The 40 lines within Exp. 1 represent a broad range of Central European forage maize germplasm, and were extremes within a larger collection of >300 maize inbreds with respect to stover cell-wall digestibility (unpublished data). Thirty-five lines originated from the current breeding program of KWS Saat AG and five lines were from the public domain (AS01, AS02, AS03, AS39, and AS40, identical to F7, F2, EP1, F288, and F4, respectively).

The inbred lines were evaluates in Grucking (sandy loam) in 2002, 2003, and 2004, and in Bernburg (sandy loam) in 2003 and 2004. The experiments consisted of 49 entries in a 7 × 7 lattice design with two replications. Nine entries from this design were consisting of not sequenced checks (such as knock-out mutants of brown midrib loci with known high cell wall digestibility). Therefore, only 40 lines were analyzed for cell wall properties and sequenced. The single row plots were 0.75 m apart and 3 m long with a total of 20 plants. The ears were manually removed and the stover was chopped 50 days after flowering. Approximately 1 kg of stover was collected and dried at 40°C. The stover was ground to pass through a 1 mm sieve. Quality analyses were performed with near infrared reflectance spectroscopy (NIRS) based on previous calibrations on the data of 300 inbred lines (unpublished results). Four traits were analyzed: Water Soluble Carbohydrate (WSC), Organic Matter Digestibility (OMD), Neutral Detergent Fiber (NDF) and digestible Neutral Detergent Fiber (DNDF). DNDF was estimated by DNDF = 100 - (IVDMD - (100 - NDF))/NDF based on Goering and Van Soest, 1970 [43]. Since the investigated traits were highly heritable [42,44], and "trait × location" interactions were not significant, replications and locations were averaged, so that each entry was represented by mean values of the cell wall traits. The results of phenotypic analysis were published previously [38]. The heritabilities of NDF and DNDF were 86.5% and 92.2%, respectively [42].

Thirty-four inbred lines were employed in Exp. 2, including public lines and ecotypes both from Europe and U.S., covering substantial variation in forage quality. The lines were evaluated in three different years: 2006, 2008, and 2009 (unpublished data of INRA Lusignan). Values of stover in vitro cell wall digestibility was estimated by DNDF = (100 × (ES - (100 - NDF)/NDF) [45], based on enzymatic solubility (ES) of Aufrere and Michalet-Doreau,[46].

The analysis of overlapping sequences of common inbred lines between Exp.1 and Exp.2 (F2, F4, F7, F288, EP1, and W64) revealed consistency of sequences for the COMT gene across both experiments. Only one SNP each was different among overlapping sequences of F4 and EP1 inbred lines, respectively (Table 1). For CCoAOMT1, F2 and F7 sequences were identical in both experiments, whereas between two (F288) and 13 (EP1) polymorphisms were identified for the other three inbred lines (Table 1). For the CCoAMT2 gene, only two inbred lines were common between both studies. One SNP and one large Indel were identified for F288, whereas sequences perfectly matched for EP1. For each of the three genes, sequences of lines used in both experiments, were in all cases most similar across both experiments: even when comparing the EP1 CCoAOMT1 allele from Exp. 1 with all alleles in Exp. 2, the Exp. 2 EP1 allele was the most similar sequence to Exp. 1 EP1 among all Exp. 2 sequences.

Analysis of haplotype diversity for the 40 inbred lines employed in Exp. 1 revealed 13, 9, and 6 haplotypes for COMT, CCoAOMT1, and CCoAOMT2, respectively (Table 2). The range of haplotype means for all four stover quality traits was larger for COMT compared to the other two genes (Table 2). A total number of 26, 12, and 14 haplotypes was previously identified by Guillet-Claude et al. [40], for COMT, CCoAOMT1, and CCoAOMT2, respectively. When sequences from both experiments were jointly analyzed for haplotype numbers, 38 haplotypes were discriminated for COMT, and 15 haplotypes each for CCoAOMT1, and CCoAOMT2 (data not shown).

Changing alignments parameters for gap opening and gap extension costs leads to the creation of alignments with different Indel sizes and frequencies [58]. Different studies investigated consequences of different alignments parameters settings in the outcome of phylogenetic studies [59-62]. However, little is known about the effect of different alignment settings in association analysis. Our results demonstrate that different sets of polymorphisms are identified by association analyses, depending on alignments settings (Table 3). This is especially true for sequences with multiple Indels as observed for COMT, where choosing optimal alignment parameters might be difficult. Comparison of different alignments varying for gap costs may be used to identify the most consistent significantly associated polymorphic sites.

In this study, cell wall traits of two relatively small populations (40 and 34 lines) were tested for associations with two large sets of polymorphic sites (SNPs and Indels). The large number of independent variables (SNPs/Indels) in relation to number of dependent variables (lines/phenotypic data) increases the chances of false positive associations due to overfitting [63]. However, we believe that high heritability values of the cell wall traits analyzed [42,44] and stringent significance levels estimated by 10.000 permutations limit the rate of false positives [64].

The control of systematic differences due to population structure and kinship has also been suggested to avoid spurious associations and for greatest statistical power in association studies [53,65]. In this study, two of our models (GLM and Logistic Regression) were controlling for population structure, and the mixed model (MLM) approach was controlling for both population structure and kinship. Most of the lines evaluated in this study could be clearly identified as Flint or Dent lines (see Figure five: Andersen et al. [38]).

Across the investigated genes in this study, most of the significantly associated polymorphisms were detected by Logistic Regression, followed by GLM and MLM. A few previous studies compared results of association analyses from these methods. Andersen et al. [66] in an attempt to validate associations of the Dwarf8 locus identified by Thornsberry et al. [52], identified a larger number of polymorphisms associated with flowering time when using GLM (without control of population structure) as compared to Logistic Regression (with control of population structure). According to the authors, GLM identified more significant polymorphisms, because variation of causal alleles was closely correlated with population structure. In corroboration with our results, Andersen et al [42] observed a larger number of significant polymorphisms by GLM analysis accounting for population structure as compared to MLM accounting for both population structure and kinship corrections. Fewer significant polymorphisms identified by MLM are expected, as MLM also corrects for kinship, and significant polymorphisms identified in GLM but not in MLM might thus reflect identity of polymorphisms by descent. Investigating three QTL in 277 diverse maize inbred lines, Yu et al. [53] concluded that the model accounting for both population structure (Q) and relative kinship (K) gave the best approximation to the cumulative distribution of p-values when compared to models containing only K, only Q and a simple model (in which family structure is ignored).

Across the three genes and four traits within Exp. 1, 13 polymorphisms were consistent across the three methods, while one polymorphism was identified by both GLM and MLM, 3 by both GLM and Logistic Regression. No significant polymorphic sites were consistent between both MLM and Logistic Regression. Logistic Regression was the method detecting the highest number of significant polymorphisms in our study. With this method, permutation tests are performed for individual markers, not controlling for experiment-wise error rates. Consequently, a larger number of false positives are expected for this method of analysis when compared to analyses that control for experiment-wise error, like GLM in our study. Logistic Regression (and GLM in one occasion) revealed significant associations between the three OMT genes with WSC, with none of these three genes being involved in biosynthesis of soluble carbohydrates (Tables 4 and 5). However, we were able to identify one SNP in Exp.1 by Logistic Regression (but not by the other two methods) that was consistent with Exp. 2 (SNP at position 1948 of the COMT gene, Table 6). Moreover, many polymorphisms identified by Logistic Regression were also identified by GLM and MLM when considering p-values between 0.05 and 0.10, especially for associations between COMT with NDF, OMD, and WSC (data not shown). It can be inferred that these QTN/Indels are at the threshold of being detectable, and even small changes in statistical methods or parameters might lead to (non-) significance of these sites. Therefore, significant polymorphisms identified by association analyses, need to be validated by independent experiments (ongoing).

Two SNPs were identified as significantly associated in both datasets (Exp.1 and Exp.2). The SNP at position 1948 bp in the COMT intron was significantly associated with DNDF. At this C/T polymorphic site, a sequence motif for a DOF3 zinc finger transcription factor was identified for the allele with the T base present. The second SNP consistent across experiments was located in the site 944 bp at the fourth intron of CCoAOMT1. This SNP did not cause motif or splicing alterations, but it was significantly associated with both NDF and OMD in Exp.1. The consistency of these two SNPs across experiments makes them excellent candidates for development of FMs.

In addition, three polymorphic sites identified in Exp. 2 were also observed in Exp.1, but not significantly associated with the investigated traits. The relative low number of common polymorphic sites between Exp.1 and Exp.2 can in part be explained by the different regions of genes that each experiment investigated. The overlapping region of each gene common to both studies was not representing the whole gene. Moreover, for some lines overlapping sequences were not completely identical in both studies (Table 1), most likely due to residual heterozygosity as the "same" lines in both studies had in all cases the most similar sequences. In addition, most of the significantly associated polymorphic sites identified in Exp. 2 were singletons, which have a low probability of being indentified across datasets. Finally, Exp. 1 and 2 were conducted in quite different environments in Germany and France, respectively, which might explain for the different performance of the "same lines" with regard to CWD in both experiments.

Most of the significant SNPs/Indels identified in Exp. 1 were identified for COMT (44 significant polymorphisms) while a smaller number of sites was identified for CCoAOMT1 and CCoAOMT2 (three and two significant polymorphic sites, respectively). Guillet-Claude et al [40] also observed a larger number of polymorphisms for COMT compared to CCoAOMT1 and CCoAOMT2. In both experiments, the alignment sequence size of COMT was approximately two times larger than CCoAOMT1 and CCoAOMT2, whereas the number of significantly associated polymorphisms for COMT were substantially higher than 2-fold.

Andersen et al. [42] investigated the same set of 40 lines for six other "lignin genes" in relation to NDF, IVDOM, and DNDF, and identified significant associations for 4CL1, C3H, and F5H genes. The DNDF range between haplotypes within these three genes was largest for C3H (10.0%), followed by F5H (7.7%) and 4CL1 (2.3%). COMT haplotypes differed for DNDF by 17.8% (Table 2). Larger differences between haplotypes suggest that COMT is a promising candidate gene to derive markers affecting stover cell-wall digestibility. Moreover, down regulation of COMT in Arabidopsis, poplar, alfalfa and maize affecting the lignin content/structure support the strong effect of this gene on variation for cell wall digestibility [30-36].

The association analyses revealed a considerable number of SNPs and Indels associated with cell wall traits for the genes analyzed. Across experiments, methodologies and genes, 11 significant QTNs/Indels were located in coding regions and 52 significant polymorphisms were located in non-coding regions. Some of the significant associations identified in this study may reflect a statistical artifact. The probability of false-positive associations is higher in case of rare alleles and cases where a polymorphic site was only detected with one of the alignments. In addition, several of the identified significant associations can likely be explained by high LD to closely linked causal QTN/QTINDELs. Within an unresolved linkage block, likely only one or few polymorphisms are causative (Figs. 1 and 3). Polymorphisms causing amino acid changes are promising candidates for causative QTN/QTINDELs. The SNPs identified to cause amino acid changes in positions 2178 and 2185 bp of COMT (Table 4) were identified by all statistical methods, being, therefore, strong candidate sites to cause DNDF variation.

Several significantly associated polymorphisms located in non-coding regions have been reported. Andersen et al. [42] identified two respective SNPs in the F5H gene as significantly associated with NDF and DNDF. Wilson et al. [67] identified significant intron polymorphisms associated with kernel composition traits for sh1 and sh2 genes. Polymorphisms located in introns can be "causative" if affecting transcript abundance (expression and stability) of genes [68]. The analysis of motif alterations revealed eight significant associated Indels in COMT introns disrupting the sequence of five motifs, which represent binding sites for transcription factors RAV1, GAmyb, and DOFs 1, 2, and 3. DOF-type transcription factors are well known for their function as transcriptional activators or repressors of tissue-specific gene expression. More recently, they have been identified as putative regulators of lignin biosynthetic genes, based on the microarray analyses of Arabidopsis ectopic lignification mutants [69]. Therefore, although polymorphisms in positions 1439, 1920, 1948, 1952, 1953, and 1954 bp are located in the COMT intron, they might play an important role as cis-regulators of gene expression, causing changes in the lignin content and variation in DNDF values. intron polymorphisms might also be part of splicing signals that could cause differential splicing encoding structurally and functionally distinct protein products. However, no divergent patterns of exon inclusion were observed for the three genes analyzed (data not shown).