Contrasting genetic regulation of plant development in wild barley grown in two European environments revealed by nested association mapping

Plant development in a barley nested association mapping population is controlled by interacting genotypic (wild donor alleles) and environmental (geographical location) effects.


Introduction
Barley (Hordeum vulgare), a model species for temperate cereals, is an important crop in marginal environments (Baum et al., 2007;Rollins et al., 2013), which are characterised by abiotic stresses such as heat, drought, and nutrient deficiency. These poor and stress-prone environments offer the biggest opportunities for substantially increased yields on a global scale (Tester and Langridge, 2010). In addition, in temperate climates most climate-change scenarios predict increasing average temperatures and elevated risks for extreme weather conditions such as drought and heat (Settele et al., 2014). These ecosystem changes may also affect current high-yielding environments. In comparison with other cereals, barley has a higher adaptability to drought and heat because its early development and extensive root system permit drought escape (Fischer and Maurer, 1978;Lopez-Castaneda and Richards, 1994). Despite the relatively high stress tolerance of barley there is still room for it to be increased further, for instance through fine-tuning of plant development.
The basic reproductive life-cycle of annual plants such as barley can be summarised as germination, vegetative growth, flowering, and seed maturation. In the life-cycle of a plant, flowering time, i.e. the switch between vegetative and reproductive growth, is a key component to adapt and respond to environmental conditions and to optimise the source-sink balance (Cockram et al., 2007), and it is closely linked to agronomic performance and yield (Cuesta-Marcos et al., 2009). The flowering behaviour of barley depends on temperature and photoperiod. Based on photoperiod, barley varieties can be classified into photoperiod-sensitive (early flowering under long-day conditions) and photoperiod-insensitive (day-length neutral, late flowering). Based on vernalisation response, barley varieties can also be classified into winter (vernalisation requirement) and spring (no vernalisation requirement) growth-habit types. In addition, facultative barley varieties exist that exhibit a reduced (or null) vernalisation requirement, which is associated with an acceleration of flowering (Hemming et al., 2009), and generally combined with frost hardiness. The genetic control of the flowering pathway has been extensively studied by mutation studies, quantitative trait locus (QTL) studies, and by exploiting natural genetic variation, resulting in the identification of a number of major genes controlling flowering time (Karsai et al., 2005;Campoli et al., 2012;Comadran et al., 2012;Faure et al., 2012).
In barley, the response to photoperiod under long-day conditions is primarily controlled by the Photoperiod 1 (Ppd-H1) gene. The photoperiod-sensitive and dominant Ppd-H1 allele promotes flowering (Turner et al., 2005) whereas the photoperiod-insensitive and recessive ppd-H1 allele delays flowering and maturity, enabling an extension of the growing season towards cooler climates (Jones et al., 2008). The delaying effect of the ppd-H1 allele is transmitted through a reduction in the expression of FLOWERING LOCUS T (HvFT1) (Turner et al., 2005). Ppd-H1 is repressed by circadian clock genes such as HvELF3, HvLUX, and PHYTOCHROME C during the night, and a dysfunctional mutation of a repressor may cause a day-neutral phenotype (Faure et al., 2012;Campoli et al., 2013;Pankin et al., 2014). The circadian clock represents an internal timekeeper that synchronises physiological and molecular processes with the diurnal cycle (Johansson and Staiger, 2015). Under short-day conditions the Ppd-H2 gene affects flowering (Kikuchi et al., 2009).
The response to cold temperatures is mainly controlled by interaction of the two vernalisation loci in barley, Vrn-H1 and Vrn-H2. Vrn-H2 is a strong inhibitor of flowering under long-day conditions, functioning as an antagonist of Ppd-H1 to prevent flowering during cold and harsh conditions (Yan et al., 2004). The expression of the MADS-box gene Vrn-H1 is only induced after extended periods of cold exposure (Oliver et al., 2013), resulting in down-regulation of the flowering repressor Vrn-H2 (Chen and Dubcovsky, 2012).
The photoperiod and vernalisation pathways converge in the expression of Vrn-H3, an orthologue of HvFT1, the key integrator of transition from the vegetative to the reproductive phase (Yan et al., 2006). In Arabidopsis, the FT protein moves from leaves to the shoot apical meristem and induces the development of flower primordia. It has been shown that this process is conserved across monocots and dicots (Tamaki et al., 2007;Turck et al., 2008).
In addition, there is a further, temperature-dependent growth regulation pathway active, which is more relevant for short-term ambient temperature changes (Wigge, 2013). This thermo-sensory regulation is poorly understood in barley. However, in Arabidopsis it has been shown that exposure to elevated temperatures may induce early FT expression and, subsequently, early flowering independently from day length (Balasubramanian et al., 2006). In this process, increasing temperatures inhibit the 'evening complex' of the circadian clock, consisting of ELF3, ELF4, and LUX (Thines and Harmon, 2010;Herrero et al., 2012). In this way, the expression of PHYTOCHROME-INTERACTING FACTOR4 (PIF4) is induced, which promotes FT expression (Box et al., 2015). To date, the integration of thermal signals in the flowering pathway is not fully understood in cereal crops. However, Hemming et al. (2012) determined that a HvFT1 independent signalling pathway exists in barley.
The perception of external signals is important in adapting to environmental conditions and in synchronising optimal plant development. This phenological adaptation, which is required for 'climate-smart' agriculture (Olesen et al., 2011), occurs on a genetic level (Zhao and Xu, 2012;El-Soda et al., 2014). In barley, extensive and genetically diverse collections of landraces and wild barley accessions are available (Newton et al., 2011;Russell et al., 2011) that could be used to improve the adaptation potential of cultivated barley and to counteract the reduction of genetic diversity in modern varieties that has resulted from continued selection for favourable agronomic traits.
Barley has become a model for understanding the potential of exotic germplasm to further improve modern elite crops (Ellis et al., 2000;Dawson et al., 2015). Nested association mapping (NAM) is one strategy that places exotic alleles into an elite genetic background in order to investigate their potentially valuable effects in improving complex agronomic traits (Yu et al., 2008;Buckler et al., 2009;Jordan et al., 2011;Nice et al., 2016). The wild barley NAM population HEB-25 was developed for this purpose (Maurer et al., 2015). HEB-25 consists of 1420 BC 1 S 3 barley lines, comprised of 25 families, each derived from crosses between the cultivar 'Barke' (H. vulgare ssp. vulgare) and one of 25 diverse wild barley (H. vulgare ssp. spontaneum, Hsp) genotypes. A number of studies have been conducted with HEB-25 that confirm the power of the NAM design to detect barley QTLs controlling plant developmental and agronomic traits and to select favourable exotic alleles for subsequent utilisation in barley breeding programs. In this way, favourable exotic alleles have been identified that, for example, improve leaf-rust resistance, flowering time, grain size, and salt-stress tolerance (Schnaithmann et al., 2014;Maurer et al., 2016Maurer et al., , 2017Saade et al., 2016). The aim of the present study was to investigate plant development in HEB-25 on a location-specific and a familyspecific basis and thus to examine the potential ability of the 25 wild barley donors of HEB-25 to improve adaptation to future environmental changes. In this regard, field trials were conducted at two locations, Halle (Germany) and Dundee (Scotland), which differ by 5° in latitude. Latitude is an important environmental driver in determining growth and development (Craufurd and Wheeler, 2009), as it is associated with differences in ambient temperature, photoperiod, and radiation intensity. Additionally, we aimed to study the function of plant developmental genes, segregating in HEB-25, and their impact on plant architecture traits such as height and lodging under diverse environmental conditions.

Plant material
The NAM population 'Halle Exotic Barley' (HEB-25;Maurer et al., 2015) was developed by crossing the German spring barley elite cultivar 'Barke' (Hv) with 25 highly diverse exotic wild barley accessions, namely 24 Hsp accessions from the 'Fertile Crescent' and Afghanistan and one H. vulgare ssp. agriocrithon (Hag) accession from Tibet (China). After backcrossing the resulting F 1 s with Barke, three consecutive rounds of selfing were conducted by means of single-seed descent. The resulting BC 1 S 3 plants were propagated as individual bulks to produce 1420 lines, sub-divided into 25 HEB families of 22 to 75 lines. For detailed information about population development, see Maurer et al. (2015). In this study, the bulk propagated progeny of BC 1 S 3 had already undergone several extra generations of selfing and the BC 1 S 3:7 generation was used for the 2014 experiments. Bulk BC 1 S 3:8 seed was then used to sow the 2015 experiments.

HEB-25 field trials
In 2014 and 2015 two field trials in Halle, Germany, and Dundee, Scotland, were carried out under two nitrogen (N) fertilisation treatments to gather phenotypic data. In Halle, the field trial was conducted in randomised complete blocks of 1420 HEB lines for each N treatment. Plots consisted of two rows of 50 seeds with a row length of 1.4 m and a spacing of 0.2 m between rows and 0.5 m between plots. In Dundee, 1371 HEB lines were grown following the same experimental set-up as in Halle. In Dundee, plots consisted of two rows, 40 seeds each, 2.0 m in length, and a distance of 0.25 m between rows and 0.75 m between plots. All field trials were sown in spring, i.e. the end of March in Halle and mid-April in Dundee. Field and disease management were in accordance with local practice.

Nitrogen treatment
At both locations, HEB-25 was cultivated under two N fertilisation treatments, namely without fertiliser (N0) and with fertiliser (N1). The targeted available nitrogen in N1 was set to 100 kg N ha -1 . The amount of added fertiliser was calculated after measuring plant available nitrogen in a soil bulk sample before sowing. In Halle, this resulted in fertilising the N1 block with 60kg N ha -1 in 2014 and 70 kg N ha -1 in 2015. Calcium ammonium nitrate was applied as inorganic N fertiliser at the beginning of the shooting stage. In Dundee, the N1 block was fertilised in 2014 and 2015 with 60 kg N ha -1 , applied as a 22:4:14 NPK compound mineral fertiliser at sowing. The estimated soil nitrogen content one month before sowing was 34 kg ha -1 in 2014 and 62 kg ha -1 in 2015, resulting in total available nitrogen in the order of 94 kg ha -1 in 2014 and 122 kg ha -1 in 2015.

Environments
The field trials in Halle were conducted at the 'Kühnfeld Experimental Station' of the Martin Luther University Halle-Wittenberg (51°29´46.47´´N; 11°59´41.81´´E). In 2014 and 2015 the moderate-to-continental atmospheric growing conditions were characterised by low average temperatures of 6 °C at the beginning of the vegetation period in March and high average temperatures of up to 21 °C in July and August. Almost 50% of the annual precipitation of 514 mm fell during the months July and August (see Supplementary  Fig. S1 at JXB online). The maximum day length was reached on June 20th with 16.63 h from sunrise to sunset.
In Dundee, the trials were conducted at the Experimental Station of the James Hutton Institute (56°28´53.71´´N; 3°6´35.17´´W). Due to its coastal location, the atmospheric conditions during the trial period can be described as maritime. With regard to the growing season, on average July was the warmest month (16 °C average temperature) and April the coldest (9 °C average maximum). The annual precipitation of 856 and 839 mm in 2014 and 2015, respectively, was equally distributed throughout the year with no discernible maximum. Both years were greater than the 30-year long-term average of 665 mm. The maximum day length was reached on June 20th with 17.45 h of daylight (see Supplementary Fig. S1).

Phenotyping
Phenotyping of eight agronomic and plant developmental traits was carried out under both treatments as described in Table 1. In Dundee, no lodging was observed and in 2014 the time of shooting was not recorded. Data for maturity in Dundee 2014 were transformed from an ordinal scale into days after sowing. Raw data are archived in the public data repositories e!DAL (Herzig et al., 2017).

Statistical analyses
Heritability analyses were performed with SAS 9.4 (SAS Institute Inc., Cary, NC, USA). For each location, broad-sense heritabilities across years were calculated within treatments as: within N or N G G R 0 1 2 = + and for each location, broad-sense heritabilities across both N levels were calculated as: where t and y are the number of N treatments (t=2) and years (y=2). V G , V GY , V GT , and V R correspond to the genotype, genotype × year, genotype × treatment, and error variance components, respectively. All effects were assumed to be random when estimating the variance components using the VARCOMP procedure. Best linear unbiased estimates (BLUEs) for each HEB line were computed across years for each N level and location and across N levels separately by using the MIXED procedure assuming fixed genotype effects. BLUEs were used for calculation of Pearson's correlation coefficients (r) between traits across N levels, using the R 3.2.4 (www.r-project.org) software with the CORRGRAM package (Friendly, 2002). We also conducted an ANOVA within each location including fixed main effects (year, treatment, and genotype) and all possible single interactions to check for significant genotype × treatment (G×T) effects for each trait.

Genotyping of HEB-25
Genotyping of 1420 BC 1 S 3 HEB-25 lines and their 26 parents was conducted by TraitGenetics, Gatersleben, Germany, with the Illumina Infinium iSelect HD 9k chip consisting of 7864 singlenucleotide polymorphism (SNPs; Comadran et al., 2012). After quality checks, 5398 informative SNPs remained for further analyses, as described in Maurer et al. (2015). The SNP scores were translated into a quantitative genotype matrix where the state of the homozygous Barke allele was coded as 0, while HEB lines showing a homozygous wild barley genotype were assigned a value of 2. Consequently, heterozygous HEB lines were assigned a value of 1.

Genome-wide association study (GWAS) analysis
For GWAS, the multiple linear regression Model-A was applied following Liu et al. (2011) and Maurer et al. (2017). Associations between molecular markers and quantitative traits were determined by Y = µ + ∑X c b c + ε, where Y is the phenotypic means vector (i.e. BLUEs), µ is the intercept (equivalent to the overall mean), X c is the vector containing quantitative marker scores for each individual HEB line at marker c, b c is the allele substitution effect (α) at marker c, and ε is the vector of residuals. The GWAS analysis was performed by applying forward-backward regression investigating all SNPs simultaneously, implemented in PROC GLMSELECT in SAS 9.4. SNPs were determined to enter or leave the model if the Schwarz Bayesian Criterion (Schwarz, 1978) was reduced. All SNPs included in the final model were determined as significant. The SNP effect is defined as 2α, where two Barke alleles (genotype 0) are substituted against two exotic alleles (genotype 2) at the respective SNP in the final model. Note that all significant SNP effect estimates are modelled at the same time in the final model. To estimate the explained phenotypic variance (R 2 ) for each marker-trait association, the significant SNP was modelled solely in a linear model.
To assess the accuracy of the prediction of all traits by GWAS, a cross-validation (CV) approach was applied following Maurer et al. (2017). A five-fold cross-validation was conducted 20 times (100 CV runs) where 80% of the HEB lines, randomly selected within each family, formed the estimation set (ES). This ES was used for detection of QTLs and effect estimation. The remaining 20% of the lines represented the prediction set (PS), wherein the phenotype was predicted based on the SNP effects estimated in the ES. The explained phenotypic variance (R 2 adj ) was determined as the modelled fit of the predicted phenotypic value of all significant markers simultaneously regressed on the observed phenotypic value within each ES. Prediction ability (R 2 pred ) was evaluated by predicting the phenotype of the PS by means of detected SNP effects from the ES. R 2 adj and R 2 pred were averaged across all 100 CV runs to calculate the final values. The number of significances for each marker was counted across all CV runs and was defined as the detection rate (DR). The presence of a reliable QTL was assumed next to a peak marker with a DR of ≥25. Single-marker P-values, R 2 , and effect estimates were taken across all runs where the respective marker was significant. To identify potential candidate genes we used BARLEYMAP (Cantalapiedra et al., 2015).

Calculation of family-specific donor allele QTL effects
As a consequence of the NAM population design, we expected a variation of allelic QTL effects among the 25 wild donor-derived NAM families. Therefore, we cumulated significant SNP marker effects and estimated family-specific donor allele QTL effects following the approach of Maurer et al. (2017). The cumulating procedure started with the identification of the peak marker, which is defined as the SNP with the highest DR across all CV runs. This peak marker formed the centre of a 26-cM interval, which was scanned for further significant SNPs. All Model-A SNP effect estimates within this interval were then cumulated for each of the 25 donors separately, as ∑ n i SNP(donor) i × α i , where i iterates through all significant SNPs (n) in the same interval, SNP(donor) i represents the donor genotype (0 or 2) of the i-th significant SNP, and α i is the SNP effect, obtained from Model-A. This method is based on the fact that different donors show different segregation patterns across significant SNPs, ultimately leading to different donor-specific effect estimates for each QTL. The family-wise cumulating was done for each CV run and the mean of them was taken as the final familyspecific QTL effect estimate. To test whether family-specific QTL effects were significantly different between locations a simple paired t-test was performed.

Grouping different QTLs into QTL regions
Peak markers significant for different traits and located within an interval ≤10 cM were combined to one QTL region, which was considered to regulate multiple traits simultaneously. Results and discussion

Descriptive statistics
The HEB-25 population was cultivated under two contrasting nitrogen treatments, in both Halle and Dundee. Overall, we observed a wide phenotypic variation for all studied traits throughout the trials (see Supplementary Fig. S2), which resulted in relatively high coefficients of variation (CoV) that were >10% for the traits SEL, HEI, LOD, and SHO in Dundee (Table 2; see Table 1 for abbreviations). In contrast, MAT revealed the lowest CoV (<4.5%). Significant (P<0.05) genotype effects were observed for all traits in both locations. Likewise, treatment effects were significant for all traits in both locations, except for SEL in Halle (Supplementary   Table S1). In contrast, significant genotype × treatment (G×T) interactions were only observed for HEA and MAT in both locations and for RIP in Halle (Supplementary Table  S1). The marginal difference between nitrogen treatments for all developmental stages may indicate that nitrogen fertilisation at this magnitude exhibited only minor effects on plant development, which is also confirmed and summarised in Hall et al. (2014). In contrast to common literature (Takeno, 2016) and our expectations, non-fertilised plants were, on average, not affected in plant development, indicating that no severe nitrogen stress was induced in our trials. In addition, the lack of strong G×T interactions in Dundee and in Halle indicated that the HEB lines generally reacted similarly to the increase of N supply from the N0 to the N1 treatment. For simplicity of data interpretation, we thus based the following GWAS analysis on phenotype data averaged across treatments (BLUEs were used), excluding SNP × treatment effects.
Differences in plant development were observed between locations. During the early stages of the growing period, HEB lines developed faster in Dundee and reached SHO on average 4.7 d earlier than in Halle (Table 2). Later on, this trend was reversed, with plants in Dundee remaining in the SEL phase on average 21.5 d longer than in Halle, so HEA occurred on average 9.0 d later than in Halle. Corresponding delays to maturation were also seen for Dundee (average +19.9 d relative to Halle). One explanation for different plant development in Halle and Dundee may be the variation of photoperiod between the two locations (Villegas et al., 2016). In total, 355 (25%) out of 1420 HEB lines carry a photoperiod-sensitive allele of Ppd-H1 (Maurer et al., 2015), the main driver of photoperiod response under long-day conditions in barley (Turner et al., 2005). In Halle, long-day conditions (>12 h light) prevailed during sowing, with a progressive increase in day length. In Dundee, day length had already reached 13.5 h at the time of sowing and the daily increase was more pronounced than in Halle (see Supplementary Fig. S3). This might explain the faster plant development until the SHO stage in Dundee. However, it is unlikely that only 25% of the population carrying the photoperiod-sensitive exotic allele of Ppd-H1 could cause an average location difference of 4.7 d across all 1420 HEB lines. Therefore, we assume that other abiotic factors such as soil temperature, ambient temperature, and precipitation may have also contributed to the generally accelerated plant development in Dundee. For example, Supplementary Fig. S1 shows that the average temperature in Halle was lower in winter than in Dundee, suggesting that the soil temperature was lower and therefore took longer to increase in spring, and hence early growth could have been delayed in Halle compared to Dundee. Hemming et al. (2012) observed faster reproductive plant development under elevated temperatures, which might explain the shift in developmental differences between Halle and the comparatively cooler location of Dundee (Supplementary Table S1). Furthermore, in Halle the SEL phase was less than half as long as in Dundee, with only minor differences in HEI between both locations, indicating a slower growth rate in Dundee (Table 2).
In Halle, the occurrence of lodging (LOD) was relatively low, with 2.6 in N0 (low N) and 2.7 in N1 (high N). All traits showed higher broad-sense heritability across than within nitrogen treatments (Table 2). In general, traits showed heritability estimates >50%, except for RIP in Dundee and Halle (N0), and MAT in Dundee. In addition, higher trait heritabilities were observed in Halle compared to Dundee, with the maximum value being observed for HEA across treatments, with h 2 =93.1%.

Trait correlations
Pearson's correlation coefficients between traits were calculated across treatments for each location separately (Table 3). In addition, the cross-correlation of a trait between locations was determined. In Halle, the highest correlations were observed between SHO and HEA and SHO and MAT, indicating that early-shooting genotypes tended to be early in HEA and MAT and late-shooting genotypes tended to be late in HEA and MAT. Maurer et al. (2016) observed similar relationships between developmental traits in Halle. In Dundee, this relationship was diminished with medium correlations between SHO and HEA (r=0.61) and HEA and MAT (r=0.54) and a low correlation between SHO and MAT (r=0.31). The traits RIP and SEL were calculated as the differences between two other stages of development, and hence correlations can provide some indications for their main determinant. For example, in Halle RIP was predominantly determined by HEA (-0.59), and SEL was equally determined by SHO and HEA. In contrast, in Dundee RIP had a medium correlation with both determinants, and HEA, MAT, and SEL were mainly determined by SHO. This suggests that MAT was strongly influenced by environmental factors such as heat and drought in Halle, as already mentioned in Maurer et al. (2016). In Dundee, equally distributed precipitation throughout the year and lower temperatures did not induce heat or drought stress during the maturity phase, so the vegetation period in total lasted 20 d longer than in Halle. Thus, in Dundee MAT might have been under stronger control of genetic factors than in Halle. The difference in plant development between the two locations was further supported by the fact that SEL and HEI showed a positive correlation (r=0.50) in Halle, but no detectable correlation (r=-0.08) in Dundee. This suggests that in Dundee plant growth was maintained after flowering, while in Halle it was essentially terminated during flowering. Moreover, as expected, LOD and HEI showed a relatively high correlation (r=0.65) in Halle, whereas tiller thickness (TCK) in Dundee was independent from HEI (r=0.14).
When comparing trait performances between Halle and Dundee, we observed strong cross-correlations for HEI (r>0.83) and HEA (r>0.88), while for SHO, MAT, and RIP medium cross-correlations between r=0.46 and r=0.53 were observed. This finding indicates that HEI and HEA were controlled largely by environmentally unresponsive QTLs while the remaining traits responded to location-specific QTLs.

QTL detection through GWAS
A GWAS was carried out for all traits to compare the genetic control of plant development between Halle and Dundee. A multiple linear regression model was used, followed by a HEB family-specific estimation of the QTL effect to exploit the full potential of the NAM design. In total, 491 significant marker-trait associations were identified with a detection rate ≥25 (cross-validation) for eight traits, across Dundee and Halle (Supplementary Table S2). Subsequently, these markertrait associations were converged into 122 and 164 QTLs (for all traits) and placed on 60 and 66 QTL (across traits) regions regulating eight plant development traits in Dundee and Halle, respectively (Table 4, Fig. 1, and Supplementary Tables S3-4). We also re-analysed the HEB-25 phenotype data using growing degree-days (GDD) as outlined by Maurer et al. (2016). However, we detected similar differences in growth rates for Halle and Dundee, and the QTLs listed in Table 4 were almost identical using GDD and regular days (data not shown). We thus conclude that the extended growth phase in Dundee was primarily a location-specific effect rather than a thermal effect. The phenotypic variance explained by a full QTL model (R 2 adj ) reached higher percentages in Halle, with a range from 62.6% (RIP) to 82.8% (HEA), than in Dundee, which had a range from 19.6% (SEL) to 81.1% (HEA, Table 4). The calculation of the prediction ability (R 2 pred ) revealed a similar result. In Halle, R 2 pred values were relatively high, affirming the robustness of the method  and the reliability of the detected QTL results. In Dundee, R 2 pred values were low in general, except for HEA and HEI. For SHO, SEL, and TCK the results seemed to be less reliable. A major cause of the decreased explained phenotypic variances of the latter traits was probably the phenotypic data, which were based on only one year in Dundee. In addition, mathematically derived traits in general have less power in QTL studies (Wang et al., 2011), potentially reducing the explained phenotypic variances of SEL and RIP.
Among the traits, we observed both decreasing and increasing Hsp (wild barley) allele QTL effects compared to the elite Barke allele ( Supplementary Tables S3-4). Most of the major QTL regions showed a consistent, location-independent effect for all developmental traits, with a low CoV across families. This was true, for instance, for Ppd-H1, sdw1, and HvCEN, which were originally described by Laurie et al. (1995). Furthermore, we obtained a broad range of family-specific QTL effects (Supplementary Tables S3-4). For instance, in Dundee the Hsp allele at the Vrn-H3 locus (QTL-7H-3) increased HEA by 5.4 d in the HEB family 04 but reduced HEA by 2.0 d in family 24. Quite often, QTL regions exerted effects on several traits simultaneously. Therefore, in the following sections we discuss in detail trait-specific QTL effects that were present in selected prominent QTL regions. These regions are summarised in Table 5. Details for all QTLs and QTL regions can be found in Supplementary Tables S3-4 and Fig. S4.

QTL region 2H-2 (Ppd-H1)
This well-known QTL turned out to be a main driver for developmental traits under long-day conditions (see Supplementary Table S4).All developmental traits studied except LOD were regulated at this QTL. The most significant marker in this QTL region was SNP BK_16, which was directly located within the Ppd-H1 gene and explained up to 42% of the phenotypic variance for HEA ( Supplementary  Table S2). Ppd-H1 mainly controls sensitivity to long photoperiods (Turner et al., 2005), accelerating reproductive development when subjected to long days. We compared the effect of the Hsp allele of Ppd-H1 on later phases of plant development between locations and observed that, on average, MAT was accelerated by 6.3 d and RIP was extended by 2.4 d in Halle, whereas in Dundee MAT was accelerated by 3.0 d and RIP was extended by 4.2 d. These effects were detectable in all HEB families. The less-pronounced accelerating effect in Dundee deviated from the expectation with regard  Fig. S3). During sowing, the day length in Dundee (13.8 h) was longer than in Halle (12.3 h) and the difference slightly increased, reaching a maximum at 75 d. According to the literature, extending the day length leads to a faster life-cycle in photo-sensitive genotypes (Roberts et al., 1988;Paynter et al., 2001). We thus expected a stronger Hsp effect in Dundee than in Halle. We conclude that the response of the photoperiod-sensitive exotic Ppd-H1 allele to an increase in day length in Dundee may have been counteracted by other environmental cues, for instance the effect of ambient temperature on plant development. Halle is characterised by higher temperatures during the growing season than Dundee, in particular during early development (April) and maturity (July) ( Supplementary  Fig. S1). Hemming et al. (2012) and Ford et al. (2016) found out that high temperatures resulted in rapid progression of reproductive development under long-day conditions through a gene × environment interaction. This interaction indicates not only a day length-dependent but also a temperaturedependent reaction of Ppd-H1. High temperature could have caused faster development in Halle and counterbalanced the accelerating day-length effect in Dundee. Interestingly, in addition to its strong effects on reproductive development, the Hsp alleles of Ppd-H1 also exerted strong effects on HEI (-6.8 cm) and TCK (-1.4 score). These effects may be attributed to the faster development induced by the Hsp alleles of Ppd-H1, which accelerates progression through plant development, thereby reducing terminal height and tiller thickness. Furthermore, the exotic donor allele present in HEB family 13 produced the smallest plant development effects across all the HEB donors (Supplementary Table S3). This observation may indicate that the donor Ppd-H1 allele in HEB family 13 has a higher similarity to the Barke allele. Preliminary exome capture SNP data support this assumption since we found only two SNPs for Ppd-H1 in HEB family 13 (Supplementary  Table S5).

QTL region 2H-3 (HvCEN)
This association was identified as a robust and stable pleiotropic effect controlling all the traits studied except LOD. The direction of Hsp effects at QTL-2H-3 were consistent    with Ppd-H1, although expressed to a lesser extent, except for TCK where the magnitude of effects were similar to Ppd-H1. The most likely candidate gene is HvCEN (Comadran et al., 2012). The Hsp allele seemed to accelerate early reproductive stages, both in Dundee and in Halle, whereas the effect on MAT was very low and only detected in Halle. The difference between locations was largest for SEL and RIP, with Hsp reducing SEL and extending RIP in Dundee. Strong familyspecific and location-specific effects were observed for RIP. The Hsp allele present in HEB family 11 produced a strong effect in Dundee, where RIP was extended by 6.4 d compared to the Barke allele. The effect of the donor allele in HEB family 11 was at least 4.0 d stronger than for all other Hsp alleles present in the HEB-25 population. The donor allele is derived from the Hsp accession HID109, originating from Syria (Maurer et al. (2015). We found a unique HvCEN SNP at position 523 378 414 bp, which was only present in HEB family 11 (Supplementary Table S5). This SNP may explain the strong effect on extending the ripening phase in Dundee (Supplementary Table S3). It would be interesting to mine further alleles from Syrian wild barleys, which are potentially useful to adjust the length of the ripening period. In contrast, the Hsp allele effect was strongly reduced in Halle, where RIP was only extended by 2.3 d in HEB family 11. It turned out that the extension of RIP was achieved at the expense of SEL. The same Hsp allele in family 11 exerted the strongest effects in the HEB-25 population, reducing SEL to 1.7 and 2.9 d in Dundee and Halle, respectively. To conclude, the family-specific and location-specific Hsp effect of donor HID109 might be of particular interest for barley breeders to enable a considerable extension of the ripening phase in Northern European environments.

QTL region 3H-5 (uzu)
We observed a QTL region that showed a significant effect on tiller thickness (TCK). Family-specific Hsp effects ranged from 0 (HEB family 01) to a score of +1.9 (HEB family 24). Thus, this QTL represents the main positive driver to increase tiller thickness, a trait supportive to increases in stem strength and lodging resistance. The most likely candidate in this region is the uzu gene. The uzu mutant induces a semi-dwarf, upright plant architecture with an improved canopy structure and sturdy straw due to its brassinosteroid deficiency (Chono et al., 2003;Saisho et al., 2004). The uzu gene is mainly observed in six-rowed barley varieties originating from East Asia, including Japan, Korea, and China (Takahashi, 1951;Dockter and Hansson, 2015). Interestingly, the donor of family 24, a H. vulgare ssp. agriocrithon accession originating from Tibet, showed the strongest positive effect on TCK (Supplementary Table S3).

QTL region 3H-11 (sdw1)
QTL region 3H-11 produced the strongest effect on plant height, with the Hsp allele increasing HEI on average by 20.8 cm in Dundee and 18.1 cm in Halle. The most likely candidate gene explaining this effect is the sdw1 locus (also called denso), encoding a GA-20 oxidase (Jia et al., 2015). The semidwarf allele, present in the Barke cultivar, contributed to the 'Green Revolution' in the 1960s by reducing plant height and, at the same time, increasing lodging resistance, harvest index, and grain yield in modern cultivars (Milach and Federizzi, 2001). Two exotic donor alleles at the sdw1 locus, present in HEB families 08 and 24, still produced increasing plant height effects in Dundee (+15.82 and +16.37 cm, respectively) and Halle (+13.38 and +13.88 cm, respectively) relative to the Barke sdw1 allele. However, these effects were much less pronounced than in the remaining HEB families, which had maximum exotic allele effects of +21.91 and +18.99 cm in Dundee and Halle, respectively (see Supplementary Table S3). We found a unique exotic SNP for sdw1 at position 634 073 570 bp, which was only present in HEB families 08 and 24 (see Supplementary Table S5). This exotic SNP may explain the less-pronounced effect of these alleles in comparison with the remaining HEB donor alleles. In addition, wild alleles seem to have a positive effect on thousand-grain weight . The Hsp effects on developmental traits differed clearly between Halle and Dundee. In Dundee, the effect of Hsp in reducing SHO (-2.6 d) was less pronounced than in Halle (-5.6 d). SEL was reduced in Dundee (-0.9 d), while it was prolonged in Halle (2.5 d). Finally, RIP was extended by 1.4 d in Dundee, whereas no clear Hsp effect was observed in Halle. The sdw1 locus might be a further example of a day-length × temperature interaction, with changing impacts on different developmental traits. Wang et al. (2010) found suggestions for such a relationship by varying sowing time. A further interesting aspect is that HEB family 24, developed from a H. vulgare ssp. agriocrithon accession, showed less-pronounced effects at QTL region 3H-11 for most of the studied traits. This may indicate the presence of a sdw1 allele that is more related to Barke at this position than the Hsp alleles present in the remaining HEB families.

QTL region 4H-4
This QTL region showed minor Hsp effects on most developmental traits. However, contrasting family-specific Hsp effects were obtained in Dundee; for example, Hsp effects on SHO ranged from -4.1 d (HEB family 17) to +3.4 d (HEB family 08). This finding indicates that contrasting wild barley donor alleles were present in the HEB-25 population, which may be particularly suited to support contrasting breeding goals -in this case early versus late juvenile development. A couple of candidate genes were located adjacent to the peak marker of QTL region 4H-4 (SNP 11_10261 at 50.85cM). One candidate, phytochrome B (PhyB), is a red/far-red light receptor, which induces photoperiod-dependent flowering (Szucs et al., 2006;Karsai et al., 2008). The different day lengths and the associated light qualities at Dundee and Halle may cause location-specific effects. PhyB also plays a role in the frost-tolerance pathway in barley (Novak et al., 2016). A second candidate, HvCO16, belongs to the CONSTANS-like (COL) gene family. COL genes possess CCT domains and are known to promote flowering under long-day conditions Downloaded from https://academic.oup.com/jxb/advance-article-abstract/doi/10.1093/jxb/ery002/4816162 by Library -Duncan of Jordanstone user on 13 February 2018 in plants (Putterill et al., 1995;Griffiths et al., 2003;Cockram et al., 2012). Two additional candidate genes, HvPRR59 and HvPRR73, belonging to the pseudo-response regulator family (PRRs), are also located in QTL region 4H-4. PRRs are part of the circadian clock, a pathway active in plants to perceive and respond to day-length and temperature cues (Song et al., 2015). These two candidate genes are Arabidopsis orthologues, indicating that functions of the clock may be conserved (Campoli et al., 2012;Calixto et al., 2015).

QTL region 4H-8 (Vrn-H2)
QTL-4H-8 showed significant effects on HEI, SHO, and HEA at both locations and, in Dundee only, on SEL and RIP. SHO and HEA were delayed by the Hsp allele, with a maximum extension of SHO by 2.7 d in Halle in HEB family 23. The QTL effects can presumably be explained by the candidate gene Vrn-H2. This relates to the fact that the pre-vernalisation flowering repressor Vrn-H2 (Yan et al., 2004) is deleted in spring barley cultivars such as Barke (von Zitzewitz et al., 2005). In spring-sown field trials without exposure to severe cold such as that found in Dundee and Halle, the initiation of inflorescence is delayed under the presence of the functional Hsp allele of Vrn-H2. In addition, broad variations of familyspecific Hsp effects were observed for SHO and HEA. This was particularly true for Dundee, where some Hsp alleles seem to have had no delaying effect on SHO and HEA whereas other Hsp alleles delayed them by 2.0 d (HEB families 01, 10, and 12) and 2.3 d (HEB family 03), respectively. Families with a strong delaying, effect especially for SHO in Dundee, seem to have originated predominantly from coastal areas of the 'Fertile Crescent' (Supplementary Fig. S5). Interestingly, Vrn-H2 possibly exerted a pleiotropic effect on HEI in a location-specific manner. In Dundee, HEI was increased by up to 6.4 cm in HEB family 04, whereas in Halle it was reduced by up to 4.5 cm in HEB family 09. The effect of Vrn-H2 on plant height was also found in a barley RIL population developed at ICARDA in Syria (Rollins et al., 2013). In addition, the rice VRN2-like gene Ghd7 was shown to exhibit pleiotropic effects on plant height and other yield-related traits (Xue et al. 2008;Greenup et al., 2009). To date, very little is known about the direct or indirect molecular effects of flowering-time genes on plant architecture. The contrasting effects of Hsp in HEB families 04 and 09 may be a good starting point for a deeper study of the potential HEI effects caused by Vrn-H2.

QTL region 6H-1
This unknown QTL is mainly characterised by strong familyspecific and location-specific effects in the HEB-25 population. For example, the Hsp allele present in HEB family 15 increased HEI by 6.7 cm in Dundee (with no effect in Halle) and in HEB family 24 it increased SHO and HEA in Halle by 3.5 and 4.3 d, respectively (with no effect in Dundee). No candidate gene explaining these Hsp effects has so far been located in the QTL-6-1 region. We intend to fine-map and, ultimately, clone these family-specific and location-specific Hsp effects in follow-up studies. Segregating high-resolution sub-populations derived from informative HEB lines, which were genotyped as heterozygous in the respective QTL regions, are available and can be used as starting points for such studies.

QTL region 7H-3 (Vrn-H3)
QTL region 7H-3 showed significant effects for most developmental traits at both locations except for HEI and RIP, for which effects were only detected at one location. A number of QTLs showed family-specific effects. For example, the Hsp alleles decreased HEA by 2.0 d and 1.5 d in HEB families 24 and 13, respectively, but increased HEA by 6.6 d in family 12. For most traits markers, SNP 12_30894 and SNP 12_30895, which are directly located in the Vrn-H3 gene, showed the highest significance. Vrn-H3 (also named HvFT1) is a homologue of the Arabidopsis FLOWERING LOCUS T gene (Yan et al., 2006) and is the main integrator of photoperiod and vernalisation signals, which facilitate the switch from the vegetative to reproductive phase of the plant. The allelic diversity in terms of family-specific effects in the HEB-25 population corresponds to previous studies, which also found ample natural variation in the promoter and first intron of HvFT1 (Loscos et al., 2014). In addition, different promoter haplotypes of HvFT1 have been identified and associated with late-and early-flowering phenotypes in a collection of barley landraces (Yan et al., 2006;Casas et al., 2011). Moreover, Loscos et al. (2014) and Nitcher et al. (2013) have pointed out that a broad range of copy number variation is present at the HvFT1 locus. An increased copy number of the transcribed region (without promoter) was associated with an acceleration of flowering time. We assume that the HEA variation observed in HEB-25 noted above was also caused by sequence variation among the Hsp alleles of HvFT1 (Vrn-H3). We are currently extending exome capture sequencing based on Mascher et al. (2013) to include all the HEB-25 lines in order to better explain the causes of the family-specific effects of HvFT1. This effort may help to relate sequence variation among HvFT1 alleles to Hsp effects on HEA and other developmental traits in the HEB-25 population.

Plant development and its impact on lodging and tiller thickness
Plant architectural characteristics, including tiller thickness (TCK) and resistance to lodging (LOD), are important traits for breeding. The heavy spikes of potentially high-yielding cultivars may lead to poor harvestable yields and grain quality if the LOD and TCK characteristics are unfavourable (Caierão, 2006;Dockter and Hansson, 2015). In Halle, we observed negative correlations between LOD and plant development traits (r LOD×SHO =-0.52, r LOD×HEA =-0.34, r LOD×MAT =-0.40) and a positive correlation with HEI (r LOD×HEI =0.65). The latter effect can be partially explained through pleiotropic effects of sdw1 containing QTL 3H-11, where the Hsp allele increased HEI by 18.1 cm and LOD by a score of 1.1 in Halle. Likewise, the Vrn-H2-containing QTL 4H-8 also showed pleiotropic effects, with the Hsp allele decreasing HEI by 3.1 cm and LOD by a score of 0.2 in Halle. Overall, there was a high similarity between the LOD QTL described here in this study and that of Tondelli et al. (2013). For TCK, a further parameter used to characterise robustness of plant stems, reliable major QTLs were detected. The Hsp alleles of Ppd-H1 (QTL 2H-2) and HvCEN (QTL 2H-3), both producing accelerating effects on plant development, showed negative effects on TCK, reducing it by scores of 1.4 and 1.6, respectively, in Dundee. We conclude that earliness of development seems to have a decisive negative impact on TCK. However, the overall effect of Hsp alleles at the uzu locus (QTL 3H-5) increased TCK by a score of 1.3 in Dundee without showing any crucial effect on plant development. We therefore conclude that a further, so far unknown, physiological process may also contribute to the expression of TCK in the HEB-25 population. Since lodging did not occur in Dundee, no direct correlation between LOD and TCK could be estimated. However, since only one QTL region (4H-1) showed corresponding QTLs for LOD and TCK, we assume that both traits are primarily controlled independently, at least under low to medium lodging pressure.

Conclusions
The ultimate goal of GWAS is to identify gene alleles that control phenotypic diversity in breeding pools . In classical SNP-based bi-parental QTL approaches only two alleles can be distinguished. In our study we showed that cumulative significant SNP effects  can be applied to identify family-specific QTL effects of exotic alleles and to make use of the full potential of the multi-parental barley NAM population HEB-25. We have found several exotic alleles that showed reliable family-specific effects for all investigated traits, even with differing effect directions. This finding provides the ability to fine-tune traits for plant development including flowering time, an important co-determinant of yield (Jung and Müller, 2009). The selection of exotic alleles to fine-tune yield-related traits may ultimately result in cultivars that are better adapted to specific environments.
In the present study, the HEB-25 population was studied under two contrasting climatic environments, Halle (Germany) and Dundee (Scotland), to investigate location-specific QTL effects on plant development. The longer growing season in Dundee is presumably caused by lower temperatures and slightly higher rainfall and day length during the summer, resulting in a delayed plant development compared to Halle. This effect is reinforced in Halle through higher temperatures in July, which accelerate plant development and maturation in a long day-dependent manner (Hemming et al., 2012). Several QTL regions seemed to interact significantly with the contrasting environments. These QTL ×environment interactions are essential for sessile plants to regulate physiological processes and to ensure that the optimal developmental stage occurs at the right seasonal time. This adaptation capacity may be valuable with regard to climate change, where higher temperatures and drought are predicted to reach Northern Europe (Lobell et al., 2011;Olesen et al., 2011). This raises the question as to how adaptable are cultivars with an insensitive ppd-H1 allele, which is prevalent in Central and Northern European cultivars? Photoperiod-insensitive ppd-H1 alleles have been favoured in regions with long growing seasons and good water availability, such as Central Europe, due to the ecological advantages they confer. In large parts of Southern Europe and south-west Asia, however, the sensitive Ppd-H1 allele is predominant (Cockram et al., 2007;Jones et al., 2008). The current long growing season in Northern Europe might increasingly change towards Mediterranean conditions and the ecological advantages of ppd-H1 could thus disappear in some regions. Our results suggest that the sensitive exotic Ppd-H1 alleles not only perceive day length but also interact with temperature to regulate plant development in barley. In this regard, Borràs-Gelonch et al. (2011) and Ejaz and von Korff (2017) have also underlined the interaction of Ppd-H1 and temperature in barley in different genetic backgrounds and environments. To overcome future agricultural challenges in a rapidly changing environment, the Ppd-H1dependent adaptation potential, which requires a shortened plant cycle under dry summer cultivation, may gain more importance in Central Europe. The observed latitudinal cline of the Ppd-H1 allele in Europe could thus move further north after sensitive alleles have been introduced, for example by using the Hsp alleles of the multi-parental HEB-25 population that have been characterised here.

Supplementary data
Supplementary data are available at JXB online. Fig. S1. Climate data for both locations during the experimental period. Fig. S2. Frequency distribution of BLUEs for all traits, plotted as density histograms. Fig. S3. Frequency distribution of BLUEs for SHO, HEA, and MAT as a function of day length. Fig. S4. Family-specific effects of all the QTLs discussed. Fig. S5. Geographic origin of HEB-25 donors with regard to QTL effects of the exotic Vrn-H2 alleles on shooting. Table S1. ANOVA results of phenotypic data within locations. Table S2. GWAS output of all significant SNPs across all cross-validation (CV) runs. Table S3. Family-specific effect estimation based on the cumulating method. Table S4. List of QTLs, extreme family effects, and candidate genes calculated for eight traits. Table S5. Presence of SNPs inside QTL candidate genes in HEB families based on exome capture sequencing.