Repeated local emergence of carbapenem-resistant Acinetobacter baumannii in a single hospital ward

We recently reported a dramatic increase in the prevalence of carbapenem-resistant Acinetobacter baumannii infections in the intensive care unit (ICU) of a Vietnamese hospital. This upsurge was associated with a specific oxa23-positive clone that was identified by multilocus VNTR analysis. Here, we used whole-genome sequence analysis to dissect the emergence of carbapenem-resistant A. baumannii causing ventilator-associated pneumonia (VAP) in the ICU during 2009–2012. To provide historical context and distinguish microevolution from strain introduction, we compared these genomes with those of A. baumannii asymptomatic carriage and VAP isolates from this same ICU collected during 2003–2007. We identified diverse lineages co-circulating over many years. Carbapenem resistance was associated with the presence of oxa23, oxa40, oxa58 and ndm1 genes in multiple lineages. The majority of resistant isolates were oxa23-positive global clone GC2; fine-scale phylogenomic analysis revealed five distinct GC2 sublineages within the ICU that had evolved locally via independent chromosomal insertions of oxa23 transposons. The increase in infections caused by carbapenem-resistant A. baumannii was associated with transposon-mediated transmission of a carbapenemase gene, rather than clonal expansion or spread of a carbapenemase-harbouring plasmid. Additionally, we found evidence of homologous recombination creating diversity within the local GC2 population, including several events resulting in replacement of the capsule locus. We identified likely donors of the imported capsule locus sequences amongst the A. baumannii isolated on the same ward, suggesting that diversification was largely facilitated via reassortment and sharing of genetic material within the localized A. baumannii population.


Introduction
Over the past three decades, Acinetobacter baumannii has become increasingly implicated in outbreaks of multiply antibiotic-resistant hospital-acquired infections worldwide (Antunes et al., 2014;Bouvet & Grimont, 1987;Perez et al., 2007Perez et al., , 2011. The organism is now universally recognized as one of the six most problematic hospital-adapted antimicrobial-resistant bacterial pathogens, contributing the 'A' in the 'ESKAPE' pathogens (Rice, 2008). The majority of hospital-acquired A. baumannii infections are caused by two globally disseminated clones, known as global clones GC1 and GC2. In intensive care units (ICUs), ventilator-associated pneumonia (VAP) is the second most common hospital-acquired infection (Kalanuria et al., 2014), affecting up to a quarter of mechanically ventilated patients (Chastre & Fagon, 2002), and has an attributable mortality risk of *13 % (Melsen et al., 2013). VAP caused by antimicrobial-resistant A. baumannii is becoming increasingly common in industrialized countries. Antimicrobial-resistant A. baumannii have been isolated in up to 45 % of VAP cases in the ICU at the Hospital for Tropical Diseases (HTD), Ho Chi Minh City, Vietnam (Nhu et al., 2014).
A. baumannii are intrinsically resistant to chloramphenicol and florfenicol, and resistance to all other first-line antimicrobials (sulphonamides, tetracycline and ampicillin) has become largely ubiquitous amongst hospital-acquired A. baumannii since the 1980s (Bergogne-Bérézin & Towner, 1996). Resistance to cephalosporins and fluoro-quinolones emerged in the 1990s, leaving carbapenems (particularly imipenem) as the mainstay for empirical treatment (Bergogne-Bérézin & Towner, 1996). Imipenem first became available for the treatment of microbial infections in 1985 (Papp-Wallace et al., 2011). Coincidentally, imipenem-resistant A. baumannii were first isolated in 1985 from a hospital in which imipenem had never been used (Paton et al., 1993), with resistance later found to be conferred by the oxa23 gene (Donald et al., 2000). Thus, the mechanisms of resistance to imipenem already existed in the A. baumannii population well before the first use of imipenem (Paton et al., 1993;Poirel et al., 2008). Over the last decade, imipenem-resistant A. baumannii infections have been observed more widely, prompting the use of colistin for antimicrobial therapy (Garonzik et al., 2011). Although colistin is effective against most carbapenem-resistant A. baumannii, it is a last-resort treatment due to toxicity (e.g. Evans et al., 2013;Ordooei Javan et al., 2015;Spapen et al., 2011) and resistance to colistin in A. baumannii is also emerging (Cai et al., 2012;Qureshi et al., 2015). Being naturally competent (Wilharm et al., 2013) and able to form biofilms (Madsen et al., 2012;McQueary et al., 2012;Tomaras et al., 2003), homologous recombination is common in

Impact Statement
Infections caused by antibiotic-resistant bacteria pose a heavy burden on human health globally. Where antibiotics are in sustained usage, bacteria have been shown to rapidly evolve to become drug resistant. The problem of antibiotic-resistant bacterial infections in intensive care units (ICUs) is becoming increasingly common and understanding the means by which bacteria develop antibiotic resistance in the ICU may help to prevent or slow this process in the future, extending the life of new antibiotics. Acinetobacter baumannii is one of the most problematic bacterial species that cause antibiotic-resistant infections in ICU patients. In this study, we used genome sequencing to retrace the evolution of resistance to an important class of antibiotics (carbapenems) amongst A. baumannii bacteria circulating in an ICU ward in Vietnam. We found that after carbapenems were introduced as a common treatment for infections in the ICU in 2008, several different strains of A. baumannii that were susceptible to carbapenems in earlier years became resistant by acquiring new genes, which encoded carbapenemase enzymes that break down carbapenem-based drugs. These newly resistant strains became very common on the ward in later years, so that most infections could no longer be effectively treated using carbapenems. In finding strong evidence for the repeated local emergence of drug resistance in response to local usage of the drug, our findings provide a predictive framework for healthcare practitioners dealing with these types of infection.
A. baumannii populations; however, few genomic studies have investigated this in detail.
We recently investigated the aetiology of VAP in the ICU at the HTD (Nhu et al., 2014). Acinetobacter was the dominant genus implicated in VAP and we documented an increase in its isolation from 28 % of VAP during 2000to 46 % in 2010(Nhu et al., 2014. Approximately 15 % of VAP-associated Acinetobacter were resistant to imipenem during 2002-2007. Imipenem was approved for empirical treatment of VAP in this ICU in 2008, following which the frequency of imipenem resistance increased to 46 % of A. baumannii VAP isolates in 2009 and 86 % in 2010 (Nhu et al., 2014). Multilocus VNTR analysis (MLVA) and PCR profiling of carbapenemase genes indicated that this increase was largely attributable to A. baumannii sharing the same MLVA profile (MLVA6) and carrying the oxa23 gene (Nhu et al., 2014). Thus, we hypothesized that the dominant VAP-associated A. baumannii subtype constituted a single clone that had been introduced and then maintained on the ICU ward, likely due to sustained imipenem usage. However, MLVA data could not resolve ancestor-descendent relationships among the samples nor identify whether GC1 or GC2 were present.
Here, we used whole-genome sequencing to study the population of imipenem-resistant A. baumannii isolated from VAP on the HTD ICU during 2008-2012 and to investigate their relationship to imipenem-susceptible A. baumannii isolated from VAP or asymptomatic carriage on the same ICU ward in 2003-2007.

Methods
Bacterial isolates and DNA sequencing. To examine the relationships between A. baumannii carbapenemresistant and -susceptible isolates, we sequenced the genomes of isolates collected during two different studies (from two different time periods) in the same ICU ward at the HTD. This ICU is largely devoted to the treatment of tetanus patients (Nhu et al., 2014). The first study aimed to characterize asymptomatic colonization (carriage) of patients by potentially pathogenic commensal bacteria (including Acinetobacter spp., Klebsiella pneumoniae, Pseudomonas aeruginosa and Staphylococcus aureus) during 2003(Schultsz et al., 2013. The following swabs were collected from patients entering the ward within 48 h of admission: nasal swab ('n' or 'N'), axilla swab ('ax'), anal swab ('an') and skin swabs at the cannula tip point of entry ('c'). These swabs yielded 70 A. baumannii asymptomatic carriage isolates available for sequencing. A total of 12 A. baumannii VAP isolates from this period were identified in a review of laboratory inventory, so these were also sequenced for comparison. The second study (described in Nhu et al., 2014) aimed to understand the bacterial aetiology of emerging drugresistant VAP on the ward and reviewed bacterial isolates cultured from specimens from intubated patients via bronchoalveolar lavage (BAL) or tracheal washes (UV). From this latter study, a total of 77 A. baumannii isolated during 2008-2012 were available for sequencing. These isolates were all considered to be causative agents of VAP; no asymptomatic carriage isolates were collected during this time period. The list of all 159 A. baumannii isolates that were retrospectively combined for the present study is provided in Table S1.
Genomic DNA was extracted using a Wizard Genomic Extraction kit (Promega) and subjected to 96-plex barcoded 100 bp paired-end sequencing via Illumina HiSeq 2500, as described previously (Holt et al., 2012). Twelve of the infection isolates failed sequencing, leaving a total of 147 genomes for analysis. Sequence reads are available in the European Nucleotide Archive under project ERP001080. Individual accession numbers are provided in Table S1.
Antimicrobial susceptibility testing. Antimicrobial susceptibilities were determined at the time of isolation by the modified Kirby-Bauer disk diffusion method, as recommended by the Clinical and Laboratory Standards Institute (CLSI, 2011): piperacillin/tazobactam (100/10 mg), imipenem (10 mg), amikacin (30 mg), ceftazidime (30 mg), ceftriaxone (30 mg), ciprofloxacin (5 mg), gentamicin (30 mg) and colistin (10 mg). Mueller-Hinton agar and antimicrobial discs were purchased from Unipath. For colistin, imipenem, ceftazidime and ceftriaxone, MICs were determined by E-test based on the manufacturer's recommendations (AB Biodisk). The results were interpreted as resistant or sensitive according to current Clinical and Laboratory Standards Institute guidelines. Escherichia coli ATCC 25922 was used as the control for these assays.

SNP identification and phylogenomic analysis.
Whole-genome sequences were inferred by mapping Illumina reads against the 3 940 614 bp reference genome of A. baumannii GC2 strain 1656-2 (GenBank accession number CP001921.1) using the RedDog mapping pipeline (https://github.com/katholt/RedDog). Briefly, RedDog uses Bowtie version 2.2.3 (Langmead & Salzberg, 2012) to map reads to the reference sequence, then highquality SNPs with Phred quality score i30 are extracted from the resulting alignments using SAMtools version 0.1.19. SNPs were filtered to exclude those with five or more reads mapped or with w2.5 times the mean read depth (representing putative repeated sequences), or with heterozygous (ambiguous) allele calls. For each SNP that passed these criteria in any one isolate, consensus allele calls for the SNP locus were extracted from all genomes (ambiguous base calls and bases with Phred quality v20 were treated as unknown alleles and represented with a gap character). SNPs with confident homozygous allele calls (Phred quality i20 calculated by SAMtools consensus) in w90 % of the A. baumannii genomes (representing a 'soft' core genome of common A. baumannii sequences) were concatenated to produce an alignment of 217 290 variant sites.
Maximum-likelihood (ML) analysis was performed using the GTR model with a gamma distribution of rate heterogeneity (GTR+G) using RAxML version 8.1.16 (Stamatakis, 2014). We modelled rate heterogeneity due to the high level of recombination known to occur in A. baumannii (Madsen et al., 2012). We did not apply an ascertainment bias correction because the variable sites under analysis were those obtained from whole-genome sequencing and not simply from typing of known variable sites. ML analyses were performed 10 times with 100 bootstrap replicates per run. Likelihood scores were compared between runs to confirm that all runs had converged on a similar likelihood; we selected the single tree with the highest likelihood as the best tree.
From the SNP alignment of 147 isolates, we extracted a SNP alignment for the 66 GC2 isolates including only the 13 950 SNPs that varied amongst these GC2 genomes. RAxML analyses were repeated 10 times with 100 bootstrap replicates per run. For these analyses we used empirical base frequencies and a GTR best-fit model of nucleotide substitution. The substitution model was estimated under the Akaike information criterion (AIC; Akaike, 1974) using MrModeltest version 2.3 (https://github.com/nylander/ MrModeltest2) in PAUP * (http://paup.csit.fsu.edu/). Temporal signal ('clock-likeness') captured in the ML trees was assessed using Path-O-Gen version 1.4 (Drummond et al., 2012), which performs linear regression of root-to-tip ML branch distances (y) on year of isolation (x). Tree topologies were rooted prior to temporal analysis using an outgroup from clonal complex CC10.
Visualization of recombination. In order to visualize clusters of SNPs that were likely imported into the A. baumannii genomes via recombination events, we generated plots of SNP density distributions along the GC2 strain 1656-2 reference genome using R (version 3.2.1). Taking the first isolate in the alignment as a baseline sequence, we calculated the number of SNPs between the baseline sequence and each other isolate, in 1 kbp windows. The resulting values, which represent density of SNPs per 1 kbp along the reference genome, were plotted as a heatmap, ordered against the ML tree and output to a PNG image file. This process was applied iteratively, taking each isolate in turn as the baseline genome. The resulting sequence of PNG files was combined into an image-sequence movie file (File S1).
Additionally, we inferred split networks to assess tree-likeness of the phylogenetic structure in the SNP alignments using SplitsTree4 (Huson & Bryant, 2006) with default settings (uncorrected-p distance, cyclic splits computed using NeighbourNet and ordinary least-squares variance).
Recovering a recombination-free phylogenetic tree for GC2. We made a number of attempts to detect recombinant sites within the GC2 population. These analyses were performed on a pseudo-whole-genome alignment for GC2 in which each isolate genome was reconstructed by copying the reference sequence and replacing the SNP sites with alleles from the isolate in question. Note that after the recombination detection analysis, invariant sites were removed and the alignment was reduced to those common genomic sites that were present in i90 % of the isolates under study.
Recombinant sites were identified and removed using: (i) ClonalFrameML (Didelot & Wilson, 2015) with default settings and 100 bootstrap replicates, and (ii) v1. 1.2 (Croucher et al., 2015) for 20 iterations with default settings. The resultant recombination-free alignments of variant sites were inspected using SplitsTree4 to build split networks. In addition, as we expected a strong temporal signal within the GC2 clone similar to that observed within other Gram-negative clones, such as Shigella sonnei (Holt et al., 2012), Vibrio cholerae (Mutreja et al., 2011) or Salmonella enterica Typhimurium (Okoro et al., 2012), we assessed the strength of temporal signal before and after the various recombination detection procedures by inferring ML trees from the variable sites in the alignments using RAxML and then analysing these in Path-O-Gen version 1.4 (Drummond et al., 2012). As the tree obtained after Gubbins analysis recovered the strongest temporal signal, we utilized the Gubbins-filtered alignment for further analysis.
We performed time-calibrated Bayesian phylogenetic inference on the post-Gubbins SNP alignment (457 sites) using BEAST version 1.8.0 with the AIC-selected GTR+G best-fit substitution model, empirical base frequencies, a strict molecular clock and a coalescent (constant size) tree prior. BEAST Markov Chain Monte Carlo (MCMC) analyses were repeated eight times starting each run from a random seed and a random tree. MCMC analyses were continued for 10|10 6 iterations, sampling every 1000 iterations. After removing a burn-in of 100 samples (i.e. 1|10 6 MCMC iterations) from each file, logs were combined using LogCombiner version 1.8 to give 72 000 post-burn-in samples (i.e. 72|10 6 MCMC iterations). Effective sample sizes (ESSs) for all parameters of interest were ww1000 as determined by analysis with Tracer version 1.6. Burn-in was 10 % of the chain length. Postburn-in trees were combined using LogCombiner and summarized using TreeAnnotator version 1.8. The common ancestor summary tree was then recovered from the set of trees (n572 000 trees) in the posterior sample so that every clade in the summary tree was assigned an age using the mean of the clade age in all posterior trees (Heled & Bouckaert, 2013).
MLST analysis and detection of acquired resistance genes from sequence reads. SRST2 (Inouye et al., 2014) was used with default settings to identify MLSTs and acquired resistance genes directly from reads. For MLST analysis, we used both the Institut Pasteur and University of Oxford MLST databases (http://pubmlst.org/ abaumannii/; downloaded 3 June 2015). All sequence types reported in the text are based on the Pasteur scheme; however, those inferred using the Oxford scheme are also provided for comparison in Table S1. For acquired resistance genes we used the SRST2-formatted version of the ARG-ANNOT database (Gupta et al., 2014), distributed with SRST2 (Inouye et al., 2014).
Genetic context and copy number of carbapenemase genes. The genetic contexts of carbapenemase genes (oxa23, oxa40, oxa58 and ndm1) were explored in detail via manual inspection of the SPAdes assembly graphs using Bandage version 0.7.0 . All isolates carrying the oxa23 gene were confirmed to carry the gene either within Tn2006 (Corvec et al., 2007) or a variant of Tn2008 (Adams-Haduch et al., 2008). These transposons are mobilized by ISAba1 (GenBank accession number AY758396) and it was not possible to determine, using the Illumina short reads, which of the multiple insertions of ISAba1 in each genome were associated with the transposon and which represented other ISAba1 insertions. However, we were able to investigate the ISAba1 insertion sites in detail using ISMapper  to identify all insertion sites for ISAba1 relative to the reference chromosome sequence (GC2 strain 1656-2; GenBank accession number CP001921.1) and using Bandage to investigate whether ISAba1 was present within additional locations, including the Aci6 plasmid (pA85-3 from a GC1 strain; GenBank accession number KJ493819.1). As oxa23 was only found in the chromosome, copy number of oxa23 was inferred from the distribution of ratios of oxa23 mapping depths to respective mean chromosome mapping depths. In the cases of isolates carrying oxa58, inspection of the assembly graphs in Bandage confirmed that all oxa58 genes were flanked by ISAba3. In most cases there was a single insertion site for ISAba3. The surrounding sequence was extracted and annotated using the automated annotation pipeline Prokka version 1.10 (Seemann, 2014), followed by manual annotation of acquired resistance genes (by comparison with the ARG-ANNOT database; Gupta et al., 2014) and insertion sequences (by comparison with the ISfinder database; Siguier et al., 2006). The genetic contexts of oxa40 and ndm1 were investigated using the same approach.
Analysis of the capsule locus in GC2. We used BLASTN version 2.2.31+ (Altschul et al., 1990) to search the assemblies of GC2 genomes for the capsule biosynthesis locus (KL) flanking genes mviN (GenBank accession number CP001182, locus tag AB57_0088) and IldP (GenBank accession number CP001182, locus tag AB57_0116). For the 12 genomes in which the KL was split across multiple contigs, we used Bandage version 0.7.0  to inspect the assembly graph and to extract the sequence for the contiguous path through the KL. For each of the KL types identified in more than one GC2 isolate (KL2, KL49 and KL58), intra-KL sequence alignments were inferred using MAFFT version 7.245 (Katoh & Standley, 2013). Intra-locus diversity was then assessed by counting the variable sites in each of the alignments. Synteny alignments and sequence comparisons between KL2 and KL58 sequences, which share extensive sequence similarity, were made using BLASTN and Mauve version 2.3.1 (Darling et al., 2004), and visualized using genoPlotR version 1.1 (Guy et al., 2010) and ACT (Carver et al., 2005). Assessment of sharing and reassortment of KLs within GC2, or between GC2 and CC10, via horizontal transfer was made after mapping reads to the KL2 reference sequence using RedDog, extracting SNPs as described above for phylogenetic analysis and visualizing recombination as described above (File S2). KLs were annotated as described in Kenyon & Hall (2013).
Ancestral state reconstruction of GC2 capsule types. To infer ancestral KL types of lineages in the GC2 tree, we performed multistate ancestral state reconstruction using BayesTraits version 2.0 (Pagel et al., 2004). We used the multistate implementation in BayesTraits to reconstruct the evolution of discrete KL types (four states: KL2, KL32, KL49 or KL58) on the set of phylogenetic trees (i.e. 72 000 pooled post-burn-in trees from the BEAST analysis). We used the MCMC implementation with a reversible jump and an exponential prior with a mean of 10. The analysis was repeated eight times, continuing each run for 10|10 6 iterations and sampling from the posterior every 1000 iterations. Run log-files were examined using R.

Results and Discussion
Population structure of A. baumannii in the ICU The genomes of 147 A. baumannii isolates from VAP (n577) and asymptomatic carriage (n570) were available for analysis. MLST analysis of the genome sequences, using the seven-locus Institut Pasteur scheme, identified 21 sequence types (Table S1). Genome-wide SNP analysis identified 217 290 core genome SNPs, which were subjected to ML phylogenetic analysis. The core genome phylogeny and MLST structure identified similar population structures, and were used to classify the isolates into six strongly supported clonal groups plus an additional group of diverse strains (n516) (Fig. 1a). The six strongly supported clonal groups were: GC2 (ST2 and its single locus variants: ST570 and ST571), GC1 (ST1), CC10 (ST10 and its single-locus variants: ST23 and ST575), CC16 (ST16 and its single-locus variant: ST452), ST25 and ST686. The isolates previously typed as MLVA6 and MLVA8 (Nhu et al., 2014) belonged to GC2 and CC10, respectively.
The temporal distribution of various genotypes is shown in Fig. 1(b). Of the sequenced isolates from 2003 to 2007 (n581), 86 % were from asymptomatic carriage. These early carriage isolates were genetically diverse, and included i10 isolates each from ST686 (32 %), GC2 (24 %), CC10 (20 %) and ST25 (12 %). The VAP isolates from this period (n511) fell within these same clonal groups, with the exception of a single VAP isolate of ST375, which was not represented within the asymptomatic carriage isolates. These data are consistent with the notion that carriage and VAP-associated isolates derive from the same population of A. baumannii on the ward, with opportunistic VAP infections arising following contact with a contaminated source (e.g. ventilator tubing). The later VAP isolates (n566), collected from 2009 to 2012, exhibited less genetic diversity and were dominated by GC2 (71 %), with smaller numbers of CC10 (9 %), CC16 (6 %) and occasional less commonly isolated strains (ST52, ST108, ST109, ST142 and three novel sequence types: ST754, ST755 and ST756).
Most isolates (75 %) were resistant to amikacin, gentamicin, piperacillin/tazobactam, ciprofloxacin, ceftazidime and ceftriaxone, including isolates from all of the common clonal groups (GC1, GC2, CC10, CC16, ST109, ST25, ST52 and ST686; see Table S1). Fewer than 6 % of isolates were susceptible to two or more of these antibiotics (Table S1). Our previous review of imipenem-resistant VAP showed 15 % of VAP-associated Acinetobacter were resistant to imipenem during 2002-2007, rising to 86 % in 2010 (Nhu et al., 2014). Eleven A. baumannii VAP isolates from 2002 to 2007 were available for sequencing in the present study and all of these were susceptible to imipenem (Fig. 2). Of the 70 carriage isolates from this period, all were imipenem-sensitive except for two GC1 isolates from 2007, which, notably, were the only GC1 isolates in this collection (MICs are shown in Fig. 2). Of the 66 post-2008 VAP isolates that were successfully sequenced, only nine (14 %) were susceptible to imipenem. The imipenem-resistant VAP isolates belonged to a variety of clones: GC2 (n547), CC10 (n55), CC16 (n51), and single isolates from ST52, ST108 and two novel singleton sequence types (ST755 and ST756) (Fig. 2). As GC2, CC10 and CC16 were also detected amongst the earlier isolates from the same ward, these data suggest that imipenem-resistant A. baumannii emerged on the ward on multiple occasions, within clones that had been previously circulating on the ICU.

Genetic determinants of imipenem resistance
We identified at least one carbapenemase gene (oxa23, oxa40, oxa58 or ndm1) in each imipenem-resistant isolate (Table S1). These were present in a variety of different genetic contexts and in variable copy numbers (Figs S1 and S2, summarized in Table 1). The oxa40, oxa58 and ndm1 genes were rare and mostly plasmid associated (Table 1, Note S1). We were able to completely assemble two novel circular plasmids carrying carbapenemase genes ( Fig. S3; p255n_1 containing oxa58, deposited under Gen-Bank accession number KT852971, and pBAL_204 containing oxa40, deposited under GenBank accession number KT946773).
The most common carbapenemase gene was oxa23, which was present in 93 % of imipenem-resistant isolates, including 47 GC2 isolates, five CC10 isolates and the two GC1 carriage isolates from 2007 (Fig. 1b). The oxa23 gene was generally encoded within a Tn2006 transposon, present in one to three copies. Four GC2 isolates carried an oxa23 variant (with a SNP leading to an amino acid substitution L125V) located within a Tn2008-like transposon (submitted as Tn2008VAR under GenBank accession number KT852972) that was flanked by 9 bp target site duplications, consistent with transposition via ISAba1 and a novel insertion sequence deposited under ISfinder as ISAba33 (Fig. 3). As the oxa23 transposons were mobilized by insertion sequences, which appear in multiple copies in the GC2 genomes, it was not possible to determine the precise insertion sites of the oxa23 transposons in each genome. For most oxa23 isolates, the gene was present in one, two or three copies (Fig. S1) and the assembly data indicated that all the ISAba1 insertions were chromosomal (Table 1). In contrast, the two GC1 isolates had low copy number for oxa23 (0.4-0.5) and harboured an Aci6-type plasmid carrying AbaR4 within which a copy of ISAba1 was present. This is consistent with plasmid carriage of the oxa23 genes in these isolates, similar to an earlier report of GC1 isolates carrying Tn2006 within AbaR4 in an Aci6 plasmid (Hamidian et al., 2014).

Emergence of imipenem resistance in A. baumannii GC2
The genomic data showed that the majority of imipenemresistant VAP isolates belonged to GC2 (82.5 %) and all Repeated emergence of carbapenem-resistant A. baumannii such strains carried the oxa23 gene, consistent with our previous work linking imipenem resistance to an oxa23positive clonal group with VNTR pattern MLVA6 (Nhu et al., 2014). This pattern could result equally from the introduction of a new resistant GC2 subclone or from de novo emergence of resistance within the previously circulating GC2 population. We aimed to use the wholegenome data to distinguish between these possibilities and also sought to investigate whether the acquisition of oxa23 occurred only once or on multiple occasions.
The genome-wide ML tree for the GC2 isolates showed that the imipenem-resistant VAP isolates fell into multiple groups, many of which were genetically distant from the earlier asymptomatic carriage isolates (File S1). However, it was clear from the distribution of SNPs in the alignment that there was substantial recombination affecting the GC2 genomes (Fig. S4a, File S1). As recombination can confound phylogenetic inference and intra-clonal relationships, we sought to recover a better estimate of vertical evolutionary signal within the GC2 isolates by first identifying and then removing recombinant SNPs prior to phylogenetic inference. Recombinant SNP detection was performed using both ClonalFrameML (Didelot & Wilson, 2015) and Gubbins (Croucher et al., 2015). The Gubbins analysis yielded an apparently reliable recombination-filtered alignment of 457 SNPs, which displayed no spatial clustering of SNPs, a tree-like split network structure and strong clock-like (linear) temporal signal (r 2 50.73, Fig. S4c; compare with the ClonalFrameMLfiltered alignment in Fig. S4b or the raw data in Fig. S4a). Based on these estimates of 457 non-recombinant SNPs positions and comparing with the 13 493 recombinant SNPs, we estimate the ratio of SNPs introduced into the local GC2 population by recombination to those introduced by substitution mutation (r/m) to be *30. On the Gubbins recombination-filtered alignment, we performed a timecalibrated Bayesian analysis using BEAST, which yielded a substitution rate of 9.7 SNPs per year [95 % highest posterior density (HPD) 8-11, ESS ww1000].
In the resulting recombination-free GC2 BEAST tree (Fig. 4), the carbapenem-susceptible carriage isolates were divided into two well-supported lineages, which * Ratio of mean read depths for the carbapenemase gene versus known chromosomal genes (see Fig. S2). D Carbapenemase gene copy number estimated from the depth ratio (see distributions in Fig. S1). d Possible insertion site(s) for the carbapenemase-carrying mobile element (details of ISAba1 sites are shown in Fig. S5).
diverged around 2002 (95 % HPD 2001-2003, effective sample size ww1000). All of the imipenem-resistant GC2 VAP isolates belonged to lineage 2; however, these were paraphyletic and formed five distinct, strongly supported subclades (labelled A-E, Fig. 4). The most recent common ancestor (MRCA) of all imipenem-resistant GC2 VAP isolates was the same as the MRCA of the imipenem-susceptible lineage 2 carriage isolates, indicating that the resistant clades emerged from within the previously circulating susceptible population of lineage 2. The ML phylogeny showed a similar topology with 100 % bootstrap support for these subclades (Fig. S4c).
All imipenem-resistant GC2 isolates carried oxa23 and each subclade had a unique clade-specific ISAba1 insertion site in the chromosome (Table 1, Fig. S5), consistent with unique insertions of Tn2006 or the Tn2008VAR transposon in each subclade. The 95 % HPDs for the MRCA date of each subclade are consistent with these acquisition events occurring around or after the approval of imipenem for empirical treatment of VAP on this ICU, in 2008 (Fig. 4). Copy number of the oxa23 gene was apparently associated with MIC: within subclade 2B (i.e. in the isolates containing the oxa23 amino acid sequence variant in Tn2008VAR), isolates with one copy of oxa23 displayed an MIC of 32 mg ml 21 and isolates with two copies displayed an MIC of 64 mg ml 21 ; outside of subclade 2B, isolates with at least a single copy of Tn2006 displayed an MIC of 64 mg ml 21 . These data indicate that imipenem resistance arose in situ within the local GC2 population on at least five separate occasions, coinciding with the introduction of imipenem for empirical use on the ICU.

Capsule replacement in GC2 A. baumannii
The regions we identified as being affected by recombination in the local GC2 population included the KL, which encodes capsule biosynthesis. Vaccines that target capsule polysaccharides are an alternative to antibiotics in the control of pan-resistant Gram-negative bacterial infections. Notably, we detected four distinct KLs amongst the GC2 isolates, consistent with capsule replacement induced via homologous recombination. Amongst the 66 GC2 isolates, 24 carried the KL2 gene cluster (Kenyon & Hall, 2013), 18 carried KL49 , 23 carried a novel gene cluster that we designated KL58 (Figs 5a and 6) and one carried KL32 (same sequence as PSgc21 in Hu et al., 2013). Variation in KLs was mirrored by allelic variation in the gpi locus, which is utilized in the University of Oxford MLST scheme (Table S1) and is located within the KL. The genetic structures of these four KLs (Fig. 6) revealed distinct saccharide processing genes in each locus; hence, they are expected to produce distinct surface polysaccharides, although to date the precise structures have been determined only for K2 and K49 Senchenkova et al., 2014;Vinogradov et al., 2014).
Nearly all of the early GC2 isolates carried KL2, which was not represented in the non-GC2 isolates. Treating KL type as a discrete trait, ancestral state reconstruction on the GC2 BEAST tree set identified KL2 as the most likely type in the MRCA of all GC2 genomes in this collection (Fig. S6) Repeated emergence of carbapenem-resistant A. baumannii did not vary markedly between our GC2 KL2 isolates, consistent with shared inheritance of the KL2 region from a common ancestor carrying this locus and flanking sequence. A de novo assembly and alignment of all KL2 sequences identified an insertion of the ISAba1 transposase next to wzc, the first gene of KL2, in the GC2 lineage 1 isolates; however, this is unlikely to affect expression of KL genes as it is located downstream of the wzc ORF (Fig. 5a).
The majority of the remaining non-KL2 GC2 isolates harboured either the KL49 or KL58 capsule genes (Fig. 5b). Notably, both KL49 and KL58 were also identified in CC10 carriage isolates and VAP isolates from 2003 to 2006, providing a potential source for these sequences within the local A. baumannii population. Detailed inspection of the KL and surrounding sequences in GC2 and CC10 isolates indicated that the KL49 GC2 strains most likely resulted from a single importation of 750 kb (spanning the KL49 locus) into the ancestor of GC2 clade 2E, derived from co-circulating CC10 KL49 strains (or from a third donor that also donated KL49 and flanking sequence to the co-circulating GC2 and CC10 clones) (Fig. 5b). Two variants of KL58 were identified: one in GC2 and the other in CC10. The GC2 KL58 sequence appears to be a hybrid of KL2 and the CC10 KL58 sequence (Fig. 5a), resulting from a local homologous recombination event, in which 24 kbp towards the 39 of the KL58 locus was imported from CC10 (or from a third donor that also donated KL58 and flanking sequence to the co-circulating GC2 and CC10 clones). The GC2 KL58 isolates formed a single monophyletic group (subclades 2A-C, MRCA around 2005; Fig. 5b) con-  Fig. 4).

Day of isolation
tical to the hybrid KL58 sequence found in GC2 2A-C isolates, suggesting it was acquired from a GC2 strain.
Capsule replacement is not well studied amongst Gramnegative bacteria, but has recently been observed in multiply antibiotic-resistant epidemic clones of A. baumannii (Kenyon & Hall, 2013;Kenyon et al., 2014Kenyon et al., , 2015, K. pneumoniae CC258 (Wyres et al., 2015) and E. coli ST131 (Alqasim et al., 2014). Potential drivers of capsule diversification include host immunity, phage predation and also antimicrobial resistance as it has recently been shown that capsule variation in A. baumannii is associated with variation in susceptibility to multiple antimicrobials (Geisinger & Isberg, 2015). Here, two of the KL replacement events resulted in novel clades, i.e. KL58 GC2 (2A-C) and KL49 GC2 (2E), that were successful in persisting within the local population for several years. The local maintenance of KL2 GC2 alongside these newer variants indicates that if there was selection for capsule variation in the local population, this was not strong enough to drive clonal replacement of KL2 GC2 in this instance.

Conclusions
Our analyses indicate that the recent surge of imipenemresistant VAP in this ICU was attributable to multiple imipenem-resistant clones, the majority of which arose from within the local A. baumannii population via the acquisition of mobile carbapenemase genes. Whilst precise details of imipenem usage are not available, these acquisition events occurred at around the same time at which imipenem was approved for empirical treatment of VAP on the ward, consistent with the hypothesis that the change in treatment strategy drove rapid changes in the local population of A. baumannii and resulted in the repeated emergence of multiple imipenem-resistant subclones from within the local population. This scenario contrasts with recent genomic investigations of epidemics of carbapenem-resistant Gram-negative clones in other hospital settings, which have reported transmission of a singular resistant clone (Mathers et al., 2015;Snitkin et al., 2012), competition between resident and incoming resistant clones (Brañas et al., 2015) or transmission of a resistant plasmid amongst diverse bacterial hosts (Mathers et al., 2011(Mathers et al., , 2015. Diversification of capsule types within GC2 appears to have commenced at around the same time as the emergence of imipenem resistance in the population and was driven by homologous recombination. The fact that we were able to identify likely donors for two of the imported capsule loci amongst previously circulating strains of the second most common A. baumannii clone on the ward (CC10) is notable. It suggests that whilst the population of A. baumannii on the ward is diverse, consisting of multiple strains, this population may nevertheless have remained 'closed', with evidence of reassortment between strains, but limited evidence for the introduction of new genetic material -a finding consistent with that observed in Enterococcus faecium by van Hal et al.  Fig. 6. Genetic structure of KLs identified in the GC2 A. baumannii isolates. Gene clusters are drawn to scale against KL2 reference GenBank accession number KJ459911.KL32, KL49 and KL58 have been deposited under GenBank accession numbers KT359615, KT359616 and KT359617, respectively. KL names are indicated on the left. Arrows indicate protein-coding genes, showing the direction of transcription and labelled with their assigned gene names; an asterisk indicates an internal frameshift. Genes are coloured by the predicted functional group of their gene product according to the legend. Grey shaded blocks indicate sequence homology. sequence analysis to tease apart fine-scale evolutionary events within bacterial pathogen populations, as our earlier MLVA and PCR analyses could not distinguish the presence of distinct subclades of GC2 that had independently evolved imipenem resistance.