To identify all prokaryotic Argonaute homologs, we performed a PSI-BLAST search against the NCBI non-redundant protein sequence database using the PIWI domain (the most highly conserved domain in the Argonaute family proteins) sequence from the Thermus thermophilus HB27 pAgo (TT_P0026, pdb: 3DLB containing; PIWI domain sequences in amino acid positions 415–685). The search was run until convergence (after the 3^rd iteration) and resulted in the identification of 100 sequences, some of which were fragmented or truncated proteins; additional searches started with some of the detected proteins showed that this sequence set represents the full complement of PIWI-domain proteins (pAgo) encoded in currently available prokaryotic genomes. For more detailed analysis, we selected 85 sequences from 80 genomes (the genomes of the bacteria Parvularcula bermudensis HTCC2503 and Halorubrum lacusprofundi ATCC 49239 encode three pAgo proteins each, and the genome of Acidobacterium capsulatum ATCC 51196 encodes two pAgos) (see Additional File 1).

Comparative sequence analysis of the identified pAgos showed that the conserved, alignable region shared by all these sequences approximately corresponded to the L2, Mid and PIWI domains, as inferred from the crystal structures of the pAgos from the hyperthermophilic bacterium Aquifex aeolicus (AaAgo; pdb: 1YVU35), Thermus thermophilus (TtAgo; pdb 3DLB3136), as well as the archaea Pyrococcus furiosus (PfAgo; pdb 1Z2533) and Archaeoglobus fulgidus (AfAgo; pdb: 1W9H37) (Figure 1; see also Additional File 2). In addition to the three conserved domains, both pAgos whose structures have been solved contain an N-terminal domain, an L1 domain, and a PAZ domain that, as in eukaryotic Argonaute, binds the 3' end of a siRNA guide and positions the middle of siRNA guide bound to the target mRNA in the catalytic pocket of the PIWI nuclease 323334. However, among the identified pAgos, more than half lack the N-terminal, L1 and PAZ domains although several instead contain an N-terminal fusion with predicted nucleases of the Sir2 family (Figure 1 and see details below).

The PIWI domain of Argonaute proteins belongs to the RNAse H fold and shares the divalent cation-binding motif DDE (aspartate, aspartate, glutamate) involved in catalysis with many other nucleases that cleave both RNA and DNA http://scop.mrc-lmb.cam.ac.uk/scop/data/scop.b.d.hh.html38. The two aspartates are essential for the slicer activity of eukaryotic Argonautes whereas the third catalytic residue can be glutamate, histidine, aspartate or lysine 34. Another conserved feature of Argonautes is the presence of a basic residue (in most instances, arginine) that is located in the catalytic site 35. Some eukaryotic Argonaute proteins appear to be inactive (hence denoted non-slicer Argonautes), especially, in nematodes 34. Apparently, non-slicer Argonautes interfere with translation through binding rather than cleavage of mRNA 39. Examination of the multiple alignment of the catalytic cores of prokaryotic PIWI domains strongly suggests that the majority of these domains are inactivated as indicated by the replacement of two or all three acidic residues required for catalysis; this apparent abrogation of the nuclease activity is particularly common in those pAgo proteins that lack the PAZ domain (Figure 2).

The AfAgo protein, which does not contain a PAZ domain, also lacks the catalytic aspartates but has been shown to bind dsRNA 3240. Structural analysis of AfAgo complexed with a siRNA-like duplex showed that in this protein a Cd²⁺ ion bound to the carboxy-terminal carboxylate and several amino acid residues in the middle (MID) domain are involved in the recognition of the unpaired 5' nucleotide of siRNA 3240. In contrast, a structural and biochemical study of AaAgo, which contains the PAZ domain and the conserved catalytic residues, showed that this protein is an active RNAse H with a preference for a DNA/RNA hybrid as a substrate, suggesting that some pAgos employ small guide DNA molecules to cleave mRNA 35. The detailed study of the Thermus thermophilus pAgo corroborated the findings on AaAgo by revealing the details of interactions with the 5'-phosphorylated 21-base DNA guide strand and the DNA-guided RNA cleavage by this protein 3136.

We constructed a phylogenetic tree of the PIWI domains from all the detected pAgos (after excluding sequences that were fragmented or truncated due to poor annotation) and a subset of eukaryotic Argonautes (Figure 3). The majority of the PIWI domains from pAgos that lack a PAZ domain form a distinct clade although a few of these short forms cluster within the other clade that consists mostly of full-size, PAZ-containing pAgos. Within the latter clade, the short proteins do not form a distinct group (Figure 3), suggesting the N-terminal part of pAgo was lost independently in several lineages. Consistent with the similarity of domain architectures and with the results of previous analyses 29, eukaryotic Argonautes belong to a well-supported clade together with a distinct subset of archaeal pAgos; in particular the structurally characterized Pyrococcus furiosus protein, that is considered to be the model for Argonaute functioning in eukaryotes 33. Other archaeal proteins are scattered in the tree, suggesting multiple horizontal gene transfers (HGT) between bacteria and archaea (Figure 3). Despite the existence of several small lineage-specific groups (alpha proteobacteria, gamma proteobacteria, bacteroides and cyanobacteria), the results of our phylogenetic analysis strongly suggest that pAgo genes mostly disseminated by HGT; the patchy distribution of these genes makes it unlikely that they perform indispensible functions in any bacteria or archaea (Figure 3).

We further examined the genomic context of the pAgo genes; analysis of genomic context has been established as a powerful approach for prediction of the biological functions of prokaryotic genes using the "guilt by association" principle 414243. In many cases, these genes form potential operons with a variety of genes encoding uncharacterized proteins (neighbor genes were predicted to be encoded in a potential operon with pAgos if they were located upstream or downstream of the respective pAgo gene on the same DNA strand and if the intergenic distances in such an array of co-directional genes were shorter than 100 nt; see Additional File 1). We performed an in-depth analysis of the sequences of the proteins encoded in the genes co-localized with pAgos using PSI-BLAST, HHpred and CDD search (see Methods). This analysis resulted in the identification of four protein families that are predicted to be co-expressed and thus functionally linked with the pAgos.

The first family is typified by the xccb100_3097 protein from Xanthomonas campestris B100, the only protein among the pAgo neighbors that, in the current sequence databases, is annotated as a "putative Sir2-family regulator" rather than a "hypothetical protein". Indeed, CDD search detected statistically significant similarity between the N-terminal domain of this protein and the SIR2 domain (cl00195, E-value = 5 × 10^-5). The Sir2 proteins, also known as sirtuins, are a well characterized family of NAD⁺-dependent histone deacetylases in eukaryotes where they play key roles in the regulation of gene silencing, DNA repair, metabolic enzymes, and life span 44454647. Representatives of this family also have been identified in both bacteria and archaea, and the structures of several Sir2 family proteins have been solved 4849. So far all experimentally characterized Sir2 family proteins have been shown to possess protein deacetylase activity 48. However, a distinct family of prokaryotic sirtuins is associated with DNA-pumping ATPases of the FtsK-HerA family 50. Because in numerous other instances the FtsK-like ATPases are associated with known nucleases, both functionally and in terms of the operon structure, it was hypothesized that this particular family of sirtuins could function as nucleases, and a conserved DxH motif was implicated in the predicted nuclease activity 50. The majority of the xccb100_3097-like proteins contain only one of these residues, namely, the aspartate in the loop between strand 7 and helix 11 (according to the crystal structure of human Sirt2 histone deacetylase, pdb: 1j8f51) but instead have an additional aspartate in the strand 2 that is conserved within this family(Figure 4A). Similarly to Sir2 proteins associated with the FtsK-like ATPases, xccb100_3097-like proteins lack the Zn-ribbon insert between strand 4 and helix 10 that is characteristic of most sirtuins, but retain all NAD⁺-binding site residues, suggesting that these proteins are active enzymes (Figure 4A).

For the C-terminal domain of xccb100_3097, we failed to detect any statistically significant similarities to known domains using CDD search or HHpred. However, PSI-BLAST search with the xccb100_3097 used as a query revealed many homologs with similar domain architectures, all of which are associated with pAgos in putative operons; moreover, several multidomain proteins (eg. GIs: 91783256, 218130589, 229435559) comprise fusions of xccb100_3097-like and PIWI domains (see the alignment of this domain in Additional File 3).

The second family of PIWI-associated proteins is typified by the mlr6203 (GI: 13475182) protein from Mesorhizobium loti. The HHpred search convincingly shows that the N-terminal domain of these proteins belongs to the Mrr family of restriction endonucleases, with the hallmark (D/E)-(D/E)XK active site 5253 (for example, the best hit is to pdb: 2ost, homing endonuclease from Synechocystis sp., E-value = 0.04; followed by a hit to pfam04471, Restriction endonuclease, E-value = 0.04). All experimentally characterized superfamily representatives are site-specific endonucleases that cleave dsDNA and possess an enormous variety of recognition sites 525354. The active site residues are conserved in all mlr6203 homologs (Figure 4B), so this domain probably is an active DNA endonuclease. As with the xccb100_3097 family proteins, no similarity to the C-terminal domain of the mlr6203 was detected in CDD and HHpred searches. However, the PSI-BLAST search identified 17 homologous proteins with the same domain architecture and predicted operon organization (see Additional File 1).

A typical representative of the third family is RHECIAT_PB0000019 (GI: 190894000) from Rhizobium etli. This protein contains an N-terminal TIR domain that was easily detected by HHpred (the best hit is to pdb: 2js7, TIR domain of myeloid differentiation primary response protein MYD88 from human, E-value of 1.1 × 10^-30). The TIR domain mediates protein-protein interactions and belongs to the STIR superfamily that includes mostly eukaryotic proteins involved in diverse signaling pathways as well as a variety of poorly characterized multidomain proteins from bacteria and archaea with large genomes (that also have been implicated in transcription regulation and signaling 555657). Notably, TIR domains play important roles in disease and stress resistance in plants 58. Similarly, in mammals, TIR-domains are key components of the immune system-based antimicrobial and antiviral response, and the programmed cell death (PCD) system 5960. Analysis of domain architectures led to the hypothesis that prokaryotic TIR-domain proteins also could be involved in PCD 61. All closely related homologs of the RHECIAT_PB0000019 protein contain the TIR domain (see Additional File 3), whereas several proteins in this family (e.g. GI: 162145848) also contain an additional N-terminal domain that belongs to the PD-(D/E)XK nuclease superfamily (a vast assemblage of nucleases that includes, among others, the restriction endonucleases) with all catalytic residues typically conserved (Figure 4B). The C-terminal domain of these proteins is not similar to any known domain, but does show a weak sequence similarity (with statistical significance difficult to demonstrate) to the C-terminal domain of the mlr6203-like family. Considering similar sizes of the corresponding domains in both families and, most importantly, the genomic association with predicted nucleases and pAgos, we strongly suspect that these domains are homologous; examination of their multiple alignment indeed shows several distinct, conserved motifs (see Additional File 3). The predicted secondary structure indicates that this is a globular domain, however, the pattern of amino acid residue conservation does not seem to suggest an enzymatic function. Given that the proteins containing this domain are found exclusively in the same neighborhoods with pAgos that lack the PAZ domain, it is tempting to speculate that this uncharacterized domain is functionally analogous to the PAZ domain, that is, involved in binding a guide nucleic acid molecule (hereinafter we refer to this domain as APAZ, after Analog of PAZ).

The fourth family of pAgo-associated proteins is linked to full-size, PAZ-domain-containing Argonaute homologs and can be typified by the protein PTH_0722 (GI: 147677057) from Pelotomaculum thermopropionicum. This protein contains a C-terminal domain that belongs to the PD-(D/E)XK nuclease superfamily (HHPred detects similarity to SfsA: Sugar fermentation stimulation protein, which contains a PD-(D/E)XK nuclease domain, with E-value = 0.022) and contains all the catalytic residues (Figure 4B); this putative nuclease is clearly distinct from and only very distantly related to the restriction endonuclease domain of the mlr6203-like family proteins. The N-terminal domain of this protein does not show similarity to any characterized domains, has a predicted predominantly α-helical structure and is present only in close homologs of PTH_0722 (see Additional File 4). In the GobsU_24486 protein of Gemmata obscuriglobus, the nuclease domain is replaced by the apparently functionally unrelated SEFIR domain of the STIR superfamily, that is only distantly related to the TIR domain, but is also involved in various signaling pathways 57.

Several other genomic neighbors of pAgos are worth mentioning (Figure 3). Two genes that encode PAZ-domain-containing but, apparently, inactivated pAgos (in the bacteria Pedobacter heparinus and Spirosoma lingual) are associated with predicted Sir2 family nucleases (Figure 4A). Furthermore, three long forms of pAgos (one inactivated, in the bacterium Dehalococcoides sp, and two apparently active ones in Microcystis aeruginosa and Clostridium bartletti) are associated with PD-(D/E)XK nucleases of a distinct subfamily related to Cas4 (COG1468), which is mostly represented within CASS 22. Most conspicuously, as noticed previously, in the archaeon Methanopyrus kandleri, the pAgo is encoded within an operon that otherwise encodes components of the CASS 22.

A potentially important pattern revealed by this analysis of the genomic context of prokaryotic PIWI-domain proteins is that, almost without exception, pAgos with an apparently inactivated catalytic PIWI domain are associated with a predicted nuclease in a putative operon (Figures 2, 3 and see Additional File 1). This observation suggests the possibility of functional complementarity between the nuclease activity of PIWI domains of pAgos and other nucleases, in particular, homologs of restriction endonucleases (see discussion below).

Considering (i) the central role of Argonaute proteins in siRNA-based antiviral response in eukaryotes, (ii) the contextual links between pAgos and nucleases (in particular, restriction endonucleases) that are involved in phage/plasmid defense in prokaryotes, and (iii) links to the TIR domain that also functions in antimicrobial response in eukaryotes, it is tempting to hypothesize that an important if not the principal function of the pAgos has to do with phage defense (or, more generally, defense against viruses, plasmids, and other mobile elements). Phage defense systems in prokaryotes are notably prone to HGT (the CASS being the prime showcase), and phylogenetic analysis of the pAgos clearly indicates that HGT shapes the evolution of pAgo-encoding genes as well (Figure 3). In addition, phage defense systems are often encoded in genomic islands 62. Therefore we sought to statistically test the hypothesis that pAgo genes are non-randomly associated with known phage resistance genes in prokaryotic genomes. To this end, we identified 4 classes of phage defense systems (some of which are also involved in a broader range of stress response reactions) in a representative set of 45 prokaryotic genomes and computed the fractions of these genes throughout the genomes and in the vicinity of pAgo genes (see Methods for details). The Fisher Omnibus test 6364 reveals a statistically highly significant enrichment of the pAgo genomic neighborhoods (see Methods for details) for different combinations of 4 classes of phage defense genes used as a target set (Table 1). As a control, we performed the same analysis for pAgo genes and typical components of the bacterial mobilome including transposases and various phage-derived genes; no statistically significant association was found between pAgos and these mobile genes (p = 0.63; see Additional Files 5 and 6).

Several convergent lines of evidence point to defense against invading mobile elements as the primary function of pAgos. (1). The analogy to eukaryotic Argonautes many of which are dedicated to the defense against viruses and transposable elements. (2). The guide-DNA-dependent nuclease activity of AaAgo and TtAgo. (3). Extensive HGT of pAgos which is best compatible with a stress-response related function. (4). Preferential location of pAgo genes in genomic neighborhoods significantly enriched in known phage-defense genes. (5). Co-localization of PIWI-domain protein genes with genes encoding other (predicted) nucleases. (6). The near perfect complementarity between the predicted nuclease and guide-binding activities of pAgos and co-localization with other putative nucleases: the inactivated pAgos that lack the PAZ domain are associated with genes encoding predicted nucleases whereas the apparently active, PAZ-containing pAgos are not (Figure 3). The latter observation suggests that pAgos function within nuclease complexes, in some cases as their catalytic subunits, and in other cases, as structural subunits interacting with the actual nucleases.

Additional functional clues allow us to tentatively propose more specific mechanisms for the functions of pAgos in the defense of prokaryotes against mobile elements (Figure 5). In eukaryotic Argonautes, the PAZ domain binds the small guide RNA and facilitates its hybridization with the complementary region of the target mRNA. Most of the pAgos that are predicted to be active nucleases also contain PAZ domains suggesting that they function via a similar mechanism, in agreement with the experimental data for AaAgo and TtAgo 31366364. The apparently inactivated pAgos lack PAZ domains but are co-localized with genes encoding predicted nucleases and the APAZ domain (Figure 1, 2). The (so far) exclusive presence of the APAZ domain within predicted operons encoding inactivated pAgos makes us speculate that, similary to PAZ domains, the APAZ domains bind guide molecules and target the putative nuclease complex to phage nucleic acids.

The PD-(D/E)XK superfamily nucleases, to which the predicted nucleases associated with the majority of pAgos are homologous, so far have been shown to cleave exclusively dsDNA. Thus, it seems most likely that the predicted pAgo-based defense systems directly target invader dsDNA genomes rather than mRNAs (Figure 5). On the other hand, as stated above, in vitro analyses have revealed that AaAgo and TtAgo are most active as DNA-guided ribonuclease, suggesting that RNA may be a target as well [REFS 3536]. The guide molecule could be either a small RNA (with the implication that the respective nuclease cleaves a RNA-DNA hybrid) or a small DNA as suggested by the study of AaAgo 6364 and TtAgo 3136.

The proposed model for the pAgo-based phage defense shows functional analogies to both CASS and the eukaryotic RNAi (Figure 5). Given the phylogenetic affinity of a distinct family of apparently active archaeal pAgos and eukaryotic Argonautes (Figure 3), this hypothetical defense system is the probable evolutionary progenitor of the eukaryotic RNAi. The spread of RNA viruses in eukaryotes that was accompanied by the displacement of the majority of DNA viruses 65 could have been the driving force behind the switch of the specificity of this defense system from DNA to RNA.

All analyzed sequences were from the non-redundant protein sequence database at the NCBI. Database searches were performed using PSI-BLAST 66, typically, with the inclusion threshold E = 0.01, and no composition-based statistics or low complexity filtering, or the HH search program available through the HHpred server 67. Multiple alignments of protein sequences were constructed by combining the results obtained with the PROMALS program 68 and the MUSCLE program 69, followed by a minimal manual correction on the basis of local alignments obtained using PSI-BLAST 66. Protein secondary structure was predicted using the PSIPRED program 70.

Maximum likelihood (ML) phylogenetic trees were constructed from the alignment of PIWI domain region (only positions with less than 30% gaps were used for reconstruction – 258 altogether), by using the MOLPHY program 71 with the JTT substitution matrix to perform local rearrangement of an original Fitch tree 72. The MOLPHY program was also used to compute RELL bootstrap values.

Only 45 completely sequenced genomes were used for this analysis; the complete genome information was obtained from FTP of RefSeq database (ftp://ftp.ncbi.nih.gov/genomes/Bacteria/;73). Proteins in these genomes were assigned to COGs using a modified COGNITOR program 74. The target sets of phages defense proteins were obtained from the following sources: restriction-modification (RM) systems related protein from REBASE 75; abortive infection (ABI) related genes from the Chopin et al. review 76; CRISPR systems related genes from 22 and toxin-antitoxin related genes from 77. Proteins of the RM and ABI systems were assigned to COG as indicated above, and for other systems, COG numbers have been already reported in the aforementioned papers (see the complete list of these COGs in Additional File 5).

In each genome, we identified the genes that belong to each of the aforementioned four well-characterized phage defense systems and computed the gene counts for each system in the entire genome (K phage defense genes in a genome containing N genes) as well as within each of windows of size ± w = 10 surrounding each pAgo gene (k genes in window). For each window, the probability to observe ≥k phage defense genes by chance was approximated using the binomial distribution:

The results obtained for multiple windows were combined using the Bailey and Gribskov's variant of the Fisher Omnibus test 63.