The consensus map spanned a total length of 1,161 cM. Chromosome sizes ranged from 147.1 cM (6H) to 194.2 cM (5H) (Figure 4). The 'bPb' DArT markers alone spanned 98.1% of the total length of the consensus map. Addition of a second set of DArT markers ('bPt' markers) increased coverage to 99.7%. The combination of all non-DArT markers resulted in a coverage of 96.9% (Additional File 7). A single 'bPb' DArT assay, therefore, provides slightly greater genome coverage than the set of 850 SSR and RFLP markers included in this study.

The map had no gap larger than 10 cM and only nine gaps between 5 and 10 cM (3H, 4H, 6H and 7H). The DArT subset of markers generated a consensus map with a single gap between 15 and 20 cM (4HS) and five gaps between 10 and 15 cM on chromosomes 3H, 4H and 5H (the set of 'bPb' markers alone had an additional gap of this size on 4H). Chromosome 4H has previously been noted by others to be less polymorphic than the others 8932. Non-DArT markers on their own resulted in a map with one gap between 15 and 20 cM (3HL) and two gaps between 10 and 15 cM on chromosome 6H (Figure 3B). The smaller number of 10–15 cM gaps in the non-DArT dataset may reflect targeted efforts to fill gaps in component maps with selected SSR or RFLP markers.

The average resolution of the consensus map was evaluated by collapsing co-segregating loci into bins and calculating the average distances between adjacent bins. The 2,935 loci of the whole dataset could be distributed into 1,629 bins with an average inter-bin distance of 0.71 ± 1.01 cM (median = 0.32 cM). This resolution was only moderately greater than the resolution obtained with DArT loci alone (1.03 ± 1.59 cM; median = 0.40 cM). The set of 'bPb' DArT markers, which were simultaneously assayed in a single reaction, provided a resolution of 1.20 ± 1.83 cM (median = 0.44 cM). Non-DArT markers on their own produced a map with a resolution of 1.91 ± 2.07 cM (median = 1.26 cM; Additional File 7).

The DArT markers were originally obtained by cloning random fragments of genomic representations 16, a process that introduces some degree of marker redundancy. The 1,546 'bPb' DArT loci could be collapsed into 959 bins, suggesting a redundancy level of 38%. Co-segregating 'bPb' DArT markers, however, were not necessarily multiple copies of a single marker because more than 100 of the 'bPb' DArT bins contained markers that were in the opposite allelic phase in some crosses (data not presented). On the other hand, a small number of genotyping errors may have prevented multiple copies of single markers from being collapsed into bins. Therefore, it may not be surprising that the redundancy estimate obtained from marker segregation analysis was quite similar to the preliminary estimate obtained by clustering the DNA sequences of DArT markers (data not presented).

Marker redundancy is a transient feature of DArT array development, which proceeds by consolidating the most informative clones in new arrays of increasing information content 21. During this process redundant markers are excluded from the final genotyping array.

Markers sometimes tend to cluster, either as a consequence of an uneven distribution of recombination events along chromosomes 33 or because they preferentially survey DNA polymorphism that is unevenly distributed along chromosomes 3435. Regions of the consensus map with high marker densities were visualized by plotting local averages of inter-bin distances and the number of loci per bin along chromosomes (Figure 5). Both DArT and non-DArT loci showed a moderate tendency to cluster around centromeres as can be deduced from the shorter inter-bin distances and the larger numbers of loci per bin in the vicinity of centromeres. This clustering tendency, however, was nowhere near as pronounced as, for example, for AFLP markers based on methylation-insensitive restriction enzymes 36. Given the different polymorphism-detection principles of DArT, SSR and RFLP markers, we suggest that the centromeric clustering largely reflects centromeric suppression of recombination 3337. Centromeric clustering, however, was less pronounced in chromosome 5H, a feature that was previously noted by others 8.

In some chromosomes the density of DArT markers also appeared to be higher in distal regions (1H, 2H, 6H and 7H). This pattern may reflect a moderate bias of PstI-based DArT markers towards gene-rich, hypomethylated areas in telomeric chromosome regions 38, a pattern we also observed in wheat 21. A preliminary analysis of the DNA sequences of the 'bPb' DArT markers indeed suggests that the majority of them are derived from the genespace (data not presented).

Markers mapping to more than one locus, if not recognized, can be a confounding factor in the process of building a consensus map. Among the 1,523 mapped 'bPb' DArT markers, only 21 (1.4%) mapped to two different loci in different populations and one (0.2%) mapped to three different loci (Additional File 8). The loci of multi-locus DArT markers were usually located on different chromosomes; two of them, however, mapped to loci within a single chromosome (1H and 4H; Additional File 4). Multi-locus markers were more common among other marker types. The set of 753 non-DArT markers contained 66 markers (8.8%) that mapped to 2 to 8 different loci each (Additional File 4, 8). The difference in frequency of multi-locus markers between DArT and other marker types reflects the fact that as a hybridization-based method DArT inherently selects against multi-locus markers. The hybridization intensities measured for such markers are a difficult-to-resolve mixture of the contributions of several loci, which makes them appear as 'monomorphic'.

A good indicator of the quality of a linkage map constructed from DH populations is the frequency of singletons (apparent double crossovers), which are often due to genotyping errors 2829. The frequency of singletons was calculated from the 1,629-bin dataset containing all types of markers by using a purpose-built perl script. Approximately 0.35% of all calls for non-DArT loci were singletons. DArT loci generated singletons at a rate of approximately 0.20% (Additional File 5). The majority of the loci with singletons introduced less than one singleton per one hundred calls (Figure 3C). Not surprisingly, high-quality DArT markers, which tend to have allelic states with more contrasting hybridization intensities, generated fewer singletons: the correlation between the across-population average of the marker quality parameter and the percentage of singletons in the concatenated segregation signature was -0.40. The Frederickson/Stander population, and to a lesser extent the Igri/Atlas68 population, contained larger numbers of singletons, presumably because some of the DNA samples got cross-contaminated during shipment as a result of insufficient sealing of microtiter plates (data not presented; see Methods and Table 1).

Singletons presumably were not only the result of genotyping errors. The comparatively large distances between adjacent loci on chromosome 4H, true double crossovers events in the RIL populations (Foster/CI4196, Frederickson/Stander and Patty/Tallon), unstable methylation patterns 16 and possibly gene conversion events 39 may have introduced some singletons. The reported singleton rates, therefore, almost certainly overestimate the error rates of marker assays. The overestimation of genotyping error rates, however, was to a degree offset by having removed low-quality markers during the construction of pilot maps (see Methods section). The frequency of DArT singletons, therefore, is in good agreement with the previous 0.2% estimate of the error rate of DArT assays 1621.

DArT assays were essentially performed as described previously 1621. Briefly, 20–100 ng of genomic DNA was digested with two units of PstI and two units of BstNI (NEB, Beverly, MA). A PstI adapter (5'-CAC GAT GGA TCC AGT GCA-3' annealed with 5'-CTG GAT CCA TCG TGC A-3') was ligated to the digested DNA with T4 DNA ligase (NEB). A 1-μL aliquot of the ligation product was used as a template in a 50-μL amplification reaction with DArT-PstI primer (5'-GAT GGA TCC AGT GCA G-3') under the cycling conditions described by Wenzl et al. 16. The resulting genomic representations were concentrated tenfold by isopropanol precipitation and denatured at 95°C for 2 min. The representations were then labelled with 0.1 μL of Cy3-labelled dUTP (Amersham Biosciences, Castle Hill, NSW, Australia) using the exo^- Klenow fragment of Escherichia coli DNA polymerase I (NEB) and random decamers for priming. Labelled representations were added to 50 μL of a 50:5:1 mixture of ExpressHyb buffer (Clontech, Mountain View, CA, USA), 10 g L^-1 herring sperm DNA (Promega, Annandale, NSW, Australia), and the carboxy-fluorecein (FAM)-labelled polylinker fragment of the plasmid that was used for library preparation 21. (The hybridization signal for the polylinker fragment was subsequently used by DArTsoft to determine for each clone the amount of DNA spotted on the array; see next paragraph).

The hybridisation mixtures were denatured and hybridized to DArT microarrays which contained 2,304 polymorphism-enriched clones from a PstI/BstNI genomic representation prepared from a mixture of barley cultivars 16. After overnight hybridization at 65°C, the microarray slides were washed according to Jaccoud et al. 15 and scanned on a Tecan LS300 confocal laser scanner (Grödig, Salzburg, Austria). The microarray images were analyzed with DArTsoft (version 7; DArT P/L, Canberra, Australia), a purpose-built software package. The program measured the relative hybridization intensities (Cy3/FAM) of each clone across all slides, identified clones with variable hybridization intensity (i.e., DArT markers), and used a fuzzy-clustering algorithm to score the corresponding DNA fragments in the genomic representations as present or absent (Cayla et al., in preparation). Individual genotype calls were classified as missing if none of the probabilities of belonging to a particular allelic state (present vs. absent) surpassed 0.8. The quality of each marker was quantified by computing the variance of the relative hybridization intensity between allelic states as a percentage of the total variance of the relative hybridization intensity.

Markers with a quality parameter and a call rate both greater than 80% were selected to construct pilot linkage maps. Markers with a quality parameter between 75 and 80% were incorporated on a case-by-case basis (see section entitled "Pilot maps" below).

The Barque-73/CPI71284-48 was typed with multiplex-ready SSR markers (Hayden, Nguyen and Chalmers, in preparation). The Dayton/Zhepi2 population was assayed with 38 SSR and STS markers as described previously (Raman et al., submitted) 53. The TX9425/Franklin and Yerong/Franklin populations were typed with four SSR markers per chromosome according to Ramsay et al. 35. The Clipper/Sahara and the Foster/CI4196 populations had previously been typed with SSR and/or RFLP markers 5052.

A total of 84 DArT assays from two populations had to be discarded because a subset of DNA samples stored in 96-well microtiter plates got cross-contaminated during shipment as a result of insufficient sealing of the plates. (These samples could be identified because of their bias toward "1" scores.) The remaining DArT assays were accepted if the relative hybridization intensities of non-polymorphic DArT markers were sufficiently correlated with the corresponding intensities in simultaneously performed assays (average correlation coefficient > 0.80). Application of this threshold led to the removal of 13 of the 966 remaining assays.

A few markers with multiple entries in a single dataset were removed because the segregation signatures of the multiple entries did not resemble each other. The segregation data of other markers with multiple entries whose segregation signatures were almost identical (>95%) were collapsed into single-marker entries with consensus segregation signatures.

DArT marker names are standardized and automatically generated by a DArT-specific Laboratory Information Management System (DArTdb; DArT P/L, Canberra, Australia). Additional File 4 contains a translation table between these names and the provisional names used in a previous paper 16. Different laboratories used slightly different names for the same SSR or RFLP markers and assigned different letters to the multiple loci of multi-locus markers. Non-DArT marker names and locus codes, therefore, were curated to an extent required to create an unambiguous nomenclature (Additional File 4).

The presence vs. absence DArT scores (0/1) of the remaining 953 lines were converted into genotype codes (A/B/C/D) by comparison with the appropriate parental DArT assays (18 in total). Some markers for which both parental assays produced unreliable data were arbitrarily assigned to one of the two linkage phases. The segregation data for DArT markers were merged with those of other markers for each population for which non-DArT data was available.

The segregation signatures of each of the ten individual datasets were imported into JoinMap 3.0 to distribute loci into linkage groups (Figure 1). The LOD threshold used to define linkage groups was necessarily dependent on the number of markers and lines in the datasets. Markers in the wrong linkage phase were identified and flipped into the opposite phase. The known chromosomal locations of a subset of the DArT markers 16 were used to assign linkage groups to chromosomes. Multi-dose markers were identified by comparing chromosomal assignment of markers across populations. Alternative loci were encoded by adding lower-case suffixes to marker names in the case of DArT markers.

Twenty redundant lines were identified with JoinMap 3.0 using a similarity threshold of 95%. They were removed from the datasets, thus reducing the total number of lines to 915.

Pilot maps for individual populations were built by ordering loci with RECORD 26 to identify potential DNA sample-tracking errors between laboratories and to remove unreliable lines and loci (Figure 1). In the case of RIL lines, A/B (instead of C/D) genotype codes were used for map construction, assuming that residual heterozygosity levels were low. Inspection of the graphical genotypes identified a clear DNA sample-tracking inconsistency in one of the datasets: the scores of 14 of the 38 SSR loci were "frame-shifted" by one DH line with respect to the DArT data. The frame shift was rectified.

Lines with an excess of singletons were identified by inspecting graphical genotypes and removed from the corresponding dataset (4 of 915 lines). Loci that introduced a large number of singletons (typically two or more) were identified using a purpose-built perl script (Additional File 13). These loci were removed from the corresponding population dataset unless visual inspection of graphical genotypes showed that they were located in map regions with low marker densities. In this way, 86 of the 2,082 DArT loci (4.1%) with a quality parameter greater than 80% and 65 of the 956 (6.8%) non-DArT loci were completely removed from the combined (all-population) data set. Unidentified DNA sample-tracking inconsistencies between laboratories are likely to have caused the removal of a small number of RFLP or SSR loci during this process. A set of 41 closely linked markers with low call rates were removed from chromosome 2H of the Foster/CI4196 dataset. A limited number of DArT loci with a quality parameter of less than 80% were added to individual pilot maps if they introduced less than two singletons, which increased the total number of DArT loci by 89.

A comparison of the pilot maps across populations revealed three additional multi-locus DArT markers that mapped to different loci within the same chromosome. The loci were encoded as described in the section entitled "Linkage group assignment". The final curated dataset contained a total of 2,935 loci (2,085 DArT and 850 other loci) partially scored across 911 lines of ten populations. The quality of this dataset was later confirmed to be sufficient for building a high-quality consensus map (see Results and Discussion).

JoinMap 3.0 was used to build an initial draft of a consensus map based on the 1,546 'bPb' DArT loci collapsed into 959 bins. The program thresholds were adjusted to LOD = 2.0 and REC = 35 according to Karakousis et al. 9. (An analysis of one of the chromosomes established that a LOD threshold of 3 did not significantly improve the results.) The consensus bin order was then improved manually in an iterative manner. The segregation data of all populations were arranged according to the consensus order in separate Excel spreadsheets and graphical genotypes were displayed in color using the conditional formatting option of the program. Starting at population one, misplaced bins with large numbers of singletons were moved within chromosomes to positions where they created less (frequently no) singletons. The modified consensus order was than imposed on the other nine datasets, and the resulting graphical genotypes were inspected to accept or reject the order. When a new position of a particular bin was rejected because it increased the number of singletons in other populations, alternative positions were tested across all populations until a better position than the original was found. Once each of the populations had been processed this way, the entire procedure was repeated until the overall number of singletons could not be reduced any further.

Other markers (separately collapsed into bins) and the 'bPb' DArT bins that JoinMap had failed to integrate were added to the DArT bin framework, using a purpose-built perl script (Additional File 13; Figure 1). The script used the DArT framework as a fixed scaffold and tested, for each new bin, all possible positions to select the one that produced the smallest number of singletons (frequently none) in the concatenated dataset. Singletons were identified by comparing each genotype call with the closest bracketing calls for the same line within a 15-bin window on either side. Once all new bins had been integrated this way, the order of multiple bins inserted at identical positions was resolved manually using the iterative procedure described in the previous section.

Map distances between adjacent bins arranged in the order of the consensus map were calculated using a purpose-built perl script implementing the multi-point algorithm described by Lalouel (Additional File 15) 40 (also implemented in JoinMap 3.0). The segregation data of each population were separately converted into two-point recombination frequency estimates between all pairs of bins. In the Foster/CI4196, the Frederickson/Stander and the Patty/Tallon populations, the cumulative recombination frequencies of the RIL were converted into per-meiosis values according to Haldane and Waddington 55. From the recombination frequency estimates, two matrices were generated for each population. The matrices contained pairwise LOD scores and Kosambi distance estimates for bin pairs separated by a recombination frequency of less than 35%. Both types of matrices were averaged and merged across populations to create two consensus matrices containing average distance estimates and average LOD scores (the numbers of lines assayed in each population were used as weights for averaging). The average distance estimates were then weighted with the squared average LOD scores and fed into a linear equation system that computed distances between adjacent bins by minimizing the overall deviation between empirical and computed average distance estimates 232440.

As outlined under "Construction of component maps" above, the distance-calculation algorithm was expected to generate negative distance estimates for some closely segregating bins, because of the statistical uncertainty inherent in the data. Bins represented by loci only scored in a single population represented an additional source of negative distance estimates because their positions were inherently more ambiguous and to some degree still unresolved by our manual bin-ordering procedure.

The consensus bin order, therefore, was further refined with a purpose-built perl script that flipped the order of bin pairs with negative distance estimates until all negative estimates were eliminated (Additional File 15; Figure 1). Inspection of the resulting graphical genotypes showed that only few singletons were introduced by this procedure; most of them were due to a limited number of bins which the flipping procedure had shifted to sub-optimal positions. The positions of these bins were rectified manually, which inevitably re-introduced a limited number of negative distance estimates that were probably due to chromosome areas in which recombination frequencies differed among populations. Finally, the bins were de-collapsed to re-integrate co-segregating loci. Because this step introduced additional segregation data, which in some cases resolved the order of hitherto (almost) unresolved loci, a final round of manual improvement of the locus order was required (Figure 1). The remaining negative distance estimates, totaling 8.8% of the length of the consensus map, were set to 0.001 cM to indicate closeness and to retain the optimized bin order.