Whole genome shotgun coverage of eight previously unsequenced Yersinia species (Table 1) was obtained by single-end bead-based pyrosequencing 34 using the 454 Life Sciences GS-20 instrument. Each of the eight genomes was sequenced to a high level of redundancy (between 25 and 44 sequencing reads per base) and assembled de novo into large contigs (Table 2; Additional file 1). Excluding contigs that covered repeat regions and therefore had significantly increased copy number, the quality of the sequence of the draft assemblies was high, with less than 0.1% of the sequence of each genome having a consensus quality score 35 less than 40. Moreover, a more recent assessment of quality of GS-20 data suggests that the scores generated by the 454 Life Sciences software are an underestimation of the true sequence quality 36. The most common sequencing error encountered when assembling pyrosequencing data is the rare calling of incorrect numbers of homopolymers caused by variation in the intensity of fluorescence emitted upon extension with the labeled nucleoside 34.

Previous studies and our experience suggest that at this level of sequence coverage the assembly gaps fall in repeat regions that cannot be spanned by single-end sequence reads (average length 109 nucleotides in this study) 34. Fewer RNA genes are observed compared to published Yersinia genomes finished using traditional Sanger sequencing technology (Additional file 1), reflecting the greater difficulty of uniquely assembling repetitive sequences with single-end reads. We assessed the quality of our assemblies using metrics implemented in the amosvalidate package 37. Specifically, we focused on three measures frequently correlated with assembly errors: density of polymorphisms within assembled reads, depth of coverage, and breakpoints in the alignment of unassembled reads to the final assembly. Regions in each genome where at least one measure suggested a possible mis-assembly were validated by manual inspection (Additional file 2). Many of the suspect regions corresponded to collapsed repeats, where the location of individual members of the repeat family within the genome could not be accurately determined. Based on the results of the amosvalidate analysis and the optical map alignment we found no evidence of mis-assemblies leading to chimeric contigs in the eight genomes we sequenced. Genomic regions flagged by the amosvalidate package are made available in GFF format (compatible with most genome browsers) in Additional file 3.

Genome sizes were estimated initially as the sum of the sizes of the contigs from the shotgun assembly, with corrections for contigs representing collapsed repeats (Table 2). We also derived an independent estimate for the genome size from the whole-genome optical restriction mapping of the species 38 (Additional file 4). Alignment of contigs to the optical maps 39 suggested that the optical maps consistently overestimated sizes (2 to 10% on average). After correction, the map-based estimates and sequence-based estimates agreed well (within 7%). Two species, Y. aldovae (4.22 to 4.33 Mbp) and Y. ruckeri (3.58 to 3.89 Mbp), have a substantially reduced total genome size compared with the 4.6 to 4.8 Mbp seen in the genus generally. The agreement between the optical maps and sequence-based estimates of genome sizes tallied with experimental evidence for the lack of large plasmids in the sequenced genomes (Additional file 5). A screen for matches to known plasmid genes produced only a few candidate plasmid contigs, totaling less than 10 kbp of sequence in each genome.

The number of IS elements per genome for the eight species (12 to 167 matches) discovered using the IS finder database 40 was much lower than in the Y. pestis genome (1,147 matches; copy numbers estimates took into account the possibility of mis-assembly and were accordingly adjusted; see Methods). Furthermore, the non-pathogenic species with the most IS matches, namely Y. bercovieri (167 matches), Y. aldovae (143 matches) and Y. ruckeri (136 matches), have comparatively smaller genomes. We also searched for novel repeat families using a de novo repeat-finder 41 and collected a non-redundant set of 44 repeat sequence families in the Yersinia genus (Table 3; Additional file 6). Interestingly, the well-known ERIC element 42 was recovered by our de novo search and was found to be present in many copies in all the pathogenic species, but was relatively rare in the non-pathogenic ones. On the other hand, a similar and recently discovered element, YPAL 43 (also recovered by the de novo search), was abundant in all the Yersinia genomes except the fish pathogen Y. ruckeri. Insertion sequence IS1541C in the IS finder database, which has expanded in Y. pestis (to more than 60 copies), had only a handful of strong matches in Y. enterocolitica, Y. pseudotuberculosis, and Y. bercovieri and no discernable matches in the other Yersinia genomes.

The sequences generated in this study provide new background information for validating genus detection and diagnosis assays targeting pathogenic members of the Yersinia genus. The assay design process commonly starts by computationally identifying genomic regions that are unique to the targeted genus ('signatures') - an ideal signature is shared by all targeted pathogens but not found in a background comprising non-pathogenic near neighbors or in other unrelated microbes. While many pathogens are well characterized at the genomic level, the background set is only sparsely represented in genomic databases, thereby limiting the ability to computationally screen out non-specific candidate assays (false positives). As a result, many assays may fail experimental field tests, thereby increasing the costs of assay development efforts. To evaluate whether the new genomic sequences generated in our study can reduce the incidence of false positives in assay development, we computed signatures for the Y. pestis and Y. enterocolitica genera using the Insignia pipeline 44, the system previously used to successfully develop assays for the detection of V. cholerae 44. We identified 171 and 100 regions within the genomes of Y. pestis and Y. enterocolitica, respectively, that represent good candidates for the design of detection assays. In Y. pestis these regions tended to cluster around the origin of replication, whereas in Y. enterocolitica there was a more even distribution. The average G+C content of the regions for the unique sequences in both species was close to the Yersinia average (47%) and there was not a strong association with putative genome islands (Additional files 7, 8, 9, 10, 11, 12, 45). For both species, most regions overlapped predicted genes (161 of 171 (94%) and 96 of 100 (96%) in Y. pestis and Y. pseudotuberculosis, respectively). Interestingly, 171 Y. pestis gene regions were spread over only 70 different genes, whereas the 96 Y. enterocolitica regions were found overlapping only 90 genes. There was no obvious trend in the nature of the genes harboring these putative signals except that many could be arguably classed as 'non-core' functions, encoding phage endonucleases, invasins, hemolysins and hypothetical proteins.

Ten Y. pestis-specific and 31 Y. enterocolitica-specific putative signatures have significant matches in the new genome sequence data (Additional files 7, 8, 9, 10), indicating assays designed within these regions would result in false positive results. This result underscores the need for a further sampling of genomes of the Yersinia genus in order to assist the design of diagnostic assays.

We performed a multiple alignment of the 11 Yersinia species using the MAUVE algorithm 46 (from here on Y. pestis CO92 and Y. pseudotuberculosis IP32953 were used as the representative genomes of their species) and obtained 98 locally collinear blocks (LCBs; Additional files 13, 14, 47). The mean length of the LCBs was 23,891 bp. The shortest block was 1,570 bp, and the longest was 201,130 bp. This multiple alignment of the 'core' region on average covered 52% of each Yersinia genome. The nucleotide diversity (Π) for the concatenated aligned region was 0.27, or an approximate genus-wide nucleotide sequence homology of 73%. As expected for a set of bacteria with this level of diversity, the alignment of the genomes shows evidence of multiple large genome rearrangements 23 (Additional file 13).

Using an automated pipeline for annotation and clustering of protein orthologs based on the Markov chain clustering tool MCL 48, we estimated the size of the Yersinia protein core set to be 2,497 and the pan-genome 49 to be 27,470 (Additional files 15, 16, 17, 18). The core number falls asymptotically as genomes are introduced and hence this estimate is somewhat lower than the recent analysis of only the Y. enterocolitica, Y. pseudotuberculosis and Y. pestis genomes (2,747 core proteins) 15. We found 681 genes to be in exactly one copy in each Yersinia genome and to be nearly identical in length. We used ClustalW 50 to align the members of this highly conserved set, and concatenated individual gene product alignments to make a dataset of 170,940 amino acids for each of the species. Uninformative characters were removed from the dataset and a phylogeny of the genus was computed using Phylip 51 (Figure 1). The topology of this tree was identical whether distance or parsimony methods were used (Additional files 19, 20) and was also identical to a tree based on the nucleotide sequence of the approximately 1.5 Mb of the core genome in LCBs (see above). The genus broke down into three major clades: the outlying fish pathogen, Y. ruckeri; Y. pestis/Y. pseudotuberculosis; and the remainder of the 'enterocolitica'-like species. Y. kristensenii ATCC33638T was the nearest neighbor of Y. enterocolitica 8081. The outlying position of Y. ruckeri was confirmed further when we analyzed the contribution of the genome to reducing the size of the Yersinia core protein families set. If Y. ruckeri was excluded, the Yersinia core would be 2,232 protein families of N = 2 rather than 2,072 (Table 4). In contrast, omission of any one of the 10 other species only reduced the set by a maximum of 22 families.

Clustering the significant Cluster of Orthologous Groups (COG) hits 52 for each genome hierarchically (Figure 2) yielded a similar pattern for the three basic clades. The overall composition of the COG matches in each genome, as measured by the proportion of the numbers in each COG supercategory, was similar throughout the genus, with the notable exceptions of the high percentage of group L COGs in Y. pestis due to the expansion of IS recombinases and the relatively low number of group G (sugar metabolism) COGs in Y. ruckeri (Figure 2).

The Yersinia proteomes were investigated for common clusters in the three high virulence species missing from the low human virulence genomes (Figure 3). Because of the close evolutionary relationship of the 'enterocolictica' clade strains, the number of unique protein clusters in Y. enterocolitica was reduced to a greater degree than the more phylogentically isolated Y. pestis and Y. pseudotuberculosis. Many of the same genome islands identified as recent horizontal acquisition by Y. pestis and/or Y. pseudotuberculosis 91315 were not present in any of the newly sequenced genomes. However, some genes, interesting from the perspective of the host specificity of the Y. pestis/Y. pseutoberculosis ancestor, were detected in other Yersinia species for the first time. These included orthologs of YPO3720/YPO3721, a hemolysin and activator protein in Y. intermedia, Y. bercovieri and Y. fredricksenii; YPO0599, a heme utilization protein also found in Y. intermedia; and YPO0399, an enhancin metalloprotease that had an ortholog in Y. kristensenii (ykris0001_41250). Enhancin was originally identified as a factor promoting baculovirus infection of gypsy moth midgut by degradation of mucin 53. Other loci in Y. pestis/Y. pseudotuberculosis linked with insect infection, the TccC and TcABC toxin clusters 54, were also found in Y. mollaretti. In Y. mollaretti the Tca and Tcc proteins show about 90% sequence identity to Y. pestis/Y. pseudotuberculsis and share identical flanking chromosomal locations. Further work will need to be undertaken to resolve whether the insertion of the toxin genes in Y. mollaretti is an independent horizontal transfer event or occurred prior to divergence of the species.

After comparison of the new low virulence genomes, the number of protein clusters shared by Y. enterocolitica and the other two pathogens was reduced to 12 and 13 for Y. pseudotuberculosis and Y. pestis, respectively (Figure 3). The remaining shared proteins were either identified as phage-related or of unknown role, providing few clues to possible functions that might define distinct pathogenic niches. Performing a similar analysis strategy between others genome of the 'enterocolitica' clade and Y. pestis or Y. pseudotuberculosis gave a similar result in terms of numbers and types of shared protein clusters.

Only sixteen clusters of chromosomal proteins were found to be common to all three high-virulence species but absent from all eight non-pathogens (Figure 3). Eleven of these are components of the yersiniabactin biosynthesis operon (Additional file 21), further highlighting the critical importance of this iron binding siderophore for invasive disease. The other proteins are generally small proteins that are likely included because they fall in unassembled regions of the eight draft genomes. One other small island of three proteins constituting a multi-drug efflux pump (YE0443 to YE0445) was common to the high-virulence species but missing from the eight draft low-virulence species.

The basic metabolic similarities of Y. enterocolitica and the seven species on the main branch of the Yersinia genus phylogenetic tree are further illustrated in Figure 4, where the best protein matches against each Y. enterocolitica 8081 gene product 15 are plotted against a circular genome map. Very few genes exclusive to Y. enterocolitica 8081 were found outside of prophage regions, which is a typical result when groups of closely related bacterial genomes are compared 55. One of the largest islands found in Y. enterocolitica 8081 was the 66-kb Y. pseudotuberculosis adhesion pathogenicity island (YAPI_ye) 155657, a unique feature of biotype 1B strains. YAPI_ye, containing a type IV pilus gene cluster and other putative virulence determinants, such as arsenic resistance, is similar to a 99-kb YAPI_pst that is found in several other serotypes of Y. pseudotuberculosis 1457 but is missing in Y. pestis and the serotype I Y. pseudotuberculosis strain IP32953 14. A model has been proposed for the acquisition of YAPI in a common ancestor of Y. pseudotuberculosis and Y. enterocolitica and subsequent degradation to various degrees within the Y. pseudotuberculosis clade. However, the complete absence of YAPI from any of the seven species in the Y. enterocolitica branch (Figure 4), as well as from most strains of Y. enterocolitica 15, argues against an ancient acquisition of YAPI, but instead suggests the recent independent acquisition of related islands by both Y. enterocolitica biogroup 1B and Y. pseudotuberculosis.

Many genes previously thought to be unique to Y. enterocolitica in general and biotype 1B in particular turned out to have orthologs in the low human virulence species sequenced in this study. These included several putative biotype 1B-specific genes identified by microarray-based screening 58, including YE0344 HylD hemophore (yinte0001_41550 has 78% nucleotide identity), YE4052 metalloprotease (yinte0001_36030 has 95% nucleotide identity), and YE4088, a two-component sensor kinase, which had orthologs in all species. Large portions of the biogroup 1B-specific island containing the Yts1 type II secretion system were found in Y. ruckeri, Y. mollaretii, and Y. aldovae. Y. aldovae and Y. mollaretii also had islands containing ysa type three secretion systems (TTSS) with 75 to 85% nucleotide identity to the homolog in Y. enterocolitica 1B. The ysagenes are a chromosomal cluster 91315 that in Y. enterocolitica, at least, appears to play a role in virulence 59. The Y. enterocolitica ysa genes are found in the plasticity zone (Figure 4) and have very low similarity to the Y. pestis and Y. pseudotuberculosis ysa genes (which are more similar to the Salmonella SPI-2 island 6061) and are found between orthologs of YPO0254 and YPO0274 9. Species within the Yersinia genus had either the Y. enterocolitica type of ysa TTSS locus or the Y. pestis/SPI-2 type (with the exception of Y. aldovae, which has both; Additional file 22). This suggested the exchange of chromosomal TTSS genes within Yersinia.

The modular nature of the islands found in the Y. enterocolitica genome was demonstrated further by two examples gleaned from comparison with the evolutionarily closest low human virulence genome, Y. kristensenii ATCC 33638T (Figure 1). The YGI-3 island 15 in Y. enterocolitica 8081 is a degraded integrated plasmid; at the same chromosomal locus in Y. kristensenii ATCC 33638T a prophage was found, suggesting that the YGI-3 location may be a recombinational hotspot. Another Y. enterocolitica 8081 island, YGI-1, encodes a 'tight adherence' (tad) locus responsible for non-specific surface binding. Y. kristensenii ATCC 33638T had an identical 13 gene tad locus in the same position, but the nucleotide sequence identity of the region to Y. enterocolitica 8081 was uniformly lower than that found for the rest of the genome, suggesting there had been either a gene conversion event replacing the tad locus with a set of new alleles in the recent history of Y. kristensenii or Y. enterocolitica or the locus was under very high positive selective pressure.

Comparison of the Y. enterocolitica genome to Y. pestis and Y. pseudotuberculosis revealed some potentially significant metabolic differences that may account for varying tropisms in gastric infections 62. Y. enterocolitica 8081 alone contained entire gene clusters for cobalamin (vitamin B12) biosynthesis (cbi), 1,2-propanediol utilization (pdu), and tetrathionate respiration (ttr). In Y. enterocolitica and Salmonella typhimurium 6364, vitamin B12 is produced under anaerobic conditions where it is used as a cofactor in 1,2-propanediol degradation, with tetrathionate serving as an electron acceptor. This study showed the genes for this pathway to be a general feature of species in the 'enterocolitica' branch of the Yersinia genus (with the caveat that some portions are missing in some species; for example, Y. rohdei is missing the pdu cluster (Table 5). Additionally, Y. intermedia, Y. bercovieri, and Y. mollaretii contained gene clusters encoding degradation of the membrane lipid constituent ethanolamine. Ethanolamine metabolism under anaerobic conditions also requires the B12 cofactor. Y. intermedia contained the full 17-gene cluster reported in S. typhimurium 65, including structural components of the carboxysome organelle. Another discovery from the Y. enterocolitica genome analysis was the presence of two compact hydrogenase gene clusters, Hyd-2 and Hyd-4 15. Hydrogen released from fermentation by intestinal microflora is imputed to be an important energy source for enteric gut pathogens 66. Both gene clusters are conserved across all the other seven enterocolitica-branch species, but are missing from Y. pestis and Y. pseudotuberculosis. Y. ruckeri contained a single [NiFe]-containing hydrogenase complex.

Y. ruckeri, the most evolutionarily distant member of the genus (Figure 1) with the smallest genome (3.7 Mb), had several features that were distinctive from its cogeners. The Y. ruckeri O-antigen operon contained a neuB sialic acid synthase gene, therefore the bacterium was predicted to produce a sialated outer surface structure. Among the common Yersinia genes that are missing only in Y. ruckeri were those for xylose utilization and urease activity, consistent with phenotypes that have long been known in clinical microbiology 67 (Table 3). Surprisingly, we discovered that Y. ruckeri was also missing the mtnKADCBEU gene cluster that comprises the majority of the methionine salvage pathway 68 found in most other Yersiniae. These genes have also been deleted from Y. pestis, but as with Y. ruckeri, the mtnN (methylthioadenosine nucleosidase) is maintained. The loss of these genes in Y. pestis has been interpreted as a consequence of adaptation to an obligate host-dwelling lifecycle, where the availability of the sulfur-containing amino acids is not a nutritional limitation 15.

Type strains of the eight Yersinia species sequenced in this study (Table 1) were acquired from the American Type Culture Collection (ATCC) and propagated at 37°C or 25°C (Y. ruckeri) on Luria media. DNA for genome sequencing was prepared from overnight broth cultures propagated from single colonies streaked on a Luria agar plate using the Promega Wizard Maxiprep System (Promega, Madison, WI, USA).

Genomes were sequenced using the Genome Sequencer 20 Instrument (454 Life Sequencing Inc., Branford, CT) 34. Libraries for sequencing were prepared from 5 μg of genomic DNA. The sequencing reads for each project were assembled de novo using the newbler program (version 01.51.02; 454 Life Sciences Inc).

Optical maps 38 for each genome using the restriction enzymes AflII and NheI (Y. aldovae and Y. kristensenii only have maps for AflII) were constructed by Opgen Inc. (Madison, WI). The newbler assemblies for each genome were scaffolded using the optical maps and the SOMA package 39 (Additional file 4). Assemblies that did not align against the optical map were tested for high read coverage, unusual GC content, and good matches to plasmid-associated genes from the ACLAME database 84 (BLAST E-value less than 10^-20) to identify sequences that could potentially be part of an extrachromosomal element.

We used two methods for detecting disrupted proteins used. In the first method clustered protein groups were used to adduce evidence for possible gene disruption events. The clusters were parsed for pairs of proteins that met the following criteria: both from the same genome; encoded by genes located on the same strand with less than 200 bp separating their frames; and total length of the combined genes was not greater than 120% of the longest gene in the cluster. The second method used was the FSFIND algorithm 85 with a standard bacterial gene model to compare the accumulation of predicted frameshifts across different genomes.

In order to rule out artifacts due to undocumented features of the newbler assemblies, new assemblies were generated for validation purposes by re-mapping all the shotgun reads to the sequence of the assembled contigs using AMOScmp 86. The resulting assembly was then subjected to analysis using the amosvalidate package 37. The output of this program includes a list of genomic regions that contain inconsistencies highlighting possible misassemblies. The resulting regions were manually inspected to reduce the possibility of assembly errors. The regions flagged by the amosvalidate package are provided in GFF (general feature format), compatible with most genome browsers (Additional file 3).

The presence of repeats is known to confound assembly programs and the newbler assembler is known to collapse high-fidelity repeat instances into a single contig. To account for the possibility of such misassemblies, we computed the copy number of contigs based on coverage statistics and used this information to correct our estimates for the abundance of classes of repeats (Additional file 3). To find known insertion sequences, the genomes were scanned for matches using the IS finder web service 40 with a BLAST E-value threshold of 10^-10 (matches to known repeat contigs were counted as multiple matches based on the coverage of the contig). In addition, we searched for common repeat sequences in the genome using the RepeatScout program 41 after duplicating known repeat contigs. The repeats found in each genome were collected (64 sequences) and transformed into a non-redundant set of 44 sequences using the CD-HIT program 87 (Additional file 6). The repeats found were then searched against all the genomes using BLAST with an E-value threshold of 10^-10 to record matches. The resultant figures for repeat content are estimations that may be lower than the true number found in the genomes.

DNA signatures for the Y. pestis and the Y. enterocolitica genomes were identified using the Insignia pipeline 44. Signatures of 100 bp or longer were considered good candidates for the design of detection assays. These signatures were then compared with the genomes of the Yersinia strains sequenced during the current study using the MUMmer package 88 with default parameters. Signatures that matched by more than 40 bp were deemed invalidated, as they would likely lead to false-positive results.

We used DIYA 89 for automated annotation, which is a pipeline for integrating bacterial analysis tools. Using DIYA, the assemblies generated by newbler were scaffolded based on the optical map, concatenated, and used as a template for the programs GLIMMER 90, tRNASCAN-SE 91, and RNAmmer 92 for prediction of open reading frames and RNA genes, respectively. All predicted proteins encoded by each coding sequence were compared against a database of all proteins predicted from the canonical annotation of Y. pestis CO92 9 as a preliminary screen for potentially novel functions. The GenBank format files created from the eight genomes sequenced in this study were combined with other DIYA-annotated, published whole genomes to form a dataset for analysis. All proteins were searched against the UniRef50 database (July 2008) 93 using BLASTP 94 and against the Conserved Domain Database 95 using RPSBLAST 96 with an E-value threshold of 10^-10 to record matches.

The annotated genome data were submitted to NCBI GenBank and the sequence data submitted to the NCBI Short Read Archive (SRA). The accession numbers are: Y. rohdei, ATCC_43380: [Genbank:ACCD00000000]/[SRA:SRA009766.1]; Y. ruckeri ATCC_29473: [Genbank:ACCC00000000]/[SRA:SRA009767.1]; Y. aldovae ATCC_35236: [Genbank:ACCB00000000]/[SRA:SRA009760.1]; Y. kristensenii ATCC_33638: [Genbank:ACCA00000000]/[SRA:SRA009764.1]; Y. intermedia ATCC_29909: [Genbank:AALF00000000]/[SRA:SRA009763.1]; Y. frederiksenii ATCC_33641: [Genbank: AALE00000000]/[SRA:SRA009762.1]; Y. mollaretii ATCC_43969: [Genbank:AALD00000000]/[SRA:SRA009765.1]; Y. bercovieri ATCC_43970: [Genbank:AALC00000000]/[SRA:SRA009761.1].

Yersinia genomes were aligned using the standard MAUVE 46 algorithm with default settings. A cutoff for 1,500 bp was set as the minimum LCB length. LCBs for each genome were extracted from the output of the program and concatenated. From the alignment nucleotide diversity was calculated by an in-house script using positions where there was a base in all 11 genomes. Because of the size of the dataset, the calculated value of Π is very robust in terms of sequence error. We calculated that 112,696 nucleotides of sequence in the concatenated core would have to be wrong to alter the estimation of P by ± 5% (Additional file 24). PHYLIP 51 programs were used to build a consensus tree of the MAUVE alignment with bootstrapping 1,000 replicates. The underlying model for each replicate was Fitch-Margoliash. The final phylogeny was resolved according to the majority consensus rule.

The complete predicted proteome from all genomes annotated in this study was searched against itself using BLASTP with default parameters. We removed short, spurious, and non-homologous hits by setting a bitscore/alignment length filtering threshold of 0.4 and minimum protein length of 30. Predicted proteins passing this filter were clustered into families based on these normalized distances using the MCL algorithm 48 with an inflation parameter value of 4. These parameters were based on an investigation of clustering 12 completed E. coli genomes, which produced very similar results to a previous study 42.

From the results of clustering analysis, 681 proteins were found that had exactly one member in each of the genomes and the length of each protein in the cluster was nearly identical. These protein sequences were aligned using ClustalW 50, and individual gene alignments were concatenated into a string of 170,940 amino acids for each genome. Uninformative characters were removed from the dataset using Gblocks 97 and a phylogeny reconstructed with PHYLIP 51 under a neighbor-joining model. To evaluate node support, a majority rule-consensus tree of 1,000 bootstrap replicates was computed.