It should be noted here that hipBA loci (found to have a role in the production of 'persister cells' in E. coli 42), as well as restriction-modification (type II) systems, can also be considered as TA systems. Nevertheless, the latter have been extensively identified and characterized elsewhere 4344, and have been excluded from our work. Because of its specific organization, the three-component TA family (ω-ε-ζ) was also excluded from the present study.

TA systems by definition consist of, at least, two genes: the 'dormant guard' role is fulfilled by the presence of a toxic and a protective protein together, although some orphan genes (for which conservation of functionality as such remains unclear) have been reported 3945. Whether or not the TA pairs are encoded by genes forming an operon, the spacer seldom extends beyond 30 nucleotides, and a small overlap (1 to 20 nucleotides in general) is the most common structure. The order of the two cooperating genes is also well conserved, with the antitoxin being upstream (Figure 3), although there is an exception: in higBA loci the toxin is upstream of the antitoxin 23.

TA genes in all prokaryotic species are small. According to Pandey et al. 39, antitoxins are 41 to 206 amino acids long and toxins 31 to 204 amino acids long, antitoxins generally being shorter than their partner toxins (Figure 4). Here too there seems to be an exception: the toxin of the HipBA system is 440 amino acids in length (not shown).

These two features have been used with success as preliminary filters to a biological search for unidentified TA pairs in E. coli 46, but this approach is too permissive to be accurate as an automatic predictor. By adding a third criterion, namely the presence of a conserved functional domain, the selectivity of the method over the input space can be improved. Furthermore, as the knowledge base of TA systems grows, sequence homology can provide further information.

To bypass the mis-annotation of TA genes, which, like many small ORFs, are easily omitted during the annotation process, the tool begins with a naïve ORF prediction. This first step is essential to ensure that the analysis leaves no possible ORF aside. RASTA-Bacteria thus starts by predicting the entire set of valid ORFs in the sequence, defined as the series of triplets occurring between one of the four accepted prokaryotic start codons (NTG), and one of the three stop codons (TGA, TAG, TAA), with no further assumption about the profile of the ORF. In the case of alternative start codons, redundancy is avoided by considering only the longest possible sequence. Although no possible ORFs should be overlooked, existing genomic information (in the case of an annotated genome, the preferred input) should not be ignored. Indeed, even if sometimes flawed, the original annotation can provide RASTA-Bacteria with valuable hints. Therefore, the tool recovers all the annotated features of the sequence, and compares the 'naïve' ORFs to the existing set of genes. If a naïve ORF overlaps an annotated gene (whose 'product' and 'confidence' fields do not display the terms 'unknown', 'putative', or 'hypothetical') by more than a threshold percentage (see parameters), then it is discarded as a spurious ORF. If the considered ORF corresponds to an annotated ORF, its score is rewarded to reflect the annotators' work, that is, the probability that this ORF actually encodes a protein. For reasons of consistency, this process also renames existing ORFs with their common designation.

Once the whole list of candidate ORFs is established, the ORFs undergo a conserved domain search. To achieve this, we use the Reverse PSI-BLAST program (RPSBLAST, part of the standalone blast archive, release 2.2.14 47), which searches a query sequence against a database of pre-computed lookup tables called PSSMs (position specific scoring matrices), originating from the Pfam, Smart, COG, KOG and cd alignment collections (the complete archive of conserved domain PSSMs can be found at 48). These profiles then need to be formatted as a usable database by the formatrpsdb tool 47. For our purposes, we thus built a dedicated TA conserved domains database (TAcddb), compiled from the existing profiles of domains known to belong to toxin and antitoxin genes (Table 1), against which all the sequences in amino acids are searched. Consequently, TA systems with unknown functionally conserved domains are unfortunately liable to be penalized. However, the combining of different criteria tempers the risk of overlooking them, and the database is able to evolve as it can be re-compiled with any new set of PSSMs.

For each candidate, the hits are analyzed to select the most likely in terms of both homology and sequence alignment length. If the candidate ORF exhibits a clear homology, namely a high score and over 80% of a full product domain aligned, but is longer than the corresponding profile, it is scanned for alternative start codons to identify any other 5' end that gives a better profile fit. If this is the case, the ORF is resized to its new coordinates. A short description of the possible domain is stored for subsequent display as a hint to the user for further classification, with an internal hyperlink to the alignment: again, no information is discarded and all the results can be visually assessed. Here, each reference domain used is levelheaded with a coefficient representing its implication in the TA kingdom: those defined by a confirmed TA family have a higher coefficient than domains found in TAs but not exclusive to them (for example, PIN versus VapB domain). This coefficient is computed together with the alignment data to yield the 'domain score'.

The candidates proceed to a size-scoring module. Based on the lengths of 1,378 TA sequences (Figure 4) described following the extensive search by Pandey's team 39, we calculated the probability for length l of a candidate to be that of a toxin or an antitoxin as follows:

P ( L = l ) = n l N MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqadeqadaaakeaacaWGqbGaaiikaiaadYeacqGH9aqpcaWGSbGaaiykaiabg2da9maalaaabaGaamOBamaaBaaaleaacaWGSbaabeaaaOqaaiaad6eaaaaaaa@3C26@

where N = 1,378. We then defined our scoring function by averaging the probability over k neighboring lengths before and after the considered length such that:

f ( l ) = 1 2 k + 1 ∑ i = − k k P ( L = l + i ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqadeqadaaakeaacaWGMbGaaiikaiaadYgacaGGPaGaeyypa0ZaaSaaaeaacaaIXaaabaGaaGOmaiaadUgacqGHRaWkcaaIXaaaamaaqahabaGaamiuaiaacIcacaWGmbGaeyypa0JaamiBaiabgUcaRiaadMgacaGGPaaaleaacaWGPbGaeyypa0JaeyOeI0Iaam4AaaqaaiaadUgaa0GaeyyeIuoaaaa@4945@

This smoothes the curb of probabilities to some extent, as it avoids accidental high or low counts of a given length to be given undue weight with respect to surrounding lengths. Several datasets were created so that the scoring function reflects the different types of organisms: general, archaea, Gram-negative and Gram-positive. The user can thus choose which model to use depending on the species being considered. Similarly, although defining size functions for each of the seven TA families is at first sight appealing, it should be emphasized that automatic classification of TA loci is risky. This is due to diverging homologies: some toxin motifs pair with antitoxin motifs, or more simply toxins/antitoxins of a given family sometimes demonstrate similarity with those of another family 39. Therefore, relying on such specific characteristics for the size criterion evaluation might lead to mis-scoring.

Finally, the method verifies that the ORFs are paired on the strand considered. To do so, the module searches for close neighbors upstream and downstream of the ORF, in agreement with the distance parameter described above: a neighbor is considered close if it lies less than 30 base-pairs away from the extremities of the ORF, and if it overlaps the ORF by less than 20 base-pairs. In practice, both values can be somewhat enlarged, so as to avoid potential loss of candidates in the case of an extended span of the ORF due to alternative start codons. Thus, if an ORF fits these criteria, its score is rewarded. Furthermore, if the neighbor exhibits a TA length and/or a TA domain, the score is given the corresponding bonus. Obviously, this diminishes the chances of fortuitous or clearly non-TA characterized operons finding themselves among the top candidates.

S. meliloti is a Gram-negative alpha-proteobacterium studied in our laboratory that is found both free-living in soil and in a symbiotic interaction with alfalfa where it forms root nodules. Its genome is made up of a 3.65 Mb circular chromosome and two essential megaplasmids, pSymA (1.35 Mb) and pSymB (1.68 Mb), all of them being GC rich (62.2% global) 50. These features (large and tripartite genome with recently acquired plasmid, free and symbiotic life ability) make S. meliloti an interesting model for the validation of RASTA-bacteria. In the 2005 search by Pandey et al. 39, 12 TA systems (2 relBE-like, 3 higBA-like, and 7 vapBC-like) were identified, but only the chromosome was considered. We analyzed all three replicons with RASTA-Bacteria, as they are all constituents of the complete genome. Of the 12 systems identified by Pandey et al., 11 were positively discriminated by RASTA, including the ntrPR operon, which was recently shown to function as a TA system 51, demonstrating the good accuracy of our software. The 12th one (higBA-2, GI15965582-15965583) was only poorly rewarded by the method described here; indeed, none of the TA domain profiles corresponding to its described classification (nor others) were matched by the members of this TA pair, which furthermore do not fit the size and distance criteria. Further sequence analysis did reveal similarity with a putative addiction module killer protein for the amino-terminal half of gene 15965582, but a second conserved domain in its carboxy-terminal half, as well as the conserved domains ('ABC transporter') found in its reported partner, are rather contradictory with the fact that this pair might comprise a valid TA system. There is thus no concrete evidence that enables us to confirm this hypothesis.

We found 14 additional putative TA loci on the chromosome (bringing the population to 25 for this replicon), 17 loci on pSymA and 11 on pSymB (Figure 5a). Hence, our approach predicts a total of 53 TA loci in the complete genome of S. meliloti, including 95 genes of which 18 are newly identified. Their distribution across the various replicons seems random, although there is an apparent alternation of rich and poor areas, in particular in the megaplasmids (Figure 6). Similarly, they are remarkably evenly distributed between lagging and leading strands (Figure 5c). Relative to the sizes of the replicons, megaplasmid A, suspected to have been acquired more recently in the genome, contains twice as many TA loci as the other replicons (Figure 5b). Interestingly, the genetic organizations are diverse, although pairs remain the most frequent (71.5 %): 12 genes in 4 triplets, 68 genes in 34 pairs and 15 solitary genes (12 encode antitoxins and 3 encode toxins, one of them being the chromosomal relE; Figure 4d).

The classification of candidates into families according to sequence homology alone is a tedious task. Nevertheless, it seems the two major families are vapBC, consistent with the findings of Pandey et al. 39, and parDE. No ccdAB locus was found, but the results indicate there may be parDE and phd/doc members (distributed on all three replicons) among the candidates, as well as one mazEF pair, situated on plasmid B.

Our tool proved to be efficient and fast for the bacterium S. meliloti, which was used for its design. The effectiveness of RASTA-Bacteria for other sequences was first assessed using 14 prokaryotes previously studied by Pandey et al. 39 (Table 2): three gamma-proteobacteria (E. coli as an AT-rich generic model, Coxiella burnetii as an obligate host-associated organism and Pseudomonas aeruginosa as a free living, GC-rich bacterium); two alpha-proteobacteria (Bradyrhizobium japonicum, which has a large chromosome with significant horizontal rearrangements, and Agrobacterium tumefaciens, which has both circular and linear chromosomes); the genome with the largest predicted set of TA loci (Nitrosomonas europeae 39); free-living firmicutes (Lactococcus lactis, Bacillus); one epsilon-proteobacteria (Campylobacter jejuni); three obligate host-associated organisms (Rickettsia prowazekii, Buchnera aphidicola, and Mycobacterium leprae for which Pandey et al. did not find any TA loci); and members of the Aquificae and Thermatogae extreme-life phylum (Thermotoga maritima, Aquifex aeolicus). Also, to assess the range of applicability of our tool, we tested the archaeum Sulfolobus tokodaii. The result files for all these species as well as for S. meliloti are available in the 'Pre-computed Data' section of our website 41.

RASTA-bacteria identified all TA loci previously predicted by Pandey et al. except for one locus in S. meliloti (see above) and one higBA system in B. japonicum, which was not retained because the confidence score was too low (although there are conserved domains, they are ambiguous and were not included in TAcddb). The absence of detectable TA genes from the three obligate host-associated organisms tested (R. prowazekii, B. aphidicola, M. leprae) was confirmed, as was the presence of a single TA locus in Bacillus sp. Our tool was more sensitive than the previously used method: in all other tested genomes, RASTA-Bacteria identified a large number of new candidate loci. This was largely due to detection of potential members of the higBA, relBE, hipBA families and especially the vapBC family. For example, even in the case of the well-documented model E. coli, RASTA-Bacteria predicts at least four new TA pairs with high confidence (yfeD/yfeC, yafN/yafO, ygjN/ygjM and sohA/yahV). In addition, the ygiT/b3022, ydcQ/yncN and ydaS/ydaT loci have at least one member with a conserved domain commonly found in antitoxins, and ranked higher than published TA genes. Finally, YbaQ demonstrates near perfect identity with the profile corresponding to VapB antitoxins, but has no physically close partner, so it most likely is a solitary antitoxin, the first such to be reported in E. coli.

Ten previously undescribed TA systems were identified in the four replicons of A. tumefaciens (Table 2), although only the two chromosomes were previously studied. RASTA-Bacteria confirmed the 14 systems previously reported and identified 5 additional (orphan) loci on the circular chromosome, 1 full pair and 1 orphan gene on the linear chromosome, and 2 TA systems on plasmid AT. It revealed plasmid Ti carries no plasmid addiction systems, although it does have a gene resembling hipA (Atu6158, GI|17939291). However, this candidate is substantially shorter than its reference, such that it is unlikely to be functional, and it is almost 60 kb away from any possible hipB candidate.

We also assessed the sensitivity of our tool by examining genomes containing many TA loci, including that of N. europaea, reported to have no less than 45 TA loci, representing 88 genes. The RASTA-Bacteria scan of the genome of N. europaea yielded high confidence scores for 76 of these previously identified genes (86%), a confidence score between 50% and 70% for 11 (12.5%) and an unranked score for 1. It identified 11 additional TA loci on the N. europaea chromosome, if the hipBA locus is taken into account. Three are clearly vapBC pairs, although one is made of two relatively short and possibly disrupted genes, raising doubt about whether this pair is functional. The NE2103/NE2104 pair gave an intermediate confidence score, but has characteristics consistent with it being a TA system. NE1375/NE1376 may well define a new MazEF-like system. Finally, three orphan vapB and two orphan higA genes were found: it would be interesting to determine whether they are silent relics of ancient systems or are still active and responsible for a function. Remarkably, all these newly identified loci map in the same regions as the previously discovered systems, reinforcing the observation that TA loci in N. europaea cluster in particular regions of the genome.

We also applied our tool to organisms where no TA loci had been found previously, including L. lactis, in which we predict ten possible TA loci, eight of which consist of an orphan gene containing a region encoding the same HTH_DNA-binding (for helix-turn-helix) profile.

Finally, the archaeum with the most TA loci was S. tokodaii, with 32 TA loci 39. RASTA-Bacteria confirmed 52 of the 61 genes at these 32 TA loci (3 singletons): the STS188/ST1628 and ST2136/37 pairs gave low scores because of an extreme overlap or because of an alternative start codon causing a bias in the size scoring process. The results for five other genes cannot be interpreted with certainty, but observations in other organisms where orphan TA genes do not seem uncommon suggests that some loci predicted to be pairs might in fact belong to the single-gene loci category. Nevertheless, four additional TA loci were identified, two of them being standard pairs of the VapBC family. The TA loci are unevenly distributed through the chromosome: two regions of approximately 240 and 440 kb seem to be devoid of TA loci, and the loci appear to be clustered in particular regions (data not shown).

We performed a second round of analyses of newly selected, mostly recently published genomes (Table 3). At least three genomes from each phylogenetic branch and lifestyle were examined, except where only less complete genome sequences were available.