Plasmid Classification in an Era of Whole-Genome Sequencing: Application in Studies of Antibiotic Resistance Epidemiology

1Plasmid genomes generally include a ‘backbone’ of core genetic loci, which are somewhat conserved amongst broadly related plasmids of the same family, and associated with key plasmid-specific functions such as replication and mobility. Accessory genes may also be present, and often confer clinically relevant traits such as virulence and antibiotic resistance. Plasmids can act as efficient vectors of horizontal gene transfer (HGT). Notably, during conjugation, a plasmid promotes its own transfer (and/or that of a co-resident plasmid) from one bacterial cell to another. Accessory genes are therefore frequently spread by virtue of being located on transmissible plasmids; moreover, they are commonly associated with smaller mobile elements such as transposons, facilitating intracellular mobilization amongst plasmids, or to the chromosome. Due to their ability to transmit genes encoding adaptive traits across bacterial populations, plasmids can enable bacteria to evolve rapidly under environmental pressure. A striking example of bacterial adaptive evolution is that of antibiotic resistance, which is driven, in part, by dissemination of resistance plasmids (plasmids conferring antibiotic resistance), and now threatens modern medicine.

2Classifying plasmids according to a typing scheme provides useful insights into the epidemiology of plasmid-mediated antibiotic resistance: for example, studying the composition of plasmid types can indicate whether an antibiotic resistance epidemic is driven by diverse plasmids or one dominant plasmid type. In addition, hypotheses about resistance transmission during outbreaks can be refined according to the relatedness of resistance plasmids harbored by clinical strains. The principal plasmid classification schemes are replicon and MOB typing, based on backbone loci encoding plasmid replication and mobility functions, respectively. Whilst these single-locus typing schemes have been widely and successfully applied, they provide limited resolution, restricting epidemiological inference: in an outbreak context, if two patients are infected by unrelated strains harboring resistance plasmids of the same type, this raises the possibility of plasmid transmission, but plasmid transmission cannot be conclusively ruled-in using single-locus plasmid typing alone; further higher-resolution investigation would be required. If resistance plasmids are unrelated, plasmid transmission can be ruled-out, though a transmission link via resistance gene transposition is possible.

3Plasmid typing may provide a stepping-stone to higher resolution analyses; identifying shared (‘core’) genes amongst related plasmids can inform development of plasmid multi-locus sequence typing (pMLST) schemes, or allow phylogenetic relationships to be reconstructed based on core gene single nucleotide polymorphisms (SNPs). Unfortunately, determining high-resolution plasmid relationships is challenging: the tendency of plasmids to gain, lose and rearrange genetic content means sets of plasmids – even if of the same type – will tend to share few phylogenetically concordant core genes, impeding subtyping and phylogenetic analysis. Even backbone genes may not be well conserved across all plasmids of the same type, and sometimes show mosaic phylogenetic origins.

4Whole-genome sequencing (WGS) data can now be obtained for many bacterial isolates, at relatively low cost, within short timescales. Whilst sequencing reads from a bacterial isolate represent plasmid(s) as well as the chromosome, WGS-based studies have often focused on the host strain chromosome as the unit of interest. For strain-level clinical surveillance to elucidate antibiotic resistance transmission routes, dissemination should primarily involve clonal transmission of particular antibiotic-resistant strains. However, recent analyses indicate that plasmids may transmit between strains frequently, even over short timescales. Therefore, the chain of transmission no longer simply corresponds to strain transmission; resistance plasmid dissemination across strains recruits different recipient strains into the outbreak too, resulting in a ‘plasmid outbreak.’ Although insight may be limited by difficulties in determining high-resolution plasmid relationships, these dynamics mean that plasmid analysis across a variety of strains is important, including for short-term surveillance studies.

5Conventional PCR-based plasmid typing methods are commonly used, but in silico approaches for classifying sequenced plasmids are also available. WGS datasets from short-read sequencing projects offer exciting opportunities for large-scale plasmid analysis, while presenting the additional challenge of assembling reads to resolve individual plasmid structures. After summarizing current plasmid classification schemes (replicon and MOB typing), this review discusses the opportunities and challenges of conducting in silico plasmid typing on WGS datasets to gain insight into plasmid-mediated resistance epidemiology. We highlight novel tools for WGS-based plasmid analysis, and examine gene-sharing networks as a complementary approach for analyzing plasmid relationships. This review focuses on WGS datasets from cultured rather than metagenomic samples; for the latter, see recent reviews.

6Replicon typing schemes exploit genetic elements of the replicon region (encoding replication machinery) (Table 1). typed plasmids according to Southern blot hybridization, using replicons from plasmids of different incompatibility groups as probes. However, this method is limited by probe cross-hybridization amongst closely related replicon sequences. PCR-based replicon typing (PBRT) – where plasmids are typed according to PCRs targeting various replicon sequences – is less laborious, and shows higher specificity in detecting replicons. For gram-negative bacteria, PBRT schemes targeting replicons found in Enterobacteriaceae and Acinetobacter baumannii plasmids are available. A PBRT scheme for plasmids of gram-positive bacteria has been developed, focusing on enterococcal and staphylococcal plasmids. For common Enterobacteriaceae replicon types, pMLST schemes have been devised for subtyping. Availability of WGS data has motivated the development of in silico replicon typing and subtyping tools, which have been validated for Enterobacteriaceae plasmids. For plasmids from taxa not represented by existing in silico tools, ad hoc in silico methods have been derived from PBRT schemes.

7MOB typing exploits the conserved N-terminal sequence of the relaxase proteins encoded by transmissible plasmids. As with replicon typing, both PCR-based and in silico approaches are used for MOB typing (Table 1). Compared with replicon typing, MOB typing classifies plasmids at lower resolution.

8A drawback of replicon typing is that individual plasmids can contain multiple replicons, complicating classification, whereas usually just one relaxase is encoded. However, due to its finer resolution, replicon typing provides more detailed information on plasmid relatedness, particularly if a pMLST subtyping scheme is available. Even within relatively well-studied taxa, neither scheme classifies all plasmids, likely reflecting diversity in plasmid backbones. assessed in silico typing, and found that the proportion of Enterobacteriaceae plasmids that could be replicon typed was 75%; for Acinetobacter plasmids the proportion was 67%. Only around half of plasmids from major gram-positive taxa could be replicon typed (51% Firmicutes plasmids, 49% Actinobacteria plasmids), although the proportion was higher for enterococcal and staphylococcal plasmids (83 and 85% respectively) (Supplementary Table S1 in ). also highlight gaps in replicon typing of Firmicutes plasmids. MOB typing only types transmissible plasmids (∼50% γ-Proteobacterial plasmids; ∼35% Firmicutes plasmids).

9When analyzing whole genomic DNA, limited information can be derived from plasmid typing alone: bacterial cells may contain multiple different plasmids, and a single plasmid may contain multiple replicons, obscuring correspondence between detected replicons and the set of plasmid types within a host cell. Therefore, in PCR-based studies – where genomic context of an amplicon remains unknown – plasmids are commonly isolated first, before being individually characterized by replicon and resistance typing (see Table 11 in ). This is time-consuming, restricts the number of isolates that can be analyzed, and has inherent limitations: if plasmids from the same isolate are of similar size they cannot be separated by pulsed-field gel electrophoresis, and isolation by transfer to recipient cells is not always achieved.

10Potentially, WGS enables in silico analyses in which plasmid typing and analysis of loci of interest, such as resistance genes (and their plasmid or chromosomal genetic context), are performed in a unified way. Consequently, much larger isolate collections can be analyzed – a key advantage given that plasmid studies have indicated a need for including more strains in analyses (e.g., environmental strains, or isolates exhibiting only low-level resistance) to uncover more complex transmission routes. Furthermore, a thorough mechanistic understanding of clinically important aspects of plasmid biology, such as host range and phenotypic effects, may require analysis of the wider plasmid genome rather than specific loci.

11Unfortunately, there are significant obstacles to in silico analysis of WGS data. Popular high-throughput sequencing technologies (e.g., Illumina) produce short (∼100–300 bp) reads, for which assembly is inherently challenging. Isolating individual plasmids prior to sequencing simplifies assembly, potentially enabling complete plasmid reconstruction, but is laborious. Long-read sequencing vastly simplifies assembly, but high costs restrict its use. Therefore, a major challenge lies in extracting useful information from short sequencing reads derived from different sources (different co-resident plasmids, the chromosome).

12Having obtained short reads from WGS of isolate DNA, reads are generally mapped to a reference and/or assembled de novo using a de Bruijn graph assembler (for detailed workflows see; ). Reference-based read mapping is a fast and accurate method to characterize SNPs and detect loci of interest. Identified core genome SNPs can be used to construct strain phylogenies. In addition, rapid epidemiological surveillance of replicons and resistance genes can potentially be achieved with read mapping tools such as SRST2. However, the read mapping approach is limited when structural information is of interest: is a detected resistance gene located on chromosome or plasmid, and if the latter, which plasmid type is it associated with? One approach to overcome this is de novo read assembly, which can be less sensitive for SNP or locus detection, but provides reference-free structural information, and can identify loci not represented on available references. Table 2 summarizes key in silico tools for plasmid analysis.

13Sometimes, short reads can be assembled into complete plasmid structures, and plasmid-localized resistance genes can be identified using a combination of in silico plasmid typing and resistance gene typing methods (e.g., PlasmidFinder and ResFinder). However, complete plasmid assembly is frequently not possible. Notably, the presence of multiple copies of the same repeat structure – a common situation in plasmid genomes – introduces assembly ambiguity, which can fragment assemblies. Paired-end sequencing data can resolve repeat location, but only if the paired reads span the length of the repeat. If contigs contain sufficient informative sequence, plasmid and chromosomal contigs can be distinguished by BLAST searching against the Genbank nucleotide database.

14Resistance genes are frequently flanked by repetitive mobile elements, and are therefore prone to poor assembly, obscuring their genetic context. There are several tools which can help resolve the location of specific loci of interest. Bandage allows visualization and annotation of the assembly graph; for example, users can zoom to unresolved repeat regions and BLAST search connecting contigs. If all connecting contigs match either plasmid or chromosomal references, then an ambiguously linked region can be assigned accordingly. For example, Bandage helped to reveal diverse plasmid contexts for the mcr-1 colistin resistance gene in UK clinical isolates. However, this manual approach is unfeasible when analyzing large datasets. In some cases, the genetic context of resistance genes may be inferred for large datasets using the ISMapper tool, as demonstrated by a study of 1832 isolates belonging to the successful H58 Salmonella Typhi lineage.

15As well as detecting loci of interest, reference-based mapping has been used to track plasmids during short-term outbreaks. Specifically, ‘index’ plasmids of an outbreak are fully assembled, often through long-read sequencing. Short reads or contigs from subsequent isolates are then mapped to the index plasmid, which is deemed present if homology is demonstrated across a given length of the reference sequence. This approach implicitly assumes that the index plasmid is important throughout the study period, and that plasmid structures are relatively conserved in the short-term. In some cases, these assumptions may hold, but other studies show major structural changes can occur, including recombination of large segments as well as mobilization of resistance genes. Crucially, demonstrated that reference-based mapping can be misleading if plasmid plasticity is high. Specifically, a reference blaKPC resistance plasmid from the index isolate was detected across diverse strains by a contig-alignment approach, leading to the initial interpretation that resistance had spread via the original blaKPC plasmid. Instead, long-read sequencing showed that often the blaKPC gene was actually present on a co-resident plasmid, suggesting that mobilization of blaKPC had recruited diverse plasmids to the outbreak. It was usually not possible to determine the genetic context of blaKPC without long reads due to long repetitive flanking sequences.

16Approaches described so far aim to address specific questions, such as the location of a resistance gene, or the short-term transmission of particular plasmids. It is yet more challenging to generate complete structures of diverse plasmids from fragmented assemblies, although several algorithms attempt to do this. The Plasmid Constellation Network (PLACNET) method takes the assembly graph and adds reference genomes as nodes; references are linked to homologous contigs, and the assembly graph is reconfigured according to the additional links. Finally, manual reconfiguration is required to retrieve disjoint connected components that should represent distinct genetic units (chromosome/plasmid), with plasmids identified by presence of replication or relaxase proteins. However, poor assembly of repetitive sequences remains a challenge for resistance gene localization, and reliance on reference sequences to order the network can lead to large-scale errors in structure. For example, used long-read sequencing to validate PLACNET reconstructions; in one case, plasmid contigs had been incorrectly reconstructed as two distinct plasmids rather than one, probably because the plasmid was a fusion of two previously-observed reference plasmids.

17Alternative algorithms such as Recycler and plasmidSPAdes are entirely automated, and independent of reference sequences. Instead, read coverage is used to reconstruct plasmids, with the assumption that contigs from the same genetic unit should share similar coverage. However, read coverage will not distinguish different plasmids if they maintain similar copy numbers within the same host cell, nor distinguish chromosomal from plasmid contigs if a plasmid is maintained at a copy number of one across sampled cells. A recent assessment shows that Recycler and plasmidSPAdes fail to accurately reconstruct all plasmid structures from short-read WGS datasets, though the goal of identifying plasmid-derived sequence (regardless of structural accuracy) is more attainable.

18Long-read sequencing technologies, notably single molecule real-time sequencing (SMRT, Pacific Biosciences) and nanopore sequencing (Oxford Nanopore) promise to revolutionize plasmid analysis. The accurate plasmid structures generated allow for a detailed picture of plasmid epidemiology and evolution. However, current cost considerations have so far restricted analyses to small isolate collections. showed how analysis of short-read data can guide economical use of SMRT sequencing: to examine plasmids during an outbreak, short reads were mapped to a reference, whilst unmapped reference regions, PCR amplification of marker genes, and excess reads were used as indicators of structural change, justifying investigation with long-read sequencing. However, minor structural variation (which may include resistance gene mobilization) is unlikely to be detected by this approach. Hybrid (short/long-read) assembly, using reduced long-read coverage, may help to partially mitigate current costs.

19Current plasmid typing schemes exploit a relatively small number of loci thought to best reflect the vertical (tree-like) component of plasmid evolution. However, plasmid genomes tend not to conform to tree-like evolution: co-integration events and genetic exchanges amongst plasmids mean that different parts of a plasmid genome may have different evolutionary origins. A complementary way to assess plasmid diversity involves ordering plasmids into dendrograms or networks according to their gene content similarity, irrespective of whether gene-sharing stems from vertical inheritance or horizontal acquisition. In a gene-sharing network, plasmids are nodes that are linked to other plasmids if they share genes at a given sequence identity threshold. The network topology can identify plasmids with interesting gene-sharing patterns. For example, some plasmids may act as ‘bridges’ in the gene-sharing network, straddling different groups that respectively share few genes (in graph theory terms, these plasmids have high ‘betweenness centrality’). Such plasmids may represent co-integrate plasmids, or may be important in shuttling genes across the network.

20Making well-supported inferences from gene-sharing networks is challenging since gene-sharing might result from vertical inheritance, horizontal acquisition, or acquisition from a source not represented within the network. However, there are various ways to infer HGT events. For example, finding genes with very high sequence identity from plasmids that otherwise share few genes has been used to identify putative recent HGT events. Overall, networks are a powerful complementary tool for visualizing relationships across diverse plasmids, and generating hypotheses about the horizontal component of plasmid evolution. Gene-sharing networks only require assembly of genes, not complete plasmids, so are suited to analysis of fragmented assemblies. Since network topology is determined by constituent nodes, large unbiased plasmid datasets will produce the most informative network analyses.

21In future, optical mapping of intact plasmids could complement sequencing-based analysis. Specifically, AT-rich plasmid sequence can be fluorescently labeled, after which plasmids are elongated within nanofluidic channels and visualized with fluorescence microscopy. This allows a course-grained optical ‘barcode’ to be obtained. Optical plasmid barcoding could be used for classifying plasmids, and the course-grained structural information from the barcodes could guide plasmid assembly.

22Meanwhile, current algorithms fail to accurately reconstruct all plasmids from short-read WGS datasets. Long-read sequencing will be increasingly used to determine accurate plasmid structures. Use of replicon and MOB typing for plasmid taxonomy will probably continue, but future developments could include methods to incorporate plasmid structural variation into a phylogenetic framework. Networks are a powerful tool for assessing plasmid relationships, from a functional rather than a phylogenetic perspective. Perhaps future plasmid databases could be structured as networks; when novel plasmids are added, information about their potential importance (e.g., centrality within the network) could be determined, and putative HGT events could be continually inferred. Future advances in plasmid metagenomics will enhance our knowledge of plasmids across a range of environments, and improve understanding of resistance gene reservoirs.

23AO undertook literature searching, and wrote a draft manuscript. AO, HP, AS, NS, AW, MA, ME, MD, and NW suggested/implemented revisions. HP, AS, NS, AW, MA, and MD helped in planning the manuscript. AO, NS, MA, MD, ME, TP, DC, NW, AW, HP, and AS read and approved the manuscript.

24The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.