In bacteria, replication fork arrest is mainly repaired by homologous recombination 1 . When such a recombination event occurs an odd number of times in one DNA replication event of circular chromosomes, the replicated chromosome is not properly segregated into two daughter chromosomes but instead produces a concatenated dimer 2 3 . Therefore, many bacteria harbor highly conserved chromosome dimer resolution (CDR) machinery to separate the dimer chromosome into two monomer daughter chromosomes.

In Escherichia coli, chromosome dimers are resolved by two tyrosine recombinases, XerC and XerD, by the addition of a crossover at a specific 28-bp sequence called the dif site, which is located in the replication termination region of the chromosome 4 5 . The dif sequence contains a pair of palindromic sequence motifs that correspond to the binding domains of XerC and XerD. The reaction is coordinated to the last stages of cell division by an essential cell division protein, FtsK, which functions as a septum-located DNA translocase 6 7 8 9 10 . FtsK moves along the chromosome unidirectionally towards the dif sequence, thanks to polar and orientated sequences, the KOPS 11 12 13 . CDR is initiated when FtsK reaches dif and its extreme C-terminal domain directly interacts with the C-terminal domain of XerD 14 15 16 17 18 . The dif/XerCD chromosome dimer resolution system seems widely conserved. In vivo experimental evidence for its conservation has been obtained in Xanthomonas campestris, Caulobacter crescentus and Vibrio cholerae 19 20 21 . In vitro characterization of Xer recombinases and dif sites has also been carried in Haemophilus influenzae and Bacillus subtilis 22 23 . However, the importance of dif/XerCD for the fitness of bacteria has only been demonstrated in E. coli and V. cholerae 20 24 . In some other bacteria, like Lactococci and Streptococci, chromosome dimer resolution is resolved by single tyrosine recombinases that act at specific dif site 25 26 . In this case, dimer resolution still depends on FtsK and dif is still located opposite the origin of replication between oriented polar sequences 27 . Several filamentous phages are known to hijack this site-specific recombination machinery of dif/XerCD for their integration into the host chromosome, containing pseudo-dif sequences within these phage genomes 28 29 30 31 32 33 34 . However, the dif sequence remains intact during such recombination process to ensure the integrity of chromosome dimer resolution machinery 35 36 . The dif-like sequences in phages often contain more variable central region that is longer than the canonical 6 bp 31 33 34 , and the XerD binding arm is considerably degenerate 28 .

Because there is only one origin of replication on bacterial circular chromosomes, replication generally terminates in a specific region of the chromosome. This can be followed by the existence of a GC skew on the two replichore arms of the chromosomes with a shift-point opposite the origin of replication 37 . Based on the observation that dif sites are generally located at or near the GC skew shift-point, Hendrickson and Lawrence proposed that replication might generally terminate at dif, which coordinate replication and chromosome dimer resolution 38 . In E. coli, the replication process usually terminates at a narrow region that includes approximately 5% of the genome length and is located directly opposite the replication origin 39 40 41 . This is partly due to the existence of the Tus/Ter replication fork trap 41 . dif is located within the replication fork trap but termination occurs precisely at the Tus site, not at dif 42 and dif is active when displaced outside of the replication termination region if it is still within the zone where KOPS converge 24 . Nevertheless, the lack of universal conservation of the Tus protein may suggest that replication terminated at dif sites until the relatively recent takeover by the Tus-Ter system 43 . We reasoned therefore that the comprehensive identification of dif sites and of their location with respect to the GC skew shift-point in hundreds of complete genomes might provide clues to the evolution of the CDR machinery and its possible link with the replication termination mechanism in bacterial species.

Prediction of the dif sequences has been reported by several groups with different approaches. Hendrickson and Lawrence showed that sequence skew can be used to predict the location of dif sites, and they identified putative dif sequences in 25 bacteria based on sequence similarity 38 . Le Bourgeois and colleagues reported a new type of tyrosine recombinase, named XerS, which is responsible for CDR in Streptococci and Lactococci and this recombinase targets a 31-bp sequence element named dif _SL 25 . For comparison, they predicted dif sequences in 22 Firmicutes based on their similarity to that of B. subtilis with Megablast 44 and on the fact that the dif sequence occurs only once per genome. Val and colleagues identified that V. cholerae chromosome II, whose many features are plasmid-like, has an original dif sequence independently, and therefore it has FtsK-dependent CDR 20 . For this purpose, they predicted dif sequences in five α-Proteobacteria and ten β-Proteobacteria that harbour multiple chromosomes, and discussed a conserved FtsK-dependent CDR on multiple chromosomes based on the close relative distance of the position of dif sequences and the GC skew shift-points. Their prediction method is based on a HMMER 45 score (<1.0e-05) with a profile built from 27 aligned dif sequences in the largest chromosomes of γ-Proteobacteria species, with manual checking for 6-bp spacing between two XerC and XerD binding motifs.

Carnoy and Roten reported the most comprehensive predictions to date, identifying putative dif sequences in 204 chromosomes in 137 Proteobacteria strains, discussing the high conservation of dif/XerCD systems and the possible loss of dif sequences in endosymbionts, with suggestions for other CDR mechanisms 46 . Here, the prediction was based on BLAST searches and YASS alignment 47 with the dif sequences of E. coli and B. subtilis, and candidates were selected based on their proximity to the GC skew shift-points and a single occurrence per chromosome. Previous predictions were therefore limited to three bacterial phyla: Proteobacteria, Firmicutes and Actinobacteria.

To this end, we describe comprehensive predictions for dif sequences based on a machine learning approach, tracing the phylogenetic conservation patterns of XerCD recombinases and using an iterative hidden Markov modeling method. Furthermore, we observed the relationship between predicted dif sequence positions and GC skew shift-points, and investigated whether replication termination occurs at the dif site.

In this study, we first demonstrated that the conservation of XerCD genes follows phylogenetic conservation patterns that are specific to each bacterial phylum (Figure 1). Based on this principle, we comprehensively predicted the dif sequences in hundreds of completely sequenced genomes using a recursive strategy that iteratively models and predicts these sequences using profile hidden Markov models. As a result, we obtained unique candidate dif sequences in 715 chromosomes in 641 strains that were validated through multiple means, resulting in the largest collection of predicted dif sequences assembled to date. In comparison to previous work by Carnoy and Roten, which predicted dif sequences in 228 genomes, our predictions coincided with their results in 208 genomes and we added 507 genomes, including Aromatoleum aromaticum str. EbN1, which Carnoy and Roten reported to lack the dif/XerCD system. Excluding strains or chromosomes we could not predict, namely, A. tumefaciens str. C58, Burkholderia sp. 383 chromosome I, II, D. psychrophila LSv54, N. multiformis ATCC 25196, P. denitrificans PD1222 chromosome I, II and S. denitrificans DSM 1251, the predicted dif sequences in this study differed in 12 chromosomes in comparison to the results of Carnoy and Roten: C. crescentus CB15, Granulibacter bethesdensis CGDNIH1, Pseudoalteromonas haloplanktis TAC125 chromosome II, Ralstonia eutropha H16 chromosome II, Rhodobacter sphaeroides 2.4.1 chromosome I, R. sphaeroides 2.4.1 chromosome II, Rickettsia bellii OSU 85-389, R. conorii, R. felis URRWXCal2, R. prowazekii, R. typhi Wilmington, and Shewanella sp. ANA-3. For R. eutropha H16 chromosome II and P. haloplanktis TAC125 chromosome II, both studies predicted positions that were symmetric from the origin of replication, and although experimental confirmation is required to confirm which candidates function in vivo, the palindromic structures of the XerCD binding sites are more conserved in the candidates predicted by our method. Therefore, overall, our results were identical with those of Carnoy and Roten for 92% of the genome analyzed (208/228), and 11/12 mismatch resulted in candidates with more conserved XerCD binding sites, with the addition of 507 genomes among numerous phyla. Carnoy and Roten noted that some Vibrio species contain two dif sites both located at the vicinity of the GC skew shift-points. Therefore, we further tested whether the predicted dif sites in multiple chromosomes are all located near the GC skew shift-points. Using 5% genomic distance as a threshold, 45 out of 54 strains with two chromosomes, including Vibrio species, and 6 out of 9 strains with three chromosomes showed such agreement of the positions, (Additional file 2, Table S1).

There are four factors that may explain the advantages of our results. First, the selection of bacterial strains in the study by Carnoy and Roten was limited to genomes harboring XerCD that were identified by their similarity to those of E. coli, whereas we used all genomes with XerCD orthologs as identified by the KEGG Orthology database. While there is a little time-delay until the sequences are annotated and incorporated into the KEGG Orthology database, use of this database provides a more generic and comprehensive starting point. Second, similarity searches using software tools such as BLAST are not suitable for short sequence motifs that undergo mutation, and the difficulty in identifying only those dif sequences with sequence similarity has been shown for C. crescentus 50 and several classes of Proteobacteria 20 . Third, dif sequences require two binding motifs of XerC and XerD to be functional 51 ; therefore, the conservation of palindromic structure at the 7-12-bp and 17-22-bp positions should be confirmed for each predicted candidate. Finally, the use of iterated HMM allowed dif sequence prediction using the profiles of closely related species for each iteration, following the phylogenetic conservation pattern of XerCD.

The high predictability shown in this study suggests that the dif/XerCD system of chromosome dimer resolution is highly conserved among bacterial species and that dif sequences are almost always conserved when XerCD is present within the genome. In fact, according to the KEGG Orthology database, XerC and XerD are conserved in approximately 60-70% of bacterial species, which is a higher percentage than is found for the replication termination protein Tus 52 and for universal genes such as the SOS response repressor LexA 53 . In light of the remarkable conservation of the dif/XerCD system, although it is beyond the scope of this study, explorations of alternative CDR machinery in species that lack the dif/XerCD machinery would be an interesting area of future research. Chromosome dimer resolution pathways are suggested to be present in species that lack the dif/XerCD system, and several alternative pathways have been reported and suggested. Le Bourgeois et al. reported an unconventional CDR pathway involving only one recombinase (XerS) in Streptococci and Lactococci, along with a 31-bp dif sequence 25 . Similarly, through computational analysis, Carnoy and Roten suggested the existence of another pathway, termed XerH, in ε-Proteobacteria in place of XerCD and XerS and discussed the likelihood of the existence of dif analogues in these species 26 46 . The basic strategy of iterated HMM should be applicable in predicting dif analogues in these species when defined seed sequences and detailed positions of recombinase binding sites are elucidated.

Although we limited our analysis to strains containing XerCD orthologs, our predictions failed in several species. In Proteobacteria, we could not identify dif sequences in five organisms and seven chromosomes, including species with single chromosomes (Nitratiruptor sp. SB155-2 and S. denitrificans DSM 1251) that are ε-Proteobacteria, where an alternative CDR mechanism involving XerH is suggested 46 , and species with multiple chromosomes (P. denitrificans PD1222 chromosome I, P. denitrificans PD1222 chromosome II, B. phytofirmans PsJN chromosome I, and Burkholderia sp. 383 chromosome I and III). Among these, B. phytofirmans PsJN and Burkholderia sp. 383 contained dif sequences in other chromosomes, indicating that the dif/XerCD system is conserved in these strains. Similarly, in Firmicutes, we could not determine dif sequence in L. helveticus DPC 4571, C. perfringens str. 13 or C. beijerinckii NCIMB 8052. Among these strains, L. helveticus DPC 4571 has an alternative CDR recombinase XerS in its genome, indicating that the dif/XerCD system may not be functional. This is an intriguing example of possible evolutionary intermediate with the co-existence of two systems, presumably resulting from a horizontal gene transfer event. While we are unable to find a dif sequence corresponding to the XerS machinery, xerS gene in this species is located close to the GC skew shift-point (xerC: 1031814-bp, xerD: 1055574-bp, xerS: 1228715-bp, and GC skew shift-point: 1225733-bp), which is indicative of its functionality as shown in previous works 25 26 46 . C. perfringens str. 13 and C. beijerinckii exhibit highly biased GC contents (28.57% and 29.86%, respectively), and hidden Markov profiling of AT-rich dif sequences may have failed due to the background AT-richness of the genome. Comparative studies of dif/XerCD systems using close relatives of these genomes may provide evolutionary clues regarding the acquisition and loss of CDR machinery. For example, mapping the types of CDR machinery to the phylogenetic tree of ε-Proteobacteria obtained using 16S rRNA sequences with the dnaml program in the Phylip package shows that a XerH type of CDR machinery may have diverged at an early stage within this phylum. The XerCD type of CDR seems to be absent in the Campylobacter and Helicobacter genera, except for Helicobacter hepaticus, which suggests the existence of the XerH type of CDR in the common ancestor of these species (Figure 3). The dif candidate in H. hepaticus was predicted with iterated HMM only marginally above the threshold, with a score of 10.2 and an e-value of 5.5e-05. Further analysis is required to identify whether this species actually contains dif/XerCD or XerH-type machinery.

Predictions failed in all species belonging to the phylum Cyanobacteria. Although XerCD is present in these species, the sequence similarity distance of XerCD in Cyanobacteria to those of other phyla was high (average of 0.358 ± 0.0159, N = 540), with a minimum distance of 0.322 to Actinosynnema mirum (Actinobacteria), which exceeded the 0.3 threshold that was shown in Figure 4. Therefore, this divergence of XerCD in Cyanobacteria from those of other phyla implies low applicability of the iterated HMM approach, which utilizes the phylogenetic conservation pattern of XerCD. One possible explanation for the prediction failure in this phylum is that the dif sequences and XerCD are highly divergent in Cyanobacteria, preventing their identification with sequence profiles. The replication origin in Cyanobacteria is yet to be identified, and GC skew is weak in these species, implying low degree of replicational mutation/selection pressures, which could also be a reason for the failure of prediction in these species.

Predicted dif sequences largely existed in non-coding regions (93.92%). More than half of these coding regions that contained dif sequences were hypothetical, with no functional annotation. Furthermore, we found two dif sequences included in phage ORF in Vibrio and Xanthomonas. While these sequences may be integrated with the phages by their hijacking of the host recombination machinery, these sequences are speculated to be the functional dif sites, due to 1. their unique occurrence within the genome opposite of the replication origin, and 2. their similarity as identified by our phylogenetic modeling approach. As previously shown in Proteobacteria 46 , the XerC binding site is more variable and the XerD binding site is more conserved in all phyla (Figure 5), both for genomes with single chromosomes and for those with multiple chromosomes, presumably due to the interaction between XerD and FtsK for the initiation of first strand exchange 14 . The dif sequences in α-Proteobacteria with single chromosomes showed higher variation compared to these of other classes and phyla, but this variation was correlated with variations in genomic GC content (Additional file 1, Figure S4). These differences between variations are partly explains the failure of our prediction in extremely AT-rich genomes, such as those found in C. perfringens and C. beijerinckii.

Although dif sequences are expected to be located near the shift-point of the GC skew, we did not use this feature to predict and validate dif sequences with iterated HMM; therefore, using the comprehensively predicted dif sequences across numerous phyla, we were able to directly compare the positions of predicted dif sequences with those of the GC skew shift-points to analyze their relationships. As expected, these two positions are highly correlated in terms of genomic loci, confirming a previous work 38 . In this respect, because GC skew is the cumulative result of replicational selection/mutations, the degree of conservation of the CDR machinery is presumably in concordance with the degree of replication selection/mutation pressures (i.e. GC skew), which is partly characterized by the difference in the replication machinery and partly characterized by the growth rate 54 . On the other hand, as shown in Figure 2B, the differences in the positions of the GC skew shift-point and the strength of the GC skew, as quantified by GCSI, were not correlated. If replication termination occurs at the dif site, as proposed by Hendrickson and Lawrence 38 , a stronger GC skew that is generated by a larger number of replication events and/or a higher mutation rate should statistically bring the GC skew shift-point closer to the dif site by the central limit theorem and law of large numbers. In fact, the overall correlation of these loci leads to the proposal that the dif site is the replication termination point. However, because a stronger degree of replication mutation/selection pressures does not bring these two loci closer to each other, they are not in a causal relationship. Therefore, although the dif sequence is located near the replication termination site for efficient CDR, the replication termination site is suggested to be at a site other than the dif site, as was recently shown in vivo 42 . On the other hand, the dif sequences in Firmicutes are more conserved in various phyla because the profile of Firmicutes was the best suited as the initial profile of iterated HMM in Chlorobi, Acidobacteria, Gemmatimonadetes, Nitrospirae, Elusimicrobia, Tenericutes, and Spirochaetes, where initial seed sequences were not available, and those in Proteobacteria were more variable, as shown by the requirement to predict by iterated HMM in classes instead of phyla. Tus proteins, which are shown to terminate replication in vivo, are more conserved in Proteobacteria and are not widely conserved in other, partly supporting the possible change in replication termination mechanism by a relatively recent takeover by the Tus-Ter system 43 . On the other hand, to the best of our knowledge, Tus analogues have not been comprehensively searched in other phyla, and therefore further analysis is required in order to fully support this hypothesis.

By taking the phylogenetic iterated HMM approach and validating predicted candidates through a combination of HMMER score thresholds, conservation of palindromic structure, and cross-validation, we achieved a comprehensive identification of unique dif candidates in hundreds of genomes. As the result, we obtained unique candidate dif sequences in 715 chromosomes in 641 strains that were validated through multiple means, resulting in the largest collection of predicted dif sequences assembled to date. All of the predicted dif sequences described in this study, as well as visualizations of dif locations on circular genome maps, are freely available in an online database at http://www.g-language.org/data/repter/. The locations of dif sequences can be useful for studies of the regions surrounding the replication terminus, for phylogenetic studies of the replication termination and chromosome dimer resolution mechanisms, and can serve as supporting evidence for GC skew analyses.

Furthermore, we compared the positions of predicted dif sequences with those of the GC skew shift-points to understand the relationship between dif sequence and replication terminus using GCSI. As the result, although these two positions were highly correlated in terms of genomic loci, the differences in the positions of the GC skew shift-point and the GCSI were not correlated. Therefore, despite the dif sequence is located near the replication termination site for efficient CDR, the replication termination site is suggested to be at a site other than the dif site.

NK carried out the analysis, and NK and KA wrote the manuscript. MT supervised the work, and all authors read and approved the final manuscript.