Linkage mapping and whole-genome predictions in canola (Brassica napus) subjected to differing temperature treatments

Additional keywords: breeding methods, climate change, predictive breeding.

Breeding for heat-stress tolerance in cool-season crops such as canola (Brassica napus L.) is necessary to ensure continued genetic gain in a warming environment. The impact of increased temperatures on canola was demonstrated through an analysis of growers’ canola yields over 34 years in Saskatchewan, Canada, which found that for every 1°C increase in temperature, yield declined by 75.5 kg ha^–1(Kutcher et al. 2010). The threshold maximum temperature at which canola yields have been found to decline was reported to be 29.5°C (Morrison and Stewart 2002). In 2016, canola was planted on 8.23 Mha in Canada (Statistics Canada 2016). Using this information, a 1°C increase in mean temperature could translate to an annual loss in canola production of >0.5 Mt in Canada. Progress has been made with respect to heat tolerance of crops such as broccoli (Brassica oleracea L.), cowpea (Vigna unguiculata (L.) Walp.) and potato (Solanum tuberosum L.) through conventional breeding (Driedonks et al. 2016). Considering this progress and the increased availability of genomic data, breeders may exploit this information to decrease the time it takes to discover and integrate a new trait such as tolerance to increased temperatures into their germplasm and cultivars.

Genetic linkage mapping when combined with phenotypic data enables the discovery of quantitative trait loci (QTLs) (Xu 2010). These QTLs enable the implementation of marker-assisted selection (MAS) or marker-assisted backcrossing breeding strategies to retain the trait of interest through generations without costly phenotypic screening (Collard et al. 2005). For QTLs to be highly effective within breeding programs, they must explain a significant proportion of the variation and be stable across environments and populations (Collard et al. 2005). QTLs discovered for abiotic stress traits such as waterlogging and drought have been mapped in B. napus (Li et al. 2014), although the maximum amount of phenotypic variation explained by any one QTL was relatively low, with a maximum of 12.56%. A review of abiotic stress QTLs by Collins et al. (2008) lists a multitude of QTLs; however, attempting to utilise many QTLs with minor effects for genetic gain of a single trait is inefficient and difficult to manage within a breeding program (Bernardo 2016).

A review by Jha et al. (2014), lists numerous QTLs responsible for conferring heat tolerance within many species including Arabidopsis thaliana L., barley (Hordeum vulgare L.), maize (Zea mays L.), potato, rice (Oryza sativa L.), sorghum (Sorghum bicolor (L.) Moench), tomato (Lycopersicon esculentum Mill.) and wheat (Triticum aestivum L.), thereby demonstrating the quantitative nature of heat-stress response. With increased access to whole-genome marker data, breeders are able to employ whole-genome prediction (WGP) for highly quantitative traits (Jannink et al. 2010). Unlike QTL discovery using linkage mapping, WGP models do not have stringent statistical significance limits and are therefore able to utilise all genetic signals detected through the analysis (Meuwissen et al. 2001). With the use of WGP, breeders are able to predict performance of untested genotypes and therefore decrease the time necessary to identify favourable allelic combinations for highly quantitative traits within the germplasm (Heffner et al. 2010). In maize, WGP has been more effective for improving grain yield under drought stress than phenotypic selection alone (Vivek et al. 2017). WGP has been applied to B. napus doubled-haploid (DH) populations for yield and quality traits with varying degrees of success, depending on the population structure, number of individuals within the population, marker density and trait heritability (Würschum et al. 2014; Zou et al. 2016). With respect to heat stress, Bita and Gerats (2013) discuss the difficulty in using low-effect QTLs for improving quantitative traits such as tolerance to increased temperatures and suggest WGP as a promising alternative.

This research utilised a DH mapping population of 148 genotypes derived from two canola inbred lines that differed in their ability to set seed under heat stress (Koscielny et al. 2018b). The objectives of this research were to: (i) identify QTLs responsible for conferring tolerance to increased temperatures, and (ii) utilise this population to assess the prospects of WGP for improving stress tolerance. Both MAS and WGP are useful breeding strategies, but selecting the optimal strategy is imperative in a competitive research area such as plant breeding. This research applies both approaches across control and high-temperature environments and discusses strengths and challenges of each strategy.

Seed from the F₁ of a cross between two B. napus DH genotypes, PB36 (heat tolerant) and PB56 (heat susceptible), as determined during a prior growth-chamber experiment (Koscielny et al. 2018b) were used as donors to create a DH population. These 148 microspore-derived DH genotypes were seed-increased in a greenhouse in Carman, Manitoba, Canada, to provide sufficient seed for field trials.

Field experiments were set up in a split-plot randomised complete block design with two replications. The main effect was planting date with an early and a late planting date, whereby the reproductive period of the second planting was shifted to a later time in order to align with increased summer temperatures. The subplot was genotype. The field experiments were set up in Carman, Manitoba (49.4903, –98.0027) in 2015 and 2017, and in Viluco, Chile (–33.7965, –70.8075) in 2015. Carman soil was of sandy loam texture, and both treatments received the same fertiliser blend of nitrogen, phosphorus, potassium and sulfur (N-P-K-S, 89-34-0-22 kg ha^–1) applied before planting. Viluco soil was of a loam texture, and both treatments had the same fertiliser blend of N-P-K-S (125-61-69-0 kg ha^–1) applied through the irrigation at 20 and 40 days after planting. The genotypes were seeded in a 3-m row with a 0.5-m spacing between rows. The 2015 and 2017 Carman sites were rainfed, and volumetric water content to a depth of 0.18 m was collected weekly using a TDR 100 Soil Moisture Meter (Spectrum Technologies, Aurora, IL, USA). The 2015 Viluco site was irrigated with drip-tape irrigation supplying 288 mm water; in addition, rainfall of 83.5 mm and 6.2 mm occurred for the first and second planting dates, respectively. Irrigation applications were made every 5 days to ensure that no differences in treatment were due to soil moisture.

Temperatures were collected every 30 min with a 2000 Series Watchdog weather station (Spectrum Technologies). For the second planting date, the average daily mean, maximum and minimum temperatures were higher at all environments, and therefore this planting date will be referred to as the stress treatment (ST), with the first planting date being referred to as the control treatment (CT) (Table 1). Number of days to start of flowering (from planting date to date of first flower), duration of flowering (from date of first flower to date of last flower), days to physiological maturity (from planting date until seed in the bottom third of the main raceme had turned brown), seed yield from the middle 1 m of the row (g), and 1000-seed weight (g) were collected on all plots. Seed-quality parameters were estimated with an NIRSystem 6500 (Foss, Hillerød, Denmark); oil and protein content are presented as a percentage of dry weight, glucosinolate content as μmol g^–1, and saturate content as the sum of C12:0, C14:0, C16:0, C18:0, C20:0, C22:0 and C24:0 expressed as a percentage of total fatty acids (Daun et al. 1994).

Table 1.  Location planting dates with temperature and soil-moisture variables during the entire growing season and daylength at the beginning of flowering

Statistical analysis of the phenotypic data was completed by using ASReml 4 (VSN International, Hemel Hempstead, UK) to estimate the best linear unbiased predictions (BLUPs) for all entries (Gilmour et al. 2009). The model used was y = Xb + Zu + e, where y was the observed phenotype, b the fixed terms for matrix X, u the random terms for matrix Z, and e the error matrix N(0,Σ) where Σ was an AR1 × AR1 structure (Gilmour et al. 1997). For the single-site analysis, an overall mean and treatment (CT, ST) were considered as fixed effects. Random effects were row and column nested within treatment, entry, and treatment × entry. The multi-location model had overall mean, treatment, location and treatment × location as fixed effects. Random effects were row and column nested within location and treatment, entry, treatment × entry, location × entry, location × treatment × rep, and location × treatment × entry. The analysis to generate the BLUPs for the WGP was a multi-location analysis with the control and heat treatments being analysed independently. Here, overall mean and location was treated as a fixed effect. Random effects were row and column nested within location, entry, and location × entry. The fitted WGP models are thus ‘treatment-specific’. Heritability (H²) was calculated within treatment by using the following model (Cullis et al. 2006):

where s.e.d. is the average standard error of the differences between entries, and is the sum of the genotypic and genotype × environment variance assuming a compound symmetry structure (Smith et al. 2005). R 3.3.1 (The R Foundation, Vienna) was used to calculate the Pearson correlation coefficients (r) and the t-test for the trait relationship matrices.

The two parental genotypes had previously been fingerprinted with 1788 and 1663 single-nucleotide polymorphisms (SNPs) for PB36 and PB56, respectively. This information was used to select 526 polymorphic TaqMan SNP markers for the DH population to be genotyped by using an array tape (Douglas Scientific, Alexandria, MN, USA). Once SNPs were assessed for missing data and segregation distortion, 368 SNPs were used to create the final linkage map in MapDisto 1.7 (Lorieux 2012) with an average distance of 4.7 cM between SNPs and a total genetic distance of 1725 cM. The genetic map, and the phenotypic BLUPs, were used to identify QTLs through composite interval mapping using 1000 permutations at a P-value of 0.05 in Windows QTL Cartographer V2.5_011 (WinQTL) (Wang et al. 2012), and only QTLs having a significant LOD score based on the permutation test are reported. QTLs present in all three locations were reported through a combined analysis and defined as ‘stable’ QTLs, whereas those present in only one or two locations are reported on a location-specific basis as ‘putative’ QTLs. Genomic BLUP (GBLUP) was used as a WGP model (Meuwissen et al. 2001), fitted with the rrBLUP R package (Endelman 2011) in R 3.3.1. Specifically, the model was y = μ + g + e where y is the vector of phenotypic values (BLUPs), obtained from the phenotypic analysis described before. The intercept is denoted by μ. The genotypic values in vector g (i.e. the genomic estimated breeding values, or GEBV) were assumed to be distributed as g and the residuals as e , with I being the identity matrix. The matrix G is the genomic relationship matrix of the entries and was calculated as described by Endelman (2011) with the rrBLUP package, from the same 368 SNPs used for the QTL mapping. The rrBLUP package uses restricted maximum likelihood (REML) for estimating the variance components and . Prediction accuracy was evaluated in a random cross-validation. The training set used for fitting the model comprised a random sample of 104 genotypes (70%), with the remaining 44 genotypes (30%) used as the validation set (Jan et al. 2016). Two prediction scenarios were considered: prediction within a treatment (‘within’, e.g. using a ST training set to predict the performance of the validation lines under ST), and prediction across treatments (‘across’, e.g. using a ST training set to predict the performance of the validation lines under CT). The prediction accuracy (i.e. correlation between predicted and true genetic values) was calculated by dividing the Pearson correlation coefficient between predicted and observed trait values of the validation genotypes by the square root of the heritability estimated for the corresponding treatment (Legarra et al. 2008). The cross-validation was replicated 500 times for each treatment and the results summarised as the mean and standard deviation of the prediction accuracies.

The mean, maximum and minimum temperatures for the second planting date were on average 1.6°C, 1.9°C and 1.2°C warmer than those of the first planting date (Table 1). The number of days >29.5°C was greater for the second planting date, with Carman 2015 having four more days >29.5°C, Viluco 2015 having 14 more and Carman 2017 having three more. These two factors of increased mean temperatures and increased number of days above the critical threshold temperature provide evidence that genotypes sown on the second planting date would have experienced increased temperature stress. Soil moisture was monitored at both Carman locations, with the CT having slightly higher water content in 2015 and slightly lower water content in 2017 than the ST, and Viluco was irrigated to ensure moisture availability was equal across planting dates.

There was a significant effect of treatment, with a decline in seed yield (YIELD) for the ST at all locations ranging from 14.6% to 18.2%, with the combined analysis having a 16.7% decline (Supplementary Materials table S1, available at the journal’s website). All traits for both treatments at all locations had significant differences within population (table S1). The DH population exhibited transgressive segregation for all traits and treatments, and at all locations. The number of days to flower initiation (DYSFLB), duration of flowering (FLWDUR), and days to maturity (MAT) all declined within the ST, whereas glucosinolate content (GLU) increased (Fig. 1).

Fig. 1.  Violin charts of the phenotypic best linear unbiased predictions from the combined analysis of Carman 2015, Viluco 2015 and Carman 2017. Parent PB36 (circle) and PB56 (triangle) are plotted and the significance level of the paired t-test between the parents and the l.s.d. (P = 0.05) for all of the double-haploid genotypes are labelled for: (a) DYSFLB, no. of days to beginning of flowering; (b) FLWDUR, flowering duration; (c) MAT, no. of days to maturity; (d) TSW, 1000-seed weight; (e) YIELD, seed yield; (f) OIL, oil content; (g) PRO, protein content; (h) GLU, glucosinolate concentration; and (i) SAT, saturated fatty acid content.

The combined analysis for FLWDUR, 1000-seed weight (TSW), protein percentage (PRO), and GLU had a significant negative correlation with YIELD in both treatments, whereas oil percentage (OIL) had significant positive correlation with YIELD (Fig. 2). MAT was negatively correlated with YIELD in both ST and CT, although the negative correlation was only statistically significant in ST.

Fig. 2.  Correlation matrix of the traits from the combined analysis including Carman 2015, Viluco 2015 and Carman 2017 for the (a) control treatment and the (b) stress treatment (see next page). Upper right-hand triangle contains the correlations, the diagonal contains the histogram of the trait data, and the lower left-hand triangle contains the scatter plot of the trait comparisons. DYSFLB, No. of days to beginning of flowering; FLWDUR, flowering duration; MAT, no. of days to maturity; TSW, 1000-seed weight; YIELD, seed yield; OIL, oil content; PRO, protein content; GLU, glucosinolate concentration.

Linkage mapping analysis identified 66 QTLs across the nine traits and two treatments (table S2). Eight of these QTLs were stable throughout the three environments and the combined analysis. The remaining 58 putative QTLs were found in one or two of the locations. The maximum amount of variation explained with a single QTL was 45.2% in the CT and 54.1% in the ST for DYSFLB (Table 2). This QTL was at the same location on linkage group A02 in both treatments, but only at a single location (Viluco 2015). However, no stable QTLs were found for DYSFLB. The greatest amount of variation explained by a single stable QTL was for FLWDUR (17.7%) in the ST. A QTL associated with YIELD was found in the ST at Viluco 2015 on linkage group A02 and explained 11.4% of the variation. This QTL overlapped with QTLs for FLWDUR and TSW found within the Carman 2015 environment and a QTL for saturated fatty acid content (SAT) within the same environment. Seed-quality traits accounted for 28 putative and four stable QTLs across both treatments. OIL had a stable QTL in the ST compared with the CT, whereas PRO had a single stable QTL present in both treatments. Although OIL and PRO were highly correlated, the QTLs did not overlap; however, they were adjacent on linkage group A09.

Table 2.  Number of QTLs with a LOD score ≥2.5 for all traits and control (CT) and stress (ST) treatments from individual location analyses
Inconsistent QTLs that did not appear at all locations (Carman 2015, Viluco 2015 and Carman 2017) are listed as putative; QTLs that were significant within all environments are listed as stable with the corresponding range of phenotypic variance explained (R², %). DYSFLB, No. of days to beginning of flowering; FLWDUR, flowering duration; MAT, no. of days to maturity; TSW, 1000-seed weight; YIELD, seed yield; OIL, oil content; PRO, protein content; GLU, glucosinolate concentration; SAT, saturated fatty acid content

The broad-sense heritability ranged from a low of 0.64 for YIELD in the CT to a high of 0.97 for GLU in the ST (Table 3). For the ‘within’ prediction scenario, the average prediction accuracy ranged from 0.14 for YIELD in the ST to a high of 0.66 for TSW. In general, the ‘within’ prediction accuracy was higher for the ST than for the CT for all traits except YIELD and SAT (Table 3). Using the ST training set to predict performance in CT decreased the prediction accuracy for DYSFLB and YIELD compared with use of a CT training set (i.e. ‘within’ CT) (Table 4). However, when the CT training set was used for predicting performance in the ST, the accuracies declined for all traits relative to the ‘within’ ST prediction, with the exception of YIELD. Traits related to plant development (DYSFLB, FLWDUR and MAT) showed the greatest decrease in accuracy, with a drop of 0.20, 0.17 and 0.15, respectively. Only a minor decrease in accuracy from ‘within’ to ‘across’ was observed for TSW and the seed-quality traits (OIL, PRO, GLU, SAT).

Table 3.  Summary of variance components (σ²_g, genotypic variance; σ²_g×e, genotype × environment variance), average standard error of the differences between entries (s.e.d.), heritability (H²) and the mean and standard deviation (s.d.) of whole-genome prediction accuracies for the ‘within’ control (CT) and stress (ST) treatment prediction scenario
DYSFLB, No. of days to beginning of flowering; FLWDUR, flowering duration; MAT, no. of days to maturity; TSW, 1000-seed weight; YIELD, seed yield; OIL, oil content; PRO, protein content; GLU, glucosinolate concentration; SAT, saturated fatty acid content

Table 4.  Mean and standard deviation (s.d.) of whole-genome prediction accuracy for the ‘across’ control (CT) and stress (ST) treatment prediction scenario
DYSFLB, No. of days to beginning of flowering; FLWDUR, flowering duration; MAT, no. of days to maturity; TSW, 1000-seed weight; YIELD, seed yield; OIL, oil content; PRO, protein content; GLU, glucosinolate concentration; SAT, saturated fatty acid content

Yield losses of 75.5–133 kg ha^–1 have been reported with a 1°C rise in mean temperatures (Nuttall et al. 1992; Kutcher et al. 2010). Morrison and Stewart (2002), indicated that 29.5°C was the threshold temperature at which canola yields decline. Considering that the ST had an increase >1°C in mean temperature and an increased number of days >29.5°C, these increased temperatures played a role in the decreased yield within the ST. Breeding is the genetic gain through selection, with breeders utilising tools to increase the response to selection by increasing selection intensity and heritability while minimising the amount of time required to describe a phenotype accurately (Cooper et al. 2014). For success, breeders therefore need to allocate finite resources strategically to maximise genetic gain. MAS processes using QTLs and WGP are tools that have shown promise in enabling increased genetic gain (Varshney et al. 2016). This research attempted to identify specific QTLs responsible for conferring tolerance to increased temperatures, and although there was a putative QTL discovered for YIELD within the ST that was absent from the CT, the instability across environments and low amount of variation explained limit the practical utility. Further investigation across populations and environments would be required before the implementation of a MAS strategy. The relatively high prediction accuracies found, and the high number of putative QTLs discovered, suggest that a WGP breeding strategy would be more promising than developing a MAS or marker-assisted backcrossing program for the traits collected within this experiment. We evaluated the utility of a popular WGP strategy of using phenotypic data from a subset of a DH population to train WGP models for predicting untested genotypes. This strategy has the potential to increase the response to selection by utilising larger populations without the need for increased resources because the increased resource requirement for genotyping could be offset by a reduction in phenotyping (Riedelsheimer and Melchinger 2013).

The parameter DYSFLB did not have any relationship with YIELD, and although there were several QTLs explaining a larger portion of the variation within specific environments, these QTLs did not respond differently across treatments. The QTL that explained the most variation (45.3% CT, 54.0% ST) at Viluco 2015 was located on linkage group A02, which aligns with the existing literature for this trait within B. napus (Chen et al. 2010; Nelson et al. 2014; Javed et al. 2016). Given that this QTL was found within the environment that had shorter days, it may be a candidate QTL if germplasm is being transferred from long-day to short-day environments. The time to initiation of flowering in B. napus has been shown to be influenced by photoperiod and thermal time (Nelson et al. 2014). Although the difference in photoperiod was minor between these two treatments, this difference, combined with the increased temperatures in the ST, could potentially strengthen the genetic signal and thereby increase the amount of genetic variation explained by the QTL in the ST. In the absence of photoperiod differences or vernalisation requirements, days to initiation of flowering has proven to be a quantitative trait (Javed et al. 2016). The results in this study show that DYSFLB can be predicted accurately with WGP. This suggests that utilising WGP to select for earlier flowering has the potential to minimise interaction of higher summer temperatures with the reproductive stage of B. napus and improve the avoidance of heat stress (Jung and Müller 2009; Franks 2011). Previous research has shown that traits with few QTLs that explain more variation can be used with higher prediction accuracy (Riedelsheimer et al. 2012). The higher DYSFLB accuracy in the ST compared to the CT may be due to the increased amount of variance explained by the QTLs within the ST.

A stable QTL for FLWDUR in the ST but not the CT requires further investigation because this QTL was stable across environments and explained 17.7% of the phenotypic variation. The negative correlation between FLWDUR and YIELD (shorter flowering duration leading to higher yield) within this population suggests that the ability to develop a large sink strength earlier in the flowering period was linked to higher yield. The ability to predict FLWDUR by using WGP with greater accuracy in the ST than the CT might again be due to the presence of a stable QTL with relatively large effect within the ST. Numerous studies have reported on days to the initiation of flowering, but none could be found reporting on the flowering duration within B. napus. The association of FLWDUR with YIELD, the stable QTL and WGP results warrant further investigation of the importance of this parameter.

The QTL mapping of MAT resulted in the discovery of a stable QTL within the CT. Similar to reports in previous literature, there were several putative QTLs demonstrating the quantitative nature of this trait (Shi et al. 2009). Although the stable QTL was discovered only within the CT, the highest WGP accuracy was found within the ST. This might be because of the considerable increase in heritability, which is expected to increase accuracy (Daetwyler et al. 2010). The reduction in prediction accuracies of DYSFLB, FLWDUR and MAT for the ST with a CT training set demonstrates the unique impact of the ST on these traits. These reductions were not unexpected given the influence of temperature on B. napus plant development (Koscielny et al. 2018a).

The stable QTL on C03 within both treatments for TSW has potential for breeding applications if it can be confirmed across additional populations. Previous QTL mapping of seed weight within B. napus has discovered several QTLs, but there are no reports of the C03 QTL discovered in this DH population (Fan et al. 2010; Ding et al. 2012; Fu et al. 2015). Given the number of QTLs found within this population, the previous literature and high prediction accuracy, future research could focus on using WGP when attempting to select for large seed size. Selection for large seed size will require careful assessment of yield because large seed can be negatively related to seed yield, and this covariance must be considered during selection. The ST had very little impact on QTL mapping and WGP enabling flexibility in phenotyping TSW.

A single QTL for YIELD was found at Carman 2015 across both treatments. There was a single QTL for YIELD found in the ST, but it was present only within a single environment and, thus, may not be useful within a breeding program. It is already accepted that yield is a highly quantitative trait; however, research has also shown the quantitative impact of abiotic stresses such as heat and drought on yield (Pinto et al. 2010; Jha et al. 2014; Thudi et al. 2014). No stable QTLs were discovered in this population for YIELD in the ST, which indicates that WGP is the only viable marker-based approach for selecting for YIELD under heat stress. The discovery that prediction accuracy in the ST was considerably lower than in CT may be due to factors such as heritability. Thus, the lower heritability in the ST was likely partially responsible for the lower accuracy. The lower heritability for YIELD in the ST was not unexpected and has been reported in previous studies (Witcombe et al. 2008). In addition, the quantitative nature of the heat-stress response and the absence of any major QTLs suggest that yield under increased temperatures has an even more complex genetic architecture than under the CT, leading to a negative impact on prediction accuracy (Daetwyler et al. 2010). The decrease in prediction accuracy for YIELD in the CT when the ST was used as the training set and the increase in prediction accuracy in the ST when the CT was used as training set is possibly a consequence of these effects on heritability.

The results demonstrating lower OIL and increased PRO on average within the ST agree with the findings of Daun (2007) in which significant variation was found across eight varieties in the relationship of annual temperatures and oil or protein content. Utilising MAS for improving seed-quality traits within well-adapted inbreds would be challenging given that there are several reports of QTLs for seed-quality traits within B. napus (Yan et al. 2009; Chen et al. 2010; Javed et al. 2016) and the large number of minor-effect QTLs discovered within this population. The seed-quality traits had high heritability and accuracies ranging from 0.38 to 0.56. These values were lower than reported previously (Jan et al. 2016; Zou et al. 2016); however, this can be explained by the larger training population sizes utilised in those studies. Würschum et al. (2014) reported on prediction accuracies within DH families consisting of few individuals and reported accuracies similar to, or slightly lower than, the results reported here. Previous studies on seed-quality traits have reported both a significant genotype × environment interaction (Shafii et al. 1992; Delourme et al. 2006) and minimal impact of genotype × environment interaction (Gunasekera et al. 2006; Aslam et al. 2009). The similarity of accuracy for ‘within’ and ‘across’ predictions reported in this study suggests minimal genotype × treatment interaction for the seed-quality traits (OIL, PRO, GLU and SAT). Therefore, because of the high heritability and the seemingly low impact of environmental stresses such as high temperature, data on seed-quality traits could be collected from a limited set of environments, providing flexibility in the collection of these data and potentially offering the ability to reduce required resources.

These results provide insight and direction for use of linkage mapping and WGP for tolerance to increased temperatures in canola. It is prudent to point out the limitations and directions of further study. Several studies have published QTLs that are population-specific (Bernardo 2016). Further research on the stable QTLs need to be performed in additional populations. The population size in this study limited the size of the training and prediction sets; thus, an increase in population sizes would improve the power of the experiment. Previous research has shown that increased phenotypic heritability has a greater impact on WGP accuracies than training set population size and marker density (Zhang et al. 2017). Given this information future research would benefit from ensuring sufficient resources are deployed to ensure accurate phenotyping once a minimum marker density is available on the germplasm of interest. This information could provide future cost savings by enabling breeding programs to phenotype a subset of germplasm continually for creation of robust training sets while predicting large numbers of genotypes with low-density genotypic information. Several studies in bi-parental maize populations have indicated that increasing marker density beyond 300 with an average distance of 5 cM did not noticeably increase WGP accuracies (Combs and Bernardo 2013; Technow and Totir 2015; Zhang et al. 2015).

Linkage mapping has provided valuable insight into qualitative traits, but its practical use in breeding for quantitative traits such as yield or abiotic stress tolerance, in which many minor QTL effects are involved, has proven impractical (Bita and Gerats 2013). This research has shown that although it is possible to discover stable QTLs that differ across treatments, this same set of data can be deployed in a WGP strategy. With the exception of yield, all traits had prediction accuracies in the ST similar to or higher than in the CT, suggesting that phenotyping within higher temperature environments enables the generation of phenotypic data that can be used for WGP while providing the opportunity to screen for future warming environments. Further investigation into the ability to utilise WGP for making genetic gain under high-temperature stress should be conducted to enable gain within this complex abiotic stress trait ensuring genetic gain can be maintained or increased in a warming climate.

The authors declare there is no conflicts of interest.

Concept and design of this experiment: CBK, SW, and RW; acquisition of the data: CBK; analysis and interpretation of the data: CBK, SWG and FT. Manuscript drafted by CBK and revised by SWG, FT and RWD. All authors approved this version of the manuscript for publication.