Improved multi-trait prediction of wheat end-product quality traits by integrating NIR predictions

doi:10.21203/rs.3.rs-1822279/v1

Historically, end-product quality testing has been costly and required large flour samples and therefore, was generally implemented in late phases of variety development imposing huge cost on the breeding effort and effectiveness. High genetic correlations of end-product quality traits with higher-throughput and non-destructive testing technologies, such as near infra-red (NIR), could enable early-stage testing and effective selection on these highly valuable traits in a multi-trait genomic prediction model. We studied the impact on end-product quality trait accuracy of adding NIR phenotypes as a second trait in genomic best linear unbiased prediction (GBLUP) to increase training population size of six quality traits (crumb yellowness, water absorption, texture hardness, flour yield, grain protein, flour swelling volume). 1400-1900 bread wheat lines were measured during 2012-2019 in laboratory assays along with NIR records of approximately 27,000 lines. All lines were genotyped with the Infinium™ Wheat Barley 40K BeadChip and imputed using exome sequence data. End-product and NIR phenotypes were genetically correlated (0.5–0.83, except for Flour Swelling Volume 0.19). Prediction accuracies of end-product traits ranged between 0.28–0.64 and increased by 30% through inclusion of NIR predicted phenotypes compared to single trait analysis. There was a high correlation between the magnitude of multi-trait prediction accuracy and genetic correlations between end-product and NIR phenotypes (0.69–0.77). Our forward prediction validation revealed a gradual increase in prediction accuracy when adding more years to the multi-trait model. Overall, we achieved genomic prediction accuracies at a level that enables selection for end-product quality traits early in the breeding cycle.

genomic prediction

multi-trait model

wheat breeding

genomic best linear unbiased prediction

NIR-predictor

forward-prediction

end-product quality traits

Including NIR predicted data into multi-trait prediction models, rises prediction accuracy of genetically correlated end-product quality traits in wheat supporting selection of desirable lines in early breeding cycles.

An exponentially growing human population, rapidly changing and more variable climate presents major risks to food security. Wheat is the most important grain food source for humans and is used for diversity of products ( FAO 2018). World wheat production will need to increase by 60% by 2050 to feed over 9.5 billion people (Gerland et al. 2014) which has to be achieved under an increasingly variable environment, with more extreme weather conditions, and land scarcity (Ceccarelli et al. 2010; Guo et al. 2020).

Current major objectives in wheat breeding include enhanced grain yield, improved agronomic performance and durable disease resistance, all production-oriented traits. While of considerable value, breeding for end-product quality traits is generally a secondary target met through application of quality thresholds for release and consumer acceptance (Battenfield et al. 2016). As such, ensuring the breeding focus is from a whole of value chain perspective through incorporation of end-product quality traits into early generation selection will have a major impact of value creation through breeding. A major obstacle in measuring end-product quality traits, such as flour quality characteristics, is that they often involve time-consuming, labour-intensive, and costly assays that require large grain samples. Thus, sophisticated quality tests have to be postponed to the later phases of variety development. It is common that candidate wheat lines, for which considerable investment in testing has been made, do not pass quality thresholds and are therefore not released to growers. Overcoming these significant limitations requires tools to enable discrimination against undesirable lines earlier in the breeding cycle, make gain in these traits and to save both time and resources.

In complex genetic phenotypes the application of marker-assisted selection (MAS) is limited due to its low power to detect minor quantitative trait loci (QTL), genotype-by-environment interactions and genotype-by-genotype interaction (interaction of QTL and plant genetic background). Numerous linkage mapping studies have shown that a large number of QTL with small effects control most end-product quality traits and, therefore, MAS is unlikely to be able to capture sufficient variance of these traits to be useful in breeding programs (Carter, Arron H., K. Garland-campbell, C. F. Morris 2012; Jernigan et al. 2018; Kristensen 2018; Yang et al. 2020). Genomic selection (GS) can now be routinely implemented through establishment of prediction models based on a suitable training population with both phenotypic and genotypic data. The trained model is used to predict genomic estimated breeding values (GEBVs) of individuals with only genotypic information (Meuwissen et al. 2001). The effectiveness of GS in complex trait breeding, in accelerating breeding cycles and in improving genetic gains per unit of time has been proven (Battenfield et al. 2016; Cericola et al. 2017; Michel et al. 2017; Belamkar et al. 2018). The use of GS is particularly advantageous in early generations, which could support breeders in the faster development of new bread wheat varieties that efficiently combine superior baking quality with higher grain yield.

To overcome difficulties in end-product quality assessments and limited phenotype availability, selection decisions for baking quality improvement have applied indirect phenotypic selection based on correlated traits (Michel et al. 2018) or predictors for grain quality (Hayes et al. 2017a). These require smaller quantities of flour and hence can be measured in earlier cycles of breeding programs. By incorporating benefits from genetic characterizations of correlated phenotypes, multiple trait genomic prediction may be applicable for improving prediction accuracy in genetic evaluations. Montesinos-López et al. (2016) have found higher predictive ability when using correlated traits compared to uncorrelated traits in a multi-trait prediction model. This approach exploits information gained from correlated indicator traits (Tsai et al. 2020). Michel et al (2018) suggested that including information from the vast number of additional lines phenotyped for correlated traits into multi-trait prediction models will improve predictions.

When compared to single-trait analysis, multivariate and/or multi-environment predictive models generally have improved accuracies, especially for traits with high genetic correlations (Guo et al. 2014, 2020; Hayes et al. 2017; Okeke et al. 2017; Azizinia et al. 2020). High genetic correlations are common for quantitative traits such as grain yield and nutritional content in cereal crops (Jia and Jannink 2012; Lozada and Carter 2019; Guo et al. 2020). The advantage of multi-variate models could be further extended where the selection population was already phenotyped for a correlated trait (Jia and Jannink 2012; Hayes et al. 2017a). The use of correlated traits is especially useful for predicting expensive or difficult to measure traits (Lado et al. 2018; Lozada and Carter 2019) such as many quality traits (Battenfield et al. 2016; Bhatta et al. 2020). Predictors from near infrared resonance (NIR) can be considered a correlated trait in multi-trait models (Dowell et al. 2006; Osborne 2006; Hayes et al. 2017a) and, as they are higher throughput, non-destructive and require only a small quantity of whole grain when compared to end-use quality assays, can be easily deployed at scale. Most empirical studies have shown that increasing the size of training population improves prediction accuracy (Heffner et al. 2011; Lorenz et al. 2012; Hoffstetter et al. 2016). We investigate the impact of adding correlated NIR phenotypes to increase training population size on prediction accuracy of wheat quality traits.

Cross-validation schemes can be implemented that mimic real breeder circumstances when predicting lines in environments/years that have not been observed in the field. In forward prediction, previous years are used to predict the following years progeny. This method represents a common breeding situation where previous data are available, but no data is available for phenotypes or environments in the future year/s. Predicting performance of lines for future years is a significant missing data problem and is challenging (Hoffstetter et al. 2016; Jarquín et al. 2017; Juliana et al. 2018). Here we explored the potential of genomic selection to predict future performance using data from 2012–2019 for end-product quality traits in five-fold cross-validation and forward prediction.

The main objectives of this study were to 1) estimate the genetic correlations of end-product quality and NIR derived phenotypes, 2) evaluate the influence of adding NIR derived phenotypes in improving prediction accuracies of end-product quality traits using five-fold cross-validation, and 3) investigate forward prediction of line performance across years in a multi-environment context.

Plant materials and phenotyping

Phenotype records for six quality traits from hexaploid bread wheat breeding lines measured in laboratory assays along with NIR records of approximately 27,000 lines were used in this study. Wheat lines were evaluated for eight years from 2012–2019 grown in NSW, VIC, WA, and SA in more than 150–197 trials of end-product traits and 409–450 trials for NIR records. Lines were evaluated for crumb yellow/blueness (b*, colour), flour water absorption (Wab, %), hardness /particle size index of flour (PSI, %), flour yield (FlrYld, %), protein (Protein, %) and flour swelling volume (FSV, ml/g). A summary of the number of records and lines in each trait and their range and means is provided in Tables 1 and 2.

Laboratory assays were conducted on 2–4kg composite samples assembled by blending all replicated plots of each line per trial. The samples were conditioned at 16% moisture content for 24 hours prior to milling using a Buhler Laboratory Mill (MLU 202). End-use quality tests were performed using approved methods of the American Association of Cereal Chemists International (AACC International 2008). Particle size index (PSI, %), a measure of wheat hardness, was determined after grinding and sieving of grain samples (AACC Method 55-30.01). Flour yield (FlrYld, %), is the percentage by weight recovered of the total products as straight-grade white flour. Flour yellowness (b*) was analyzed with a Minolta Chroma Meter (C-100, Minolta Camera Co. Ltd, Osaka, Japan) (Oliver et al. 1992). Flour swelling volume (FSV, ml/g) test was performed using AACC Method 56-21.01 and water absorption (Wab, %) was measured using a Farinograph (Brabender) following AACC Method 54-21.02. Near-infrared spectroscopy data was also acquired for each of the end-use quality traits. NIR predictions were generated by loading 100 grams of sample into the XDS Rapid Content Analyzer (FOSS). AACC Methods 39-25.01 and 39-70.02 were used to determine NIR predicted protein content and PSI of whole grains using a local calibration.

Genotyping

Genomic DNA was extracted from 6 seeds per sample using a modified CTAB method (Tibbits et al. 2006), where a magnetic bead clean-up step replaced isopropanol precipitation. In brief, the modifications consisted of mixing 120 µL of upper aqueous phase with 120 µL of 10X diluted AMPure XP beads (Beckman Coulter, USA) (diluted with a solution containing 20% PEG and 2.5M NaCl), followed by one wash with 200 µL of 50% DNAzol® ES (Molecular Research Centre, USA) and 42.5% ethanol, and two washes with 70% ethanol. DNA was eluted in 15 µL of 10 mM Tris-HCl, pH 8.0.

The reference panel were genotyped with the Illumina Infinium Wheat Barley 40K XT SNP array v1.0 (Keeble-Gagnère et al. 2021) according to manufacturer’s instructions (Illumina Ltd., CA, United States), with modifications detailed in Keeble-Gagnère et al. 2021. Genotype calling was performed using a custom pipeline (Maccaferri et al. 2019; Keeble-Gagnère et al. 2021).

Genotypes were imputed to exome density following the procedure described in Keeble-Gagnère et al. 2021. Briefly, sporadic missing data was filled in with Beagle v4.1(Browning and Browning 2016), before converting SNP coordinates to positions in the IWGSC RefSeq v2.0 (O’Leary et al. 2016) assembly and imputing to 435,404 SNPs with Minimac 3 (Das et al. 2016) using the reference haplotypes described in Keeble-Gagnère et al. 2021, together with 102 exome-sequenced historical lines from InterGrain’s breeding program. This target set of 435,404 SNPs was defined as the set of SNPs with ${r}^{2}>0.7$ to the set of genotyped SNPs, based on the LD in the InterGrain historical lines. This set of SNPs was further reduced to 330,169 after selecting the SNP in common with the transcriptome genotyping-by-sequencing (tGBS) genotypes after imputation.

Broad-sense heritability

Phenotypic records were edited for possible outliers (mean ± 4SD). Phenotypes were also corrected for fixed effects using a linear mixed model (Eq. 1) in ASReml (Butler et al. 2017), where y was a vector of quality phenotypes, $\mu$ was the trait mean, Trial was fixed effect for environment, a group effect of year, location and nursery.

(1)

Line was fitted as fixed to estimate BLUEs as corrected phenotypes in GEBV estimation and random to estimate the variance due to lines in order to determine the broad sense heritability using:

where ${\sigma }_{g}^{2}$ and ${\sigma }_{e}^{2}$ are line and residual variances, respectively. T and R are the mean number of trials and replications per line.

Genomic predictions

Genomic Restricted Maximum Likelihood (GREML) was used for estimating GEBVs and variance components in the MTG2 software (Lee and Van Der Werf 2016).

Single trait genomic prediction

The single trait GREML model used in this study was a linear mixed model described as y = 1_nµ + Zu + e, where y is the vector of adjusted BLUEs for the trait, µ is the overall mean, 1n is a vector of ones, Z is a design matrix relating records to breeding values, u is a vector of GEBVs, and e is a vector of residual effects. It was assumed that u ~ N(0, Gσ²_g), where σ²_g is additive genetic variance and G is the genomic relationship matrix calculated from the 330,169 SNP markers (Yang et al. 2010), and e ~ N (0, Iσ²_e), where σ²_e is the residual variance and I is the n × n identity matrix. We performed a Principal Component Analysis (PCA) of G using the prcomp function in R (R Core Team 2020) to investigate the population structure in the sample. A plot of first two components was visualised using the ggplot2 package (Wickham 2016) in R. Narrow-sense heritability (h²) was calculated as the ratio of the additive variance to the total phenotypic variance using the GREML model.

Single trait genomic prediction accuracy was evaluated with 5-fold random cross validation. In each cross validation run, one-fold was used as validation set with their phenotype data masked in the analysis (Fig. 1). The accuracy of genomic prediction was calculated as the Pearson correlation coefficient between corrected phenotypic values and GEBVs in the validation subset. This process was repeated 10 times and average prediction accuracy across all folds (50) was reported.

Multi-trait genomic prediction

A basic multi-trait mixed model was used as follows:

$$\left[\begin{array}{c}{\mathbf{y}}_{\text{E}\text{P}}\\ {\mathbf{y}}_{\text{N}\text{I}\text{R}}\end{array}\right]=\left[\begin{array}{cc}{1}_{\text{E}\text{P}}& 0\\ 0& {1}_{\text{N}\text{I}\text{R}}\end{array}\right]\left[\begin{array}{c}{{\mu }}_{\text{E}\text{P}}\\ {{\mu }}_{\text{N}\text{I}\text{R}}\end{array}\right]+\left[\begin{array}{cc}{\mathbf{Z}}_{\text{E}\text{P}}& 0\\ 0& {\mathbf{Z}}_{\text{N}\text{I}\text{R}}\end{array}\right]\left[\begin{array}{c}{\mathbf{u}}_{\text{E}\text{P}}\\ {\mathbf{u}}_{\text{N}\text{I}\text{R}}\end{array}\right]+\left[\begin{array}{c}{\mathbf{e}}_{\text{E}\text{P}}\\ {\mathbf{e}}_{\text{N}\text{I}\text{R}}\end{array}\right]$$

2

where y_EP and y_NIR are the vectors of adjusted phenotypes, 1 is vector of ones, µ _EP and µ_NIR are general means, Z_EP and Z_NIR are the design matrices of breeding values, ${\mathbf{u}}_{\text{E}\text{P}}$ and ${\mathbf{u}}_{\text{N}\text{I}\text{R}}$ are the vectors of genomic breeding values, and e_EP and e_NIR are the vectors of random residual effects, for traits “End-Product” and “NIR”, respectively. Residuals ($\mathbf{e}=\left[{\mathbf{e}}_{\text{E}\text{P}},{\mathbf{e}}_{\text{N}\text{I}\text{R}}\right]$), are assumed to follow a normal distribution, e | R₀ ~ N(0, R₀ ⊗ I) and ${{R}}_{0}=\left[\begin{array}{cc}{\sigma }_{{e}_{EP}}^{2}& {\sigma }_{{e}_{EP-NIR}}\\ {\sigma }_{{e}_{NIR-EP}}& {\sigma }_{{e}_{NIR}}^{2}\end{array}\right]$.

Wheat end-product quality traits are costly to assay directly and therefore cannot be rapidly determined for large number of lines in breeding programs. NIR data can be generated easily across many lines and used to predict end-use traits, making them an interesting test case for examining bivariate models. The benefit of using NIR to predict end-product quality traits was investigated using three different cross-validation scenarios.

Cross validation scenario S1: End-product quality phenotypes of validation lines were masked and predicted with a reference population with equal numbers of end-product and NIR phenotypes. This aimed to assess the effect of multi-trait analysis in improvement of prediction accuracy of end-product quality traits where both traits are available on the same lines (i.e., their assessment is costly and time consuming).
Cross validation scenario S2: Validation lines were masked from both end-product and NIR traits, however, the reference set included all additional lines with only NIR. This aimed to study the effect of extra NIR data on improving prediction accuracy.
Cross validation scenario S3: Assessment of adding NIR on validation lines to increase prediction accuracy, assuming that NIR are easy and cost effective to generate on all lines. In this scenario, end-product phenotypes were masked only in the validation set, while their corresponding NIR measurements were included in the reference set.

Forward genomic prediction

In breeding programs, the aim is to predict future years/environments using previous years’ data. These independent predictions were performed by training the model on previous years data and predicting future environments using GREML. 2012–2015 data was used to predict 2016, 2012–2016 to predict 2017, and 2012–2018 to predict 2019.

Trait heritability and correlation

Phenotype records of wheat lines from different breeding cycles (2012–2019) were used to estimate broad- and narrow-sense heritabilities as well as genomic prediction analyses. Records were unbalanced across years and trials with, in total, 1,400–1,900 end-product wheat lines and 27,000 NIR lines were available. For both end-product and NIR records, the lowest number of individuals were recorded in 2012, while in 2018 and 2019 the largest number of lines were measured. Phenotypic records by trait and year are summarised in Table 1. Phenotypic values exhibited variation suitable for breeding with H² being high for B*, Wab, PSI, moderate for FlrYld and Protein and relatively low for FSV (Table 2). SNPs also captured a high proportion of the phenotypic variance, as demonstrated by h².

Table 1

number of lines used for genomic analysis
Traits	Years	b*	Wab	PSI	FlrYld	Protein	FSV
End-Product	2012	29	29	29	28	18	29
	2013	188	189	189	189	189	188
	2014	169	169	169	169	169	166
	2015	219	223	177	251	219	147
	2016	161	161	194	209	156	155
	2017	166	164	304	309	114	131
	2018	308	308	233	379	308	302
	2019	401	352	469	400	407	284
Sum		1641	1595	1764	1934	1580	1402
NIR predicted	2012	29	29	29	29	29	16
	2013	80	80	80	80	80	5
	2014	92	86	86	86	86	84
	2015	114	114	114	114	114	114
	2016	249	249	249	249	249	249
	2017	439	439	439	439	439	439
	2018	8375	8375	8375	8374	8375	8358
	2019	17,653	17,653	17,651	17,653	17,653	17,653
Sum		27,031	27,025	27,023	27,024	27,025	26,918
b: crumb yellow/blueness; Wab: water absorption; PSI: texture hardness (particle size index); FlrYld: flour yield; Protein: grain protein; FSV*: flour swelling volume

Table 2

Range of quality traits (Minimum: Min; Mean and Maximum: Max), variance components (genetic variance:${\sigma }_{g}^{2}$ and residuals:${ \sigma }_{e}^{2}$), broad-sense heritability (H²) and narrow-sense heritability (h²) estimated with the SNP of lines analysed from 2012 to 2019.
Trait	Min	Mean	Max	${\sigma }_{g}^{2}$	${\sigma }_{e}^{2}$	H²	h²(SNP)
*b (colour)**	5.96	9.62	18.70	0.86	0.13	0.87	0.68
Wab (%)	48.90	60.74	71.10	1.9	0.91	0.68	0.67
PSI (%)	7.00	15.68	36.80	3.64	1.47	0.71	0.68
FlrYld (%)	65.03	74.14	80.61	0.93	0.51	0.65	0.54
Protein (%)	7.55	11.47	15.80	0.19	0.30	0.40	0.40
FSV (ml/g)	7.14	12.31	29.00	8.42	8.31	0.50	0.27

For use in genomic prediction genetic correlations between end-product quality and NIR predicted traits were estimated using a multi-trait model (Table 3). All end-product traits were strongly correlated with NIR traits, with the exception of FSV, which had a low correlation (0.19), and Protein, with a moderate correlation. Phenotypic correlations for the trait sets were high for most traits, except for FlrYld and FSV (Table 3).

Table 3

Phenotypic and genetic correlations between end-product and NIR predicted quality traits.
Trait	Genetic correlation	Phenotypic correlation
b*	0.83 ± 0.02	0.46 ± 0.01
Wab	0.86 ± 0.02	0.67 ± 0.01
PSI	0.82 ± 0.03	0.69 ± 0.01
FlrYld	0.67 ± 0.04	0.26 ± 0.01
Protein	0.50 ± 0.06	0.85 ± 0.01
FSV	0.19 ± 0.08	0.12 ± 0.02

Population structure

A principal component analysis was used to investigate population structure. The two first components jointly explained 56.2% of the variation (Fig. 2). The plot shows the relationship between groups of individuals used in the analysis. There is slightly denser concentration of lines in the lower right corner of the chart for three groups (i.e., end-product only, NIR only, both phenotypes). However, there is a complete overlap of lines in all groups. In other words, lines form diverse backgrounds have been recorded for both NIR and end-product performance. This close genetic relationship between lines with end-product assay data and those with NIR data are expected to improve prediction accuracy and result in more accurate genomic breeding values.

Single trait prediction

Single trait BLUE prediction accuracies for end-product quality traits were highly variable ranging from 0.20 for FSV to 0.56 for b* (Table 4). Correlation between trait heritability and prediction accuracies was observed with traits of lower heritability tending towards lower prediction accuracies. The correlations of broad- and narrow-sense heritabilities with single trait prediction accuracies were 0.62 and 0.54, respectively.

Multi-trait prediction

We assessed three scenarios that included NIR phenotypes to predict end-product quality traits in a different manner. In the first scenario (S1), NIR data were limited to the lines which had been assessed for end-product quality traits, while in the second scenario (S2) a large number of additional lines with NIR predicted records were added to the training population. In both S1 and S2, end-product measured phenotypes of the validation set and their corresponding NIR data were removed from training, however in the third scenario (S3), NIR predicted data of lines in the validation set were included. The aim of this approach was to investigate whether early NIR phenotyping of candidate lines improves prediction accuracy of end-product traits. Five-fold cross validation within population was used to estimate prediction accuracy from GBLUP. In S1, adding NIR data increased the accuracy of predictions in all end-product quality traits compared to single trait prediction, except for Protein, in which there was a reduction in prediction accuracy and b*, with no significant change (Fig. 3). In this scenario, PSI showed the highest improvement followed by Wab.

To identify whether the inclusion of additional NIR phenotypes in the model could further improve the predictive performance of end-product quality traits, S2 and S3 scenarios were studied. Adding more information through inclusion of NIR data improved mean prediction accuracy of all traits (Fig. 3), although traits responded differently depending on the scenario. For most multi-trait scenarios with inclusion of NIR data prediction accuracy increased. Protein content was an exception with multi-trait prediction inferior to the single trait model in all scenarios. Wab and PSI showed the highest accuracy increase in S2 and S3 when compared to single trait scenarios, with an average accuracy of 0.40 and 0.50 for Wab and PSI, while their corresponding single trait accuracy was 0.26 and 0.27, respectively. FlrYld and b* also positively responded to including more NIR data, where in S3 an accuracy of 0.41 (Flour Yield) and 0.64 (b*) were observed compared to the single trait model (0.35, 0.56, respectively).

The scale of the genetic correlation between end-product and NIR predicted traits was reflected in the level of genomic prediction accuracy. For example, genetic correlations and prediction accuracies were high for Wab and PSI. There was a high correlation between the magnitude of prediction accuracy in multi-trait scheme and genetic correlation of end-product and NIR derived data (0.69–0.77). Correlations between broad- and narrow-sense heritability with multi-trait prediction accuracies in different scenarios ranged between 0.53 to 0.60 and 0.43 to 0.47, respectively. Broad sense and narrow sense heritability also had a correlation of 0.58 and 0.45, respectively, with mean prediction accuracies across all three scenarios.

Forward prediction

Despite some fluctuations, there was a gradual increase in prediction accuracy by increasing the size of the training population through the addition of phenotypic data over years, though PSI accuracy was variable (Fig. 4). All trait prediction accuracies increased by 13–30% compared to mean accuracy across years between 2016–2019 emphasising the effect of training population size. The highest forward prediction accuracy was again observed for b*, which was similar to single trait and multi-trait scenarios.

Forward prediction in a univariate scheme showed similar average prediction accuracy across years with a lower upward trend than the multi-trait, demonstrating more accurate predictions were achieved when using correlated NIR predictor traits to increase training size (Table 4).

Table 4

Mean forward prediction accuracies across 2016–2019 in single and multi-trait analysis.
	b*	Wab	PSI	Flour Yield	Protein	FSV
Multi-trait	0.54	0.09	0.42	0.17	0.21	0.16
Single trait	0.49	0.07	0.42	0.17	0.19	0.15

Table 5

prediction accuracies for end-product quality traits with and without using NIR predicted data in reference set.
	Multi-trait prediction			Single trait prediction
Trait	S1	S2	S3	Single trait prediction
b*	0.56	0.62	0.64	0.56
Wab	0.31	0.38	0.41	0.26
PSI	0.41	0.54	0.45	0.27
FlrYld	0.36	0.42	0.41	0.35
Protein	0.28	0.31	0.30	0.35
FSV	0.23	0.26	0.28	0.20

Wheat end-product quality traits have been difficult to include in early selection as testing is labour intensive, costly and requires large grain samples which are usually not available until late in the breeding cycle. These limitations cause breeding program inefficiency with high yielding lines, having incurred significant field trial resources, often being discarded late in the breeding cycle based on quality testing results. Genomic selection offers a new way to include end-use quality traits into the early breeding cycle, however, genomic selection success requires having training populations of sufficient size for accurate prediction and some of the limitations listed above also make the accumulation of adequate training populations problematic. In this paper we demonstrate two effective strategies to increasing the power of genomic selection. First, by increasing the training population size using correlated NIR predictions of end-product quality in a multi-trait model and, second, by adding observations to the training population in a breeding relevant forward validation scenario across years. Overall, the multi-trait models were effective in increasing prediction accuracy.

Heritability and genetic correlations

The broad-sense heritability of end-product quality traits evaluated varied from 0.40 to 0.87, with most of them having a value above 0.60. These heritability estimates for quality traits are consistent with values reported in other studies (Carter, et al. 2012; Jernigan et al. 2018; Michel et al. 2018; Kristensen et al. 2019). The lower heritability of traits like FSV may be representative of a complex and polygenic underlying architecture (Hayes et al. 2017b). Intermediate to high heritability estimates suggested that most of the variation of the trait is genetic (Tsai et al. 2020). Genomic selection is considered suitable for predicting highly heritable traits as the models are generally able to capture much of the additive genetic variation given suitable training populations. Multi-trait models can improve prediction performance of traits with lower genetic correlations, where correlated predictors have higher heritability (Jia and Jannink 2012; Guo et al. 2014). However, for very complex polygenic traits, the multi-trait model did not show advantage over single trait model even with high heritability (Jia and Jannink 2012; Lado et al. 2018)

End-product and NIR predicted traits showed a high genetic correlation. This is consistent with other studies, where genetic correlation of end-quality traits and their corresponding NIR prediction were notably high (Hayes et al. 2017a). High correlations have been shown to have an impact on increasing prediction accuracy of traits (Hayes et al. 2017b; Michel et al. 2018; Azizinia et al. 2020; Tsai et al. 2020; Guo et al. 2020). Studies also have shown that closely related training and validation populations result in accurate genomic breeding values (Habier et al. 2007; Daetwyler et al. 2008; Meuwissen 2009).

A lower improvement in prediction accuracies were obtained for traits with lower genetic correlations of end-product and NIR data, a result similar to the those reported by Hayes et al. (2017a). Our results also align with previous studies that reported that intermediate to low correlations of traits would not positively affected prediction accuracy in multi-trait schemes (Calus and Veerkamp 2011; Lado et al. 2018). Jia and Jannink (2012) indicated that, for very complex traits with low heritability, multi trait models would have small advantage over other models. However, considering the positive effect of an increased NIR training population on prediction accuracy of traits with lower genetic correlation, our hypothesis was that using a larger reference population may increase prediction power in the validation population. Overall, our results suggest that multi-trait models using correlated attributes does improve the accuracy of genomic prediction (Guo et al. 2014; Rutkoski et al. 2016) and this increases the potential uses of NIR derived data to predict wheat end-product quality traits (Dowell et al. 2006; Hayes et al. 2017a). Increasing the size of the reference set through the addition of both end-product quality and NIR phenotypes is recommended to improve end-product predictions, especially for traits with moderate and low correlations (e.g., FlrYld, Protein, and FSV).

Multi-trait prediction

Multi-trait models combine information of correlated traits to deliver higher prediction accuracy thus are useful to enhance prediction accuracy of favourable traits which are difficult to be assessed in early cycles of the programme. We showed higher prediction accuracy across quality traits in multi-trait models (S1, S2 and S3), which is consistent with Lado et al. (2018) who reported increased prediction ability of quality traits using information of correlated attributes compared to single trait prediction model.

Using NIR predicted data in multi-trait scheme, prediction accuracies showed on average a 0.05 to 0.2 (8–73%) rise across all scenarios and traits (excluding Protein). This is comparable to Lado et al. (2018) who reported a 60–100% increase in prediction accuracy of target quality traits using correlated traits in different multi-trait models. Hayes et al. (2017a) also reported an increase of 0.03 to 0.45 (6 to 300%) in soft and hard wheat for grain hardness, grain protein, b* and water absorption in different multi-trait scenarios, although the strategy used to predict performance was different. Overall, the addition of NIR predicted data substantially improved trait prediction accuracy and it is worthwhile perusing as strategy in breeding programmes.

Higher accuracies of different multi-trait scenarios also suggested that an expanded reference set of predictors would improve prediction accuracy of correlated traits (S2 and S3). While using NIR data of the reference population (S1) showed an 11% increase in prediction accuracy over the single trait model, increasing NIR measures as a correlated trait in the second scenario (S2) had a 33% increase over the single model. Including NIR measures of candidate lines in the training population (S3) also showed a 30% increased prediction accuracy across all traits but did not considerably affect prediction accuracy compared to S2. The additional NIR information in S3 compared to S2 involved adding only a small number of observations. It could be that addition of a relatively small number of NIR predicted phenotypes would be unlikely to substantially affect the prediction accuracy. particularly, focusing on improving genetic correlation of NIR predictions with end-product assays and therefore increasing NIR records will positively affect prediction accuracies in target traits. The findings of the present study can be potentially applied in plant breeding to achieve more accurate and improved predictions compared to single trait predictions for end-product quality traits of high importance.

Forward prediction

Predicting the performance of new individuals lacking phenotypes is always a challenge in breeding programs. We demonstrated continuous increases in forward prediction accuracy across years as the size of the training population increased. The importance of large training population sizes in increasing forward prediction accuracy has been previously demonstrated (Battenfield et al. 2016; Yao et al. 2018; Michel et al. 2019a). Jarquín et al (2017) indicated difficulty in predicting future years for grain yield, with lower prediction accuracy compared to cross validation. In our study, cross-validation in all scenarios, had higher prediction accuracy compared to mean forward prediction accuracy. Higher accuracies from cross-validation arise from over-inflation in the method and the fact that the likelihood of assignment of close relatives in test and training sets can cause inflation of prediction accuracy (Rutkoski et al. 2015). Lower average forward prediction accuracies compared to cross validation can be due to random selection of training and test sets in the latter method, which results in better estimation of environmental variation (Rutkoski et al. 2015; Battenfield et al. 2016). In forward prediction genotype by environment interactions may play a major role since the training set is not representative of the test set. Prediction accuracy of 2019 in forward prediction, on the other hand was higher suggesting the importance of including bigger training sets and genotype-environment interaction which has been highlighted in other research studies (Crossa et al. 2014). Prediction ability could be improved by optimizing different factors affect GS accuracy (Daetwyler et al. 2008). Increasing training population size and updating genomic selection models every year is recommended for the best prediction accuracy (Michel et al. 2019b). Furthermore, integrating GxE and spatial effects in the model can improve prediction accuracy. Extending multi-trait model to the multi-trait and multi-environment model which takes into account the interaction of trait-genotype-environment is the approach to enhance prediction ability of complex traits (Montesinos-lópez et al. 2016; Tsai et al. 2020). Given the promising results of forward prediction for the quality traits, implementing GS will enable breeders to make selection for those quality traits in a larger population at earlier stages, thereby enhancing selection efficiency (Fiedler et al. 2017). Additionally, eliminating less favourable lines can be done in early cycles of breeding programs based on GEBVs alone. This approach will remove costs from breeding programs by only keeping potential candidate lines that will be of sufficient end-product quality for relevant markets.

We investigated the potential benefit of including NIR predicted data into multi-trait models for predicting six end-product quality traits in a wheat breeding program. In general, the multi-trait models combining direct measures of end-product quality traits with their NIR predictions had higher predictive accuracy than their respective single-trait models. The increased prediction accuracy was likely driven by the increased reference set size that arose from the addition of NIR predictions. The prediction accuracies we achieved using the combined data were at a level that breeding selections could be confidently applied earlier in the breeding process to increase breeding efficiency and genetic gain.

Acknowledgments The authors are thankful to all InterGrain staff who collected the data used in this study. The authors express their appreciation for Agriculture Victoria team members for genotyping. The first author is grateful to Dr. Majid Khansefid for his support in using the MTG2 program.

Authors contribution statement Author contributions: SA analysed data and wrote the manuscript; DM, AR, JG, HR, DMo, provided samples and phenotypes, KF, GK, MH, genotyped samples; MH, JT, and HD supervised the study, HD edited the manuscript; all authors have read and approved the manuscript.

Conflict of interest The authors declare that there is no conflict of interest.

Competing interests The authors declare no competing interests.

Data Availability The datasets analysed during the current study are not publicly available due to third party commercial restrictions but will be made available from the corresponding author on reasonable request and with permission of Intergrain Pty Ltd.

AACC International (2008) Approved Methods of Analysis. 11th Edn. Paul, MN AACC Int. St. 11th Edn. Paul, MN AACC Int. St
Azizinia S, Bariana H, Kolmer J et al (2020) Genomic Prediction of Rust Resistance in Tetraploid Wheat under Field and Controlled Environment Conditions. Agronomy 10:1843. https://doi.org/10.3390/agronomy10111843
Battenfield SD, Guzmán C, Gaynor RC et al (2016) Genomic Selection for Processing and End-Use Quality Traits in the CIMMYT Spring Bread Wheat Breeding Program. Plant Genome 9. https://doi.org/10.3835/plantgenome2016.01.0005
Belamkar V, Guttieri MJ, Hussain W et al (2018) Genomic Selection in Preliminary Yield Trials in a Winter Wheat Breeding Program. 8:2735–2747. https://doi.org/10.1534/g3.118.200415
Bhatta M, Gutierrez L, Cammarota L, et al (2020) Multi-trait Genomic Prediction Model Increased the Predictive Ability for Agronomic and Malting Quality Traits in Barley(Hordeum vulgare L.). G3: Genes|Genomes|Genetics 10:1113–1124. https://doi.org/10.1534/g3.119.400968
Browning BL, Browning SR (2016) Genotype Imputation with Millions of Reference Samples. Am J Hum Genet. https://doi.org/10.1016/j.ajhg.2015.11.020. 98:
Butler DG, Cullis BR, Gilmour AR et al (2017) ASReml-R Reference Manual Version 4. 1–176
Calus MPL, Veerkamp RF (2011) Accuracy of multi-trait genomic selection using different methods. Genet Sel Evol 43:1–14. https://doi.org/10.1186/1297-9686-43-26
Carter AH, Garland-campbell K, Morris CF KKK (2012) Chromosomes 3B and 4D are associated with several milling and baking quality traits in a soft white spring wheat (Triticum aestivum L.) population. Theor Appl Genet 124:1079–1096. https://doi.org/10.1007/s00122-011-1770-x
Ceccarelli S, Grando S, Maatougui M et al (2010) Plant breeding and climate changes. J Agric Sci 148:627–637. https://doi.org/10.1017/S0021859610000651
Cericola F, Jahoor A, Orabi J et al (2017) Optimizing Training Population Size and Genotyping Strategy for Genomic Prediction Using Association Study Results and Pedigree Information. A Case of Study in AdvancedWheat Breeding Lines. 1–20. https://doi.org/10.7910/DVN/QUTXAP.Funding
Crossa J, Pe´rez P, Hickey J et al (2014) Genomic prediction in CIMMYT maize and wheat breeding programs. Heredity (Edinb) 112:48–60. https://doi.org/10.1038/hdy.2013.16
Daetwyler HD, Villanueva B, Woolliams JA (2008) Accuracy of Predicting the Genetic Risk of Disease Using a Genome-Wide Approach. PLoS ONE 3:e3395. https://doi.org/10.1371/journal.pone.0003395
Das S, Forer L, Schönherr S et al (2016) Next-generation genotype imputation service and methods. Nat Genet 48. https://doi.org/10.1038/ng.3656
Dowell FE, Maghirang EB, Xie F et al (2006) Predicting wheat quality characteristics and functionality using near-infrared spectroscopy. Cereal Chem 83:529–536. https://doi.org/10.1094/CC-83-0529
FAO (2018) FAO Cereal Supply and Demand Brief. In: Crop Prospect. Food Situat. http://www.fao.org/worldfoodsituation/wfs-home/csdb/en/
Fiedler JD, Salsman E, Liu Y et al (2017) Genome-Wide Association and Prediction of Grain and Semolina Quality Traits in Durum Wheat Breeding Populations. 1–12. https://doi.org/10.3835/plantgenome2017.05.0038
Gerland P, Raftery AE, Ševčíková H et al (2014) World population stabilization unlikely this century. Sci (80-) 346:234–237. https://doi.org/10.1126/science.1257469
Guo G, Zhao F, Wang Y et al (2014) Comparison of single-trait and multiple-trait genomic prediction models. BMC Genet 15:1–7. https://doi.org/10.1186/1471-2156-15-30
Guo J, Khan J, Pradhan S et al (2020) Multi-Trait Genomic Prediction of Yield-Related Traits in US Soft Wheat under Variable Water Regimes. Genes (Basel) 11:1–25. https://doi.org/10.3390/genes11111270
Habier D, Fernando RL, Dekkers JCM (2007) The impact of genetic relationship information on genome-assisted breeding values. Genetics 177:2389–2397. https://doi.org/10.1534/genetics.107.081190
Hayes BJ, Panozzo J, Walker CK et al (2017a) Accelerating wheat breeding for end-use quality with multi-trait genomic predictions incorporating near infrared and nuclear magnetic resonance-derived phenotypes. Theor Appl Genet 130:2505–2519. https://doi.org/10.1007/s00122-017-2972-7
Hayes BJ, Panozzo J, Walker CK et al (2017b) Accelerating wheat breeding for end-use quality with multi-trait genomic predictions incorporating near infrared and nuclear magnetic resonance-derived phenotypes. Theor Appl Genet 130:2505–2519. https://doi.org/10.1007/s00122-017-2972-7
Heffner EL, Jannink J-L, Sorrells ME (2011) Genomic Selection Accuracy using Multifamily Prediction Models in a Wheat Breeding Program. Plant Genome 4:65–75. https://doi.org/10.3835/plantgenome2010.12.0029
Hoffstetter A, Cabrera A, Huang M, Sneller C (2016) Optimizing Training Population Data and Validation of Genomic Selection for Economic Traits in Soft Winter Wheat. 6:2919–2928. https://doi.org/10.1534/g3.116.032532
Jarquín D, Lemes C, Gaynor RC et al (2017) Increasing Genomic-Enabled Prediction Accuracy by Modeling Genotype ´ Environment Interactions in Kansas Wheat. https://doi.org/10.3835/plantgenome2016.12.0130
Jernigan KL, Godoy JV, Huang M et al (2018) Genetic Dissection of End-Use Quality Traits in Adapted Soft White Winter Wheat. 9:1–15. https://doi.org/10.3389/fpls.2018.00271
Jia Y, Jannink J-L (2012) Multiple-Trait Genomic Selection Methods Increase Genetic Value Prediction Accuracy. Genetics 192:1513–1522. https://doi.org/10.1534/genetics.112.144246
Juliana P, Singh RP, Poland J et al (2018) Prospects and Challenges of Applied Genomic Selection — A New Paradigm in Breeding for Grain Yield in Bread Wheat. 1–17. https://doi.org/10.3835/plantgenome2018.03.0017
Keeble-Gagnère G, Pasam R, Forrest KL et al (2021) Novel Design of Imputation-Enabled SNP Arrays for Breeding and Research Applications Supporting Multi-Species Hybridization. Front Plant Sci 12:1–16. https://doi.org/10.3389/fpls.2021.756877
Kristensen PS (2018) Genome-Wide Association Studies and Comparison of Models and Cross-Validation Strategies for Genomic Prediction of Quality Traits in Advanced Winter Wheat Breeding. 9. https://doi.org/10.3389/fpls.2018.00069
Kristensen PS, Jensen J, Andersen JR et al (2019) Genomic prediction and genome-wide association studies of flour yield and alveograph quality traits using advanced winter wheat breeding material. Genes (Basel) 10:1–19. https://doi.org/10.3390/genes10090669
Lado B, Vázquez D, Quincke M et al (2018) Resource allocation optimization with multi – trait genomic prediction for bread wheat (Triticum aestivum L.) baking quality. Theor Appl Genet 131:2719–2731. https://doi.org/10.1007/s00122-018-3186-3
Lee SH, Van Der Werf JHJ (2016) MTG2: An efficient algorithm for multivariate linear mixed model analysis based on genomic information. Bioinformatics 32:1420–1422. https://doi.org/10.1093/bioinformatics/btw012
Lorenz AJ, Smith KP, Jannink J-L (2012) Potential and Optimization of Genomic Selection for Fusarium Head Blight Resistance in Six-Row Barley. Crop Sci 52:1609–1621. https://doi.org/10.2135/cropsci2011.09.0503
Lozada DN, Carter AH (2019) Accuracy of Single and Multi-Trait Genomic Prediction Models for Grain Yield in US Pacific Northwest Winter Wheat. Crop Breeding. Genet Genomics 2:1–23. https://doi.org/10.20900/cbgg20190012
Maccaferri M, Harris NS, Twardziok SO et al (2019) Durum wheat genome highlights past domestication signatures and future improvement targets. Nat Genet 51:885–895. https://doi.org/10.1038/s41588-019-0381-3
Meuwissen TH (2009) Accuracy of breeding values of “unrelated” individuals predicted by dense SNP genotyping. Genet Sel Evol 41:1–9. https://doi.org/10.1186/1297-9686-41-35
Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-wide dense marker maps. Genetics 157:1819–1829. https://doi.org/doi: 10.1093/genetics/157.4.1819
Michel S, Ametz C, Gungor H, Akgöl B (2017) 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. https://doi.org/10.1007/s00122-016-2818-8
Michel S, Kummer C, Gallee M et al (2018) Improving the baking quality of bread wheat by genomic selection in early generations. Theor Appl Genet 131:477–493. https://doi.org/10.1007/s00122-017-2998-x
Michel S, Löschenberger F, Ametz C et al (2019a) Simultaneous selection for grain yield and protein content in genomics-assisted wheat breeding. Theor Appl Genet 132:1745–1760. https://doi.org/10.1007/s00122-019-03312-5
Michel S, Löschenberger F, Ametz C et al (2019b) Combining grain yield, protein content and protein quality by multi-trait genomic selection in bread wheat. Theor Appl Genet 132:2767–2780. https://doi.org/10.1007/s00122-019-03386-1
Montesinos-lópez OA, Montesinos-lópez A, Crossa J et al (2016) GENOMIC SELECTION A Genomic Bayesian Multi-trait and Multi-environment Model. 6:2725–2744. https://doi.org/10.1534/g3.116.032359
O’Leary NA, Wright MW, Brister JR et al (2016) Reference sequence (RefSeq) database at NCBI: Current status, taxonomic expansion, and functional annotation. Nucleic Acids Res 44. https://doi.org/10.1093/nar/gkv1189
Okeke UG, Akdemir D, Rabbi I et al (2017) Accuracies of univariate and multivariate genomic prediction models in African cassava. Genet Sel Evol 49:1–10. https://doi.org/10.1186/s12711-017-0361-y
Oliver JR, Blakeney AB, Allen HM (1992) Measurement of flour color in color space parameters. Cereal Chem 69:546–551
Osborne BG (2006) Applications of near Infrared Spectroscopy in Quality Screening of Early-Generation Material in Cereal Breeding Programmes. J Near Infrared Spectrosc 14:93–101. https://doi.org/10.1255/jnirs.595
R Core Team (2020) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. In: R A Lang. Environ. Stat. Comput. R Found. Stat. Comput. Vienna, Austria. http://www.r-project.org/
Rutkoski J, Poland J, Mondal S et al (2016) Canopy temperature and vegetation indices from high-throughput phenotyping improve accuracy of pedigree and genomic selection for grain yield in wheat. G3 Genes, Genomes. Genet 6:2799–2808. https://doi.org/10.1534/g3.116.032888
Rutkoski J, Singh RP, Huerta-Espino J et al (2015) Efficient Use of Historical Data for Genomic Selection: A Case Study of Stem Rust Resistance in Wheat. Plant Genome 8:1–10. https://doi.org/10.3835/plantgenome2014.09.0046
Tibbits JFG, McManus LJ, Spokevicius AV, Bossinger G (2006) A rapid method for tissue collection and high-throughput isolation of genomic DNA from mature trees. Plant Mol Biol Report 24. https://doi.org/10.1007/BF02914048
Tsai H-Y, Cericola F, Edriss V et al (2020) Use of multiple traits genomic prediction, genotype by environment interactions and spatial effect to improve prediction accuracy in yield data. PLoS ONE 15:e0232665. https://doi.org/10.1371/journal.pone.0232665
Wickham H (2016) ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag, New York
Yang J, Benyamin B, Mcevoy BP et al (2010) Common SNPs explain a large proportion of heritability for human height. 42:565–569. https://doi.org/10.1038/ng.608
Yang Y, Chai Y, Zhang X et al (2020) Multi-Locus GWAS of Quality Traits in Bread Wheat: Mining More Candidate Genes and Possible Regulatory Network. 11:1–19. https://doi.org/10.3389/fpls.2020.01091
Yao J, Zhao D, Chen X et al (2018) Use of genomic selection and breeding simulation in cross prediction for improvement of yield and quality in wheat (Triticum aestivum L.). Crop J 6:353–365. https://doi.org/10.1016/j.cj.2018.05.003

Improved multi-trait prediction of wheat end-product quality traits by integrating NIR predictions

Status:

Version 1

Abstract

Figures

Key Message

Introduction

Materials And Methods

Plant materials and phenotyping

Genotyping

Broad-sense heritability

Genomic predictions

Single trait genomic prediction

Multi-trait genomic prediction

Forward genomic prediction

Results

Trait heritability and correlation

Population structure

Single trait prediction

Multi-trait prediction

Forward prediction

Discussion

Heritability and genetic correlations

Multi-trait prediction

Forward prediction

Conclusion

Declarations

References

Status:

Version 1