Background

Helicobacter pylori is a bacterial pathogen associated with gastritis, peptic ulcers, gastric adenocarcinoma, and rare lymphomas [1]. It has a highly panmictic population structure in which homologous recombination makes the predominant contribution to sequence differences within a highly diverse population structure [2]. The acquisition of genes from other strains and species is by far the most rapid evolutionary process. This occurs frequently without loss of existing functions, is central to the evolution of niche-adaptive and pathogenic characteristics of bacteria, and greatly influences inter-strain differences in gene complement [3–5]. In this context, it is notable that none of the traits typically used to differentiate E. coli from Salmonella can be attributed to point mutation genes but are broadly attributable to horizontal exchange [6]. H. pylori is relatively unusual in that it is a naturally transformable Gram-negative species that does not appear to have a species-specific DNA uptake sequence and appears to rely upon its niche separation as a transformation barrier [7]. Disease associated H. pylori strains have been divided into two types, type I being those that carry the cag pathogenicity island [8] (cag PAI), which has a foreign species origin, and are associated with more severe disease.

Dinucleotide composition is highly stable within a genome and can distinguish between sequences from different species. Based upon its constancy the species composition is referred to as a 'genome signature' [9, 10]. This characteristic has been applied to assessments of DNA metabolic processes such as methylation and base conversion, DNA structure, and evolutionary relationships. It has also become established as a method for the identification of sequences that have been acquired by inter-species horizontal transfer. For example, lateral transfer has recently been shown using these methods for a tryptophan pathway operon [11], the gain of additional metabolic functions in Pseudomonas putida [12], a determination that many gain of function genes have been acquired by E. coli rather than lost from S. typhi [13], and more recently developed Bayesian methods based upon similar premises have been used to assess global signatures and determine the origins of some lateral transfer events [14, 15]. However there are problems associated with this and other methods that use progressive 'walking windows', and the larger the window the greater the problems. These result from the inclusion of intergenic sequence, the inability to distinguish divergences due to a single highly divergent gene from that from a cluster of less divergent ones, and an inability to identify the limits of the abnormal regions. In practice additional features are necessary to determine the ends of such regions, such as the location of repeats typical of pathogenicity islands in H. pylori [16], or comparisons with other sequences as in N. meningitidis strain MC58 [17]. In addition, divergence scores are influenced by the size of the sampling window used such that sampling effects limit analysis of sequences shorter than about 800 bp (data not presented), and the need to use fixed window sizes prevents gene by gene studies.

We describe the use of a linear implementation of signature analysis that can efficiently address a range of walking window sizes using dinucleotide signatures (DNS) and longer signatures. In addition, use of a new approach based upon classical text analysis that allows analysis of genomes gene-by-gene is described. Analysis of H. pylori sequences, combined with comparisons of the identified genes between genomes, reveals complex changes that influence both niche-adaptive and core functions illustrating a previously unpredicted range of functions which are continuously undergoing variation and selection.

Results and discussion

Genes were ranked on the basis of their divergence from the mean genome composition. The degree of divergence that is indicative of acquisition from other species is not an absolute. The frequency with which genes are acquired, the untypicality of the donated material, and the rate at which they are ameliorated to the host sequence composition influence it. Strains J99 and 26695 had 53 (Table 1) and 60 (Table 4) genes respectively with DNS that were >2 SD from the mean. Those with annotated functions included genes from the cag pathogenicity island (6 and 5), vac and related toxins (3 and 4), and restriction-modification genes (2 and 4). On the basis of the similarities determined in the H. pylori strain J99 sequence annotation, 7 of the most divergent genes as determined by DNS are not present in strain 26695. Likewise, 2 of the 50 most divergent genes in strain 26695 are not present in strain J99. This is consistent with the identification of genes acquired from other species that have not extended to both sequenced strains. It also suggests that a significant proportion of the 6 to 7% of genes unique to one or other strain [18] are inherent to the Helicobacter gene pool, but are variably present in different strains rather than reflecting recent foreign origins. Comparisons of a selection of identified orthologous genes in the two strains are shown in Figure 1.

Table 1 The 53 most divergent (>2 SD) genes in H. pylori strain J99 by DNS showing their ranking in strain 26695 and in TNS and HNS analysis
Full size table
Table 4 Top 60 most divergent (>2 SD) genes by DNS in H. pylori strain 26695 plus those additional genes in the top 50 genes from TNS and HNS
Full size table
Figure 1
figure 1

Comparisons using LAlign between a representative selection of orthologous genes with divergent DNA present in both H. pylori strains J99 and 26695 (presented in descending order of divergence as determined in strain J99).

Full size image

It cannot be assumed that all genes identified in this manner have been recently acquired. It is necessary to assess the nature of the sequence to determine if its divergence might be accounted for on the basis of features of the encoded protein. For example, JHP0476/HP0527, JHP1300/HP1408 and JHP0074/HP0080 include repetitive sequences likely to account for their DNS divergence. This type of analysis cannot be used to determine the possible foreign origin of such genes. Notably, the most divergent cag PAI gene (the 1st and 2nd most divergent gene in the whole genomes of strain 26695 and J99 respectively, JHP0476/HP0527) has a highly complex repetitive structure and the size of the large divergent peak associated with this island using previous methods is largely due to the presence of this gene.

While a significant proportion of the genes identified in this analysis are associated with regions including several such genes and which share characteristics of islands of horizontal transfer or pathogenicity islands, this is far from universally true. There are many instances of single genes or small numbers of genes that are present that are not associated with any features that might otherwise have been used as indicators of horizontal acquisition such as transposases and flanking repeats.

Our initial goal was to identify recently acquired and exchanged genes as candidates likely to be important in niche-adaptation, host interactions, and alterations in bacterial fitness. It has been argued that essential genes are unlikely to be transferred successfully since recipient taxa would already bear functional orthologues, which would have experienced long-term co-evolution with the rest of the cellular machinery. In contrast, it is proposed that those under weak or transient selection – like those associated with nonessential catabolic processes, new operons, and those providing new niche-adaptive changes are likely to be successfully transferred and retained [19]. This leads to a model in which a stable 'core genome' comprised of essential metabolic, regulatory, and cell division genes provides a stable context for the more labile non-essential and niche adaptive genes. On this basis such genes are used for phylogenetic studies and are thought to provide a relatively constant background in which species evolution occurs. Many of the genes identified for which functions are known affect virulence or niche adaptive genes, including: the vacuolating cytotoxin and related toxins (2 and 3), urease and flagellar components, and genes involved in iron acquisition. However, we also find clear evidence, confirmed by differences between the two genome sequences, that recent, and therefore relatively frequent, horizontal transfer is not limited to genes associated with niche adaptation and virulence. Amongst the core function genes identified were mut S, fts K, xer D, and pol A. The comparisons of the latter three between the sequence strains are shown in Figure 1f,g &1j. These comparisons support the results suggesting that these genes have been the substrates for horizontal exchange between species.

Tetranucleotide composition has been used for the consideration of the presence of palindromic sequences that might be substrates for restriction systems and Chi sites and the presence of unstable repeats mediating phase variation [10], but the use of longer component signatures has not been used to identify horizontally acquired regions in bacterial genomes. Following analysis of eukaryotic sequences it was concluded that DNS captures most of the departure from randomness in DNA sequences and that longer component lengths correlate highly with the DNS results [20]. Also, analysis of dinucleotides separated by no, one, or two other nucleotides showed that separated pairs are more nearly random than adjacent pairs and were concluded to be relatively uninformative [9]. However, in preliminary analyses, while results using the typically long walking windows gave concordant results as previously reported, we found that the use of smaller walking windows generated progressively more different patterns of divergence with other length components. Using tetranucleotide (TNS) and hexanucleotide (HNS) signature analysis we find that, while in some instances there is significant overlap between the genes identified using the different component lengths, there are substantial differences that indicate additional horizontally transferred genes not identified by DNS alone (Tables 2 to 6).

Table 2 Top 50 most divergent genes by TNS in H. pylori strain J99 plus those additional genes > 2 SD greater than the mean by DNS and the 50 most divergent by HNS
Full size table
Table 3 Top 50 most divergent genes by HNS in H. pylori strain J99 plus those additional genes >2 SD greater than the mean by DNS and top 50 by TNS
Full size table
Table 5 Top 50 most divergent genes by TNS in H. pylori strain 26695 plus those additional genes > 2 SD greater than the mean by DNS and the 50 most divergent by HNS
Full size table
Table 6 Top 50 most divergent genes by HNS in H. pylori strain 26695 plus those additional genes > 2 SD greater than the mean by DNS and the 50 most divergent by HNS
Full size table

The 50 most divergent J99 ORFs by HNS included 26 (52%) that were not in the 53 (>2 SD) most divergent by DNS, these included 11 restriction-modification system genes and 6 others that were not annotated within the strain 26695 genome sequence. The identification of genes of a type known to be horizontally exchanged, and different between the gene complements of the strains, is strong corroboration for the foreign origin of the additional genes identified by HNS. In several instances (Tables 2 to 6) the DNS did not detect these genes at all e.g. restriction enzymes that were the 3rd, 13th and 41st most divergent genes by HNS, were 319th, 857th and 750th most divergent by DNS, respectively. In some instances the TNS gave intermediate results and in others identified other genes as more divergent than the other methods. The TNS was most sensitive for the detection of rpoB (HP1198 / JHP1121) which is associated with a significantly different gene length in the two strains (Figure 1h). One explanation for this observation is that while the DNS may initially be the most sensitive indicator of horizontal exchange it may become ameliorated to the new sequence characteristics more rapidly that the longer component features, which are probably detecting qualitatively different sequence characteristics.

The differences in the analyses using different length components, and a comparison of the results from the two sequenced strains, suggest a complex evolutionary history for the cag pathogenicity island. These suggest that it probably has mosaic structure including sequences from more than one species background, in addition to sequence that is entirely typical of H. pylori.

It is normally impossible to determine the chronology of events to distinguish insertions and deletions when comparing strains. In strain 26695 there are two open reading frames that are both good candidate coding sequences. There is only one gene in this location in strain J99 composed of the 5' gene from strain 26695 and the 3' end of the subsequent gene. This could have arisen from either a deletion or an insertion event. However, the normal DNS of the J99 gene (JHP0073, 799th in divergence) and the 5' 26695 gene (HP0079, 751st in divergence), and the high divergence of the 3' 26695 gene (HP0078, 68th in divergence), indicate that the most likely event is an insertion into strain 26695 (Figure 1l). Likewise HP0119 is likely to contain an insertion and JHP1113 probably reflects the original sequences (Figure 1k).

The inclusion of two DNA metabolism genes associated with recombination and repair is notable. Both mutS and recN were identified in both strains (22nd and 35th, and 45th and 51st most divergent genes by DNS in strains 26695 and J99 respectively). When the homologous genes were compared between the strains, extensive divergences were evident between more than one region of each protein. That these genes have divergent signatures in both strains suggests that neither has a wholly native composition. This observation is consistent with the models of rapid evolution which suggest that transient competitive advantages are enjoyed by organisms that are hypermutators under conditions of environmental stress and transitions, and that these states which can be produced by mutations in DNA repair genes [21–26]. However, such states have to be reversed so that an unsustainable mutational burden is not attained, and it has been proposed that this reversal is mediated by repair following horizontal transfer and homologous recombination, and that such strains are hyper-recombinogenic [27–29]. The untypicality of mutS and recN suggest that H. pylori is another species that can make use of this strategy for diversification under stressful conditions.

The identification of RNA polymerase genes, with associated differences between the strains, is striking. The divergence of phylogenetic trees based upon different sequences has been highlighted, and particularly the differences between the trees associated with RNA polymerase genes and rRNA [30, 31]. It has been argued that RNA polymerase is as essential to cell function as is rRNA and that there is no compelling reason to chose rRNA as the more reliable marker [32]. While the DNS analysis does not address the stability of rRNA (and specifically excludes the rRNA sequences because their differing coding requirements and evolutionary pressures generate a divergent signature for other reasons), it does indicate that RNA polymerase can be a substrate for horizontal transfer, and that trees based upon this gene, or other essential genes, need not necessarily be considered a challenge to rRNA based phylogenies.

Conclusions

The spectrum of recently horizontally acquired sequences identified emphasizes the two driving forces of horizontal exchange: the transfer of a phenotype which alters or enhances bacterial fitness resulting in increased competitive fitness or altered niche adaptation, and the presence of a substrate for homologous recombination. Because of the focus upon, and relative ease of identifying, large islands associated with readily identifiable features and phenotypes, the importance of the latter component has perhaps been underestimated. The genes that have been considered to code for 'core metabolic' 'house-keeping' functions are amongst those most likely to be changed by horizontal transfer events because of the presence of homologous substrates, and changes are likely to persist even when the change is phenotypically neutral. Equally, changes in the genes involved in core functions such as gene expression and DNA metabolism may have pleotropic effects and there may be significant differences in strain behaviour, that are not simply the consequence of differences in their respective gene complements. The selection of genes for phylogenetic analysis on the basis of their coding for conserved core functions is also problematic because these are also frequently the genes most likely to share the high homology that facilitates recombination and horizontal exchange.

Methods

A traditional nucleotide signature is generated by segmenting a sequence of DNA into k equal-sized subsequences (or 'windows'). The mathematical basis for the signature is an odds ratio – p i – calculated by dividing the frequency of a length-L oligonucleotide by its expected frequency. The odds ratios for each of the 4Loligonucleotides in each window (w) are compared with the odds ratios for the overall sequence (s) [9, 10, 33]. The normalized difference δ is plotted and thus a nucleotide signature consists of a k-length sequence of δ values: δ(w,s) = (1/4L)Σ(4L,i:x)|p i (w) - p i (s)|, where x is the set of all permutations of length L and i is one such permutation.

There are interesting parallels between signature-style genome analysis and stylometric techniques previously used to determine the authorship of controversial literary texts. This is analogous with the biological problem and it is from this that our method is derived. Rather than using a fixed-window signature, signature scores are calculated for each coding open reading frame (ORF) and weighted with variance estimates so that the scores for shorter ORFs confer with their longer counterparts. Bissell's weighted cusum (cumulative sum) [34], , is modified so that n denotes the number of ORFs in the genome, X i the number of oligonucleotides in ORF i, and w i the number of nucleotides in ORF i. The results are scaled according to ORF size using the standard error σ = √(*#ORF). In this way false positives are abrogated by normalizing for over-representation of lower order peptides.

The method is implemented in Java and efficiency is maintained through an O(N) (N = sequence length) refinement: probabilities for the complete sequence are calculated in O(N) steps for any length-L oligonucleotide, and maintain O(N) when 4L>N through a hashing function; the second part of the program calculates σ for each ORF using a loop flattening technique, thereby avoiding the program having to recalculate overlapping sub-expressions. The program is available from ftp://ftp.dcs.warwick.ac.uk/people/Stephen.Jarvis/ and http://www.molbiol.ox.ac.uk/~saunders/.

Sequence alignments, as shown in Figure 1, were performed and displayed using the programs: Lalign and viewed using Lalignview [35].