Three parental winter rye elite inbred lines (Lo90-N, Lo115-N, Lo117-N) were chosen for this study. Two segregating populations, each consisting of 220 F₂ lines were generated by crossing (1) inbred line Lo115-N with inbred line Lo90-N (Pop-A) and (2) inbred line Lo115-N with inbred line Lo117-N (Pop-B). Parents belong to the Petkus gene pool (seed parent) and possessed normal cytoplasm (N). F₂ plants were randomly forwarded to F₃ generation by single-seed descent and tested in F₄ generation (named F_3:4 lines in the following) The 440 randomly taken F_3:4 lines were crossed to an unrelated cytoplasmic-male sterile (CMS) single-cross tester of the Petkus gene pool (seed parent pool) by open pollination between isolation walls. The resulting three-way crosses of the type (A · B) × F_3:4 line will be addressed as testcrosses throughout the paper. They consisted only of non-restorer materials (Petkus × Petkus), which required the supply of external pollen for fertilization during field trials. For this, a 1:1-mixture of the two pollen-shedding open-pollinated varieties ‘Danko’ and ‘Recrut’ was planted in the alleys and in stripes in a regularly distance of ten plots through the whole experiment. These two population cultivars were chosen to result in an extended flowering period. All plant materials used in this study (except ‘Danko’) were kindly supplied by KWS LOCHOW GMBH, Bergen, Germany.

Field experiments were conducted in the years 2010 and 2011 at five locations: (1) Wohlde (WOH), Germany, N52.8°, E10.0°, 80 m above sea level; (2) Beckedorf (BEK), Germany, N52.5°, E10.3°, 80 m above sea level; (3) Petkus (PET), Germany, N51.6°, E13.2°, 130 m above sea level; (4) Stuttgart/Hohenheim (HOH), N48.4°, E9.1°, 400 m above sea level, and (5) Walewice (WAL), Poland, N52.6°, E19.4°, 184 m above sea level. The location × year combinations were referred to as environments in the following.

Both populations with each of 220 lines were evaluated in field trials together with their parents (repeated 9 times) and arranged in an incomplete 24 × 10 lattice design with two replications. Plot size ranged from 5 to 6 m² and seeding rate varied from 150–200 kernels m^-2. Standard production practices of mineral fertilizer, herbicides and fungicides applications were used following local standards. To avoid lodging, growth regulators were applied three times. Caused by insufficient rainfall during both growing seasons irrigation was applied at WOH, PET, and WAL by a drip irrigation system. Five drip lines per plot were established to ensure a consistent irrigation in each plot depending on the particular local condition. Data were recorded for plant height (cm), grain yield (Mg ha^-1), 1000-kernel weight (TKW, g), test weight (kg), single ear-weight (g), falling number (sec.), total pentosan (%), soluble pentosan (%), protein (%), and starch contents (%). For the first five traits data across ten environments and for the latter five traits data across six environments (WOH, PET, WAL in 2010 and 2011) were available. For falling number, data from one environment (WAL10) was missing. Plots were machine-planted and combine-harvested and grain yield was adjusted to a moisture concentration of 140 g H₂O kg^-1. The traits total pentosan (%), soluble pentosan (%), protein (%), and starch contents (%) were determined by near-infrared reflectance spectroscopy (NIRS, calibration by KWS SAAT AG, Einbeck) taking two samples per plot. Test weight refers to the weight of 100 l of grain.

We used a two-step procedure to analyse the phenotypic data. In a first step, ordinary lattice analyses of variance were performed for each environment and population separately 14 . In a second step, adjusted means of each genotype were used to estimate variance components across environments for each of the two segregating populations 15 . We assumed in this model that genotype and environment effects were random. Heritability on an entry-mean basis was estimated from the variance components as the ratio of genotypic to phenotypic variance 16 :

h 2 = σ G 2 σ G 2 + σ GxE 2 E + σ e 2 ER ,

where σ² _G denotes the genotypic variance, σ² _GxE the genotype × environment interaction variance, and σ² _e the error variance, E and R are the numbers of environments and replications, respectively. Test of skewness and kurtosis was performed for each trait 15 , pp. 79–81]. All statistical analyses were performed with the PLABSTAT software package 17 .

Genomic DNA was extracted from leaf samples of the two segregating populations at seedling stage using the procedure described by 18 . Pop-A (219 lines) was genotyped by simple sequence repeat (SSR) and single nucleotide polymorphism (SNP) markers, Pop-B (202 lines) by SSR- and diversity array technology (DArT)-markers, because SNP markers were not available. For SSR genotyping we followed established procedures 3 . Briefly, PCR products were separated on an ABI3130×l Genetic Analyzer (Applied Biosystems) according to manufacturer’s instructions. Allele identity was assigned using the GENEMAPPER software (Applied Biosystems). The Illumina iSelect Rye5k SNP array 7 was used for genotyping Pop-A. Briefly, a total of 200 ηg genomic DNA per plant was used for SNP genotyping on Illumina’s iScan platform using the Infineon HD assay for Pop-A and the GoldenGate assay for Pop-B following the manufacturer’s protocol. Raw hybridization intensity data processing, clustering, and genotype calling were performed using the genotyping module in the BeadStudio package (Illumina, San Diego, CA, USA). Data of the Rye5k SNP array for Pop-A were generated by Eva Bauer’s group at the Technische Universität München, Freising, Germany. SSR genotyping for both populations was done by KWS LOCHOW GMBH. DArT genotyping with the current rye array was performed by Triticarte Pty Ltd, Yarralumla, ACT, Australia (http://www.triticarte.com.au).

For the construction of the genetic linkage maps, polymorphic markers in each population were transformed into genotype codes according to the parental score. For quality checks a pre-selection with regard to their deviation from expected segregation ratio was performed. The genetic linkage maps were constructed with the software JoinMap® 4.0 19 . Seven linkage groups were established and chromosome names and orientation were assigned to linkage groups based on a subset of markers for which the positions have been published previously 5 . Locus order and genetic distances in centimorgan (cM) along the chromosome were calculated with the maximum likelihood algorithm using Kosambi’s mapping function 20 . After each run post mapping quality tools provided by JoinMap 4.0 were used excluding markers which do not fit into the genetic linkage map. To estimate average marker densities, markers with the same position were considered as single marker (unique locus) to avoid an overestimation of the marker density.

For each trait and each population, QTL analyses were based on the genetic linkage map of each population and on adjusted entry means using the software PLABQTL 21 . Markers with a distance below 1 cM were excluded by the software. Our QTL study is based on testcross performance of F_3:4 families. The genetic make-up of our populations allows only the detection of main effect QTL contributing to the additive genetic variation. Therefore, we assumed an additive model as outlined by 22 . QTL analyses were based on Composite Interval Mapping (CIM) with a multiple regression approach 23 . Cofactors were selected by stepwise regression according to Miller 24 , pp. 49] with an “F-to-enter” and “F-to-delete” value of 3.5. Estimates of QTL positions were attained at the position where the LOD score assumed its maximum in the region under consideration. The critical LOD threshold was analysed empirically for each trait using 1,000 permutation runs 25 . The proportion of the phenotypic variance explained by the QTL was determined by the estimator R ² _adj as described by 22 . Additionally, five-fold cross validation (CV) was applied to determine the bias of R ² explained by detected QTL resulting in R ² _CV . For this, the entire data set (DS) was split into five genotypic subsamples. Means from four out of five subsamples served as estimation set (ES) for QTL detection, localization, and estimation of genetic effects. The remaining subset forms the test set (TS) in which predictions derived from the ES are tested for their validity by correlating predicted and observed data. By permuting the respective subsets used for the ES and TS, five different cross-validation runs were used 22 . As a measure of reliability we give for each QTL the frequency of occurrence, i.e. the percentage of runs in which the QTL was detected within 1,000 cross-validation runs. The proportion of genotypic variance (p _G ) explained by the model was calculated: p _G  = R ² _adj /h ² , where R ² _adj is the adjusted proportion of phenotypic variance explained by the model and h² is the heritability of the trait 26 .

For all traits significant (P < 0.01) genotypic variation was observed in both populations (Table 1, Additional file 1). Variances due to genotype × environment interaction effects were also significantly (P < 0.05) greater than zero with the exception of starch content. Intensive phenotyping at up to ten environments led to moderate to high heritability estimates with the exception of falling number and soluble pentosan content that had lower estimates (0.3-0.4). Although parents of Pop-A only slightly differed in their testcross performance, the segregation variance of their progenies was similar to that of Pop-B. In contrast, parents varied widely for Pop-B: Lo117 was lower yielding, shorter and had a lower TKW than Lo115. Progeny means resembled parental means well. Both populations differed in their medians for all traits, but the genotypic variances indicated by the 50% quantile were rather similar (Figure 1). All phenotypic data did not significantly deviate from normal distribution.

Means of parents (1 = Lo115, 2 = Lo90 in Pop-A and Lo117 in Pop-B, respectively) and each of 220 testcross progenies of two populations (Pop-A, Pop-B), estimates of variance components (genotype, σ _G ²; genotype x environment interaction, σ _GxE ²; error, σ _e ²), and entry-mean heritabilities (h²) for agronomic and quality traits evaluated across six to ten environments.

^aAnalysis across 5 environments.

Coefficients of phenotypic correlation among traits were low to moderate in both populations (Table 2). Significant (P < 0.01) correlations were found between single-ear weight and grain yield or TKW in Pop-A (r = 0.54; r =  0.51). Plant height was significantly (P < 0.01) correlated with TKW and starch content (r = 0.5; r = −/+0.3) in both populations and additionally with grain yield, test weight and single ear weight in Pop-A. Among the quality traits, consistent significant coefficients were found between starch content and test weight (r = 0.2 to 0.5) or protein content (r = −0.6 to −0.7).

Estimated for each of 220 testcross progenies of Pop-A (below diagonal) and Pop-B (above diagonal) across ten (agronomic traits) and six (quality traits) environments.

*,** Significant at the 0.05 and 0.01 probability level, respectively.

^aAnalysis across 5 environments.

In this study, we used either a newly developed SNP marker array 7 or DArT markers for successfully generating high-density genetic linkage maps for both populations (Additional files 2 and 3). Finally, 813 and 921 markers were used for genetic linkage map construction for Pop-A and for Pop-B, respectively. The genetic linkage maps possess a total length of 980 cM for Pop-A and 2,349 cM for Pop-B. We found for Pop-A an average marker density of 1.48 cM and for Pop-B of 2.58 cM. The seven rye chromosomes were covered with each of 71 to 164 markers in Pop-A and 111 to 152 markers in Pop-B. The majority of neighbouring markers had a distance of less than 1 cM, only a few of the pairs had a genetic map distance larger than 10 cM (Figure 2). Map length and average marker density of the latter population are comparable with a rye population mapped by DArTs recently 5 . Obviously, the combination of DArT and SSR markers resulted in longer maps than that of SNP and SSR markers. However, for QTL mapping a marker distance of 15–20 cM is sufficient 27 . The maps, therefore, provide a solid basis for identifying genomic regions underlying the analyzed traits.

QTL analyses yielded one to nine QTL per trait, for soluble pentosan and protein contents no QTL were found in Pop-A (Table 3, for details see Additional files 4 and 5). Cross-validated phenotypic variances R² _CV were of similar order than the original R² _adj for all traits indicating a high quality mapping. QTL were distributed across all chromosomes (Figures 3 and 4). The only QTL for grain yield detected in Pop-A co-segregated with a QTL for TKW on chromosome 1R. For TKW, in each population two QTL with large effects were found on chromosomes 6 and 7 in Pop-A and on chromosomes 5 and 6 in Pop-B, respectively (Additional files 4 and 5). Some quality traits agreed in some of their QTL positions, especially in Pop-B.

Number (no.) of QTL, proportion of explained phenotypic variance before (R ² _adj ) and after cross validation (R² _CV) and genotypic (p _G) variance for each of 220 testcross progenies in two populations (Pop-A, Pop-B) estimated across ten (plant height and yield-related traits) and six (quality-related traits) environments.

^aAnalysis across 5 environments.

The explained genotypic variance of the individual QTL (p _G ) ranged from 5 to 55% (Figure 5). The highest values of p _G were observed for TKW, test weight, and falling number. For these traits and starch content several QTL with large effects and a frequency of recovery of about 90% were identified in both populations (Additional files 4 and 5). Parents of the two crosses contributed equally to the detected QTL.

Most QTL revealed significant (P < 0.01) QTL × environment interaction variances as expected for quantitative traits (Additional files 4 and 5). The QTL × environment interaction variances, however, were in all cases smaller than QTL variance (data not shown). The identified QTL had a high stability across locations as illustrated for grain yield and TKW QTL (Table 4). They had high additive effects in at least nine out of ten environments.

Effects of all detected quantitative trait loci (QTL) for grain yield (Mg ha^-1) and TKW (g) for each of 220 testcross progenies in two populations (A, B) at ten individual environments and combined across environments.

The precise estimation of phenotypic values is an important prerequisite for detecting QTL with a high power. We used F_3:4 lines and consequently, only 1.5 times the additive genetic variance was exploited and not 2 times the additive genetic variance as by using for instance recombinant inbred line populations. Due to the high inbreeding depression in the outcrossing species rye, testcross progenies of two elite line populations were used in this study. This allows exploiting only half of the total additive genetic variation, but takes into account that line per se performance is only of limited predictive value for hybrid performance in yield-related traits in rye 28 . Dominance is a serious obstacle when testcross performance is used in mapping biparental populations. When the tester contributes a strong dominant allele, the effect of the allele contributed by the other parent (an inbred line) is undetectable. The limitations on the use of strong elite testers were discussed by 29 30 . However, despite these limitations we observed significant genetic variation (P < 0.01) for all measured traits. The significance of genotype-by-environment interaction for all traits illustrate that multi-environmental phenotyping is indispensible. Entry-mean heritabilites reached 0.9 for plant height, TKW, and test weight in both populations and even grain yield had a heritability of 0.7. Lower heritabilities of some traits were mostly associated with a small, although significant amount of genotypic variation. Consequently, a high power to map QTL should be possible.

The parents in the Pop-A did not differ much in their testcross performance, which underscores a typical situation in elite rye breeding, which relies on crosses of the “best × best” type. In contrast, in Pop-B a superior (Lo115) and a lower performing (Lo 117) parent were crossed. Despite this, genotypic variance in Pop-A was similar (Figure 1) pointing towards a high importance of transgressive segregation. Obviously, both parents contributed different positive alleles at several loci for each of the traits. This is in accordance with quantitative-genetic theory assuming a large number of segregating loci with mainly additive effects for complex traits 31 . The contrast of the mean testcross performance of the parents and the average testcross performance of their segregating progenies is a test for net epistasis across the genome 32 . Consequently, the lack of differences between the mean performance of the parents and the progenies suggests that epistasis was not important in these populations. This can be explained by prevalence of additive gene action but might also be due to the fact that negative and positive effects at individual loci may cancel each other out.

Plant height is regulated to a large extent by dwarfing genes in elite wheat and barley populations 33 34 , because they have positive effects on elite wheat grain yield 35 . The genetic architecture of plant height in rye detected in this study, however, is in sharp contrast to the genetic architecture of plant height in elite wheat and barley. Five to nine QTL located on all seven chromosomes were identified for plant height in both elite rye populations. None of these QTL had a large genotypic effect and the recovery rates in permutation tests were low although the trait displayed a maximum heritability (0.9) and high proportions of explained genotypic variances (61 and 77% in Pop-A and Pop-B, respectively). This clearly illustrates that plant height in rye is a typical quantitative trait with a lot of segregating loci. Accordingly, 36 detected 11 genomic regions significantly contributing to plant height among testcrosses with two rye introgression line libraries. The main cause for the difference between wheat and barley on one hand, and rye on the other is that in rye the stem is used as reservoir for water and carbohydrates, especially when abiotic stress occurs. Very often grain yield is associated with tall plant stature as demonstrated by the significant correlation between plant height and grain yield in Pop-B (r = 0.30, P < 0.01). Therefore, no dwarfing gene has been successfully used in commercial rye cultivars until now although such a gene has been described and mapped 8 . Interestingly, we observed in Pop-B a QTL for plant height in the telomeric region of chromosome 5RL (QTL #8) where the Ddw1 dwarfing gene from a Russian source was located (Figure 4, 8 ). This QTL had in our population, however, a much smaller effect (R² = 18.9) than expected from a major gene although the recovery rate was 96%, perhaps reflecting a multi-allelic series at this locus. In conclusion, quantitative inheritance of plant height in rye is an example for a highly crop-specific trait.

Segregation at major effect QTL underlying grain yield in rye is not expected as any large-effect QTL would have been long fixed in the course of breeding. In agreement with this expectation, we observed for Pop-A where two high-yielding elite parents were crossed only one QTL explaining 5% of the genotypic variation (p _G ). In Pop-B deducted from a high- and a lower-yielding parent, seven QTL for grain yield were detected with p _G varying from 10 to 24% and low recovery rates. High effects in most of the tested individual environments reveal their environmental stability (Table 4). Accordingly, cross-validated variance explained by all QTL amounted to R² _CV = 51%. Here, linkage blocks due to limited resolution of QTL mapping in biparental populations may have lead to detection of clusters of linked QTL, thus underestimating the number of QTL involved in complex traits and overestimating their effects as shown by 37 in a comparison of a conventional F₃ and an intermated F₃ population.

Population size used was about 220 lines per population. This is similar to the US-nested association mapping (NAM) population 38 in maize, but an even greater size might be valuable for quantitative traits. For grain yield, 26 detected two QTL by analyzing 244 testcross progenies of maize, but up to seven QTL when regarding 976 progenies. We can, therefore, expect that the number of QTL estimated here might represent the lower limit of QTL segregating.

In conclusion, grain yield follows an infinitesimal model as already proposed by 39 and MAS for individual QTL seems not to be a realistic option. Yield components might identify better candidates. Indeed in Pop-A, where grain yield had only a minimum cross-validated R² of 3% only, TKW resulted in 64%. In Pop-B this value was similar high for both traits (>50%). Interestingly, in both populations individual QTL for TKW with large effects were found: QTL on chromosomes 6 and 7 in Pop-A as well as on chromosomes 5 and 6 in Pop-B.

The most prominent QTL for TKW on chromosome 7R had a remarkably high p _G value of 48% and a recovery rate of 87% (Additional file 4). Interestingly, a major QTL on this chromosome was already reported in a Petkus population associated with the marker SCM40 40 . This marker was located in the centromeric region of the chromosome 41 and might correspond to our QTL #5 for TKW in Pop-A. A second locus for this trait 40 was associated with SSR marker WMS5 located in the telomeric region of chromosome 5RL (renamed as GWM5, 4 ) and might correspond to QTL #2 in Pop-B. The two other loci with high p _G values and recovery rates >90% were located in a similar region on chromosome 6 in both populations but contributed by different parents. These results illustrate that some QTL for TKW (#4 and #5 in Pop-A and #3 in Pop-B) have such high effects that they might be caused by single genes. Further fine mapping of multiple alleles per QTL in this region will be needed to test this hypothesis.

A similar localization of QTL was observed for some yield-related traits. Two QTL for grain yield on chromosome 1R (Pop-A) and 4R (Pop-B) had similar positions like each of one QTL for TKW and test weight reflecting the significant correlations between the traits. The corresponding QTL were contributed by the same parent in both instances. Thus, an indirect improvement of grain yield by MAS of individual QTL for yield components that have a higher recovery rate might be feasible.

For yield and yield-related traits we tested our mapping populations at ten environments including locations from North, East and South Germany as well as Poland. Compared to other QTL studies, this is a high number and broad range of environments and we report here only QTL with significant effects across all environments. This surely restricts the number of QTL, but enables us to detect only environmentally stable QTL that should be valuable also in environments not tested here. QTL for TKW and grain yield detected in the combined analysis had similar effects in most of the individual environments.

In summary, QTL effects for grain yield were mainly small as expected from theory, three QTL with high effects, however, were detected for TKW. They were highly stable across environments and had a high recovery rate in the cross validation and, thus, should be investigated further.

Two to nine QTL were detected for each of the quality traits. Their p _G values were mostly >60%, starch content amounted even to 84% in Pop-B, although testing intensity was lower than for the yield-related traits. It should be noted that until now selection in practical breeding programs where the elite parents were derived from was solely based on falling number as a measure for pre-harvest sprouting. Despite its high heritability reported in earlier studies 42 it is a complex inherited trait 10 11 . In this study, we found each of two QTL for falling number in both populations and the QTL on chromosome 4 RL might correspond to a QTL already described 11 .

Generally, quality traits are regulated by a complex genetic network resulting in phenotypic and pleiotropic interactions among the traits as shown for the maize nested association mapping (NAM) population 38 . Also in hybrid rye, significant correlations were detected among traits ranging from −0.2 to +0.6 (Table 1). In particular, QTL for starch content overlapped with QTL for grain yield on chromosomes 3R, 4R and 7R in Pop-B. Accordingly, both traits showed a significant (P < 0.01) phenotypic correlation (r = 0.32). On chromosomes 1R, 3R and 5R a QTL for test weight had a similar position than a QTL for starch content indicating that plumber kernels might have higher starch content. Significant (P < 0.01) correlations were indeed found between starch content and test weight. The negative correlation between starch and protein content in both populations (r = −0.6 and r = −0.7) is well known from other studies in cereals (Jansen, pers. commun.), but could not be explained by co-localization of QTL in our populations. A negative correlation between starch and TKW was also observed in durum wheat 43 . The authors state that end-use of the kernel is clearly influencing the physio-chemical kernel characteristics implying properties such as milling quality, kernel hardness, and kernel protein content. In rye, a high selection pressure is on large, plump kernels, because rye as a crop tends to low kernel size limiting flour yield.

In conclusion, each of one QTL with large effects (p _G  > 20%) and high recovery rates (>90%) was found for test weight (QTL #2), falling number (#2) and starch content (#2) in Pop-A and starch content (#1) in Pop-B that should be considered for MAS.