models are actually employed for conducting selection decisions in applied plant breeding programs. All models for the single-trait genomic and marker-assisted selection were fitted with the package rrBLUP (Endelman 2011) as implemented in the R programming environment (R Development Core Team 2016).
Multi-trait genomic selection
Phenotypic data of dough rheological and correlated traits, i.e., protein content is only available for the training population.
The validation/selection population has already been phenotyped for protein content.
A large number of additional lines has been genotyped and phenotyped for protein content, though were not advanced for further testing. Nevertheless, these lines are an integral part of training populations in breeding programs with genomic selection, and have the ability to double the training population size of a correlated trait such as protein content in multi-trait prediction models.
Phenotypic data of dough rheology is merely available for the training population, but protein content is phenotyped for a large number of additional lines as well as the validation and training population.
Forward prediction and response to selection
This estimate was averaged over all independent validation populations and dough rheological traits and compared with the proportion of correctly selected lines by genomic and random selection for every selection intensity investigated in this study. Finally, an example dataset and R Code illustrating the implementation of single and multi-trait models for genomic selection with the R package sommer (Covarrubias-Pazaran 2016) was made available for the interested reader (Online Resource 2, Online Resource 3).
Quantitative-genetic parameters and trait correlations
Variance components, heritability, mean and range of the dough rheological parameters and the protein content for the entire population of dough rheologically analyzed lines from 2009 to 2016
Water uptake (%)
Quality number (× 10 min)
Protein content (%)
Phenotypic (upper triangle) and genetic correlation (lower triangle) as estimated by a multi-trait G-BLUP model between dough rheological parameters and the protein content in the entire population of 401 dough rheological analyzed lines from 2009 to 2016
Genomic selection with candidate loci and de novo found marker-trait associations
Integrating these de novo found markers as fixed effects into a RR-BLUP model for genomic selection followed the same pattern (Fig. 2d–f), and in the independent validation the prediction accuracy of the according W-BLUP models was approximately halved (r = 0.18) in comparison to the baseline RR-BLUP model (r = 0.37). Nevertheless the cross-validation and independent validation revealed some merit of including the known Glu-1 loci markers as fixed effects into the RR-BLUP model for constructing a W-BLUP model, as the prediction accuracy could be slightly increased even for the naïve approach when using all Glu-1 loci markers irrespectively of the explained genetic variance for a dough rheological trait. Independent validation resulted, e.g., in an average prediction accuracy across all dough rheological traits of r = 0.41 that was slightly higher than the estimated baseline prediction accuracy of r = 0.37 obtained with a standard RR-BLUP model. The increase in prediction accuracy by the W-BLUP model was furthermore independent of the population structure that was investigated via principal component analysis (Fig. S1 Online Resource 4).
The general advantage of a W-BLUP model largely disappeared at thresholds higher than 10% as few of the three assayed Glu-1 loci markers explained such a huge amount of genetic variance of the dough rheological traits. The threshold values were however based on a cross-validation within the basis population 2009–2013 and thus stricter than utilizing the often biased, i.e., over-optimistic values derived from linear models without such a scheme. Finally, we set a threshold of 5.00% explained genetic variance for including the Glu-1 loci markers as a compromise between explained genetic variance and marker number, as, e.g., a clear advantage was already seen by modelling Glu-D1 as fixed effect for predicting resistance to extension in the validation set of the 20% of lines that were left out for the GWAS. These markers would otherwise been excluded by using a stricter threshold of 10.00% as suggested by Bernardo (2014). The average prediction accuracy of the forward prediction was increased by 13% from r = 0.39 to r = 0.44 when integrating this prior knowledge about the genetic architecture of wheat storage protein composition into the genomic prediction models.
Multi-trait prediction models and response to selection for baking quality
Notwithstanding, the prediction accuracy might be increased by including a large number of additional lines that have been genotyped and phenotyped for protein content in multi-environment trials though were not selected for further testing. This scenario is common to most wheat breeding programs and the retained phenotypic data of the additional lines could be efficiently salvaged by doubling training populations size of the protein content, i.e., the correlated trait in multi-trait prediction models, which resulted in, however, no increase of prediction accuracy (r = 0.44) compared with the single-trait prediction model (r = 0.44) trained merely with phenotypic data of the dough rheological traits (Fig. 4).
Another interesting scenario in plant breeding is the availability of an easy to score correlated trait to the actual traits of interest for the validation population. Possessing prior information of the protein content enabled a marked increase of the prediction accuracy of dough rheological traits under this condition (r = 0.48), while a combination of this and the previous scenario for increasing the training population size did not further effect the average prediction accuracy (r = 0.48). However, we observed higher prediction accuracies the multi-trait prediction model that melted the genomic estimated breeding values of the protein content and dough rheological traits by the model-based selection index method described in this study. The prediction accuracy of a model including a large number of additional lines phenotyped for protein content was elevated by this method (r = 0.46) as was it when prior information of the protein content was available (r = 0.50). The model-based selection index method was furthermore computationally far less demanding and did not lead to convergence problems that often occurred with the completely unstructured multivariate linear mixed models, hence we used this method for all further multi-trait predictions.
Explained genetic variance estimated by cross-validation of the Glu-1 loci in the basis population 2009–2013, and prediction accuracy for each dough rheological trait in the forward prediction of 2014–2016
Explained genetic variance (%)
Model prediction accuracy
This study concentrated on genomic selection for labour-intensive and costly to assess dough rheological traits in order to support the improvement of baking quality in bread wheat breeding. The prospects of this approach were examined in various cross-validation schemes and a forward prediction with three independent validation populations using a large population of bread wheat lines tested across multiple years for dough rheological traits as well as their protein content. Apart from these prediction scenarios, several extensions of the genomic prediction models were investigated including multi-trait models and the exploitation of prior knowledge about the underlying trait genetic architecture.
Integrating marker-assisted and genomic selection by knowing the trait genetic architecture
It is commonly accepted that the glutenins and gliadins strongly influence dough properties and baking quality due to the unique property of this cereal protein fraction of building a gluten network in a flour–water mixture (Shewry et al. 1995), which enables gas retention during dough formation allowing the dough to rise (Shewry et al. 2003; Weipert 2006; Rasheed et al. 2014). Apart from yield, disease resistance, and stress tolerance, the genetic improvement of quality traits is an important part of many national and international wheat breeding programs (Battenfield et al. 2016; Guzmán et al. 2016a; Liu et al. 2016; Würschum et al. 2016). Hence, great efforts were undertaken to decipher the genetic basis of storage protein quality of wheat, and it has been shown that amongst others the high molecular weight subunit composition of the glutenins strongly influences bread-making quality (Payne et al. 1987; Lukow et al. 1989; Rogers et al. 1989). The Glu-1 loci found on the long arm of the homeologous chromosome pairs 1A, 1B and 1D associated with dough viscosity (Payne 1987) were found to be of especially large importance by many genetic studies for conferring dough resistance to extension (Kuchel et al. 2006; Mann et al. 2009; Tsilo et al. 2011; Cooper et al. 2016; Krystkowiak et al. 2016) and the respective favourable alleles also lead to a stronger gluten network thereby improving dough-mixing properties (Tsilo et al. 2013).
Accordingly, markers associated with the Glu-1 loci have been used for predicting wheat quality traits by marker-assisted selection (Eagles et al. 2002; Kuchel et al. 2007; Oury et al. 2010) and also showed some merit for this purpose in our study. Analogous to alveograph W values (Oury et al. 2010) some rheological traits like the dough energy of the extensogram could in this way be predicted with a sufficient accuracy (r = 0.48), while other traits like dough extensibility and water uptake had a very low prediction accuracy or were not predictable at all, as the Glu-1 loci did not explain a sufficient amount of genetic variance for the respective traits. A potential explanation could be that these traits are influenced by the large number of other genes associated with quality in wheat, amongst others the known Gli-1 loci coding for gliadins associated with dough quality (Plessis et al. 2013; Sherman et al. 2014; Würschum et al. 2016) or genes that are directly influencing baking quality like wbm (Furtado et al. 2015; Guzmán et al. 2016b). Furthermore, a substantial number of small effect loci without candidate gene information have also been identified in numerous genetic mapping studies suggesting a complex inheritance of quality in wheat (Bordes et al. 2011; Reif et al. 2011; Tsilo et al. 2011, 2013; Cabrera et al. 2015), which indicates the worth of a genomic selection approach that takes both these small and medium effect loci as well as major QTL into account. Accordingly, we observed a substantial benefit of genomic over classical marker-assisted selection for dough rheological traits that is comparable to previous reports about genomic selection for milling and baking quality related traits (Battenfield et al. 2016; Liu et al. 2016).
Genomic selection is commonly used to predict complex quantitatively inherited traits with low and medium heritability in early generations of variety development when the available phenotypic information of important traits like grain yield is limited, and the usage of genomic estimated breeding values could substantially improve genetic gains by supporting breeders in their selection decisions (Auinger et al. 2016; Michel et al. 2016; Sallam and Smith 2016). Although, baking quality traits have usually a high heritability their assessment is often time-consuming, labour-intensive, costly, and too less plant material, i.e., grains are available from each selection candidate in early generations forcing breeders to postpone thoroughly quality testing into later generations of variety development. Genomic selection has on the other hand the great advantage to enable a pre-selection of high performing lines in a much broader population 2–3 years before conducting these costly tests, thereby promoting the selection of lines that combine desirable quality characteristics and grain yield. Integrating major genes as fixed effects into genomic prediction models has furthermore been shown to improve such a genomic selection approach for plant morphological and disease resistance traits (Bernardo 2014; Zhao et al. 2014; Arruda et al. 2016), and we could verify this W-BLUP method with the Glu-1 loci markers that were associated with dough rheological traits in our study. Fine-tuning these W-BLUP models includes an appropriate compromise between marker number and proportion of explained genetic variance, which most likely depends on the breeding material and can to some extent be guided by an appropriate cross-validation in the training population. A major prerequisite is of course that alleles of these QTL have not been fixed in the respective breeding population yet. Extending this idea, Spindel et al. (2016) suggested to integrate de novo mapped marker-trait associations into genomic prediction models. Liu et al. (2016) could though not find any advantage of this method in the analysis of a large hybrid wheat population for quality traits, while other studies reported a significant increase in prediction accuracy of this method (Boeven et al. 2016; Moore et al. 2017). The increase in prediction accuracy using the same population for marker-trait associations discovery and subsequent prediction model validation has been termed the inside trading effect by Arruda et al. (2016), and is the result of selecting predictors before leaving observations that are supposed to be unobserved out, leading consequently to an overfit of the respective prediction models to the training data. Accordingly, the prediction accuracy was even negatively affected when such W-BLUP models that included de novo found markers were used for predicting the three independent validation populations, which firstly showed though great promise in the cross-validation within the estimation set. However, we also did not find an advantage of de novo found markers in the independent validation, which seemed to be promising at first based on the correctly applied cross-validation using a validation set of left-out lines. Possible reasons that these latter de novo found markers could not be validated might be a too small population size for mapping combined with too low marker coverage for accurately mapping the underlying loci, linkage phase changes between training and validation population or the false positive rate. Nevertheless, these issues suggest a prudent interpretation of interesting markers identified in GWAS, and marker-trait associations should be validated with data that was left-out for mapping. We suggest thus favouring known major QTL like Fr-2 for frost tolerance (Erath et al. 2017; Würschum et al. 2017) or TaPHS1 for pre-harvest sprouting (Moore et al. 2017) when predicting complex traits with W-BLUP models in bread wheat and other species. Important genes associated with dough rheological parameters like the Glu-1 loci could nevertheless be readily identified by GWAS (Zheng et al. 2009); however, it often fails to detect rare variants like the wbm gene (Furtado et al. 2015; Bernardo 2016; Guzmán et al. 2016b). Mapping within bi-parental populations using the same marker system as employed in the respective breeding program might thus be a more appropriate strategy for finding new interesting marker-trait associations, which could subsequently be integrated into the genomic selection framework taking advantage of the vast results and experiences gained in QTL mapping during the last two decades.
Enhancing genomic selection by utilizing the association between protein quality and quantity
Apart from the wheat storage protein composition dough rheological and thus baking quality is also determined by protein quantity. Accordingly, we observed strong genetic and phenotypic correlations between protein content with different traits especially dough water uptake, development and extensibility. Whereas the glutenin loci also play a role in the expression of the latter mentioned traits (Zheng et al. 2009; Tsilo et al. 2013), they are to a larger extent controlled by loci associated with the gliadin protein fraction and the gliadin/glutenin ratio (Bekes 2012a, b; Plessis et al. 2013; Sherman et al. 2014) probably causing this strong correlation with the protein content (Osman et al. 2012; Marti et al. 2015).
Some response to indirect selection based on protein content can thus be expected for these quality traits where a marker-assisted selection with Glu-1 loci markers was not applicable, especially for dough extensibility which is highly dependent on the gliadin protein fraction. However, for optimal selection breeders should also consider the complex interaction of protein fractions, which was reflected by the negative relationship between extensibility and resistance to extension found in our and previous studies (Zanetti et al. 2001; Bordes et al. 2008). The resistance to extension generally increases if the glutenin fraction prevails, i.e., with a lower gliadin/glutenin ratio (Melnyk et al. 2012) and baking quality traits like loaf volume are thus dependent on both protein fractions. The same consideration is valid for traits related to dough processing, where genotypes with a prevailing gliadin fraction show a fast water uptake and dough development but an often insufficient dough stability (Weipert 2006). Given these interactions, a simultaneous improvement of extensibility and resistance to extension could be achieved by selecting for dough energy or area under the extensogram curve as an integrated measure that takes both resistances to extension and extensibility into account. Likewise, mixing properties and thus the stability of gluten networks could be improved by utilizing the farino quality number as an integrated index. These two traits showed also a high prediction accuracy using a genomic selection approach, which was superior to indirect selection by the protein content. Although the information of such single point values might be limited (Dobraszczyk and Morgenstern 2003), they could be used to pre-select lines with desirable trait combinations before complete dough rheological profiles can be created for a final selection decision of lines entering variety registration trials.
Notwithstanding, the protein content is an easy to measure rapid test that can be applied to a large number of samples in a short time period. Vast phenotypic information of a large number of lines can thus be expected in wheat breeding programs, which could be used for enlarging the training population size in a genomic selection approach. This is certainly beneficial for wheat breeders as a larger training population size has been shown to increase the prediction accuracy in numerous genomic selection studies (Heffner et al. 2011; Battenfield et al. 2016; Nielsen et al. 2016), and exploiting the strong correlation between dough rheological traits and the highly predictable protein content in a multi-trait prediction model gave accordingly some increase in prediction accuracy if the training population size of the correlated trait protein content exceeded the one of the main trait of interest. On the other hand, multi-trait prediction models performed often very similar to single-trait predictions or were even slightly inferior when the training population was comprised only of lines that were phenotyped in parallel for both traits (Jia and Jannink 2012; Guo et al. 2014; dos Santos et al. 2016). Hence, we suggest that information from the vast number of additional lines phenotyped for correlated traits should be included into multi-trait prediction models to fully exploit the merit of such models. This issue is especially interesting in wheat breeding programs, where breeders can strongly profit from the protein content as a highly heritable as well as easy to phenotype trait which is genetically highly correlated with traits related to baking quality. This advantage could be further extended when the validation or selection population was already phenotyped for the correlated trait (Jia and Jannink 2012; Hayes et al. 2017). This scenario gave a strong increase in prediction accuracy in our study, and could effectively be used to enhance selection in advanced generations when reliable protein content data is already available, though too less plant material for dough rheological test. Additionally, the according genomic predictions could support the choice in combination with the sedimentation value as an intermediate quality analysis step which material should actually be sent to the laboratory for in depth quality analysis. This is another important decision in quality breeding, and we suggest analysing a broad sample from all quality classes in order to avoid a bias and thereby keeping a high prediction accuracy for a successful long-term selection strategy (Zhao et al. 2012).
This study focused on the merit of genomic selection for the genetic improvement of laborious to phenotype dough rheological traits that are related to baking quality in bread wheat. Genomic selection showed a superior performance over marker-assisted and indirect phenotypic selection and could be enhanced by exploiting prior knowledge about the underlying trait genetic architecture for the estimation of genomic breeding values. It was also shown that great care must be taken when upweighting the effect of certain markers in the prediction models and a trait-specific fine-tuning by the proportion of explained genetic variance is advisable. Additional fine-tuning of the predictions was furthermore possible by employing multi-trait prediction models when increasing the training population size of the rapid to phenotype protein content that served in this case as highly heritable and correlated trait. Finally, a genomic selection approach revealed a major benefit over classical selection methods for many quality traits as it would allow a 2–3 years earlier selection for the often costly, labour-intensive and time-consuming assessment of line performance by sophisticated quality tests. This considerable advantage, combined with a higher applicable selection intensity in early generations could support breeders in developing new bread wheat varieties that efficiently combine superior baking quality with comparatively higher grain yield.
Author contribution statement
SM wrote the manuscript and analysed the data. CA and JH supported in the statistical analysis. MG performed the dough rheological analysis and CK supported in the dough rheological interpretation of the results. FL, DE, and BA designed the field trials and collected the phenotypic data in the field. FL and HB initiated and guided through the study. All authors read and approved the final manuscript.
Open access funding provided by University of Natural Resources and Life Sciences Vienna (BOKU). We like to thank Maria Bürstmayr and her team for the tremendous work when extracting the DNA of several hundred wheat lines each year, Herbert Hetzendorfer for managing the collection of the phenotypic data, and Monika Opalo for screening the germplasm with regard to the Glu-1 loci. This research was funded by the EU Eurostars projects “E! 6399 Genomic selection of wheat varieties for robustness, yield and quality” and “E! 8959 Genomic selection for nitrogen use efficiency in wheat”. We thank the anonymous reviewers for their valuable comments and suggestions for improving the manuscript.
Compliance with ethical standards
Conflict of interest
The authors declare no conflict of interest.
The authors declare that the experiments comply with the current laws of Austria.
- AACC (2000) Approved methods of analysis. AACC International, St. PaulGoogle Scholar
- AGES (2016) Österreichische Beschreibende Sortenliste 2016 Landwirtschaftliche Pflanzenarten. http://www.baes.gv.at/fileadmin/user_upload/BSL16_gesamt_internet.pdf. Accessed 6 Feb 2017
- Burgueño J, Cadena A, Crossa J (2000) User’ S guide for spatial analysis of field variety trials using Asreml. CIMMYT, MexicoGoogle Scholar
- Krystkowiak K, Langner M, Adamski T et al (2016) Interactions between Glu-1 and Glu-3 loci and associations of selected molecular markers with quality traits in winter wheat (Triticum aestivum L.) DH lines. J Appl Genet 58:37–48. doi:10.1007/s13353-016-0362-5 CrossRefPubMedPubMedCentralGoogle Scholar
- Laidig F, Piepho H-P, Rentel D et al (2016) Breeding progress, environmental variation and correlation of winter wheat yield and quality traits in German official variety trials and on-farm during 1983–2014. Theor Appl Genet 130:223–245. doi:10.1007/s00122-016-2810-3 CrossRefPubMedPubMedCentralGoogle Scholar
- Mann G, Diffey S, Cullis B et al (2009) Genetic control of wheat quality: interactions between chromosomal regions determining protein content and composition, dough rheology, and sponge and dough baking properties. Theor Appl Genet 118:1519–1537. doi:10.1007/s00122-009-1000-y CrossRefPubMedGoogle Scholar
- Michel S, Ametz C, Gungor H et al (2016) Genomic assisted selection for enhancing line breeding: merging genomic and phenotypic selection in winter wheat breeding programs with preliminary yield trials. Theor Appl Genet 130:363–376. doi:10.1007/s00122-016-2818-8 CrossRefPubMedPubMedCentralGoogle Scholar
- Oury F-X, Hubert C, Annie F et al (2010) The prediction of bread wheat quality: joint use of the phenotypic information brought by technological tests and the genetic information brought by HMW and LMW glutenin subunits Franc. Euphytica 171:87–109. doi:10.1007/s10681-009-9997-1 CrossRefGoogle Scholar
- R development core team (2016) R: a language and environment for statistical computing. http://www.rproject.org/. Accessed 19 Oct 2017
- Weipert D (2006) Fundamentals of Rheology and Spectrometry. In: Popper L, Schäfer W, Freund W (eds) Future of flour: a compendium of flour improvement. ERLING Verlag, Clenze, Germany, pp 117–146Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.