All the available complete plasmids, phages and chromosome sequences were downloaded from NCBI ftp site at (http://www.ncbi.nlm.nih.gov/Ftp/, as on February the 1^st 2010). Concerning plasmids, we focused our attention only on those longer than 3 kb and harboring at least 2 ORFs, in order to be able to detect δ* and differences in GC% content among all the genes. This allowed to assemble a dataset of 1853 plasmids for a total of 128.569 ORFs. The complete list of plasmids analyzed in this work (together with other information such as their size, their accession codes etc.) is available as Additional file 1.

Classically, atypical genes detection has been pursued either by i) phylogenetic methods based on sequence alignment) and/or ii) by composition-based methods that involve alignment-free features such as GC% content, synonymous codon usage or the frequencies of overlapping short oligomers 16 . Several in silico based methods have been conceived in the past few years to identify foreign genes that were recently acquired by chromosomes 1 17 18 19 20 21 22 23 24 25 . However, it is not possible to discriminate between plasmids atypical or native genes using phylogeny-oriented methods since a set of homologous and universally shared sequences (necessary to build a reference phylogeny) are often unavailable. Furthermore, the absence of a universally shared "core" of genes (i.e. likely not subject to extensive HGT) when analysing plasmids sequences does not allow the use of classical statistics methods (such as Markov model-based approaches), which require the presence of a set of native genes in order to identify those ORFs that have a signifcant different composition. Hence, in this work, we have developed a computational strategy (Figure 1) that combines two compositional measures (GC content difference and δ*) in order to identify putative PAGs within a dataset of 1853 plasmids. Briefly, GC content is calculated as (G+C)/(A+T+G+C), where G, C, A and T is the number of guanines, cytosine, adenines and thymines, respectively. Conversely, δ* between two sequences a and b (from different organisms or from different regions of the same genome) is calculated as δ^∗(a,b) = (1/16)∑_XY|ρ*_XY(a)−ρ*_XY(b)| where the sum extends over all possible XY dinucleotides and where ρ_XY = f_XY /f_Xf_Y, with f_XY representing the frequency of the dinucleotide XY in the genome.

Hence, for each of the 128.569 ORFs of the 1853 plasmids (Figure 1a), we estimated both the δ* and the GC content difference (ΔGC) in respect to the corresponding source plasmid (Figure 1b and 1c). Since the distributions of these values did not follow a normal distribution (according to Kolmogorov-Smirnov test with a p-value threshold of 0.05), we used a distribution-independent procedure to evaluate the probability of each point of these distributions. performing a bootstrap sampling (Figure 1d and 1e) of all the obtained values. In other words, a probability was assigned to each of the values [e.g. P(A)] of these two distributions computing it as P(A) = n(A)/N, where n(A) is the number of times in which the observed value (of ΔGC and δ*, respectively) was greater than the other (128.569) values after N samplings. By doing so, two distinct p-values (Figure 1f and 1g) were associated to each sequence of the dataset: the first accounting for the probability of a gene to be atypical in terms of δ^∗ and the other accounting for the probability of a gene to be atypical in terms of ΔGC. Further on, these two distinct p-values were integrated in a single one according to the Fisher method 26 . In its basic form, the Fisher method is used to combine the results from two (or more) tests bearing upon the same overall null hypothesis. In other words, Fisher's method combines p-values into one test statistic (X²) using the formula: X² = -2∑log_e(p_i), where pi is the p-value for the i ^th hypothesis test. Accordingly, (Figure 1h) only those sequences identified at a confidence interval (CI) greater than 95% were considered PAGs. Moreover, in order to explore different CI thresholds, we also collected gene sets that were identified as atypical with lower confidence values (i.e. 70%, 80% and 90%). The whole pipeline has been implemented in Perl codes and is available upon request. As it might be expected, lower CI thresholds allowed the assembly of larger PAGs dataset, ranging from 14731 (with a CI of 90%) to almost 40000 (with a CI of 70%) (see Additional file 1).

In order to identify the most likely source molecule of identified PAGS we developed a similarity-oriented computational pipeline according to which each of the identified PAG was used as a query for a BLAST 27 search against three different databases, each of which embedding 30000 sequences retrieved from NCBI plasmid, phage and chromosome databases (see Methods), respectively. For each of the BLAST searches, only the best BLAST hit was considered, in order to reduce any possible bias due to the presence of closely related sequences in the database that would falsely increase the number of homologs for a given ORF. This strategy was repeated 1000 times for each PAG and, for each of the 1000 runs, new plasmid, chromosome and phage databases were assembled, randomly sampling 30000 sequences from the NCBI databases. Finally, the putative source molecule was identified according to the database (plasmid, phage or chromosome) that produced the highest number of best hits after 1000 BLAST probings.

All statistical tests were performed with the R package 28 . All other statistical analyses were performed using in-house developed Perl scripts.

As previously pointed out, PAGs may derive from homologous recombination or homology-facilitated illegitimate recombination between a given plasmid and/or other informative molecules, that is phages, chromosomes or other plasmids. Hence, we aimed at identifying the putative source molecule of PAGs. To this purpose, we have developed a computational approach (see Methods) that, on the basis of the number of orthologs present in different (and randomly assembled) plasmid, chromosome and phage databases allows to assign the most likely source molecule to a set of sequences (in our case the 8065 PAGs). It must be stated clear that this kind of strategy has only an explorative purpose and might be strongly influenced by the present content of public databases that, undoubtedly, represents just a glimpse of the real biodiversity present in nature. For this reason, the reliability of the developed approach was firstly revealed by a test on a set of 1000 likely chromosomal native sequences retrieved by Cortez et al. 10 from a set of 119 bacterial and archaeal chromosomes. Indeed, results of this preliminar screening showed that almost 90% of the probed sequences were correctly annotated, i.e. resulted to possess a putative chromosome origin. Furthermore, we sought to test the implemented strategy on a set of sequences of plasmid and phage origin. However, as already pointed out, in these cases a set of "core" sequences is very difficult (if not impossible) to be retrieved. Hence, the previously described pipeline was applied on two distinct randomly assembled datasets of viral and plasmids sequences, embedding 1000 sequences each. Results similar to those obtained with (likely) chromosome native sequences were obtained (84% and 82% of "correct" identifications in the case of plasmdis and phages, respectively) thus suggesting that, in most cases, the implemented strategy is able to detect the correct source molecule of a given sequence.

Applying the described strategy to our PAGs dataset revealed that, as shown in Figure 7, most of the identified PAGs (almost 75%) are of likely plasmid origin. On the contrary, chromosome and phage sequences appear to be much rarer, being represented by 5.7% and 4.8% of PAGs, respectively. Remarkably, almost 13% of all the probed PAGs (1087), did not retrieve a clear corresponding match in the plasmid, chromosome and phage databases and, for this reason, their origin remained "undetermined" (Figure 7). Finally, less than 1% of the PAGs did not retrieve any match in any of the databases during our BLAST probings and were labelled as "Not Found" in Figure 7. Again, also sequences retrieved with lower CI thresholds showed the same hypothetical origin, with likely plasmid orgin sequences being largely over-represented (73.8%, 72.9% and 73.5% for 70%, 80% and 90% CI thresholds, respectively). Taken together these results indicate that most of the PAGs are likely to appear only in plasmids rather than begin shared among different types of informative molecules. We speculate here that most of retrieved PAGs likely derive from plasmid-plasmid gene exchanges and that, vice versa, integrations following virus-plasmid and/or chromosome-plasmid gene exchange appear to be less frequent. Remarkably, this finding fits with previous analyses 40 on the structure of the overall DNA-exchange network and that suggested that DNA families are mostly exchanged among the same type of DNA carriers (i.e. plasmids, phages or chromosomes). Our analyses indicate that this might hold true (at least) in the case of plasmid molecules.

Chromosomal atypical genes are often organized in (more or less) tight clusters. These chromosomal regions, named genomic islands 10 , embeds, for example, genes involved in pilus and fimbriae formation, lipopolysaccharide biosynthesis or virulence and have been shown to have played a crucial role in microbial evolution 41 . Similarly, PAGs might either stand alone or be embedded in (more or less compact) clusters on their source molecule representing, for example, successfully integrated transposons and/or integrons. Hence, we investigated this issue and evaluated if retrieved PAGs lied in clusters or were scattered throughout the corresponding plasmid. We estimated gene clustering at three different gene distance thresholds, that is 100, 200 and 300 bp. Results shown below refer to genes that are not separated by more than 200 bp (results for 100 and 300 bp are provided as Additional file 8, together with clusters distribution retrieved at different CI thresholds). Overall we found that 1653 PAGs (on a total of 8065) are embedded in 677 clusters of different size (Table 1 and Figure 8). The complete list of all identified PAGs clusters, together with their corresponding organisms and GI codes, are provided as additional material (see Additional file 9). Most of them (503) are bi-cistronic clusters, while another important fraction (119) is represented by arrays of three genes. Longer clusters were quite rare and we found only 26, 21 and 5 clusters embedding 4, 5 and 6 genes, respectively. This trend is in partial agreement with previous findings on cromosomal clusters of atypical genes 10 and suggests that PAGs clusters are quickly fragmented and/or eroded following integration. Alternatively, this could suggest that the transfer and integration of shorter clusters is somehow favoured in respect to longer ones. Finally, as shown in Additional file 10, short gene arrays seem to be more frequent also when PAGs were retrieved al lower CIs.

Among identified PAGs clusters, some possess a partially documented evolutionary history, mainly driven by HGT/recombination events as, for example, mercury resistance gene cluster(s). In fact, mercury resistance genes (mer) have been usually found embedded in a single compact operon 42 that, in turn, has been suggested to represent an aberrant mercury resistance transposon (namely TndPKHLK2) that, in some cases, has lost those genes responsible for its transposition 43 . The analysis of PAGs clusters allowed the identification of (at least) 10 different mer clusters (see, Additional file 10) that showed a different composition in respect to the source molecule, thus revealing the pivotal role of HGT in the spreading of this metabolic ability across bacteria belonging to (sometimes) very different taxonomical units and inhabiting separate ecological niches. Moreover, other gene clusters (or part thereof) coding for important biological traits (i.e. antibiotic resistance, host invasion and cobalamin biosynthesis) were retrieved. For example invasion-associated genes encoding proteins involved in invasion of mammalian cells were found in atypical clusters on plasmids retrieved from different specie of Shigella genus (see Additional file 10) providing further support to the idea that one (or more) HGT envent(s) played a role in spreading this feature within representatives of this genus 32 44 . Similarly, plasmid mediated HGT seems to have contributed to the spreading of other key metabolic traits in microbial representatives, as, for example, cobalamin biosynthesis 45 and tetracycline resistance for which very similar (atypical) gene clusters were retrieved from very distantly related microorganisms, including Geobacter, Halorubrum, Methylibium, Methylobacterium and Deinococcus representatives in the case of cobalamin biosynthesis and Escherichia, Klebsiella, Aeromonas, Serratia, Yersinia and Enterobacter in the case of tet genes (see Additional file 10).

Among plasmids, conjugative ones have been defined "vessels" of the communal gene pool 11 . Indeed, this class of plasmids possesses the ability to "visit" different cells and, in principle, undergo genetic rearrangements (such as homologous recombination) with other plasmids and/or other informative molecules (phages and chromosomes). For this reason, conjugative plasmids might be expected to possess a higher amount of atypical genes in respect to plasmids that are not (or are less) mobilizable. To test this hypothesis, we performed a comparative analysis of the amount of PAGs belonging to conjugative and non-conjugative plasmids. The conjugative/mobilizable plasmids (hereinafter referred to as CMPs) embedded in our dataset were identified by searching for their propagation-related genes adopting the following strategy: 1) a dataset of genes involved in the propagation of conjugative plasmids (tra and mob like genes) was assembled probing a search on the ACLAME database http://aclame.ulb.ac.be/; 2) a BLAST search with each of the sequence of our dataset against the tra/mob-like database previously assembled (setting a minimum e-value of 1e-20) was carried out and then 3) the number of tra/mob-like genes was determined for each plasmid. In this way we were able to divide our plasmid dataset into (693) conjugative/mobilizable plasmids (CMPs, possessing at least one gene involved in conjugation and/or mobilization) and (1160) not conjugative/mobilizable plasmids (NCMPs, i.e. plasmids that do not possess any tra or mob related genes). Data obtained revealed that, on average, CMPs harbor a higher number of PAGs than that exhibited by NCMPs and that the distribution of PAGs in these two classes of plasmids were significantly different (Mann-Whitney test, p-value < 0.001). Indeed, CMPs possess, on average, 5.7 PAGs, whereas NCMPs posssess only 2.9 atypical ORFs for each plasmid. Data obtained are summarized in Figure 9 which shows the different trends for CMPs and NCMPs. This holds true also when PAGs dataset retrieved at lower thresholds (i.e. 70%, 80% and 90%) were analyzed (Additional File 11)

The high number of PAGs retrieved in CMPs provide strong support to the idea that plasmids have played (and are still playing) a central role in microbial evolution 11 . In fact, our data suggest that, by visiting different cells, CMPs can undergo recombination event(s) with the host's DNA molecules more frequently than NCMPs; consequently, they can probably acquire pieces of exogeneuos DNA that, in turn, can be further spread within the whole microbial communities. This idea is also partially supported by the finding that CMPs, on average, possess more and longer PAGs clusters in respect to NCMPs as reported in Additional file 12.