Shotgun sequencing of 17 soybean BACs selected for containing retained duplicate loci yielded a total of 36,873 sequence traces and a total of 2,028,159 bp of assembled soybean genomic sequence (Table 1). Six BACs (768,449 bp) have previously been shown to represent homeologous regions of the soybean genome anchored by either N-hydroxycinnamoyl/benzoyltransferase genes (HCBT; gmw1-74i13 and gmw1-52d3; 8 or ω-6 fatty acid desaturase genes (FAD2; gmw1-105h23, gmw1-15k6, gmw1-11j16, gmw1-45m6; 25. The 11 additional sequenced BACs were anchored by either RFLP clones (A711; UMb001-24d13 and Umb001-5f5) or by the duplicate transcripts cellulose synthase (gmw2-133d1 and gmw1-93l19), galactinol synthase (gmw1-5g16 and gmw1-103e11), raffinose synthase (gmw1-13o17 and gmw1-8g7) and caffeoyl-CoA O-methyltransferase (gmw1-58k3, gmw1-57d24 and gmw1-27d20). To date, this is the largest analysis of homeologous regions from the soybean genome. Although most of the BACs were sequenced to completion (phase III), seven remaining BACs contained a small number of ordered contigs with fewer than three gaps (phase II) and one BAC (gmw1-27d20) was phase I with five ordered contigs (Table 1).

With the exception of BACs UMb001-24d13 and Umb001-5f5 that were already mapped by an RFLP marker (A711), all but two of the remaining BACs were mapped by either BLAST-based identity of predicted coding sequence (CDS) to previously mapped transcript-based single nucleotide polymorphisms (SNPs) 27 or simple sequence repeats (SSRs) identified from each BAC sequence. Eight SNP markers were identified. Six of these markers confirmed already known map positions for gmw1-105h23, gmw1-15k6 25, gw1-74i13 8, UMb001-24d13, UMb001-5f5 (RFLP marker A711) and gmw2-133d1 (mapped by SSR as described below). The final two SNPs provided map positions for gmw1-57d24 and gmw1-27d20 (Table 1). In addition to SNPs, SSRs derived from BACs were identified, tested for polymorphisms and mapped. Only two BACs, gmw1-8g7 and gmw1-45m6 showed no polymorphisms in the mapping population or any matches to mapped transcript-based SNPs 25. Although there are multiple BACs on linkage groups I and O, eleven linkage groups are represented in this analysis (Table 1).

A total of 238 genes were predicted across the ~2.03 Mb of soybean sequence for an average gene density of 1 gene/11.1 Kb (Table 1) slightly less than previous estimates 2829825. All gene structure predictions as well as the annotations, ab initio predictions and EST-based support for each structure can be viewed at the following website 30. On average, 59.06% of the predicted gene structures had either EST or cDNA based support, regardless of whether coverage was normalized for gene size (average EST coverage) or not (ratio of EST coverage; Table 1).

Levels of gene conservation between BACs varied from being gene for gene in both order and orientation, with the exception of an eight-gene block inversion, for BACs gmw1-15k6 and gmw1-105h23 25 to very weak homeology anchored by only a single gene (gmw1-13o17 and gmw1-8g7; Table 1; Figure 1). While both of these extremes were observed, more often, homeologous BACs showed mid-range homeology; i.e. approximately 25 to 50% of genes in overlapping regions are retained. In those cases, most retained homeologs had 90% or greater sequence identity (Table 2) with a few extremes. The average nucleotide identity between homeologs ranged from 53.7 to 97.4% with an average of 86.6% while average protein similarity ranging from 53.3 to 99.0% with an average of 88.8% (Table 2). It should be noted that when homeologs were also tandemly duplicated on a BAC, they were not included in these estimates due to the inability to accurately determine which gene copy was the true ancestral homeolog between BACs.

To visualize the level of nucleotide identity between BACs, VISTA plots for BACs anchored by the RFLP A711, cellulose synthase, galactinol synthase, raffinose synthase and caffeoyl-CoA O-methyltransferase (COMT) were generated [see Additional Files 1, 2, 3, 4, 5]. VISTA identity plots as well as values for nucleotide identity, protein identity and protein similarity for HCBT and FAD2-anchored BACs have been previously reported [8,25 respectively]. Nucleotide identity between BACs is strongest in the coding regions and extends both 5' and 3' from predicted genes before dropping to below 50% between BACs with more duplicate gene conservation 825. This is likely due to retained non-coding sequences such as promoter elements between homeologous regions. However, as the level of gene conservation drops, so does the nucleotide identity beyond duplicate genes.

In a number of cases, homeologs appear to have varying gene lengths such as the selenium-binding protein found on BACs UMb001-24d13 and UMb001-5f5 (Figure 1, third homeolog) [see Additional file 1]. The exon number for this gene varies and a stop codon in the first exon of the UMb001-24d13 encoded selenium-binding protein truncates the resulting transcript (Table 2). There is however, EST-based support for the mRNA on UMb001-24d13 extending further 3' but the alignment is not a perfect match (92% identity). Other cases of variation in exon number between duplicate genes are observed (Table 2). Most of the differences can be accounted for in two ways: 1) ab initio based prediction of gene structures with little to no EST support vary between BACs and/or 2) truncation of one of the predicted genes due to an encoded stop codon. Reliance on ab initio predictions for gene structures combined with the lack of EST-based support can lead to differences between homeologs in exon number. In many cases, even alignment to putative orthologs could not verify the gene structure.

Synonymous (Ks) and nonsynonymous (Ka) substitutions between all of the duplicate genes were calculated (Table 2). The average Ks value was 0.42398 and average Ka value was 0.05775. Again, the Ks and Ka values for HCBT and FAD2 BACs are previously reported 825. All Ks values gave an average divergence estimate of 34.75 Mya. This value likely is inflated due to the extensive divergence between the duplicate genes identified on gmw1-57d24, gmw1-58k3 and gmw1-57d24 and between raffinose synthase on gmw1-13o17 and gmw1-8g7. When these duplicate genes were excluded from the calculation, the average divergence estimate was 9.665 Mya, similar to previous estimates 25 but still more recent than EST-based estimates 1920. When only the most divergent duplicate genes are used for coalescence estimates, a date of 153 Mya was obtained. Two caveats to divergence estimates should be noted: 1) The Ks values for the most divergent duplicate genes were for the most part well past saturation (greater than 1) and 2) in the most divergent regions, we cannot be certain that we are comparing homeologs and not paralogs (segmental or single gene duplications) without the context of the whole genome or more sequence in these regions. Only two pairs of homeologs showed evidence for positive selection; a ribonuclease HII encoding gene on gmw1-15k6 and gmw1-105h23 with a Ka/Ks ratio of 2.078 25 and the RAD-like encoding gene from gmw1-103e11 and gmw1-5g16 with a Ka/Ks ratio of 1.023. All other retained homeologs appear to be under purifying selection for retained function.

To quantify the potential confounding effects of paleopolyploidy on soybean whole-genome shotgun sequence assembly (WGS), all of the sequencing traces for the 17 BACs discussed above were used in large-scale or batch assemblies. The goal was to determine what effect homeology between duplicated regions will have as the soybean genome is reconstructed. Base-calling and assemblies were performed using Phred and Phrap, respectively 313233 with default parameters and viewed in Consed 34.

To first test if standard assembly parameters could distinguish between the most conserved homeologous BACs, sequence trace files for gmw1-105h23 and gmw1-15k6 were combined into a single "batch" assembly. Figure 2 shows that there is no cross assembly and no inclusion of sequencing traces between BACs. Assemblies were analyzed both manually and based upon BAC-specific tags to determine that sequence traces were assembled into the correct BAC contig. There are obvious regions with high levels of sequence identity between the BACs as determined by Crossmatch (Figure 2). Even with upwards of 97% sequence identity in exonic regions, sequence traces resolved into their correct "original" BACs. Quantification of the "batch-based" reassemblies against the original single-BAC assemblies was done using Vmatch 35. The three reassembled contigs for gmw1-105h23 had 99.58% sequence identity with 99.06% coverage to the original BAC assembly. Likewise, for gmw1-15k6 the resulting reassembly contigs had 99.80% sequence identity with 99.44% sequence coverage. As these results show, the assemblies were nearly identical to the original BAC assembly with the exception of small sequence gaps between the contigs, although clone pair ends clearly order and orient the contigs (Figure 2). Extrapolated to a whole-genome scale assembly, this shows that for soybean, unless there are regions of the genome that have higher levels of homeology than has been observed, the conserved paleopolyploidy of soybean will not have a substantial effect on the genome assembly.

All of the 38,673 traces from all 17 BACs were then combined into a single assembly using both standard assembly parameters as well as various other parameter sets. Assemblies were quantified using three measures: 1) the number of contigs containing greater than 100 traces versus the original 35 contigs from individual BAC assemblies 2) average percent coverage of the reassembled contigs to original contigs and 3) average percent nucleotide identity of the reassembled contigs to the original contigs (Table 3). These last two values were determined by Vmatch analysis that performed a global pair-wise alignment between all of the reassembled contigs and original assembly contigs as described in materials and methods. Under all of the parameter sets, some contigs were split into multiple contigs thereby increasing the contig number to greater than the original 35.

Experimental parameters were varied in an attempt to increase the percent coverage and percent nucleotide identity of the batch assemblies. The first parameter, revise_greedy, split initial contig assemblies at weak joins (regions that may be misassembled between duplicate regions due to sequence identity) and then attempted to reattach them for a higher overall alignment score. While only barely increasing the percent identity score, the percent coverage score was reduced by just over 7%. The forcelevel flag specifically reduced the stringency during the final contigs merge pass with 0 being most stringent and 10 least stringent, standard parameters using 0. When the forcelevel was relaxed slightly to 3, the percent coverage was nearly the same with only a slight drop in percent identity. However, increasing forcelevel to 5 decreased the percent coverage by just over 2% but increased the percent identity by over a full percent. It also had the effect of reducing the number of contigs from 44 at forcelevel 0 to 40 at forcelevel 5. Finally, the minmatch value was adjusted from 14 (standard) to 30 to increase the assembly stringency, a modification that dramatically increased the number of contigs to 50, as expected, and dropped the overall percent coverage. Combinations of these parameter changes also were investigated and the results are given as assemblies 6 and 7. Overall, it appears that standard Phred/Phrap assembly parameters return the greatest percent coverage out of all assemblies as well as the nearly best percent identity to the original contig assemblies.

Two potential sources of assembly error were identified in this analysis. First, under the last three assembly conditions (assemblies 5–7, Table 3) a contig from gmw1-27d20 and from GM_UMb-5f5 were incorrectly merged at a large (TATA)_n simple sequence repeat region. The resulting contig clearly shows the transition from one BAC to the other across the TA repeat with low quality sequences and low sequence coverage flanking the repeat. Lower quality sequences are not uncommon with simple sequence repeats that are large in length as these regions are difficult to sequence through. Secondly, the assembly of BAC gmw1-103e11 was especially troublesome in both the "batch" assembly of all of the BACs and on an individual assembly scale. Table 3 shows how the inclusion of the 103e11 contigs (which in most cases did not meet the Vmatch parsing criteria as is noted in Table 3) lowers both the average percent coverage and percent identity across the assembly.

Under standard assembly conditions, the 89,397 bp BAC gmw1-103e11 is fragmented into two contigs, a 19,452 bp contig with clone pair matches to the middle of the larger 69,905 bp contig. Clearly, a region from the middle of gmw1-103e11 is misassembled into a separate contig. This region can be partially resolved without manual reassembly by changing the forcelevel to 3 and minmatch to 30. The assembly still results in two contigs, but this is due to a gap in the middle of the contig and not exclusion of a region in the middle of the contig as with standard assembly parameters. The overall sequence coverage is 84.7% and sequence identity of 82.49% to the original BAC sequence. When this parameter set is used to reassembly all of the BACs however, it reduces the percent coverage by just over 5% but does increase the percent identity by almost 2% (Table 3).

This then raised the question as to what in the gmw1-103e11 sequence could be causing the re-assembly (both individual BAC and in the context of all BACs) to generate a second contig from the middle of the BAC. Utilizing Vmatch to identify sequence matches within the region being misassembled, non-retroelement, highly identical unique repeats (blue rectangles on Figure 3) were identified. Two major repeats occur in tandem in this region; a 566 bp repeat that is 96% identical (labelled as A and A' on Figure 3) and a 1, 198 bp repeat that is 95% identical (labelled as B and B' on Figure 3). Repeat A is present in the first unknown gene, repeat B in the pentatricopeptide repeat (PPR)-like 1 gene and both of the secondary repeat copies, A' and B' are contained within the PPR-like 2 gene (Figure 3).

GeneSeqer alignments 36 were generated of each predicted gene structure from this region realigned to the gmw103e11 BAC sequence. A portion of the PPR-like 2 gene aligns to the region predicted to contain the PPR-like 1 and unknown genes (Figure 3; orange gene structures). Similarly, the PPR-like 1 gene aligns to a portion of PPR-like 2. All of these alignments were using the "moderate" stringency function of GeneSeqer. The two predicted PPR-like genes in this region vary greatly in their structures and lengths. As discussed above, often there is little to no EST support and ab initio predictions must be relied upon. For this region, the first unknown gene has 7 ESTs with only 90% sequence identity that support the last exon, the rest of the gene is based upon ab initio predictions. The phosphotransferase and second unknown gene have nearly full EST support. Both of the PPR-like genes, however, are completely ab initio predicted.

Although there is variation in the predicted structures of the PPR-like genes, BLASTP annotation identified conserved petatricopeptide repeat (PPR) repeats in both. PPR repeats are a degenerate ~30 amino acid motif that occur tandemly multiple times within a protein 37. To identify potential PPR repeats across this region, MEME and MAST were used to generate PPR motifs and search the gmw1-103e11 BAC sequence for all possible occurrences of the motif 38. Two PPR repeats were found in the first intron of the predicted unknown gene, at least six PPR repeats were identified in the PPR-like 1 gene and eleven repeats were identified in the PPR-like 2 gene. These PPR repeats are 81–99 nucleotides in length that range from 25–100% similar at the amino acid level and 33–95.8% similar at the nucleotide level (within and between both PPR-like genes). The black lines on Figure 3 show the start location of the PPR domains that are located end to end within the coding sequence. These repeats account for the Vmatch identified repeat sequences A/A' and B/B'. The similarity of a portion of PPR-like 2 to both the first unknown gene and PPR-like 1 suggests two scenarios: 1) PPR-like 2 is incorrectly predicted and should be two separate genes or 2) PPR-like 2 is incorrectly predicted and should be fused with the first unknown gene. In either case, these PPR containing genes and repeats are the source of assembly error, as discussed below.

Identified repeats A/A', B/B' and all of the predicted genes from this region of gmw1-103e11 were re-aligned using GeneSeqer to the Phred/Phrap re-assembled gmw1-103e11 contigs. Both of the PPR-like gene structure predictions as well as the repeat A containing unknown gene align to a ~3,500 bp region in the middle of the 69,905 bp major contig. This region also contains clone pair matches to both ends of the 19,452 bp secondary contig. What has occurred is the PPR-containing regions are above the threshold of distinguishing one copy from another and have collapsed into a single structure in the larger contig. The phosphotransferase gene and second unknown gene are excluded from this region and placed in the separate contig. These results show that highly identical tandemly duplicated genes, especially those genes that themselves contain repetitive domains will be a potential source of assembly errors. In this case, the structure of the PPR repeats across the PPR-like genes cannot be resolved without manual curation of the assembly.

To determine how well our assemblies were screening for highly repetitive sequence, a preliminary assembly using standard Phred/Phrap parameters of 80,000 randomly chosen JGI trace files was done. Contigs containing greater than 15 traces were considered highly represented even after initial trace screening against known repetitive sequences. Each of these contigs was subject to a BLAST-based annotation against the NCBI nonredundant database and then clustered into groups based upon that annotation (Figure 4). Surprisingly, 23% of the JGI contigs showed no sequence identity to any anything in the NCBI nonredundant database. However, when the contigs comprising this 23% are BLASTed against the repetitive database generated by Gill et al. 39 only 5 contigs out of 44 had no match and 7 contigs had a bit-score less than 90 and were considered poor matches. Forty thousand randomly chosen JGI trace files were combined with the 36,978 BAC generated trace files in a standard Phred/Phrap assembly. The addition of the JGI whole-genome shotgun generated trace files had no effect on either the percent identity of the reassembled contigs (99.07%) or on the percent coverage (98.52%).

BACs gmw1-74i13 and gmw1-52d3, corresponding to duplicate loci anchored by N-hydroxycinnamoyl benzoyltransferase (HCBT) genes, were identified, sequenced and annotated by Schlueter 8. Four BACs, gmw1-15k6, gmw1-105h23, gmw1-11j16 and gmw1-45m6 anchored by ω-6 fatty acid desaturase (FAD2) genes were identified, sequenced and annotated by Schlueter 25. BACs anchored by the RFLP probe A711 with known cytogenetic information 24. GM_UMb-24d13 and GM_UMb-5f5 were used to construct shotgun libraries for sequencing and assembly as described previously 56.

Retained duplicate transcripts corresponding to isoflavone synthase/cellulose synthase, galactinol synthase, raffinose synthase and caffeoyl-CoA o-methyltransferase were identified with TBLASTX (default parameters) using a reference sequence against all soybean ESTs 58. Identified ESTs were aligned into contigs using Sequencher v.4.5, also with default parameters (Gene Codes Corp., MI). PCR primers were designed to distinguish between copies using Oligo 6.82 (Molecular Biology Insights, Cascade, CO) [see Additional file 6]. Multidimensional pools of the Williams 82 G. max BAC library (gmw1) were PCR screened. BAC DNA was isolated using a Plasmid Midi kit (Qiagen, Valencia CA) and reverified with PCR as previously described 8.

BACs gmw1-13o17 and gmw1-8g7 were subcloned and assembled as described in Schlueter 8. Subclones were sequenced at the Iowa State DNA Sequencing and Synthesis Facility (Ames, Iowa). Sequence for BACs gmw2-133d1, gmw1-93l19, gmw1-5g16, gmw1-103e11, gmw1-58k3, gmw1-57d24 and gmw1-27d20 was generated at the University of Oklahoma using conditions previously described 5960616263. Accession numbers for all sequenced BACs can be found in Table 1.

BACs were mapped using two methods. First, already mapped EST-based SNPs were identified by BLASTN of annotated genes from each BAC against mapped ESTs 10. Only ESTs that match to BAC-derived genes with an e-value of 0.0 (near identical match) were considered. In addition, each EST was aligned to the BAC to confirm that it corresponded to one homeolog (or paralog) versus the other. Secondly, each BAC that was not previously mapped was scanned for di- and tri-nucleotide repeats using Sputnik (Espresso Software Development, Seattle WA). Primer pairs flanking the potential SSR markers were designed using Oligo 6.82 (Molecular Biology Insights) and tested against various soybean parents of mapping populations. PCR reactions were 10 μl in volume and contained 1 × PCR buffer, 1.5 mM magnesium chloride, 5 mM dNTPs, 0.5 μM each primer, 50 ng Glycine max parental DNA, and 0.025 U of Taq DNA polymerase (Invitrogen). PCR cycling conditions were 94°C for 2 min, 35 cycles of 94° for 45 sec, 60° for 30 sec, 72° for 45 sec, followed by a final extension of 72° for 3 min. Resulting bands were run on either a 3% agarose 1 × TAE (Tris, Acetic Acid, EDTA) gel for larger (greater than 250 bp) products or 6% polyacrylamide 0.5 × PBE gel for smaller fragments. Polymorphic SSRs from each BAC were mapped in the Glycine max A81-356022 X Glycine soja PI 468.916 population 6413. Genetic map positions of these SSRs were determined using MapMaker/Exp 3.0 with a minimum lod score of 3.0 6465. Sequences for these SSRs are available [see Additional file 7].

Gene prediction was done using a combination of ab initio and EST-alignment based methods as previously detailed 825. Annotation was completed using yrGATE and viewed as part of the xGDB system 6667. A database with annotations was created called GmaxGDB 30. Each predicted gene was subjected to a BLASTP query of the NCBI nr database with default parameters to assign a putative function. An e-value threshold of 1^e-10 was used to assign putative function.

Alignment of homeologous BACs used shuffle-LAGAN 68 with default parameters anchored by predicted gene structures producing a VISTA plot 69. The nucleotide and protein percent identity and similarity of homeologs, was calculated using WATER, a pairwise alignment program (gap penalty of 10; extension penalty of 0.2; EMBOSS)70. Synonymous and nonsynonymous distances were calculated using PAML, default parameters 71. Coalesence estimates were calculated as in 20.

Trace files for all of the assembled BACs were combined into a single assembly utilizing 36,978 sequence reads. Base calling and sequence assemblies were performed using the Phred 3132 and Phrap 33, respectively. Assemblies were viewed using the Consed viewer and Cross-Match 34. All assemblies were run with standard Phred/Phrap parameters unless otherwise noted in the text or table. Briefly, parameters that were varied were: 1) revise_greedy that splits initial contig assemblies at weak joins (regions that may be misassembled due to high sequence identity) and then attempts to reattach them for a higher overall alignment score. 2) forcelevel reduces the stringency during the final contig merge pass and 3) minmatch which is the minimum length of a matching word in sequence comparisons during assembly. Further explanation of each parameter is found in the Phrap documentation 33.

Previously characterized repetitive sequences from soybean available at the time of assembly were included in prescreening during assembly (Marek et al. unpublished results) 39. Quantification of assemblies was done using Vmatch for large-scale sequence matching (a large-scale global sequence alignment)35. This program returns the percent nucleotide identity as well as the start and stop position for each contig alignment to allow for the calculation of percent coverage. Only contigs that contained greater than 100 traces were included in the analysis.

Trace files from the soybean whole-genome shotgun sequencing effort were downloaded from the NCBI trace archive 72. These files are reads all uploaded from August 9–10, 2006 (ti's range from 1397334945 – 1399236113) to for a total of 80,000 sequencing reads. To determine the sequence composition of the JGI-only assemblies, contigs contained greater than 15 traces were blasted against the nr database to assign a putative annotation. These contigs were assumed to represent what will be observed at a high frequency in the whole-genome assemblies.