Overall plasmid diversity is visualized in plasmid dendrograms.

Overall, the ten ST131 genomes analyzed contain 39 plasmids (including one potential ICE), which can be assorted by their relative sizes and MOB groups [53], as shown in Table 1. The most conspicuous group was that of MOB_F12/IncF plasmids (11 plasmids), present in all ten sequenced ST131 genomes. Other relevant plasmid backbones belong to the MOB_P (RIP groups IncI1/K, IncI2, IncX1, IncX4, ColE1), MOB_Q (Qu, Q12) and MOB_C (C12) REL families. The non-F plasmids comprise a total of 20 plasmids belonging to eight plasmid groups. Two plasmids were phage-like and belong to the Rep-3 RIP family. Finally, 5 plasmids corresponded to the no-MOB category. The E35BA genome (ST43/H30 virotype B) showed a MOB_P11 relaxase within a 234 kb chromosomal contig, implying the presence of an ICE (integrative and conjugative element). Detailed inspection of this contig identified a 14.2 kb IME (integrative and mobilizable element) (see below and Suppl. Mat.). Once plasmids were identified by PLACNET and contigs assorted, the next step in plasmid analysis consisted in the construction of a dendrogram that summarizes plasmid gene content, allowing a visualization of relatedness between individual plasmids. The dendrogram of the 44 plasmids found in ST131 genomes (39 described in this report plus five plasmids already published) is shown in Fig. 3. The figure shows how the plasmids divide into branches that coincide with backbone MOB groups. There are 14 plasmid groups, according to the dendrogram, shown in the figure by different color backgrounds. Since each dendrogram group links related plasmids, they can be now analyzed individually, by comparing them either among themselves (Fig. 3) or with selected references (Fig. 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Hierarchical clustering dendrogram of ST131 plasmids.

The UPGMA dendrogram was based on protein cluster analysis using 60% sequence identity and 80% coverage. Plasmid names are colored according to their clade, taking into account ST, fimH allele and virotype, following the color code shown at the upper right. The five plasmid names in black correspond to previously sequenced plasmids from ST131 strains. Different color backgrounds are shown to emphasize branches of related plasmids. To the right of the dendrogram, four columns show, respectively, plasmid size, MOB type, RIP type and Inc type.

https://doi.org/10.1371/journal.pgen.1004766.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Hierarchical clustering dendrogram of ST131 plasmids and relevant references.

The left dendrogram shows the complete tree, with references. Dendrogram construction and color codes are as in Fig. 4. The right dendrogram expands the MOB_F12/IncF branch, with new background colors highlighting plasmid groups within this branch that are mentioned in the text.

https://doi.org/10.1371/journal.pgen.1004766.g004

Analysis of IncF plasmids.

Representatives of the largest plasmid group, formed by 15 MOB_F12/IncF plasmids, were found in each of the analyzed ST131 strains (Table 1). Included in this set are four plasmids lacking MOB and tra regions but containing RIP and other backbone genes related to IncF plasmids. The dendrogram of Fig. 3 clearly indicates that these plasmids belong to the IncF plasmid family, underscoring the usefulness of this step in plasmid reconstruction. Judging from their position in the dendrogram, it seems IncF plasmid genes are scrambled as if the precise constitution of each individual IncF plasmid could not be predicted at all for isolates of each specific ST131 cluster or phylotype. This is even clearer in Fig. 4. In this figure, ST131 plasmids are represented together with the reference sequences that were used for PLACNET reconstruction and analysis. Besides, BRIG comparison of IncF plasmids (Fig. 5) reveals high heterogeneity between them, with not a single completely conserved gene (confirmed by the fact that not a single plasmid gene was found to belong to the ST131 core genome). KClust software [54] at 30% identity and 50% coverage was used to group all proteins coded by IncF plasmids into 354 reference clusters. Manual curation was used to classify these 354 clusters in three groups (Fig. 5): (i) plasmid backbone (i.e., conjugation, RIP and maintenance genes) and metabolic protein genes, (ii) antibiotic resistance and virulence genes, and (iii) other protein genes such as ISs, transposases or hypothetical proteins. Conjugation proteins represent 30 of the 53 backbone proteins and constitute the most conserved set. As mentioned above, 4/14 plasmids do not keep a complete backbone. Table 2 contains the functional annotation of the IncF plasmids. As shown, 11/15 of the MOB_F12/IncF plasmids contain antibiotic resistance genes (conferring resistance to beta-lactams, in all cases, but also to sulfonamides, aminoglycosides, trimethoprim, chloramphenicol, tetracycline and macrolides, in some of them). In addition, nine of the ten antibiotic resistance-plasmids confer a multidrug-resistance (MDR) phenotype (equal or more than four antibiotic families). Besides, 10/15 MOB_F12/IncF plasmids presented putativevirulence genes (Table 2). As previously noted, there was an apparent trade-off between antibiotic resistance and virulence, genes coding for these adaptive traits being located in different plasmids [2]. Finally, a DNA modification gene (adenine-specific DNA methylase) was conserved in all IncF plasmids except in pHVH177_1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. MOB_F12/IncF plasmid analysis.

Protein cluster analysis was performed with kClust software (parameters: 30% identity, 50% coverage) on the set of 14 plasmids shown in Table 4. Plasmid pGUE-NDM [119] was excluded from this comparison since it is only distantly related to the others (see dendrogram in Fig. 5). A total of 354 protein clusters were obtained and annotated versus the NCBI protein database (Blastp). Manual inspection was carried out to classify the reference proteins of each cluster into one of these three groups (comparative analysis shown with BRIG): (i) Backbone and metabolic proteins (panel A); (ii) Virulence and Antibiotic resistance proteins (panel B); and (iii) ISs and hypothetical proteins (not shown).

https://doi.org/10.1371/journal.pgen.1004766.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Resistance genes and virulence determinants in MOB_F12 plasmids.

https://doi.org/10.1371/journal.pgen.1004766.t002

The ST131 IncF plasmids belong to four different branches of the dendrogram, as shown in Fig. 4 inset. Group I includes four plasmids similar to the well-known virulence plasmids pAPEC-ColV like (also called pS88-like), which are commonly detected among avian pathogenic E. coli (APEC) [2], [55]. A suitable reference is the ST131 plasmid pJIE186-2, coming from a ST131 strain previously recovered in Australia in 2006 [56]. As shown in S11A Fig., group I IncF plasmids share two large homologous regions: a 70 kb region containing virulence genes iss, iroBCDEN, iucABCD, iutA, cvaBC and sitC and the cassette ompT-hlyF-mig14, eventually also linked to estABCDE [55] and a 40 kb region containing the tra region and other backbone genes. Group II contains 10 MDR plasmids, 8 of which are ST131 plasmids with characteristic F2:A1:B- replicons and multiple antibiotic resistancegenes. A suitable reference is the ST131 plasmid pJJ1886-5, coming from a ST43/fimH30 lineage from USA. As shown in S11B Fig., group II IncF plasmids share most of their genomes. It should be noted that three of these plasmids (pFV9873_5, pEK499 and pEK516) have extensive deletions within their tra regions, as seen in the figure. Groups III and IV are just represented by one plasmid each. None of them contains antibiotic resistance genes and they are poor in virulence genes. While group III plasmid pHVH177_1 is not similar to any reference outside the backbone genes (S11C Fig.), the group IV plasmid pECSF1 is extensively similar to various large E.coli plasmids, (S11D Fig.). A more comprehensive comparison of F plasmids recovered from ST131 with previously described F-like plasmids is given in Suppl. Mat.

Analysis of other ST131 plasmid groups.

Besides MOB_F12/IncF plasmids, 28 other plasmids were represented in ST131 isolates (Table 1). Ten were large, presumably conjugative plasmids (>18 kb), while 17 were small plasmids (<12 kb) and there was one IME.

Among the large plasmids, a most remarkable branch is composed by two almost identical 109 kb plasmids (pBIDMC20B_2 and pBWH24_2) present in two ST43/H30 isolates of virotype C, for which only RepFIB (Rep3-superfamily) and the maintenance protein ParB could be identified as plasmid backbone genes. No conjugative genes were detected. On the other hand, they code for an integrase protein and several phage-typical proteins. The plasmids are highly similar to pECOH89, recently recovered from a CTX-M-15 producer E. coli isolate from Germany [57]. Closest reference hits were the adherent invasive E. coli (AIEC) plasmid pLF82, isolated from a patient with Crohn's disease [58], the STEC plasmid p09EL50 [5], the Salmonella plasmid pHCM2 [59] and the Salmonella bacteriophage SSU5 [60]. These are all cryptic plasmids isolated from pathogenic enterobacteria that have been barely analyzed and thus are poorly annotated. The possibility arises of these elements being similar to lysogenic phages that are stably maintained as plasmids, analogous to phage P1 [57]. S12 Fig. compares this branch of related plasmids, using plasmid pECOH89 as a reference. As can be seen, both ST131 plasmids share most of their sequences with this 111 kb plasmid, including several phage-like protein genes. Significantly, none of the cryptic plasmids described in this study or those mentioned in the references, except pECOH89, harbor a resistance gene.

The 98.3 kb MOB_P12/IncK plasmid pE2022_1 is most similar to pCT [61]. pE2022_1 contains a bla_CTX-M-14 gene identical to that in pCT, a plasmid carrying bla_CTX-M-14 that is globally spread among humans and animals and particularly prevalent in clinical isolates form Spain [61], [62]. The backbone of plasmid pE2022_1 is homologous to that of the reference IncI1 plasmid pEK204. These plasmids are described in S13. Fig. Despite their different Inc names, IncI1, IncK and IncB/O have similar backbones, belonging to different branches of the IncI complex (an analogous case to IncF plasmids).

MOB_P6 is a large plasmid family, as can be observed in the phylogenetic tree of the MOB_P6 relaxase family (S14A Fig.). The two ST131 MOB_P6/IncI2 plasmids are rather different, as judged by the distant positions of their REL. Plasmid pBWH24_3 (60.3 kb) is similar to the IncI2 prototype R721 [63], but most similar to the APEC plasmid pChi7122_3 [64]. In turn, pE61BA_7 (37.9 kb) is most similar to Salmonella agona plasmid SL483 and the enterohemorrhagic E. coli (EHEC) plasmid pO157_Sal [65]. They were recovered from isolates identified as H30_virotype C from the USA and H324_virotype D from Spain, respectively.

Another ST131 important group is MOB_P3/IncX, composed by three ST131 plasmids (pFV9873_4 and pE2022_3, as well as the reference pJIE143). Plasmids pFV9873_4 and pE2022_3, obtained from different H30 subgroups, are rather different between them and belong to different plasmid groups (IncX1 and IncX4), as shown their relatively distant positions in the MOB_P3 phylogenetic tree of S15A Fig., even if showing similar sizes of about 34 kb. Their coding capacity is shown in the BRIG representations shown in S15B and S15C Fig. The IncX1 plasmid pFV9873_4 is most similar to the EC plasmid p2ESCUM [66], although genetic similarity is constrained to their backbone genes, occupying about half of the reference plasmid sequences. Conversely, the IncX4 plasmid pE2022_3 is most similar to pSH696_34 and the ST131 reference plasmid pJIE143 all over its sequence length (S15C Fig.).

Two plasmids and the IME belong to the MOB_P11/IncP1 family, as shown in the MOB_P11 relaxase phylogenetic tree of S16A Fig. Plasmids pJJ1886_4 and pE61BA_4 showed widely different sizes (56 and 18 kb, respectively). Plasmid pE61BA_4 is only distantly related in its backbone genes to environmental plasmid pMBUI2, isolated from an uncultured bacterium [67]. Plasmid pJJ1886_4, on the other hand, is similar to the E. coli plasmid pHS102707 (GeneBank Acc. N° KF701335). These two plasmids thus represent new additions to the ST131 plasmidome (see S16B Fig.). The IME_E35BA is a 14.2 kb insertion within a 234 kb chromosomal contig. S16C Fig. shows some detail on the genetic structure of the IME and its insertion site in the ST131 core genome.

The last potentially conjugative plasmid is the 24.5 kb MOB_C12 plasmid pE61BA_2, also not closely related to any reference, as seen in the REL phylogenetic tree of S17A Fig. This plasmid resulted in a single contig, so it could be closed. The closest homolog is the Yersinia pestis plasmid pCRY, with which it shares all backbone genes. Its RIP belongs to the Rep_FIIA superfamily, although this plasmid group is not represented in the classical PBRT method [8]. It does not contain any gene with known adaptive function, except a protease and a putative secreted thermonuclease. As in the case above, this plasmid represents a new addition to the ST131 plasmidome (S17B Fig.), its relaxase being only 70% identical to its closest homolog, the pCRY relaxase.

Among the small plasmids, the first group is composed by four MOB_P5/ColE1-like plasmids (11.8, 6.9, 6.6 and 5.6 kb). Three of the plasmids are relatively different, as judged by the MOB_P5 phylogenetic tree of S18A Fig. The large plasmid (pBIDMC38_1) contains a type II restriction-modification system (Cfr10I) and is almost identical to the ST131 reference pJJ1886_3 (S18B Fig.). The two MOB_P5/ColE1 plasmids of strain E61BA (plasmids pE61BA_5 and pE61BA_6) contain colicin ColE and ColK genes, respectively (S18C and S18D Fig.). Colicins are considered both as virulence factors as well as traits that influence bacterial fitness and survival in the presence of competitors [68].

Besides, there were four almost identical MOB_Qu plasmids of around 4.1 kb (S19A Fig.), which populate all H30 subgroups (two of virotype A, one of virotype B and one of virotype C). Nothing remarkable could be distinguished it their genetic constitution, besides a common MOB region and a pIGWZ12 -like Rep protein (S19B Fig.). Four very similar MOB_Q12 plasmids (around 5.2 kb) are also represented in H30 (two of virotype A, one of virotype C) and H22 (one of virotype D). They contain RIP and REL proteins but, as in the case of the MOB_Qu plasmids, no phenotype could be pointed out (S20 Fig.). MOB_Qu and MOB_Q12 plasmids have received little attention because they are cryptic and remain unnoticed in most typing schemes. The present ST131 plasmidome analysis suggests they can be surprisingly abundant in E. coli. Finally, there were five no-MOB cryptic plasmids (four of them were 1.5 kb long and the other 5.0 kb). They all contain distinguishable Rep proteins (Rep_HTH_36_superfamily), without assignment in the PBRT method. Four of them are almost identical among themselves (S21 Fig.), while the fifth (pFV9873_3) was unique and unrelated to any reference. A detailed analysis of MOBF11/IncN plasmid family is detailed in S22 Fig.

Unassigned contigs and reference sets.

After plasmid reconstruction, occasionally, one or a few contigs may remain unassigned. In the set of ST131 genomes analyzed in this work, there were only two unassigned contigs (>200 bp), both in E61BA (Fig. 2). The fact that only two unassigned contigs appeared in the analysis of eight E. coli genomes suggests that this is not a quantitatively important problem. As could be expected, unassigned contigs are more frequent in genomes for which there are fewer references available. The lack of a suitable reference set results in poor quality clustering and an increased fraction of contigs without references. It is obvious from the preceding discussion that bacterial taxons for which not enough references exist will be more problematic for plasmid reconstruction. Thus, any such project should start by the generation of a sufficiently ample set of plasmid references. In this respect, E. coli constitutes probably the best choice, due to the large reference set available.

Repeat sequences and difficult plasmids.

Usually PLACNET works well because contigs belonging to individual plasmids pair with different selected references and thus cluster in disjoint connected components in the Cytoscape representation after a single pruning step. The pruning step consists in identifying the bridging contigs (network hubs) as repeat sequences (RS). Two sets of evidence were used: (i) homology to known ISs or transposons, and (ii) existence of three or more scaffold links. Contigs fulfilling these two criteria were assumed to be in fact repeated in the connected network. Thus, the pruning operation consisted of duplicating the alluded nodes and splitting their scaffold links. In the tested set of E. coli genomes that were used to validate PLACNET (the set of eight ST131 genomes analyzed here, the 32 genome set analyzed by de Been et al. (2014) < submitted to Plos Genet together with this work >, plus another set of 10 other ESBL genomes obtained from clinical strains of bioprojects PRJNA186205 and PRJNA202876), there was only one case in which RS pruning operation was not sufficient to obtain disjoint components. It was the case of genome E35BA, where two coexisting MOB_F12/IncF plasmids (pE35BA_2 and pE35BA _3) could be inferred, but repeated pruning did not result in disjoint components. The evidence for the existence of two plasmids was the finding of two sets of contigs containing REL and other plasmid backbone genes. PLACNET failed in discriminating both plasmids probably because network links to references were interlocking, since several PLACNET-selected references established best hits to different components of each set. Besides, the assembly program could not distinguish among parts of both sequences and considered them as RSs. Closely related plasmids that coexist in a given cell poses the most serious problem we encountered in the application of PLACNET.

Plasmidome description.

We describe a remarkable heterogeneity of plasmids among the E. coli ST131 genomes analyzed, with the identification of 39 plasmids to add to the 11 plasmids in the ST131 lineage already sequenced (Table 1). Interestingly, these plasmids encompass 8 out of 17 main MOB plasmid groups found in the whole class of γ-proteobacteria [11], [53], namely F12 (IncF), P3 (IncX), P5 (ColE1), P6 (IncI2), P11 (IncP), P12 (IncI/K/BO), Q12 (Rep_pSC101-like), Qu (Rep_pMG828-2/IGWZ12-like), several of them undetectable by PBRT. Besides conjugative or mobilizable plasmids, there were other plasmids, lacking REL, but identifiable by their RIP proteins. An in-depth analysis of each plasmid group identified in this study has been diverted to a Supplementary Discussion (though exciting for clinical epidemiology or plasmid biology, it is outside the mainstream goal of this work). Our findings substantially enlarge the repertoire of plasmids identified among E. coli ST131 isolates, which now reflect a genome widely open for plasmid infection. It is of note that this scenario has also been described for E. coli clones of different pathovars [47], [70]–[73]. Comparative genomics of the few E. coli lineages comprehensively analyzed to date suggests that this species is a generalist able to colonize and infect humans. It also suggests that phages and plasmids make an important contribution to specialization by accessorizing the genome with new adaptive traits and tools that modify genome structure and, eventually, by modifying transcriptional regulation [70], [71], [74], [75]. It should also be emphasized that almost all available studies on ST131, included this one, focused on strains involved in the spread of antibiotic resistance genes, which constitute, undoubtedly, a biased fraction of the ST131 plasmidome and thus preclude an accurate view of its evolutionary history [76] (see also below).

Plasmids and E. coli diversification.

Specific ExPEC lineages have scarcely been analyzed in the context of multilevel population genetics with the exception of punctual cases involving clonally unrelated isolates [70]. A recent phylogenomic analysis of 95 ST131 isolates from different geographical areas identified the same three clusters studied in the present work [16]. This analysis concluded that point-mutations and recombination events associated with diverse MGEs, including prophages and genomic islands, determined the diversification of this ExPEC lineage. However, the diversity of plasmids was only inferred by searching for incompatibility regions based on PBRT schemes. The role of plasmids in genomic versatility were not further analyzed [16]. Even though our study analyzed a smaller number of isolates, some observations can be drawn about the role of plasmids in the diversification of the ST131 lineage.

The rate of mutagenesis of E. coli has been roughly estimated in one mutation per genome per year [77]. Although this number is no doubt controversial, such study and those addressing the role of recombination in the ST131 lineage add context to understand its evolution as represented in Fig. 1. Compared to the limited sequence divergence among the ST131 core genomes (the ST43/H30 branch includes just about 600 SNPs), plasmids represent a very active fraction of ST131 adaptive evolution, as can be concluded from the analysis of Table 1. Such plasmid variability suggests that independent plasmid acquisitions and losses frequently occur within and between ST131 sublineages. Within the H30 cluster, the presence of antibiotic resistance plasmids is remarkable, specially the identification of structurally similar F2:A1:B- plasmids carrying genes conferring antibiotic resistance, since early acquisition of bla_CTX-M-15, mainly associated with F2 plasmids, is considered a key event in the selection of the ST131 cluster C/H30 subclone [13], [16], [17]. The modular structure of F2 plasmids, containing multiple copies of ISs, facilitates gene rearrangements and the interchange of antibiotic resistance platforms linked to resistance to first line antibiotics between plasmids of the same and different families. This notion, inferred from our present analysis, has already been proposed [17], [78]–[80] and is of great concern nowadays because of the increasing risk of encountering E. coli isolates carrying bla_KPC of bla_NDM genes predominant in Klebsiella (http://www.cdc.gov/drugresistance/threat-report-2013/) [81]. Beyond F2 plasmids, the presence of other plasmid groups (N, I1/K/BO, I2, A/C, X) carrying antibiotic resistance genes is observed at variable rates in this and other studies, clearly influenced by local ecology. Most of these antibiotic resistant non-F plasmids occur only rarely in E. coli isolates [82]. This could be due to an intrinsic lack of fitness of these plasmids in E. coli under natural conditions. Alternatively, they could represent cryptic indigenous plasmids now identifiable because of the acquisition of antibiotic-resistance cassettes. Nevertheless, the acquisition of mosaic regions carrying multiple antibiotic resistance genes by broad host plasmids (e.g. N, I2, A/C) increases the risk to spread resistance to first line antibiotics to different bacterial species in and outside hospitals [79], [83]. Besides antibiotic resistance plasmids, an outstanding finding of this work was the frequent detection other plasmid groups, generally considered cryptic, that are clearly underrepresented in previous ST131 studies, as ColE1, MOB_Q, and phage-like Rep3 plasmids. All of them are highly heterogeneous plasmid groups able to acquire adaptive traits or contribute to the mobilization of other plasmids (See supplementary dataset for details).

Members of the ST131 plasmidome such as MOB_F12/IncF, MOB_H/IncA/C and Rep3/phage-like plasmids can also shape the E. coli chromosome by facilitating mobilization in trans of genetic islands or integrating new genetic material [32], [70], [84]. Interestingly, recombination of large chromosomal regions occurring at the sites of insertion of either prophages or transferable genomic islands seems to have contributed to the split of the ST131 lineage in different clusters [16]. Although experimental studies on ST131 did not yet associate plasmids with genome structure, the hypothesis is plausible taking into account the frequency at which such events occur in other B2 E. coli populations.

On a more general note, our study identifies antibiotic resistance plasmids, which are favored in high density environments, such as the human gastrointestinal tract, and under antibiotic selective pressure, that are predominant in hospitals [85] together with phages (or phage-like cryptic plasmids), apparently predominant in low density environments [86], [87], and other cryptic plasmids (frequently very small and devoid of any possible adaptive gene). This mixed constitution, which is difficult to understand on purely selective grounds, highlights potential roles of plasmids in the context of multilevel selection, a recurrent issue in evolutionary biology. The plasmid flux in ST131 strains occurs while disseminating genes coding for resistance to extended spectrum beta-lactamases (ESBL). The results presented here find a complement in the study of de Been et al. (accompanying paper), where the authors document the dissemination of ESBL-carrying epidemic plasmids from animal to human clonally unrelated E. coli lineages. Thus, many plasmids appear in a clonal lineage, and many lineages can be infected by a single predominant plasmid. These conceptual notions have relevance in Public Health as they deal with the hierarchical units of selection that contribute to increase the population size of antibiotic resistance genes in human and animal pathogens [69], [88], [89].

In summary, our study reveals the utility of PLACNET in multilevel population genetics analysis, critical to understand the evolutionary processes and dynamics of both bacterial and plasmid lineages. Its application to E. coli ST131 allowed us to infer the roles of plasmids in the dissemination of globally spread antibiotic resistance and virulencegenes, some of them being underrepresented in Genbank. It is probable that these plasmids are critically relevant to understand the adaptive evolution of E. coli populations and their bacterial exchange communities. Armed with this new tool for plasmid analysis, future scrutiny of a larger number of significant strains will allow us to understand the interplay among different plasmid associations that often appear in bacterial pathogens.

Assembly.

Velvet assembly software [99] and its script VelvetOptimiser.pl were used to determine the best assembly and to scan the optimum parameters. Velvet provides also coverage information for each contig, which adds useful information for network interpretation.

Scaffold links.

Although assembly programs perform scaffolding between contigs, when the assignation is ambiguous, contigs remain unbound. A method based in the mapping tool Bowtie 2 [100] was used to find all possible scaffold links. All reads were mapped using the contigs as references, using default parameters of Bowtie 2, with the option to report all possible hits. The output file was converted to SAM format [101] to give the potential adjacency information for each contig. We considered as potential PLACNET scaffold links those which comply with two rules: (i) the contigs were paired at their extremities, themselves defined as twice the read length, and (ii) the number of pair-end reads linking those two contigs had to be higher than one third of the mean of the total pair-end reads that scaffold all contigs. This procedure was implemented as an in-house Perl script.

Reference search.

At least for bacteria widely covered by sequencing projects, most contigs in any new sequence are similar to one or more previously published sequences (reference sequences). Our initial hypothesis was that, for any physical DNA unit, its contigs will “BLAST” a related set of reference sequences. Thus, in the PLACNET network representation, they will cluster around the homologous references. The more DNA databases grow, the closer the references will be to the query sequence and the better PLACNET will work. A homemade BLAST [102] database was constructed from the NCBI genomes database by joining all sequences contained in [ftp://ftp.ncbi.nlm.nih.gov/genomes/Bacteria/] and [ftp://ftp.ncbi.nlm.nih.gov/genomes/Plasmids/]. The version used in this work was from March 7^th 2013 and contains 6,432 genomes (plasmids and chromosomes). Megablast search of all contigs was carried out against the homemade BLAST-genome database with the objective of selecting a few best matches for network construction. Due to the different length of each contig, fixed thresholds by e-value or score cannot be chosen. Since the score is not a normalized parameter, and varies depending on sequence length, hits were selected by applying a dynamic threshold, based on the number of homologous sequences and the score of each hit. If the threshold is defined as 85%+2n of the mean of the n previous sequence scores, and n is the ranking position of sequence i retrieved by megablast with score S_i, then T_n+1 is the threshold for sequence n+1:

All reference sequences above the threshold were taken as nodes in the PLACNET representation.

Protein prediction of replication initiator proteins (RIP) and relaxases (REL).

Some genes are indicative of a plasmid sequence. Among them we selected REL, key proteins in the conjugation process [53], [103] and RIPs, key proteins in the replication of most plasmids [104]. Although not all plasmids contain a RIP and/or REL, their presence in a contig is diagnostic of a plasmid (or ICE) sequence. Some plasmids have more than one RIP (i.e. IncF family plasmids) [9], [105], [106] but plasmids rarely have more than one REL [107]. ORF prediction was carried out by GeneMark [108], which optimizes predictions based on GC content of DNA. The heuristic prediction implemented in this software is especially useful to predict ORFs in plasmid-containing genomes because it takes each contig individually and selects the best prediction model case by case.

To implement specific search protocols for REL and RIPs, three homemade databases (DB) were developed. A REL database (REL-DB) was constructed according to [53], [109]. Similarly, RIP-DB was constructed from all RIPs annotated in UniProt database. RIP sequences were clustered by CD-HIT [95], using 40% identity as a threshold. Next, a Hidden Markov Model was built from each cluster by hmmer3 [110]. Finally, the HMM profiles were used in a search against UniProt. The HMM search and initial dataset were joined in one database (RIP-DB). An additional step was necessary to classify RIPs according to the widely used plasmid classification protocol PBRT [8]. A homemade nucleotide database (INC-DB) was created to identify PBRT types in silico using blastn. Finally, the relevant ORFs (REL and RIPs) identified to specific contigs by using REL-DB and RIP-DB were incorporated to the network as tags. All relevant steps in network construction were implemented as a Perl script available at the following web page: http://placnet.sourceforge.net/.

Plasmid constellations.

As explained above, each plasmid is represented in PLACNET by a connected component (a constellation). Thus, different physical units (plasmids and chromosomes) should be represented by disjoint (unlinked) connected components. Cytoscape software [111] was used to visualize and analyze plasmid constellations, which incorporate all the information (similarity to reference sequences, scaffold links, and protein tags) in a single network. Node attributes such as contig size, coverage and reference description, are added to the network. At this stage, network pruning is needed to resolve individual plasmids as disjoint components. When a genome has a number of repeated sequences (e.g., insertion sequences (ISs) or transposons), or two very similar plasmids, the assembly process outputs those sequences as contigs with multiple scaffold links. In the network context, they represent hubs, that is, nodes with a high number of connections. This makes the network very dense and complicates the analysis of network connected components. In PLACNET, contigs smaller than 200 bp were directly eliminated from the analysis. Hubs were examined by blastx against protein databases (i.e. UniProtKN or NCBI nr). If there was identity to any transposase gene, the hub was duplicated, and scaffold links were partitioned among them, to maximize the number of disjoint components. Contigs that remain unbound are classified as “unassigned sequence” in the contig assignation table.

Plasmid definition, dendrograms and cluster analysis.

The final steps in plasmid reconstruction involve the definition and verification of each plasmid. This is an iterative process, as shown in Fig. 6. First, each contig was assigned to a putative plasmid (or chromosome) based on visualization of disjoint connected components in the Cytoscape representation. Assignments take into consideration additional types of evidence like the presence of REL and/or RIPs, size of the putative plasmid compared to reference plasmids and coverage of each contig (contigs belonging to the same plasmid must have similar coverage). Taking into account the information provided by related genomes within the same sequencing project can also be helpful (same sequences cluster around the same references). In this respect, PLACNET is more robust in multi-strain collections. The performance of PLACNET was validated by testing a number of previously sequenced and annotated E. coli genomes (S2 Table). The ART software (Huang et al., 2012) was used to simulate pair-end Illumina reads from those genomes, which were then analyzed by PLACNET as explained above. Results are shown in S2 Text, S3 Table and S25-S34 Figs.

After PLACNET has defined the plasmids carried in the relevant genomes, the next step in plasmidome analysis is to build a dendrogram that produces a hierarchical clustering of plasmid proteomes similar to those described in [112]–[114]. CD-HIT (thresholds: 70% identity and 80% coverage) was used for clustering references and query plasmids. Based on the output file, a presence/absence table (present or absence of each protein cluster in each plasmid) was built. Each table row represents a plasmid protein profile. Raup-Crick distance method, implemented in vegan package for R software [115], was used to calculate the distance matrix of plasmid protein profiles. The Ape package [116] was used to calculate the dendrogram bootstrapping confidence value. Finally, a hierarchical clustering dendrogram was built using the UPGMA algorithm.

Putative plasmids and references belonging to the same dendrogram branch were compared using BRIG [117] or Abacas [118]. While BRIG is not sensitive to contig arrangement, Abacas can be used to order contigs according to a given reference. With these tools, the curator is able to visualize the correspondence between reconstructed plasmids and references, or can go back to dendrograms or PLACNET in the search for missing or extra contigs. This iterative mode of analysis is represented in Fig. 6 by the backward arrow linking plasmid cluster analysis with plasmid definition.

Plasmids were mainly classified according to their REL in MOB families, as described by [53]. Classical Inc families are also given when typing them by in silico PBRT was possible. Plasmids that could not be classified one way or the other were termed no-MOB by exclusion.