Introduction

Maize is together with wheat and rice one of the three main sources of nutritional energy for humans and is extensively being used in animal feeding, either as grain or whole plant forage. The high efficiency of its C4 metabolism also makes it a resource for biofuel production, as attested by the recent development of BioGas in Germany (Herrmann and Rath 2012; Rath et al. 2013). In Europe, maize cultivation was adopted on a broad scale rapidly after the discovery of America (Rebourg et al. 2003) and a dramatic evolution of varieties occurred with the development of hybrids after World War 2. Dent lines from Northern American origin proved at that time to be highly complementary with Flint lines from European origin to combine productivity and environmental adaptation features for maize cultivation in Central and Northern Europe. These Flint × Dent hybrid varieties have proven to be extremely successful for both grain and silage production. Subsequent reciprocal selection of the two groups increased their differentiation and complementarity. However, potential of this material for biomass production remains poorly documented. Biomass quantitative trait loci (QTL) were detected in biparental crosses (Barrière et al. 2001, 2010), but to our knowledge there was no association genetics study for biomass yield on more diverse material. It is, therefore, of high interest to investigate the variability of this trait and the underlying genetic determinism within these two groups.

Panels of highly diverse materials have proven to be most useful to investigate the organization of diversity available for breeding at phenotypic and genotypic levels. The high density of molecular markers now available for many species makes it possible to discover major genes involved in the variation of traits of agronomic interest using Genome Wide Association Studies (GWAS) (Ozaki et al. 2002; Beló et al. 2007; Jones et al. 2008). Highly diverse panels have accumulated numerous historical recombination events, leading to a limited extent of linkage disequilibrium (LD), which is favorable to fine-map QTL. However, LD in association mapping panels is not only due to genetic linkage, but can also be caused by population structure, relatedness, drift, and selection (Jannink and Walsh 2003; Flint-Garcia et al. 2003). The contribution of these factors relative to linkage can be evaluated statistically (Mangin et al. 2012) and proved for instance to be substantial in grapevine and maize (Mangin et al. 2012; Bouchet et al. 2013). This component of LD due to population structure and relatedness can generate false positives and has thus to be taken into account in association mapping models to control false positives (Ewens and Spielman 1995; Thornsberry et al. 2001). Once these effects are correctly modeled, only marker-trait associations due to linkage should be detected. Efficient softwares were developed to infer population structure using genotypic data (Pritchard et al. 2000; Alexander et al. 2009), and several estimators of relatedness between individuals are available (VanRaden 2008; Astle and Balding 2009; Rincent et al. 2014). The estimated admixture (Q) and kinship (K) matrices can be included in the GWAS statistical model to control false positive efficiently (Yu et al. 2006).

The present work was conducted within the European Cornfed project (Rincent et al. 2012). Its objectives were to (1) investigate genotypic diversity in European and American Dent and Flint inbred lines (2) evaluate phenotypic variability within these two groups for traits related to biomass and flowering time, and (3) detect QTL for these traits by association mapping. For this, original Dent and Flint panels representing different periods of European maize breeding were assembled, characterized with the 50 k SNP array (Ganal et al. 2011) and evaluated per se and as hybrids using a tester line representative of the complementary group in a European field trial network. Association mapping was conducted using the approach recently developed by Rincent et al. (2014) to limit confounding between the tested marker effect and the random polygenic effect.

Materials and methods

Genetic material and genotyping data

Within the “CornFed” project, we developed two new specific Dent and Flint panels (CF-Dent and CF-Flint) aiming at analyzing more precisely the two genetic groups of interest for maize hybrid breeding in Northern Europe, as briefly described in a methodological context by Rincent et al. (2012). Both panels are composed of 300 lines aiming at best representing the diversity of these groups and different generations of genetic materials. These include the first commercially used inbred lines created from open pollinated varieties (OPVs), further referred to as first cycle lines, and more recent lines developed by public institutes or, in the case of the CF-Dent panel, private companies. The CF-Dent panel (see list in Table S1) includes 124 lines from the C-K panel (Camus-Kulandaivelu et al. 2006) determined as belonging to the “Corn Belt Dent” and “Stiff Stalk” groups with an admixture coefficient above 0.5. These were complemented by 58 from the University of Hohenheim (Stuttgart, Germany), 25 from the Misión Biológica de Galicia and the Estación Experimental de Aula Dei (CSIC, Spain), 12 from Centro Investigacións Agrarias de Mabegondo (CIAM, Spain), 58 from the ex plant variety protection (ex-PVP) lines (Mikel 2006; Nelson et al. 2008), and 23 recent lines from Institut National de la Recherche Agronomique (INRA, France). Similarly, the Flint panel (CF-Flint, see list in table S2) includes 118 lines of the C-K panel determined as belonging to the European Flint and Northern Flint groups with an admixture coefficient above 0.5. These were complemented by lines derived from breeding programs of the following institutes: 70 from the University of Hohenheim (Riedelsheimer et al. 2012), 56 from CSIC, 23 from CIAM, 23 from the Eidgenössische Technische Hochschule Zürich (ETHZ, Switzerland), and 10 recent lines from INRA. Four lines (FP1, C105, F816 and EM1027) attributed by STRUCTURE to both Dent and Flint groups with probabilities close to 0.5 in Camus-Kulandaivelu et al. (2006) were assigned to both CF-Dent and CF-Flint panels.

These panels were genotyped with the Illumina MaizeSNP50 BeadChip described in Ganal et al. (2011), as presented in Rincent et al. (2012). Individuals which had marker missing rate and/or heterozygosity higher than 0.1 and 0.05, respectively, were eliminated. Markers which had missing rate and/or average heterozygosity higher than 0.2 and 0.15, respectively, were eliminated from the concerned panel. In each panel, few individuals were highly related. One individual was removed from each pair when the inbreds were identical for more than 98 % of the loci. Three Dent lines and nine Flint lines were eliminated for this reason. Missing genotypes (below 2 % in both panels) were imputed with the software BEAGLE (Browning and Browning 2009). In total 276 and 259 phenotyped individuals passed the genotyping filters for the CF-Dent and CF-Flint panels, respectively (Tables S1 and S2). The filtered markers with a Minor Allele Frequency (MAF) above 0.05 were tested for association (42,214 and 39,076 markers for the CF-Dent and CF-Flint panels, respectively).

Diversity analysis

In the diversity analysis (Q and K estimation), an additional filtering criteria was used to select the SNPs. To reduce the ascertainment bias noted by Ganal et al. (2011), we only used the markers that were developed by comparing the sequences of nested association mapping founder lines (PANZEA SNPs, Gore et al. 2009). In total, 29,418 and 28,513 markers which had a MAF above 0.01 were considered for the diversity analysis in the CF-Dent and CF-Flint lines, respectively. Genotypic data of each panel were organized as G matrices with N rows and L columns, N and L being the panel size and number of SNP loci, respectively. Genotype of individual i at marker l (G i,l) was coded as 1, 0.5, or 0 for homozygote for an arbitrarily chosen allele, heterozygote, and the other homozygote, respectively.

Kinship was estimated following Astle and Balding (2009) as follows:

\(K\_{\text{Freq}}_{i,j} = \frac{1}{L}\mathop \sum\nolimits_{l = 1}^{L} \frac{{\left( {G_{i,l} - p_{l} } \right)\left( {G_{j,l} - p_{l} } \right)}}{{p_{l} \left( {1 - p_{l} } \right)}}\), where \(p_{l}\) is the frequency of the allele coded 1 of PANZEA marker \(l\) in the panel of interest; subscripts i and j indicate the lines for which the kinship was estimated. Note that contrary to the Identity By State (IBS, the proportion of shared alleles) estimation, this formula gives a higher weight to loci with a low gene diversity. Also, similarity is higher if two individuals share rare alleles than common alleles.

Admixture was estimated in the CF-Dent and CF-Flint panels using the software ADMIXTURE (Alexander et al. 2009) with a number of groups varying from 2 to 8. This software is based on the same statistical model as STRUCTURE (Pritchard et al. 2000; Falush et al. 2003) but uses a fast numerical optimization algorithm, which permits to considerably reduce computational time. The groups identified by the software were interpreted using the available pedigree information. Differentiation among genetic groups (Fst, Nei 1973) was estimated at each locus using the R-package r-hierfstat (Goudet 2005) for each number of groups Q (from 2 to 8), using the individuals attributed to one subgroup with a probability above 0.7 (these individuals are then considered as representative of the corresponding subgroup). Gene diversity (Expected heterozygosity, He; Nei 1978) was also estimated at each marker as 2p l (1 − p l ). The parameters Fst and He were averaged on all the markers to characterize the panels more globally. A principal coordinate analysis (PCoA) was performed on the genetic distance matrices (Gower 1966), estimated as \(1_{N,N} - {\text{K}}\_{\text{Freq}}\), where \(1_{N,N}\) is a matrix of ones of the same size as K_Freq. We also represented each panel by a network, in which two individuals were linked when their relationship coefficient was above 0.2, unlinked otherwise. For this, the genomic relationship matrix was transformed in a matrix of booleans indicating if the coefficients were above 0.2 or not. These networks were drawn with a Fruchterman and Reingold’s force-directed algorithm (Fruchterman and Reingold 1991) with the package « network » in R 3.0.0 (R development Core Team 2013).

Linkage disequilibrium (LD)

To estimate the minimum number of markers needed to cover the genome for GWAS, we estimated intra-chromosomic LD using all the markers. LD was first estimated as the squared correlation between the allelic doses at two markers (denoted by r 2) located on the same chromosome (Hill and Robertson 1968). As kinship has to be taken into account in the GWAS model to control false positives, we need to take it into account to estimate the number of markers required to cover the genome. For this reason, the approach of Mangin et al. (2012) was used to correct for kinship and estimate the part of LD only due to linkage (r 2K). To visualize major trends of LD variation along each chromosome, r 2K was averaged along the genome using a sliding window of 4 Mbp. This was represented on a graph together with marker diversity (He) and differentiation (Fst) after adjusting cubic smoothing splines along the genome using the R function smooth.spline (Hastie and Tibshirani 1990).

Genetic distances between loci were taken from the map of Ganal et al. (2011) based on the cross F2 × F252. Unmapped markers were positioned according to the local ratio between physical and genetic distances. The variation of LD with the genetic distances on each chromosome was adjusted to the model of Hill and Weir (1988), using only the pairs of markers separated by <4 cM. We estimated the LD decay for each chromosome as the abscissa of the intersection between the fitted curve and the horizontal line \(y = 0.1\). Knowing the length of each chromosome (in cM), we could estimate the minimum number of markers required on each chromosome to get an average r 2 or r 2K of 0.1 between each pair of adjacent markers.

Phenotypic data

The Flint and Dent lines were, respectively, crossed to a Dent (F353) and a Flint (UH007) tester to produce hybrid progenies for phenotypic evaluation. These two lines were representative of advanced materials within their respective groups. The two hybrid panels were evaluated for flowering and biomass production-related traits in trials located in Germany, France, and Spain. Two separate experiments were conducted for the Dent and Flint hybrids, with five locations in 2010, and six (CF-Dent) and five (CF-Flint) locations in 2011 (see Table S3 for geographical coordinates). Within each panel, the hybrids were divided into two groups of precocity and each group was evaluated in a different block. Each block was composed of 9 sub-blocks of 20 plots. Within the 20 genotypes of a given sub-block, 4 were repeated in 4 other sub-blocks of the same block, and 2 were repeated in 2 sub-blocks of the other block to estimate experimental error, block and sub-block effects. Male and female flowering time, plant height (PLHT, cm), dry matter content at harvest (DMC, %), and dry matter yield (DMY, Mg/ha) were recorded for each plot. Male and female flowering times were converted into growing degree days, considering a base temperature of 6 °C, using the mean daily air temperature measured at each location (these measures were, respectively, denoted by Tass_GDD6 for male flowering and Silk_GDD6 for female flowering). The Anthesis to Silking Interval (ASI_GDD6) was obtained by subtracting Tass_GDD6 from Silk_GDD6. DMC and DMY were observed at only nine of the ten trials for the CF-Flint panel. DMC and DMY were corrected by flowering precocity (DMCcorr and DMYcorr) by regressing the raw data on Silk_GDD6 for each block for DMC or for each trial for DMY.

$$\begin{gathered} {\text{DMC}}_{ijkl} = \mu_{jk} + \alpha_{jk} \times {\text{Silk}}\_{\text{GDD}}6_{ijkl} + E_{ijkl} \,\;{\text{and}}\quad {\text{DMCcorr}}_{ijkl} = \hat{E}_{ijkl} \hfill \\ {\text{DMY}}_{ijkl} = \mu_{j} + \alpha_{j} \times {\text{Silk}}\_{\text{GDD}}6_{ijkl} + E_{ijkl} \;{\text{and}}\quad {\text{DMYcorr}}_{ijkl} = \hat{E}_{ijkl} \hfill \\ \end{gathered}$$

with subscripts i, j, k and l the indices indicating, respectively, the genotype, the trial, the block, and the repetition in the block; \(\mu_{jk}\) and \(\mu_{j}\) indicate the intercepts for block jk and trial j, respectively; \(\alpha_{jk}\) and \(\alpha_{j}\) are the block and trial-specific regression coefficients on silking for DMC and DMY, respectively.

Outlier plots with extreme phenotypes showing a discontinuity from the rest of the distribution were excluded from the study (<2.5 % of the observations were removed in both panels). Least-squares means of genotypes over the global network were calculated with the GLM procedure (SAS 9.3, SAS Institute 2011) by adjusting for block and trial effects, except for DMY adjusted means, which were not corrected by block effects. Such a correction would indeed rely on the performances of the genotypes common to the two blocks, which are likely to be affected by competition effects (early genotypes being penalized in the “late” block and late genotypes favored in the “early block”). All effects were considered fixed at this step. Considering the important difference of residual variance among trials, we took heteroscedasticity into account by estimating a residual variance for each trial. For all the traits except DMY and DMYcorr, weighted least squares were computed using the following model:

\(Y_{ijkl} = \mu + G_{i} + T_{j} + T(B)_{jk} + E_{ijkl}\), with \(E_{ijkl}\)~ N(0,\(\sigma_{j}^{2}\)), and for DMY and DMYcorr with the model: \(Y_{ijkl} = \mu + G_{i} + T_{j} + E_{ijkl}\), with \(E_{ijkl}\)~ N(0,\(\sigma_{j}^{2}\)), where Y ijkl is the phenotype of the repetition l of genotype i in block k of trial j; \(\mu\) is the global mean; \(G_{i}\) is the fixed genotype effect of individual i; \(T_{j}\) is the effect of trial j; and \(T(B)_{jk}\) is the effect of block k within trial j. Other field effects (sub-block, row and column effects) were not included in the model as they were not significant.

Trait heritability was estimated at the level of the experimental design. For traits other than DMY and DMYcorr, variance components of heritability were estimated in two steps. In a first step, genotypes were considered as fixed effect to get block effect estimates based only on the repetitions across blocks.

$$Y_{ijkl} = \mu + G_{i} + T_{j} + B_{k(j)} + {\mathbf{E}}_{ijk}$$

In a second step, phenotypes were corrected by block effects and were analyzed considering genotype and genotype x trial effects as random:

$$Y_{ijkl} - \widehat{{B_{k(j)} }} = \mu + {\mathbf{G}}_{i} + T_{j} + {\mathbf{G}} \times {\mathbf{T}}_{ij} + {\mathbf{E}}_{ijk} ,$$

where \({\mathbf{GxT}}_{ij}\) is the random interaction effect between genotype i and trial j. For DMY and DMYcorr, variance components of heritability were estimated in one step only to prevent confounding block effects with competition between early and late lines:

$$Y_{ijkl} = \mu + {\mathbf{G}}_{i} + T_{j} + {\mathbf{G}} \times {\mathbf{T}}_{ij} + {\mathbf{E}}_{ijk}$$

Heritabilities were then estimated as \(h^{2} = \frac{{\sigma_{g}^{2} }}{{\sigma_{g}^{2} + \sigma_{e}^{2} /r + \sigma_{g \times t}^{2} /L}}\), where \(\sigma_{g}^{2} , \sigma_{e}^{2}\) and \(\sigma_{g \times t}^{2}\) are the variance estimates of the random effects \({\mathbf{G}}_{i}\), \({\mathbf{E}}_{ijk}\) and \({\mathbf{G}} \times {\mathbf{T}}_{ij}\), respectively; L is the mean number of environments; and r is the average total number of repetitions per genotype across environments. We also computed adjusted means and heritabilities for each trial by simplifying accordingly the above-described statistical models.

The lines per se were also evaluated at one trial for Tass_GDD6, Silk_GDD6 (Dent and Flint lines), and PLHT (only Flint lines). The Dent and Flint lines per se were evaluated in 2011 in Saint-Martin de Hinx and Gif-sur-Yvette (France), respectively. Least-squares genotype means of per se lines were calculated with the GLM procedure by adjusting for block effects. Variances of the per se experiment were estimated with the same mixed model used to estimate heritabilities at the trial level in the hybrid experiments. Analyses were conducted using SAS 9.3 (SAS Institute 2011).

Phenotypic characterization of the genetic groups within each panel

Genetic groups defined by admixture were compared within each panel for their phenotypic performance by estimating the genetic average of each group (denoted by \(\mu_{q}\)) using the following model:\(Y_{i} = \mathop \sum\nolimits_{q = 1}^{{N_{Q} }} \mu_{q} F_{i,q} + {\mathbf{E}}_{i}\), where \(Y_{i}\) is the adjusted mean of individual i, \(F_{i,q}\) is the admixture coefficient of individual i in group q, and N Q is the number of groups. N Q = 8 was considered for both panels to characterize the groups at the finer level that is informative to breeders.

Statistical model for association mapping

Mixed models are classically used to detect QTL while controlling false positive rate in GWAS (Yu et al. 2006). Relatedness among individuals is taken into account by considering that the random polygenic effects are not independent, with a covariance matrix determined by K. A fixed structure effect (associated to a structure matrix Q) can also be included if the dataset is highly structured. Comparison of p values obtained with different (Q + K) models revealed that K was sufficient to control both structure and relatedness (Fig. S1).

We tested each SNP with a MAF above 0.05 (42,214 and 39,076 SNPs in the CF-Dent and CF-Flint panels, respectively) in the following model: \(Y = X\beta + U + E\), where Y is the vector of phenotypes (adjusted means of the per se performances, or of the hybrid performances at one trial or in the whole trial network); \(X\) includes a vector of 1 and the genotypes at the tested marker (coded as 0, 0.5, or 1 as mentioned above); \(\beta\) includes the intercept and the additive effect of the tested marker (\(\beta_{l}\)), defined as the difference between the two homozygous genotypes; \(U{ \rightsquigarrow }N(0,K.\sigma_{gl}^{2}\)) is the vector of random polygenic effects, \(K\) being the kinship estimate and \(\sigma_{gl}^{2}\) the residual polygenic variance; \(E{ \rightsquigarrow }N(0,I.\sigma_{e}^{2}\)) is the vector of remaining residual effects with variance \(\sigma_{e}^{2}\); I is an identity matrix of size equal to the number of individuals (N), U and E are supposed to be independent. We used two different estimates of K in the model: K_Freq as presented above, and K_Chr (Rincent et al. 2014) which is computed only with the markers physically unlinked to the tested SNP: \({\text{K\_Chr}}_{i,j,c} = \frac{1}{{L_{ - c} }}\mathop \sum\nolimits_{{l \notin {\text{c}}}} \frac{{\left( {G_{i,l} - p_{l} } \right)\left( {G_{j,l} - p_{l} } \right)}}{{p_{l} \left( {1 - p_{l} } \right)}}\), where c is the considered chromosome, \(L_{ - c}\) is the number of markers not located on chromosome c. This second estimator was developed to take into account the fact that including markers in high LD with the tested SNP in the kinship estimation decreased power (Listgarten et al. 2012; Rincent et al. 2014). Each marker was tested for association with the different traits using a Wald test (Wald 1943) in ASReml-R (Gilmour et al. 2009). The scripts were written in R 3.0.0 (R development Core Team 2013). The statistical significance threshold was set to 0.05/Meff, which corresponds to a Bonferroni correction on Meff tests, Meff being the number of independent tests estimated as in Li and Ji (2005). This procedure evaluated 3,638 and 3,527 independent tests in the CF-Dent and CF-Flint panels, respectively, which led to a −log10(p value) threshold of 4.9 in both panels. Significant SNPs separated by <100 kb were considered as a single QTL for interpretation of the results. QTL size was estimated as the proportion of variance explained by the most significant SNP (Bouchet et al. 2013) of the QTL region after removing the structure effect (admixture being estimated with ADMIXTURE for N Q = 6).

Results

Diversity and structure analysis

The histograms of the minor allele frequencies (MAF) of the polymorphic PANZEA markers showed a slight deficit in rare alleles in the CF-Dent panel and a slight excess in the CF-Flint panel, compared to a uniform distribution (Fig. 1). This is consistent with typical flint lines (assignment to a given group above 0.95 at the structure level N Q = 8) showing a higher proportion of monomorphic PANZEA markers compared to dent lines (18 and 15 %, respectively; data not shown). MAF was on average slightly higher in the CF-Dent (0.25) than in the CF-Flint panel (0.24), which resulted in a lower index of gene diversity (\(\overline{He}\)) in the CF-Flint than in the CF-Dent panel (0.36 and 0.37, respectively). These differences between CF-Dent and CF-Flint panels for MAF and \(\overline{He}\) were highly significant (t test). Gene diversity He was variable along the genome (Fig. 2), with generally lower values in centromeric regions.

Fig. 1
figure 1

Histograms of the minor allele frequencies of the polymorphic PANZEA markers in the CF-Dent (a) and CF-Flint (b) panels

Full size image
Fig. 2
figure 2

Differentiation among groups (Fst, estimated at Q = 4), gene diversity (He) and linkage disequilibrium along the genome (physical distance) in the CF-Dent (a) and CF-Flint (b) panels. For each parameter a cubic smoothing spline was adjusted along the genome (with a smoothing parameter of 0.9). Centromere limits, Vgt1, and Vgt2 are indicated by blue, pink, and purple lines, respectively

Full size image

The cross-validation criterion proposed by ADMIXTURE suggested the presence of at least 4 main groups in each of the panels, and the criterion always improved with the number of groups (results not shown). For an expected number of genetic groups ranging from 2 to 8, all the subgroups identified by ADMIXTURE were interpretable in terms of pedigree and/or geographical origins. The genetic groups were composed of lines sharing a common recent ancestor (for example F252 in the CF-Dent panel), or a common ancestral origin (for example Northern Flint in the CF-Flint panel). We noted that groups at level N Q could generally be related to groups at level N Q + 1 by the subdivision of one subgroup into two (see Fig. S2 for an empirical synthesis). For a given number of groups, the differentiation among groups was higher in the CF-Dent than in the CF-Flint panel (Table 1). The mean Fst over the genome increased with the number of groups in both panels, but it reached a plateau at 7 groups in the CF-Flint panel. Fst was variable along the genome; in particular, when considering four groups (Fig. 2), peaks of Fst were clearly visible in the CF-Dent panel (chromosomes 7 and 10) and in the CF-Flint panel (chromosome 8).

Table 1 Average differentiation index Fst among the genetic groups for different numbers of groups (N Q) varying from 2 to 8
Full size table

The first two axes of the PCoA explained 16.1 and 15.7 % of the variability in the CF-Dent and CF-Flint panels, respectively (Fig. 3). The different groups identified by ADMIXTURE were clearly identifiable on the PCoA graphs. The first axis separated the Iodent from the non-Iodent lines in the CF-Dent panel and the Northern Flint from the other Flint lines in the CF-Flint panel. Note that extreme positions along the axes were observed for the key founders of these groups (e.g. PH207 for Iodent, B73 for Stiff Stalk, Mo17 for Lancaster, D105 for Northern Flint).

Fig. 3
figure 3

Principal coordinate analysis (a, b) and network (c, d) representations of the CF-Dent (a, c) and CF-Flint (b, d) panels. Both representations are based on the covariance matrix K_Freq. The most representative individuals of each subgroup at N Q = 4 (admixture above 0.7) were colored. Few key individuals are indicated in each panel (PH207, D06, B73, B14, Mo17, OH43, respectively, numbered from 1 to 6 in CF-Dent and D105, F2, F7, D171, L03, and EC46, respectively, numbered from 7 to 12 in CF-Flint). EC46 is a typical line of the group “Aranga”. In the network representation, individuals are linked if their covariance is above 0.2, unlinked otherwise. In these networks, distances are not informative

Full size image

Network representations of the CF-Dent and the CF-Flint panels revealed clusters of related individuals and isolated lines (Fig. 3). The shape of the network was different in the two panels: the CF-Dent panel was composed of isolated lines having no or only limited relatedness to any other line and few clusters of related individuals. The network of the CF-Flint panel also revealed clusters of related individuals but was much looser than the network of the CF-Dent panel. Groups identified with ADMIXTURE at N Q = 4 were in good agreement with the network visualization. In each panel, one of the four groups (called “Others” in Fig. 3) was composed of more heterogeneous material including many 1st cycle lines and appeared fragmented in the network.

Linkage disequilibrium

The LD was on average more extended in the CF-Dent than in the CF-Flint panel (0.21 and 0.12 cM to reach an r 2 of 0.1 on average over all chromosomes, respectively, see Table 2). Inter-chromosomal LD was observed in both panels (Fig. S3). When considering physical distances, LD extent was highly variable between chromosomes and along chromosomes (Fig. 2), being more extended in centromeric regions. Taking relatedness into account substantially reduced the extent of LD in both panels, particularly in the CF-Dent panel (Table 2), and considerably reduced inter-chromosomal LD (Fig. S3). For intra-chromosomal LD, the decrease observed when considering relatedness was particularly strong for chromosomes 3 and 8 in both panels, and chromosomes 4 and 7 in the CF-Dent panel only (Table 2). The chromosomes 3, 4, and 8 in the CF-Dent panel had a more extended LD (0.16, 0.18 and 0.18 cM to reach a r 2K of 0.1, respectively) than the others (between 0.09 and 0.13 cM, Table 2). In the CF-Flint panel, all the chromosomes displayed similar r 2K except chromosome 8 for which LD was more extended (0.14 cM to reach a r 2K of 0.1 for chromosome 8, only 0.09–0.10 cM for the other chromosomes). Knowing the length of the chromosomes (in cM), these statistics allowed the estimation of the minimum number of markers required to cover the genome (assuming evenly spaced markers on the genetic map): more markers are needed in the CF-Flint (24,387) than in the CF-Dent panel (19,000) to get a r 2K of 0.1 between evenly spaced adjacent markers (Table 2).

Table 2 Extent of linkage disequilibrium and number of markers needed to reach an average r 2 or r 2 K of 0.1 for each chromosome (Chrom) in the two panels (CF-Dent and CF-Flint)
Full size table

Phenotypic variation

We observed a substantial variability for all the traits in both panels and in both hybrid and per se evaluations (Tables 3 and S4), with for instance least-squares means of DMY of the hybrids over the trial network varying between 11 and 20 Mg/ha in both panels. Averages and heritabilities were variable between trials for a given trait (Table S3), but high heritabilities were observed at the trial network level (over 0.73 and 0.65 in the CF-Dent and CF-Flint panels, respectively). For most traits, heritability was higher in the CF-Dent than in the CF-Flint panel. This was related to higher residual variances in the CF-Flint panel. Tass_GDD6 and Silk_GDD6 were the most heritable traits (0.96 and 0.97 in the CF-Dent and CF-Flint panels, respectively). ASI_GDD6 and yield traits (DMC, DMCcorr, DMY and DMYcorr) were the less heritable traits. The lowest heritability was 0.65 for ASI_GDD6 in the CF-Flint panel. The heritabilities of the per se evaluations were close to the heritabilities of the hybrid trial network (Table 4), although inbred lines were evaluated at only one trial. This was due to much higher genetic variances in the per se evaluation (up to 6.4 times higher). The correlations between the hybrid and the per se adjusted means were quite high for Tass_GDD6 and Silk_GDD6 (between 0.68 and 0.87), but lower for ASI_GDD6 (between 0.22 and 0.43). These correlations were higher in the CF-Dent than in the CF-Flint panels for the three traits that were common to both panels (Table 4).

Table 3 Variances in the hybrid experimental design at the network level
Full size table
Table 4 Variances in the line per se experimental design (one trial), and correlation between the line per se and the hybrid adjusted means (Corr Hyb/PerSe)
Full size table

Phenotypic characterization of the genetic groups within each panel

For hybrid performances, we observed differences between the genetic groups identified within the two panels. Adjusted R 2 explained by the groups were between 0.11 and 0.47 in CF-Dent and between 0.05 and 0.41 in CF-Flint when considering 8 groups (Table 5). In the CF-Dent panel, the lines related to D06 or to F252 displayed the earliest flowering time and the highest DMC and DMCcorr (Table 5). The Lancaster and Stiff Stalk groups displayed the latest flowering time and were also the most productive (DMY of up to 17.6 Mg/ha). In the CF-Flint panel, the Lacaune descendants (Fv7 related), the Northern Flints, and the Hohenheim Flints displayed the earliest flowering time and the highest DMC and DMCcorr. Groups from Southern Europe (related to CIAM Aranga and descendants from Italian open pollinated varieties (OPV) or from other non-Northern Flints introductions into Europe) displayed the latest flowering. The lines related to CIAM Aranga, to UH006 or to Fv7 (Lacaune) were the most productive when crossed to the Dent tester (DMY of up to 16.6 Mg/ha). Despite the negative correlation between flowering precocity and productivity in both panels (results not shown), we could observe different levels of productivity at a given flowering time in some cases. For example, lines related to B73 and those related to OH43 both displayed late flowering but the first group was more productive. In the CF-Flint panel, the group “CIAM Aranga and EC18 related” was by far the most productive, although earlier than other groups. We observed that the three Flint groups which had the highest contribution in first cycle lines, namely Italian OPVs, Pyrenean and Northern Flints, were the least productive, with DMY below 15 Mg/ha (Table 5). A similar trend was found for Dents, with most first cycle lines grouped in the “Minnesota13” group, which displayed the lowest value for DMYcorr. We also noted substantial variation within genetic groups (see, for example, Iodent and Italian OPVs in Tables S1 and S2, respectively), consistent with the limited proportion of variance explained by admixture for all the traits. Within a given group, the most typical lines (admixture above 0.98) could differ by up to 5 Mg/ha (e.g. non-admixed individuals of the “UH006 family” group ranged from 12.7 to 17.7 Mg/ha, Table S2). A formal analysis of genetic gain over breeding generations could not be conducted due to the complexity of the pedigrees but some interesting trends could be noted. For instance, within the PH207 group, most lines derived from the founder PH207 appear superior to it in terms of performance (Table S1).

Table 5 Characterization of the different genetic groups at N Q = 8 in the CF-Dent and CF-Flint panels
Full size table

Association mapping results

The complete lists of significant SNPs are presented in Tables S5 and S6, and the most significant associations (−log(p value) above 5) are summarized in Tables 6 and 7. The highest −log(p value) was 9.98 on chromosome 8 in CF-Dent and 6.71 on chromosome 1 in the CF-Flint panel, corresponding both to associations with flowering traits (Tass_GDD6 or Silk_GDD6).

Table 6 Most significant associations in the CF-Dent panel at the network level
Full size table
Table 7 Most significant associations in the CF-Flint panel at the network level
Full size table

Regarding the two statistical methods which were used, both kinship estimators (K_Freq and K_Chr) appeared efficient to control false positive rate, as revealed by QQ-plots (Fig. S1). At the chosen Bonferroni threshold, the kinship estimator K_Chr permitted the discovery of more SNPs than K_Freq for all the traits in both panels, at the trial or at the network level, except DMYcorr in the CF-Flint panel (Table 8). K_Chr permitted the discovery of 62 additional SNPs in the CF-Dent panel, and 15 in the CF-Flint panel (11 and 7 at the network level, corresponding to an increase of 41 and 39 % in the CF-Dent and CF-Flint panels, respectively). Only 1 and 3 SNPs were identified with K_Freq but not with K_Chr in the CF-Dent and CF-Flint panels, respectively.

Table 8 Statistics on the number of significant SNPs and QTL regions in the CF-Dent and CF-Flint panels evaluated on tester, for all traits, using K_Freq or K_Chr as kinship
Full size table

Comparing the two panels, the total number of SNPs (Table 8) significant in at least one environment or at the network level with one of the two methods (K_Freq or K_Chr) was more than two times higher in the CF-Dent (258 SNPs) than in the CF-Flint panel (116 SNPs). This difference was less pronounced when considering regions (QTL) instead of SNPs (173 and 108 QTL identified in the CF-Dent and CF-Flint panels, respectively). The only exception to this global trend was PLHT, for which more QTL were discovered in the CF-Flint than in the CF-Dent panel. The QTL identified in these panels explained between 1 and 15 % and between 3 and 12 % of the phenotypic variance that was not explained by the admixture groups in the CF-Dent and CF-Flint panels, respectively.

Considering traits, more SNPs were discovered for DMY, DMYcorr, Tass_GDD6. and Silk_GDD6 than for DMC, DMCcorr, and ASI_GDD6 (Table 8). However, most of the DMY and DMYcorr SNPs (96–100 %) were significant in only one environment, whereas some Tass_GDD6 and Silk_GDD6 QTL were stable across most of the environments (Fig. 4, chromosomes 2, 3, 4, 7, and 8; Fig. 5 chromosome 1). The proportion of SNP significant in only one environment was higher in the CF-Flint than in the CF-Dent panel, with the exception of PLHT. At the network level, more SNPs were declared significant for Tass_GDD6 and Silk_GDD6 in the CF-Dent panel, and for DMY and DMYcorr in the CF-Flint panel. Note that some of the significant SNPs were associated with more than one trait (Tables S5 and S6). These pleiotropic effects particularly concerned the following couples of traits: Tass_GDD6 and Silk_GDD6, Tass_GDD6 (or Silk_GDD6) and DMC (or DMY, or PLHT), PLHT and DMY (or DMC, or Tass_GDD6, or Silk_GDD6). In the CF-Flint panel, one SNP was associated with Tass_GDD6, Silk_GDD6, PLHT, and DMC.

Fig. 4
figure 4

Significant SNPs identified in the CF-Dent panel for the different traits in the different environments and for the global adjusted means. Circle diameters are proportional to the −log10 (p value), and the red color indicates the additional significant SNPs when using K_Chr as covariance matrix (the markers physically linked to the tested SNP are not used to estimate kinship). Chromosomes are separated by black lines; Vgt1 and Vgt2 (from right to left) are indicated by purple lines. The different traits are A Tass_GDD6, B Silk_GDD6, C ASI_GDD6, D PLHT, E DMC, F DMCcorr, G DMY, H DMYcorr. The trials are: a Mons 2010, b Pontevedra 2010, c La Coruna 2010, d Roggenstein 2010, e Einbeck 2010, f Mons Late Planting 2011, g Moulon 2011, h Mons Early Planting 2011, i Pontevedra 2011, j La Coruna 2011, k Pocking 2011

Full size image
Fig. 5
figure 5

Significant SNPs identified in the CF-Flint panel for the different traits in the different environments and for the global adjusted means. Circle diameters are proportional to the −log10 (p value), and the red color indicates the additional significant SNPs when using K_Chr as covariance matrix (the markers physically linked to the tested SNP are not used to estimate kinship). Chromosomes are separated by black lines; Vgt1 and Vgt2 (from right to left) are indicated by purple lines. The different traits are A Tass_GDD6, B Silk_GDD6, C ASI_GDD6, D PLHT, E DMC, F DMCcorr, G DMY, H DMYcorr. The trials are: a Mons 2010, b Pontevedra 2010, c La Coruna 2010, d Roggenstein 2010, e Einbeck 2010, f Moulon 2011, g Ploudaniel 2011, h Pontevedra 2011, i La Coruna 2011, j Pocking 2011

Full size image

When testing associations with per se adjusted means, QTL of Tass_GDD6 and Silk_GDD6 were discovered but only one QTL of ASI_GDD6 (in the CF-Dent panel) and no QTL of PLHT. Again, more QTL were found in the CF-Dent panel (25) than in the CF-Flint panel (14). Most of these QTL were located on chromosomes 3 and 8 in the CF-Dent panel and on chromosomes 1, 3, and 9 in the CF-Flint panel. Five QTL of Tass_GDD6 and four of Silk_GDD6 were found associated with both hybrid and per se performances (including ZCN8, see discussion) in the CF-Dent panel. In the CF-Flint panel, only one QTL of Tass_GDD6 and one QTL of Silk_GDD6 were found in both hybrid and per se evaluations.

Discussion

In the present study two highly diverse panels (a flint and a dent) were characterized genotypically and phenotypically for traits related to flowering time and biomass production. This allowed the evaluation of the available diversity and the detection of QTL. This is a first step to analyze the breeding potential of this material for traits related to biomass.

Genetic diversity organization

Both panels displayed a high genetic diversity attested by a high proportion of polymorphic PANZEA-markers in both panels (85 and 82 % in the CF-Dent and CF-Flint panels, respectively). Average gene diversity (\(\overline{He}\)) at polymorphic loci was high (0.37 and 0.36 in the CF-Dent and CF-Flint panels, respectively). These (\(\overline{He}\)) values are in the upper range of those reported in diversity studies based on SNPs (Hamblin et al. 2007; Lu et al. 2009; Truntzler et al. 2012; Van Inghelandt et al. 2010; Bouchet et al. 2013). The lower polymorphism and slightly lower gene diversity observed in the CF-Flint panel are consistent with the observations of Bouchet et al. (2013), who hypothesized that this could be the consequence of the severe bottleneck encountered by the Flint material when diverging from tropical germplasm. As in our study (Fig. 1), Bouchet et al. (2013) observed more rare alleles in the Flints. Based on this, they hypothesized that the initial bottleneck was followed by an expansion involving migrations from different origins, each with limited contributions. This hypothesis is consistent with the more complex population structure observed for the Flints (see below). The grouping based on the molecular information revealed the complex structure of both panels. From N Q = 2 to N Q = 8, all the identified groups could be interpreted using the pedigree information and/or known assignation to heterotic groups (Fig. S2). The groups identified in the CF-Flint panel appear to be related to the ancient history of this material. In particular, the first introductions of maize into Europe (from the Caribbean into Southern Europe by Columbus in 1493, and from Northern America into Northern Europe before 1539, Rebourg et al. 2003; Revilla et al. 2003) are still clearly visible in our results (Southern OPVs vs. Northern Flint, respectively, Fig. 3). The CF-Dent panel does not show such ancient historical patterns, consistent with the fact that this group originated from Corn-Belt Dent open pollinated varieties which displayed limited population structure (Camus-Kulandaivelu et al. 2006). Admixture groups observed in our study appear to be the result of the diverse breeding strategies which have been applied since the early development of hybrid maize in the USA. The network and PCoA visualizations revealed that the material available relates to a large extent to a limited number of key lines, in particular in the CF-Dent panel (Fig. 3). Each key line and the material derived from it generated structure groups which were also clearly visible in the network and PCoA visualizations. This clustering around key lines (B73, Mo17, and PH207) corresponds to the three main Dent groups (Stiff Stalk, Lancaster and Iodent, respectively) and was also shown for instance by Dubreuil et al. (1996), and Romay et al. (2013) using genotyping by sequencing (GBS) data. The fact that the CF-Flint panel appeared less structured by modern breeding than the CF-Dent panel is consistent with the fact that Flint genetic group was submitted to less breeding cycles. Hybrids involving Flint parents are indeed recent (1960s) compared to the first Dent hybrids (1930s) developed in the USA. Also, as contrary to Dents, hybrids between two Flint lines only had a very limited success, this group was not sub-structured into complementary heterotic groups (Revilla et al. 2002). The different history of the genetic groups resulted in higher differentiation of the groups in the CF-Dent than in the CF-Flint panel (this was true for N Q = 2 to N Q = 8, Table 1). Despite efforts made to assemble materials from different institutes, it appeared that some heterotic groups or families were common to these institutes. There are, however, some noticeable exceptions like CIAM-Aranga and Hohenheim Flints which appear specific for the institutes which created the corresponding lines.

Relatedness between individuals greatly influenced LD between pairs of markers (in particular between unlinked markers, Fig. S3) in both panels, particularly in CF-Dent. When taking kinship into account, LD remained higher in the CF-Dent panel. As observed in previous studies (Van Inghelandt et al. 2010; Bouchet et al. 2013), LD decreased with the physical (and genetic) distance. LD decay was variable between chromosomes, with a higher extent on chromosomes 3, 4, and 8 in the CF-Dent panel, and chromosome 8 in the CF-Flint panel. This is in accordance with Djemel et al. (2013), Romay et al. (2013) and with Khobragade et al. (pers. com.), who identified particularly long haplotypes on chromosome 4, in regions including important domestication genes. Other important genes related to flowering time (Vgt1 and Vgt2) are located on chromosome 8 (Chardon et al. 2005; Salvi et al. 2007, 2009; Veyrieras et al. 2007; Ducrocq et al. 2008; Van Inghelandt et al. 2012; Bouchet et al. 2013). The slight drop of diversity in the region of Vgt1 and Vgt2 in the CF-Flint panel (Fig. 2) may be due to the fixation of the early alleles during adaptation to short growing seasons. The higher LD extent in the CF-Dent panel resulted in a reduced number of SNPs required for a minimum coverage of the genome (19,000 markers in comparison to 24,387 markers in the CF-Flint). The number of SNPs available in GWAS in the panels (42,214 and 39,076 in the CF-Dent and CF-Flint panels, respectively) makes it possible to conduct a first genome-wide analysis. However, these available markers are not evenly spaced along the genetic map, and a LD of 0.1 between adjacent pairs of SNP is insufficient to detect QTL of small to intermediate effect in our panels. In the CF-Flint panel, fewer markers were available for GWAS, whereas more markers were needed to cover the genome than in the CF-Dent panel. This could lead to a lower power in the CF-Flint panel in some regions of the genome. In both panels, we expect that a substantial gain in power could be obtained by increasing the number of markers (by combining GBS, sequencing, and imputation for example).

Also, one of the main limitations in the dissection of quantitative traits is the size of the population under study, which affects GWAS power and the reliability of genomic predictions. For this reason, the panel size should be as large as possible. However, we showed in this study that, at some point, the sampling of additional individuals often results in relatedness (possibly high, Fig. 3), which may decrease GWAS marginal gain of power. This highlights the importance of screening collections of landraces and of first cycle lines, which can probably be used to increase panel size and diversity without increasing too much relatedness, and as a result increase the potential of the panels for QTL detection, and for the inference of evolutionary events.

Trait variation within and among genetic groups

All traits in both panels showed high genetic variability, which resulted in high heritabilities at the trial network levels. Male and female flowering time (Tass_GDD6 and Silk_GDD6) were the most heritable traits (above 0.96 at the network level), and Anthesis to Silking Interval (ASI_GDD6), which is highly sensitive to environmental stresses, was the least heritable trait (0.73 in CF-Dent and 0.65 in CF-FFlint). Heritabilities at the trial-network level were in the range that is expected for the observed traits. Trial heritabilities of yield traits (DMC, DMCcorr, DMY and DMYcorr) were highly variable between trials, likely because of different environmental conditions and different management practices. The hybrid heritabilities were slightly higher for the Dent than for the CF-Flint panel except for flowering time. This is mostly due to higher residual variances in the CF-Flint panel, partly explained by plant lodging in some of the trials.

The heritabilities in the per se single location evaluations are close to the heritabilities at the trial-network level in the hybrid evaluation for Tass_GDD6 and Silk_GDD6. This is due to a genetic variance 5.5–6.4 times higher, and a residual variance only 1.2–3.4 times higher than in the hybrid experimental design (Tables 3, 4). This difference of genetic variability between per se and hybrid evaluation is higher than the fourfold increase expected under an additive model. This suggests the existence of a substantial amount of non-additive genetic effects. The range of correlations between per se and hybrid adjusted means revealed the importance to evaluate biomass production potential of the lines in hybrid progenies and not per se only.

The high genetic diversity and phenotypic variability of these two panels encourage the development of more productive biomass maize. Comparison of group materials revealed by population structure analysis showed a significant effect on all traits (Table 5). It highlighted groups with original characteristics like the “CIAM Aranga and EC18 related” group in the Flints, or the Stiff Stalk lines (particularly those related to B73) in the Dents, which displayed a high productivity relative to their earliness (Table 5). High variances nevertheless exist within genetic groups. Although a formal analysis was not possible due to the complexity of pedigrees, we observed some groups for which recent materials were more productive than that of ancestral founder lines (e.g. PH207 is considered by breeders as the most representative of the related founders of the Iodent group, and displays a DMYcorr value below that of its three direct descendants PHH93, PHG50 and PHG83, see Table S1). This reveals that both Flint and Dent groups have undergone genetic progress (Tables S1 and S2). However, substantial variability remains in the more recent lines (e.g. group “CIAM Aranga” in Table S2), which is promising for further breeding.

Association mapping results

The distribution of the p values (QQ-plot, Fig. S1) illustrates that a random polygenic effect was required to control false positive rate efficiently and that both K_Freq and K_Chr were efficient for this (distribution near diagonal for p values above 0.01). However, the use of K_Chr instead of K_Freq substantially increased the number of significant SNPs (increase of around 40 % in both panels). Mixed model with K_Freq as kinship appeared indeed to be too conservative in simulations, and as a result less powerful (Rincent et al. 2014). This confirms the importance of removing markers in LD with the tested marker from the kinship estimation to limit “proximal contamination” (Listgarten et al. 2012). As expected from simulations in Rincent et al. 2014, the gain of power was less important in the CF-Flint than in the CF-Dent panel.

QTL were identified for all the traits in both panels (Tables 6, 7, 8, S5 and S6; Figs. 4, 5). Globally, more QTL were discovered in the CF-Dent than in the CF-Flint panel in the hybrid evaluation (173 and 108 QTL, respectively) and in the per se evaluation (25 and 14 QTL, respectively). This is consistent with the higher MAF, number of markers and LD extent in the CF-Dent panel (see above).

As expected based on knowledge of trait complexity and consequences on power, more QTL were found for Tass_GDD6 and Silk_GDD6 than for more complex traits (ASI_GDD6 or DMC), and these flowering QTL were more stable across environments. In particular, four regions of the genome in the CF-Dent panel (Fig. 4, chromosomes 2, 3, 4, and 8) and one region of the genome in the CF-Flint panel (Fig. 5, chromosome 1), were associated with flowering time in most of the environments. Polymorphism in the vicinity of ZCN8 gene appeared as the most significant in both hybrid and per se evaluations in the CF-Dent panel. It corresponds to the Vgt2 QTL found in numerous studies (Chardon et al. 2005; Salvi et al. 2007, 2009; Veyrieras et al. 2007; Ducrocq et al. 2008; Van Inghelandt et al. 2012; Bouchet et al. 2013; Romay et al. 2013). Note that it was not significant in the CF-Flint panel, neither in hybrid nor in per se evaluations, consistent with the quasi fixation of the early allele in Flint material reported by Bouchet et al. (2013). None of four other regions for flowering time appeared as strongly significant in Bouchet et al. (2013). Also, the strong association with days to silking corresponding to gene ZmCCT (Romay et al. 2013) on Chromosome 10 was not detected in our study, probably because the late allele at this locus (Ducrocq et al. 2009) is underrepresented in our panels, or marker density was too low in this region for capturing this effect.

PLHT was an exception to the global trend, as more QTL were found in the CF-Flint than in the CF-Dent panel (18 and 8 QTL, respectively), probably because it is the only trait (with Tass_GDD6 to some extent) which had a much higher genetic variance in the CF-Flint than in the CF-Dent panel (Table 3). We found common associations with the study of Peiffer et al. (2014), in particular in the CF-Flint panel (e.g. the QTL close to position 249 Mb on chromosome 1 near the gene brassinosteroid-deficient dwarf1, Pettem 1956). Interestingly, most of the PLHT QTL are not associated with flowering traits, as also found by Peiffer et al. (2014). As both flowering time and plant height are increasingly documented in the literature and less subject to GxE interactions than yield, a formal meta-analysis of our study and literature investigations would be highly beneficial to go beyond these preliminary trends.

For DMY or DMYcorr, many significant associations were discovered, but they were highly instable between environments (more than 96 % of these SNPs were significant in only one environment). The genetic determinism of these traits is more difficult to investigate because of interactions with the environment and/or because they are highly integrative. We noted that some associations for DMY were common to flowering time, suggesting a pleiotropic effect of the corresponding QTL. Note, however, that the QTL observed at network level for DMYcorr and DMY at position 154,077,833 bp on chromosome 1 (Table 7) in the Flint and position 190,732,112 bp on chromosome 5 in the CF-Dent panel (Table 6) do not belong to this category and, therefore, would be particularly interesting to select for biomass yield without modifying flowering time. These two QTL are all the more promising that the favorable alleles were both found with a frequency below 0.2. Note that the favorable alleles of these two QTL were absent for most of the groups, except in the Lancaster, the Stiff Stalk and the D06-related groups for the dent QTL, and in the Pyrenean group and the descents from non Northern Flint introductions for the flint QTL. Finally, DMC, DMCcorr, and ASI displayed the fewest number of detected QTL, highlighting that they are most likely affected by numerous factors of small effects and/or strong environmental effects.

Most of the significant SNPs identified with the hybrid adjusted means were different from those identified with the per se adjusted means. This could be due to the fact that lines were observed per se in only one environment, or to interactions between alleles (dominance and possibly epistasis), which was also shown by the genetic variance higher than expected in the per se evaluations. This effect was more pronounced in the CF-Flint than in the CF-Dent panel. As the proportion of SNPs significant in only one environment was also higher in the CF-Flint panel, we can hypothesize that the CF-Flint panel is probably submitted to more gene × gene and gene × environment interactions. This probably relates to the fact that flint inbreds were specifically selected in different European environments, while the dent inbreds were mainly introduced from the USA because they had good performance over a large range of environments.

Conclusions

We could illustrate, using genotypic and phenotypic information, that Dent and Flint groups have a different history and that this has strong consequences on diversity, variability, and LD extent, which in turn influence detection power. The combination of phenotypic and genotypic data permitted the identification of flowering time and biomass-related QTL in both panels. This study would probably profit from increasing the number of markers and population size. Analyzing this dataset with statistical models including interaction terms would also probably be beneficial. Although further analyses are required, the favorable biomass QTL alleles detected in this study are potentially of considerable interest, because they could be introgressed in elite material to increase productivity.

Author contributions

AC, CB, PR, JMG, AEM, MO, PD, PF, CCS conceived the experiments; PR, RAM, JMG, AEM, WS, JL, CB, AC provided genetic materials; DM, VC, FD, SN, TA, DB managed DNA extraction and genotyping; CG, EB, CCS, NM, MO, CB, PJ, LC, JMG conducted field experiments; RR, SB and SN conducted statistical analyses; RR, LM, AC wrote most of the MS with the help of JMG, PR, EB, AEM, CB, SB.