Figure 1a shows an alignment of the segment spanning the CxEQ motif for a representative set of bacterial α₂M homologs. Not all bact-α₂Ms possess the CxEQ motif. Using E. coli as the reference, YfhM is the archetype of a large group, mostly with the thioester motif, and YfaS is the archetype of a smaller, diverged group always lacking the motif. The sequences of the YfhM group are sufficiently divergent that accurate alignment proved time-consuming, but was achieved over almost the whole sequence length, other than the highly variable amino termini. We did not attempt to align together the YfhM and YfaS groups and the metazoan α₂Ms. This would only be useful if the trees would be informative, but the high divergence between the groups precludes accurate alignment, leading to unreliable tree calculation. (In future, given more YfaS sequences and α₂Ms from more metazoan lineages and a solved three-dimensional structure to guide alignment, this might be worth revisiting.) One feature apparent in many of the aligned YfhM sequences is a conserved cysteine directly following the signal peptide (Figure 1b), indicating palmitoylation. The presence of an aspartic acid residue following the palmitoylated cysteine has been shown in E. coli to dictate sorting to the inner membrane 1718, in which case YfhM will be found in the periplasmic space, attached to the inner membrane. Given the CxEQ motif, covalent trapping of proteases in the periplasmic space seems to be the most likely function (whether the covalent links are to the trapped protease or between the α₂M multimers, as in the horseshoe crab Limulus 12). The YfaS group of bact-α₂Ms lack a palmitoylable cysteine, so may be secreted, while absence of the CxEQ motif indicates the molecular function must be different, at least in part, though this does not, of itself, rule out protease entrapment, as in chicken ovostatin which also lacks the reactive thioester motif 19.

A survey of completely sequenced bacterial genomes was undertaken to establish which lineages possessed bact-α₂Ms and which did not. Representative results are summarized in Figure 2. It is clear that there is a highly inconsistent correlation of bact-α₂M possession and phylogenetic relationship, except for very closely related species.

Bact-α₂Ms are absent from the full proteomes of the following anciently diverged free-living species: the hyperthermophilic chemolithoautotroph Aquifex aeolicus, the thermophilic photolithoautotroph Chlorobium tepidum, the cyanobacteria Synechocystis, Synechococcus and Prochlorococcus, all firmicutes including Bacillus subtilis, all actinobacteria including Streptomyces coelicolor, the β-proteobacterium Nitrosomonas europaea and the δ-proteobacterium Geobacter metallireducens. Furthermore, possession of bact-α₂M is inconsistently represented within clades such as the proteobacteria, spirochetes and cyanobacteria. This is well illustrated by the two species of Helicobacter, one exploiting the acidic stomach and the other the very different environment of the liver: only the latter has a bact-α₂M. The H. hepaticus genome lacks essentially all the proposed H. pylori virulence factors and is believed to possess a quite different set, adapted to its hepatobiliary habitat 20. The irregular phylogenetic correlation suggests that bact-α₂Ms are 'lifestyle' genes, affecting which niches a bacterium is able to exploit. Although an association with colonization seems clear (Figure 2), there is a strong bias in bacterial genome sequencing in favor of pathogenic species: this currently precludes a statistical assessment and might create a misleading phylogenetic perspective.

The STRING server 21 was used to check for neighboring genes that persistently co-occur with bact-α₂Ms. Using either yfhM or yfaS as seed, STRING reported two conserved gene sets that are widely found with bact-α₂Ms. The results are summarized in Figure 2. The yfhM group always co-occurs with pbpC, which encodes penicillin-binding protein 1C (PBP1C). The gene topology is almost always consistent with pbpC and yfhM being in the same operon (or co-transcribed from a bidirectional promoter, as in Anabaena). The more strongly an operon structure is conserved across species, the more likely are the encoded proteins to have associated functions 22. Moreover, products of conserved gene pairs very often associate physically 23. Therefore, if YfhM is involved in colonizing or pathogenic lifestyles, so should be its partner. PBP1C is a paralog of the periplasmic cell-wall biosynthesis proteins PBP1A and PBP1B, though with the addition of a carboxy-terminal non-enzymatic domain of approximately 100 residues (PFAM:PF06832). The PBP1A and PBP1B peptidoglycan synthases each have two enzymatic domains, an amino-terminal transglycosylase and a carboxy-terminal transpeptidase (reviewed in 24). Although it possesses the two enzymatic domains, studies have shown that PBP1C does not substitute for these proteins in cell-wall biosynthesis during vegetative growth 25: indeed deletion of pbpC has a weak phenotype not affecting cell viability in the laboratory, although the number of peptide crosslinks is increased 25. The transpeptidase domain in PBP1C is thought not to bind to most of the β-lactams that inhibit the paralogous enzymes, nor to be a functional transpeptidase 25. One curious finding is that, in vitro, PBP1C accounts for 75% of transglycosylase activity, yet is responsible for only 3% of de novo peptidoglycan biosynthesis in the cell 25. As PBP1C does not substitute for the biosynthetic enzymes, a possible role would be in emergency repairs to the peptidoglycan, where its efficient transglycosylase activity would be appropriate.

The yfaS group of bact-α₂Ms is likewise usually found in a candidate operon, at least within the proteobacteria (Figure 2), in this case with four other gene families, defined by the E. coli yfaA, yfaQ, yfaP and yfaT genes. All these genes have signal sequences and their encoded proteins are expected to be secreted or periplasmic, but, otherwise, sequence analysis has yielded no clues to their function. It is possible that all the encoded proteins function to disrupt or resist host defenses. The YfaS-like bact-α₂Ms of the free-living and highly divergent Thermotoga, Deinococcus and Rhodopirellula (none of which is known to be invasive) are not found associated with most of these other genes.

The STRING server was also used to check for any significant coexpression of yfhM, yfaS and other members of the two candidate operons, using E. coli data from the Stanford microarray database 26. All the genes associated with those for bact-α₂Ms are present in the experiments included in the STRING database, and are expressed at levels significantly above background. However, none of the genes exhibits coordinated variation in expression levels either with each other or with any other genes in the E. coli genome under the conditions investigated.

An initial rough tree calculated from an alignment of yfhM family sequences gave strong indications that several horizontal transfers had occurred among the available set. As yfhM is always found together with pbpC, indicating that the paired genes should have a shared phylogenetic history, a quick check of the PBP1C tree was also done. The two trees, which provide controls for each other's topologies, were very similar, indicating that the apparent HGTs were unlikely to be artifacts. Therefore, we undertook a more careful phylogenetic analysis with a view to improving the phylogenetic signal-to-noise ratio and using a method that is less prone to rate variation artifacts than neighbor-joining.

Alignments were reviewed and edited by hand, then processed to remove especially noisy segments, as outlined in Materials and methods. Trees were calculated with MrBayes, a Bayesian resampling protocol that is now widely adopted 27: MrBayes approaches the quality of maximum-likelihood methods while being quicker to calculate (though still computationally demanding). Results of the tree calculations are presented in Figure 3. The two trees differ by only three branch placements, indicating that the topologies are mostly sound, except for a few branches with low support (low posterior probabilities). As the calculated trees are unrooted, the ordering of the deepest branches cannot be mapped onto time.

The number of ancestral genes required to explain an observed tree topology can be determined by embedding the sequence tree within a species tree. We prepared a species tree for the bacterial species in Figure 3 such that currently uncertain affinities were assigned in favor of the observed trees: this will provide a minimum estimate of ancestral gene number. The sequence tree topology was embedded into the bacterial species tree using GeneTree 28. The reconciled tree required six gene-duplication events and 29 lineage-specific deletions. The last common ancestor (LCA) of the full set had a minimum of three genes, the LCA of the proteobacteria had four genes, while the LCA of the α/β-proteobacteria had six genes. The tree reveals a tendency for increasing gene number over time when vertical descent has strictly occurred.

The problems of the vertical descent model are manifold. First, all sequenced extant genomes have single copies of the yfhM/pbpC genes, yet vertical descent shows a progression toward increasing gene number over time. This requires late but fully independent massive gene loss to have occurred in all lineages. Second, the observed robust sequence tree topologies would require a clear affinity between cyanobacteria and spirochetes, an affinity that has hitherto gone entirely unnoticed in the field of bacterial phylogeny. Third, the number of events (gene duplications and deletions) found to be required under a model of vertical descent is based on a species tree chosen to minimize this number (see Materials and methods.) As the species tree used is unlikely to be accurate in places where bacterial phylogeny is unresolved, the number of such events required under a vertical descent model is probably greater than described (and hence, correspondingly less likely.)

Although bizarre evolutionary scenarios can always be invoked, the given tree topologies are difficult to explain solely by vertical descent from a common ancestral eubacterium.

Difficulties in accounting for the observed YfhM and PBP1C trees disappear if it is assumed that a number of horizontal gene transfers have occurred. Vertical transmission then only occurred among some sets of quite closely related bacteria. There are four deeply diverged sets within the tree, which will be discussed in turn.

There is increasing evidence that HGT has had - and continues to have - a major role in the adaptation of organisms, especially prokaryotes, to exploiting new environments. Nevertheless, it is often hard to demonstrate HGT, and there is considerable confusion about how to do so. The default hypothesis should remain vertical transmission unless there is good evidence for HGT. The over-hasty assignment of recent bacterial-to-vertebrate gene transfers, solely on the basis of BLAST E-values 31, has been firmly refuted 3233. Such premature HGT assignments have been surveyed and used to provide guidelines for evaluating HGT 3435. Sometimes the evidence is clear-cut, as when adaptive genes are carried on phage, plasmid or transposon. Inconsistent phylogenetic distribution may be evidence for HGT but must be carefully balanced against gene-loss models, recognizing that the two processes are not mutually exclusive. Phylogenetic trees only provide good evidence for HGT when branching is robust and clearly delimited by appropriate outgroups: the HGT must carry a diagnostic molecular evolutionary signal.

One of the best paradigms for investigating recent and ongoing HGT in parasitic prokaryotes is the γ-proteobacterium Vibrio cholerae, which acquired pathogenicity late in recorded history. Free-living Vibrio species are common, harmless aquatic microorganisms. The first recorded cholera pandemic occurred in 1817, the sixth and seventh occurred recently enough to be investigated with modern molecular techniques, and the eighth is probably underway now (see 36 for details). The basic pathogenicity genes ctxAB, which encode cholera toxin, lie within the genome of the filamentous phage CTXφ 37. Other pathogenicity gene 'islands' include the toxin-co-regulated pilus, needed for colonization, and the VSP-1 and VSP-2 islands, which appeared in strains of the seventh pandemic and are suggested to have been integral to that event 38. The recent O139 serotype arose by wholesale replacement of the pre-existing gene cluster encoding lipopolysaccharide O side-chain synthesis, yielding an outer surface with a different architecture, less susceptible to pre-existing immunity 39. Thus, pathogenic V. cholerae continues to adapt to the invasive lifestyle, to a large extent through HGT-mediated acquisition of new capabilities, including, but not limited to, better avoidance of host defenses. Although many of the functions encoded by the genes within pathogenic islands are not understood, their absence from the free-living Vibrio species is good evidence that they have been incorporated, and then conserved, because of a direct or indirect role in enhancing virulence. Even though it is a γ-proteobacterium, the genomic sequence data show that V. cholerae has not (re-)acquired a bact-α₂M gene. At least, not yet.

Our unexpected finding that α₂-macroglobulins, hitherto only known from metazoans, are widely present in eubacterial genomes has provided one of the most clear-cut examples of widespread HGT between extremely divergent bacterial taxa that can be monitored by molecular phylogenetic approaches. We have been able to infer a minimum of 11 independent HGTs for the major yfhM group among 27 sequences tested. Because this group always coexists with a second gene, pbpC, shared evolutionary history means the trees are controlled for topological consistency, so that the assignment of HGT is not in doubt. This work does not address an earlier evolutionary history preceding the link-up of this gene pair.

It is striking that all four deeply diverged groups in the trees include proteobacterial species. This alone clearly indicates that HGT has occurred. Because this is the most heavily researched bacterial taxon and provides most of the sequenced genomes, it is not yet clear whether other taxa will also show multiple independent acquisitions of bact-α₂M and pbpC. Currently, the trees show a minimum of 11 independent HGT events, even if the originating (but unknown) taxon were represented here. A twelfth HGT is indicated if bact-α₂M was originally captured from a metazoan (or vice versa). Extensive gene loss is also likely to have contributed to the phylogenetic distributions in Figure 2, particularly amongst the α-,β-, and γ-proteobacteria, where possession seems the default yet both vertical and horizontal transmission occur. Quite possibly, a cycle of gain-loss-gain has repeatedly occurred as strains adapt between colonization and free-living environments. The role of gene loss cannot be quantified with current data, but this may become possible in the future with more comprehensive genome coverage.

Where pathogenic bacteria and their eukaryotic hosts share related genes that appear to be transferred from one to the other, it is believed that the direction is overwhelmingly from the eukaryote to the bacterium. The failure to find phylogenetic evidence for bacterium-to-vertebrate gene transfers is consistent with this direction 3233. We expect that bact-α₂M was transferred from a metazoan host to a pathogenic bacterium, but this is not yet demonstrable and remains supposition. Given a simple early metazoan, where the germ cells would not be physically isolated from any bacterial infection, one can see how selection could act to fix a bact-α₂M gene transferred in the opposite direction, if bact-α₂M was originally bacterial. This issue may become resolvable in future given much more extensive phylogenetic coverage.

Many bacterial taxa contain a plethora of strains adapted for free-living, symbiotic and pathogenic lifestyles. Examples include the Ralstonia and Anabaena genera adapted to plants, Escherichia and Treponema adapted to animals and pseudomonads adapted to both. Many free-living bacterial strains are also facultative colonizers. This creates some difficulty in cataloguing genes that are adapted to colonizing niches versus free-living: it is rarely certain whether an apparently free-living species never colonizes a higher organism, or is not part of a continuum of strains frequently exchanging lifestyle genes. Given this caveat, we reviewed all the currently completed genomes of bacteria that are not in any way known to have close associations with higher eukaryotes. The available set of Gram-positive bacterial genomes stand out as never possessing a bact-α₂M gene (see below). Only three apparently free-living Gram-negatives (Magnetospirillum, Caulobacter and Thermotoga) have bact-α₂Ms while seven (Aquifex, Chlorobium, Synechocystis, Synechococcus, Prochlorococcus, Nitrosomonas and Geobacter) do not. Thus this crude estimate would suggest that possession of a bact-α₂M gene is associated with colonization, not as a core colonization factor, but as an accessory that enhances fitness for the colonization environment. Further, it may imply that the three 'free-living' species possessing a bact-α₂M gene have undocumented facultative symbiotic capabilities with higher eukaryotes.

The Gram-positive firmicutes and actinobacteria stand out as always lacking bact-α₂M genes (Figure 2). However, certain Gram-positives have found a more direct way to take advantage of α₂M proteins. Pathogenic Streptococcus pyogenes directly co-opt host α₂M for defense against host proteases through the cell-surface proteins GRAB and protein G 4041. As Gram-positive bacteria do not possess an outer membrane, defensive strategies are likely to differ from those of Gram-negatives. Invasive Gram-positives are found to coat themselves in a selected set of host proteins to obstruct host defenses. Streptococcal GRAB mutants that are unable to bind α₂M have attenuated virulence 40. It seems remarkable that prokaryotes have evolved two totally independent strategies to take advantage of α₂M. On the one hand, Gram-positives are able to use the host's own protein, on the other, Gram-negatives have acquired their own gene. The clear implication is that α₂M functionality has a wide and general significance spanning many bacterial taxa.

The lipopolysaccharide (LPS) layer of the outer membrane of Gram-negative bacteria provides a first line of defense. The outer membrane barrier is sufficient to prevent the enzyme lysozyme from lysing Gram-negative bacteria in culture 42. Under attack from host immunity and antimicrobial peptides 43, LPS can be disrupted or stripped away - for example, when released into the circulation, it can lead to septic shock 44 - leaving the peptidoglycan cell wall and inner membrane exposed. There is current interest in antibacterial strategies that endeavor to enhance lysozyme activity by co-administration with agents that disrupt the outer membrane, such as EDTA 42.

The following assumptions lead us to a hypothesis for YfhM bact-α₂M/PBP1C as a periplasmic defense system. First, bact-α₂M and PBP1C form a complex, probably through the carboxy-terminal non-enzymatic domain of PBP1C. Second, the complex resides in the periplasmic space, attached by acylation to the inner membrane. Third, bact-α₂M functions to entrap attacking proteases. Fourth, PBP1C is a transglycosylase that polymerizes glycan chains. Fifth, a periplasmic defense is only needed when the outer membrane has been breached and peptidoglycan is under attack.

The role of the bact-α₂M/PBP1C system is then perceived to be defense at, and repair of, peptidoglycan breaches induced by the host (Figure 4). PBP1C provides 75% of the transglycosylase activity in vitro, but only 3% of peptidoglycan biosynthesis in vivo 25: it is a fast linear transglycosylase, ideal for traversing and repairing a breach. During repair it will, however, be exposed to attacking proteases and may be rapidly rendered dysfunctional. The role of bact-α₂M will be to entrap attacking proteases, protecting PBP1C and other periplasmic proteins such as the high-affinity lysozyme inhibitor Ivy in E. coli 45. In this way, the fate of the invading bacterial cell will depend on the relative balance of the host's attacking forces versus the bacterial defense systems. Under an optimized host attack, such defenses would be rapidly overwhelmed but when (or where) the host is not well prepared, these defenses may serve to prolong colonization.

The yfhM/pbpC gene pair in bacteria not only suggests experimental research strategies, but may have medical potential to help combat pathogenic organisms. Predicted periplasmic location and complexing of bact-α₂M and PBP1C with each other (and any other periplasmic proteins) should be straightforward to investigate biochemically. Elucidation of the host proteases entrapped by bact-α₂Ms should reveal which host defense proteases are targeted at which parasites, leading to enhanced understanding of host defense mechanisms. Bact-α₂M-inhibited proteases should be directly active against pathogen proteins - or else act indirectly as, for example, do the proteases of the complement cascade. PbpC deletions should show increased sensitivity to lysozyme treatments and pbpC/ivy double mutants, yet more so.

The bact-α₂M/PBP1C proteins also provide targets for medical intervention, for example by training host immunity, the administration of anti-bact-α₂M monoclonal antibody or in combination therapies. Antibodies to bact-α₂Ms should act not just by promoting immune clearance but also to block the bact-α₂M activity, so that the host antibacterial proteases are unhindered. This dual effect may provide an enhanced prophylactic efficacy for vaccines that are augmented with extra bact-α₂M protein (probably as an inactive variant) or be directly invoked by targeted anti-bact-α₂M antibody administration for combating acute infection. PBP1C should also be rendered dysfunctional by specific antibodies, perhaps in combination with transglycosylase inhibitors such as the antibiotic moenomycin.

Bacterial α₂Ms were clearly revealed in a search of SWISSALL 46 using BLAST2SRS 47 in which the species names are included in the BLAST output 48. Profile searches as described 49 using the EMBL Bioccelerators 50 supported and extended the findings and were used to retrieve a set of bacterial sequences. Reciprocal searches with bact-α₂M profiles reconfirmed the findings with good E-values (<1.e-25). The sets of proteomes provided by the BLAST server 5152 at the National Center for Biotechnology Information (NCBI) 53 were surveyed to determine the presence or absence of α₂Ms in bacteria and in non-metazoan eukaryotes.

The STRING server 54 is a resource for exploring genome context (for example, identifying groups of genes found in close proximity in many different genomes 21). Queries with bact-α₂Ms from E. coli or other bacteria yielded a recurring result: in most species the bact-α₂Ms cluster consistently with certain other gene families. This behavior is typical of gene sets belonging to the same operon. These families were retrieved and used for further database explorations, alignments and trees. To identify the location of these gene families in other genomes where linkage to bact-α₂Ms is less direct than those presented by STRING, we downloaded genomic database entries from the NCBI, converted the format of these files to EMBL using BioPerl 55, and assessed the location of the genes using Artemis 56. In addition, linkage of these gene families was investigated in organisms not included in STRING using the same method.

Sequences were aligned using Clustal X 1.83 57. Because many sequences are very dissimilar to each other, misaligned regions were to be expected. These were identified using the 'low scoring segments' check and either realigned using the 'realign selected range' option or were hand-edited in SeaView 58. Corrections were assessed by both improvements to conserved hydrophobic columns (indicating structurally important residues) and with the 'low scoring segments' check. Sequences excluded because they were either too divergent to be aligned or may contain sequencing errors included Deinococcus radiodurans and Bacteroides thetaiotamicron.

Preliminary trees were made by neighbor-joining 59 as implemented in Clustal X, excluding gaps and correcting for multiple substitutions with the Kimura PAM model. These initial trees indicated that HGT had occurred, warranting more careful assessment. Alignments were processed with the Gblocks server 60 (for the divergent bact-α₂Ms, the low stringency settings were used). Gblocks heuristically removes poorly conserved excessively divergent segments of alignments with low signal-to-noise ratio in order to enhance the phylogenetic signal 61. Processed alignments were used to derive tree topologies using Bayesian inference of phylogeny as implemented by MrBayes v2.01 27 with maximum-likelihood branch-length estimates provided by PUZZLE 62. MrBayes was used with four heated chains over 250,000 generations, sampling every 20 trees. The likelihoods of these trees were examined to estimate the length of the burn-in phase, and all trees sampled 20,000 generations later than this point were used to create a consensus tree using the 50% majority rule. Both MrBayes and PUZZLE were used with the JTT model of amino-acid substitution 63, assuming the presence of invariant sites and using a gamma distribution approximated by four different rate categories to model rate variation between sites, estimating amino-acid frequencies from the alignment. Trees were displayed and rooted in Njplot 64.

The program GeneTree 28 was used to evaluate the cost of embedding the YfhM sequence tree in a bacterial species tree. To compute the minimum gene number required in the last common ancestor of the given bacterial set, we set the unresolved bacterial affinities to match the YfhM/PBP1C trees (that is, cyanobacteria and spirochetes form a clade, as do bacteroidetes and fusobacteria; within the proteobacteria, the subgroup affinities were allocated to minimize the number of duplications required in the observed trees). Magnetospirillum was excluded from the analysis as its position is not stable in the YfhM and PBP1C trees. Embedding the observed tree topology in this bacterial species tree yielded a reconciled tree requiring six duplication and 29 deletion events.

STRING was used to investigate the expression patterns of genes as detected by DNA microarray. The Stanford Microarray Database (SMD) 2665 was used to verify that these genes were indeed spotted on the arrays used by STRING, and that the spots displayed intensities significantly higher than background levels.