Within-patient evolution of plasmid-mediated antimicrobial resistance

Abstract

Antimicrobial resistance (AMR) in bacteria is a major threat to public health, and one of the key elements in the spread and evolution of AMR in clinical pathogens is the transfer of conjugative plasmids. The drivers of AMR evolution have been extensively studied in vitro, but the evolution of plasmid-mediated AMR in vivo remains poorly explored. Here, we tracked the evolution of the clinically-relevant plasmid pOXA-48, which confers resistance to the last-resort antibiotics carbapenems, in a large collection of enterobacterial clones isolated from the gut of hospitalised patients. Combining genomic and experimental approaches, we first characterized plasmid diversity and the genotypic and phenotypic effects of multiple plasmid mutations on a common genetic background. Second, using cutting-edge genomic editing in wild-type multidrug resistant enterobacteria, we dissected three cases of within-patient plasmid-mediated AMR evolution. Our results revealed compensatory evolution of plasmid-associated fitness cost, as well as the evolution of enhanced plasmid-mediated AMR, in bacteria evolving within the gut of hospitalised patients. Crucially, we observed that the evolution of pOXA-48-mediated AMR in vivo involves a pivotal trade-off between resistance levels and bacterial fitness. This study highlights the need to develop new evolution-informed approaches to tackle plasmid-mediated AMR dissemination.

Introduction

Antimicrobial resistance (AMR) in bacteria has emerged as a major global threat to public health¹. AMR is particularly concerning in clinical settings, where nosocomial infections increase mortality rates among hospitalised patients and raise the costs associated with infection control and management². The gut microbiota of patients is one of the most important hotspots of AMR dissemination and evolution³, and a crucial element in this process is the transfer of conjugative plasmids – circular DNA molecules that replicate independently of the bacterial chromosome and can transfer horizontally between bacteria⁴.

Numerous studies in recent years have characterized the evolution of plasmid-mediated AMR, expanding our understanding of how AMR plasmids evolve and persist in bacterial populations. AMR plasmids dramatically enhance bacterial fitness in the presence of antibiotics, and plasmid-mediated resistance can further evolve through changes in plasmid copy number (PCN)^5–7, mutations or duplications of plasmid-encoded AMR genes^8,9 or interactions with chromosomal mutations¹⁰. However, in the absence of antibiotics, plasmid-induced physiological alterations frequently lead to a decrease in bacterial fitness, a phenomenon known as plasmid cost^11,12. This cost can be mitigated over time through compensatory mutations in the plasmid or chromosome^13–15. Remarkably, previous studies showed that the costs associated with AMR plasmids mainly arise from the expression of resistance genes^11,16,17.This insight suggests that bacteria carrying AMR plasmids probably experience a trade-off between fitness in the presence and absence of antibiotics (fitness-resistance trade-off)^18,19. Despite the importance of these earlier studies, current understanding of the evolution of plasmid-mediated resistance derives almost entirely from highly controlled experiments conducted in vitro. The lack of access to suitable bacterial collections of clinical origin, together with the arduousness of performing genetic manipulations with wild-type, multidrug-resistant bacterial isolates, has prevented study of the evolution of plasmid-mediated AMR in clinically relevant real-life scenarios.

Here, we tracked the evolutionary dynamics of plasmid-mediated AMR in the gut microbiota of hospitalised patients. We focused on the widespread pOXA-48-like conjugative plasmids, which constitute one of the most relevant plasmid groups in clinical settings in Europe^20,21. pOXA-48-like plasmids are found in the order Enterobacterales, giving rise to carbapenem-resistant enterobacteria, which were recently reported to be the fastest-growing resistance threat in Europe²². We used a previously characterized collection of 224 pOXA-48-carrying enterobacteria isolated over a two-year period from more than 9,000 hospitalised patients at the Ramon y Cajal University Hospital in Madrid, Spain (R-GNOSIS collection, Extended Data Fig. 1. For the characterization of pOXA-48-carrying isolates in R-GNOSIS, see^23–25). We studied multiple pOXA-48 variants carrying distinct mutations and elucidated the evolution of specific associations between pOXA-48 and wild-type enterobacteria in the gut microbiota of three patients. Our results revealed that the in vivo evolution of pOXA-48-mediated resistance is shaped by interplay between AMR and plasmid-induced fitness costs.

Results

Analysis of pOXA-48 plasmid variants

To identify mutations potentially associated with plasmid-mediated AMR evolution, we characterized the genomes of all pOXA-48-like plasmids in the R-GNOSIS collection. Comparison of the full sequences of all 224 pOXA-48-like plasmids, identified a total of 35 plasmid variants (PVs), defined as pOXA-48-like plasmids carrying any SNP or insertion/deletion (indel) compared with the most common variant (PV-I), which is present in ~67% of the isolates in the collection (Figure 1, Extended Data Fig.1, Supplementary Data 1 and Methods). Therefore, we focused on the study of a single plasmid, but with multiple genetic variants (35) across the collection.

Figure 1. pOXA-48 plasmid variants tested in E. coli J53.

A) Representation of pOXA-48 variants (PVs) A-N in concentric circles (for variant details see legend). The gene map in the outer circle represents the most common variant (PV-I, used as a reference and highlighted by grey shading in panels B-E). Arrows indicate the reading frames, and colours indicate gene functional classification. Gene names are indicated in the outer circle, and the names of genes showing mutations are represented inside boxes. B) Box plots showing relative fitness (w) of pOXA-48-carrying E. coli J53 relative to plasmid-free J53. Horizontal lines inside boxes indicate median values, the upper and lower hinges correspond to the 25^th and 75^th percentiles, and whiskers extend to observations within the 1.5 x the interquartile range (IQR). Individual points show independent replicates (n=6). C) Resistance to ertapenem (ERT) in plasmid-free and plasmid-carrying E. coli J53, represented as the 90% inhibitory concentration (IC90) in mg/L. Bars indicate the mean values and dots indicate individual replicates (n=10 for PV-K & PV-E and n=5 for the remaining PVs). Black bars represent the standard error of the mean. D) Conjugation rates of different PVs in E. coli J53 (in log₁₀ scale, n=14 for PV-I, n=9 for PV-J, n=8 for PV-E & PV-K, n=3 for PV-A, and n=6 for the remaining PVs), represented as boxplots as in B. E) Plasmid copy number of different PVs in E. coli J53, represented as boxplots as in B (n=6). Asterisks in panels B-E indicate significance for the comparison of each PV with PV-I (two-sided pairwise comparison Wilcoxon rank-sum exact test with FDR correction P<0.01 after two-sided Kruskal-Wallis test). F) Correlation between relative fitness (w) and resistance to ertapenem (mean IC90) in PV-carrying E. coli J53. Individual points indicate the mean value, and lines represent the standard error of the mean IC90 and the propagated standard error of the relative fitness. The size of each point is proportional to PCN in J53. The diamond represents the plasmid-free J53 values, which were not included in the correlation. Individual PVs are indicated by letters. The red dashed line indicates the regression and the gray-shaded zone covers the 95% confidence interval.

We next studied the phenotypic and genotypic effects of a selection of 14 of the PVs (Figure 1A). The 14 pOXA-48 variants were selected based on the following criteria: i) PVs carrying non-synonymous mutations/deletions covering a wide representation of different genes and functions and avoiding PVs with redundant mutations in the same genes; ii) PVs carrying insertions and large rearrangements; and iii) PVs with intergenic mutations near to housekeeping plasmid genes, such as genes involved in replication, conjugation or partition. The 14 PVs were introduced into the Escherichia coli J53 strain²⁶ (K12 derivative), used as a common isogenic bacterial host to specifically dissect plasmid effects. The genomes of J53 carrying the different PVs were sequenced to confirm plasmid presence and the isogenic nature of the transconjugants (Supplementary Data 2, see Methods). The following phenotypic and genotypic variables were examined in each transconjugant and in plasmid-free J53: i) bacterial fitness, assessed from growth curves and competition assays; ii) plasmid conjugation rate; iii) antimicrobial resistance; and iv) plasmid copy number (PCN).

The PVs produced a variety of phenotypes in J53 (Figure 1 B-E and Extended Data Fig. 2). For example, although the fitness effect of most PVs was similar to that of PV-I (the most common PV), two PVs were associated with a large decrease in cost (two-sided Kruskal-Wallis rank-sum test chi-sq.=64.37, df=13, P<0.001, followed by pairwise comparison two-sided Wilcoxon rank-sum exact test with FDR correction P<0.01, see Methods for details on statistical analysis). One of these variants (PV-A) had a deletion in the carbapenemase gene bla_OXA-48 that abolished plasmid-mediated AMR (Figure 1C). The other one (PV-B) carried two deletions: i) a small deletion affecting the IS1 element upstream of bla_OXA-48 and ii) a ~13.5 kb deletion including genes involved in conjugation, associated with a conjugation-incompetent phenotype (Figure 1D). PV-D and PV-J both had lower conjugation rates than PV-I (Figure 1D; two-sided Kruskal-Wallis rank-sum test chi-sq.=56.58, df=13, P<0.001, followed by two-sided pairwise comparison Wilcoxon rank-sum exact test with FDR correction P<0.01) and carried mutations in conjugation-related genes: a nonsynonymous SNP in traY and an insertion in trbN (PV-D), and a nonsynonymous SNP in traU (PV-J) (Supplementary Data 1). PCN was significantly elevated in one PV, PV-K (Figure 1E), which carried a mutation upstream of the gene encoding the replication initiation protein RepA (two-sided Kruskal-Wallis rank-sum test chi-sq.=42.09, df=13, P<0.001, followed by two-sided pairwise comparison Wilcoxon rank-sum exact test with FDR correction P=0.025).

Pairwise analysis of the effects of AMR, PCN and conjugation rates on plasmid-associated fitness costs in J53 revealed a clear trade-off between AMR and fitness costs (Figure 1F, Spearman’s rank correlation S=779.7, rho=-0.7136, P=0.004). PVs conferring AMR were associated with a high fitness cost (27%-34% reduction in relative fitness), whereas the two PVs which conferred low or no AMR imposed only a small fitness cost (<4 % reduction in relative fitness). Neither conjugation nor PCN showed a significant association with plasmid fitness costs (Spearman's rank correlation, P>0.25, Extended Data Fig. 3). Additionally, the results of a generalized additive model including the three variables confirmed that only AMR correlated with plasmid-associated fitness costs (GAM, AMR: df=3, P=0.043, PCN and conjugation rates: df=2 and P>0.25).

Screening within-patient pOXA-48 evolution

The above characterization of pOXA-48 variants suggested that PV mutations, and the resulting fitness-resistance trade-off, could contribute to the evolution of pOXA-48-mediated AMR in vivo. In the R-GNOSIS project, hospitalised patients colonized with carbapenemase-producing enterobacteria were sampled periodically, generating timelines of isolates that allowed us now to investigate the evolution of pOXA-48-mediated resistance in the gut of these patients. We screened the timeline of bacterial isolates collected from individual patients and selected cases suggestive of within-patient plasmid evolution (i.e. detection of plasmid mutations -change of PV- in the same bacterial host). In other words, we selected those patients from which multiple isolates of the same bacterium (based on core-genome identity, see Methods) were recovered carrying different PVs over time. Among the 121 patients colonized by pOXA-48-carrying enterobacteria, we identified three whose timelines matched these conditions: patients HKH, JWC, and WDV (Figure 2 and 3A; Supplementary Data 1; see Methods).

Figure 2. Screening the within-patient evolution of pOXA-48-mediated AMR.

A) Timelines of the isolation of pOXA-48-carrying enterobacteria from patients HKH, JWC, and WDV (see legend). Isolate features are detailed in the legend. Swab dates are indicated next to the timeline (day/month). pOXA-48-selecting antibiotic treatments are indicated in the timeline (MER, meropenem; ERT, ertapenem; AMC, amoxicillin + clavulanic acid). Each PVs is indicated by a letter. Species are indicated by letters and symbols (KPN and circle for K. pneumoniae; EC and square for E. coli). The multilocus sequence-type code is indicated next to the species label. Isolates in which the new PV was detected are indicated by blue type and larger size. The patient JWC timeline reveals the emergence of two K. pneumoniae isolates co-isolated on an agar plate supplemented with ertapenem 0.3 mg/L. B-D) Genetic relationship built using core-genome comparisons (midpoint rooted phylogenetic trees) of K. pneumoniae ST11 (n=85, B), K. pneumoniae ST307 (n=20, C), and E. coli ST10 and ST744 (n=12 & n=2 respectively, D) from the collection. Strain designation, patient codes, and PVs are indicated (see Supplementary Data 1). Isolates involved in putative cases of within-patient evolution are highlighted in blue. Bold lettering marks the isolate in which the novel variant was identified. The tree-scale indicates nucleotide substitution per site.

Figure 3. Characterization of the in vivo evolution of plasmid-mediated AMR.

A) Workflow used to investigate the within-patient evolution of pOXA-48-mediated AMR. B) Relative fitness (w) of each plasmid-bacteria combination compared with the plasmid-free strain (see Methods). Horizontal lines inside boxes indicate median values, the upper and lower hinges correspond to the 25th and 75th percentiles, and whiskers extend to observations within 1.5 × the IQR. Individual points represent independent replicates (n=18). Asterisks in panels B-D indicate significant differences (P<0.05 in pairwise comparison two-sided Wilcoxon rank-sum exact test with FDR correction in B and C, and Padj<0.05 by one-way ANOVA in D); ns, nonsignificant (Padj>0.05). C) Resistance to ertapenem (ERT) measured as IC90 (mg/L) of plasmid-free and plasmid-carrying combinations. Lines indicate median values and individual replicates are indicated by points (n=5). D) Plasmid copy number (PCN) of each PV, represented in boxplots as in B (n=6). E) Schematic representations of the trade-off between antibiotic resistance (median IC90) and relative fitness (median w) in the strains under study. Black arrows represent the trade-off and arrowheads indicate the PVs timeline (from ancestral to novel PV).

In the genomic analysis of pOXA-48-carrying bacteria from the R-GNOSIS collection, two lines of evidence strongly suggested that plasmid mutations emerged and were subsequently selected in specific clones in the gut microbiota of these three patients. First, comparison of the core genomes of all K. pneumoniae and E. coli isolates revealed a tight grouping of isolates potentially involved in within-patient evolution events (Figure 2B-D, <5 SNPs between all isolates for each clonal line, Supplementary Data 3). This result makes it very unlikely that the observations in patients HKH, JWC, and WDV were the result of independent colonization by clones carrying different PVs.Second, each of the novel PVs originating in these patients was restricted to an individual patient, and was not described in any other isolate in the collection (not even in other clones from the same patient). This result challenges the possibility that these new bacteria-PVs associations were generated by independent conjugation events.

Evolution of pOXA-48-mediated resistance in vivo

To characterize the within-patient plasmid-mediated AMR evolution, we studied the isolates carrying the novel PVs in each patient (from now on, identified with the patient code followed by an asterisk, Figure 3A). First, we cured the pOXA-48 PVs from these isolates using an in-house CRISPR-Cas9 system specifically designed to remove plasmids from multidrug-resistant enterobacteria (see Methods and Extended Data Fig. 4). Then, for each patient, we independently re-introduced both the ancestral PV (the initial PV present in the same clonal line in the same patient) and the novel PVs in these isolates. Crucially, we sequenced the genomes of the wild-type clones by combining long-read and short-read technologies, and resequenced the genomes of all strains after plasmid curing to ensure that no significant mutations occurred during the process (Figure 3A, Extended Data Fig. 4 & Supplementary Data 3). Once we had introduced the two PVs into each clone, we measured i) plasmid fitness effects, ii) antimicrobial resistance (to all beta-lactam antibiotics used for treatment in these patients), and iii) PCN for every clone (Fig. 3 B-D, Extended Data Fig. 4). The results of these analyses are presented in the following sections patient-by-patient in chronological order.

Patient HKH. Increases in PCN, AMR, and fitness costs

Four pOXA-48-carrying isolates were recovered from this patient (Figure 2A). Two isolates belonged to a K. pneumoniae sequence-type 628 (ST628) clone, and both carried the same plasmid variant (PV-E). The other two isolates belonged to a Escherichia coli ST744 clone. The first of the E. coli isolates also carried PV-E, while the second (HKH*) was recovered 8 days later and carried a different plasmid variant, PV-K, differing from PV-E by only a single base pair insertion upstream of repA. The in vitro genotypic characterization in E. coli J53 had revealed an association between PV-K and increased PCN (Figure 1E). Analysis of the effects of PV-E and PV-K in HKH* revealed that PV-K was present at a higher PCN (from 3 to 8 copies, one-way ANOVA F=61.42, d.f.=1, Padj<0.001). The high PCN of PV-K in HKH* was associated with increased AMR to ertapenem (two-sided Wilcoxon rank-sum test W=2, P=0.02) and meropenem (Extended Data Fig. 5A-B) but decreased fitness in the absence of antibiotics (two-sided Wilcoxon rank-sum exact test, W=306, P<0.001, Fig. 2 B-D, Extended Data Fig. 5A-B). Patient HKH’s clinical history revealed meropenem treatment before the isolation of HKH*, which carried the high-PCN PV-K.

Patient JWC. Loss of AMR and amelioration of plasmid cost

Over a 10-week period, six pOXA-48-carrying isolates were recovered from patient JWC (Figure 2A). Four of them belonged to a K. pneumoniae ST307 clone and the remaining two to an E. coli ST2600 clone. Five isolates carried the most common pOXA-48 variant, PV-I, but the K. pneumoniae isolate JWC* carried PV-A, which differs from PV-I by a 199 bp deletion starting 163 bp upstream the coding DNA sequence of the bla_OXA-48 gene. As in the J53 analysis, PV-A was associated not only with loss of resistance to ertapenem and amoxicillin-clavulanic acid in JWC* (Fig. 3C and Extended Data Fig. 5C-D), but also with a reduction in plasmid fitness costs compared with PV-I (two-sided Wilcoxon rank-sum test W=0, P<0.001 for both phenotypes). PV-A was also associated with a modest but significant decrease in PCN in JWC* (one-way ANOVA F=6.51, d.f.=1, Padj=0.029, Fig 3 B-D).

In R-GNOSIS, carbapenemase-producing enterobacteria were recovered by selective plating, and it was therefore difficult to understand how the carbapenem-susceptible JWC* isolate was obtained from this patient. To investigate this, we plated the original frozen JWC isolate stocks on agar with and without ertapenem. Antibiotic-containing plates inoculated with the JWC* frozen stock (but not the other isolates) contained large resistant colonies of K. pneumoniae ST307/PV-I surrounded by smaller susceptible colonies of isogenic K. pneumoniae ST307/PV-A (Figure 2A), an example of the phenomenon of cross-protection by secreted beta-lactamases known as satellitism²⁷. Sequencing of the entire genomes of three large and three satellites colonies confirmed that they were isogenic bla_OXA-48+ and bla_OXA-48-variants of the same K. pneumoniae ST307 clone. This results strongly suggests that two versions of the K. pneumoniae ST307 clone, with the different PVs, coexisted in the patient’s gut at that timepoint. Although PV-A and PV-I carrying bacteria coexisted at the time of JWC* sampling (Figure 2A), 8 days later only the fully resistant PV-I-carrying clone was detected. This shift is probably explained by an AMC treatment that began right before the isolation of JWC*, since OXA-48 confers high-level resistance to AMC.

Patient WDV. Plasmid deletion leading to loss of conjugation

Four isolates of a K. pneumoniae ST11 clone were recovered from this patient over a four-month period. The three initial isolates carried PV-N, but the last isolate (WDV*) carried PV-B, which differed from PV-N by a ~13.5 kb deletion (Figure 2A and Figure 1A). Compared with PV-I, both PV-N and PV-B carry the same small deletion affecting the IS1 element upstream of bla_OXA-48. The large ~13.5 kb deletion in PV-B affected multiple genes involved in conjugation, leading to the loss of conjugation ability in J53 (Fig. 1). In the wild-type strain, PV-B was also associated with a conjugation-incompetent phenotype, and it produced a small, marginally significant, decrease in fitness costs in WDV* compared with PV-N (two-sided Wilcoxon rank-sum exact test W=101, P=0.054). AMR and PCN were the same in the PV-N-carrying and the PV-B-carrying WDV* isolate (two-sided Wilcoxon rank-sum test W=23 & W=14.5, P>0.5). The results from patient WDV thus revealed no clear change in plasmid-associated effects, although the slight difference in fitness costs imposed by PV-N and PV-B could suggests that the large deletion in PV-B might act as a compensatory mutation (Figure 3B-C and Extended Data Fig. 5E-F).

A fitness-resistance trade-off shapes within-patient AMR evolution

In line with our observations in E. coli J53, the analysis of the in vivo evolution of pOXA-48-mediated AMR in patient gut microbiota indicated that this process is shaped by a fitness-resistance trade-off (Figure 3E). Moreover, clinical metadata from patients strongly suggests that antibiotic treatments direct the rapid, and even bidirectional, navigation of this trade-off.

In patient JWC we detected coexistence of almost isogenic K. pneumoniae ST307 populations differing only in the presence of an intact bla_OXA-48 gene in pOXA-48. The presence of the bla_OXA-48 was associated with fitness cost in the absence of antibiotics. The emergence of the bla_OXA-48-lacking PV-A followed a two-month period of no OXA-48-selecting antibiotic treatment, but this variant was rapidly depleted by a cycle of amoxicillin + clavulanic acid. This result not only supports the impact of the fitness-resistance trade-off on the evolution of AMR, but also highlights the importance of clonal diversification in the gut microbiota for this process.

In patient HKH, meropenem treatment triggered the rapid emergence (8 days between isolates) of PV-K, which conferred increased PCN and AMR. Fitness results indicated that PV-K was associated with an increased fitness cost in the absence of antibiotics. Unfortunately, no further samples from this patient were included in the R-GNOSIS collection, and we were therefore unable to investigate the fate of PV-K after the antibiotic treatment ended.

Discussion

The role of fitness-resistance trade-offs in the evolution of AMR has received considerable attention^18,28. However, despite plasmids being arguably the most important vehicle for the acquisition of AMR in many key pathogens the impact of this trade-off on the evolution of plasmid-mediated AMR in clinically relevant situations remains unclear. We anticipate that the fitness-resistance trade-off described here may affect the evolution of plasmid-mediated AMR more generally, because AMR gene expression is one of the central sources of plasmid-associated fitness costs^11,16,17,29. One naive prediction arising from this result is that, in the absence of antibiotic pressure, natural selection could favour plasmid loss or mutations that inactivate plasmid-encoded resistance genes, reversing AMR evolution. However, this prediction is challenged by at least two lines of evidence. First, we observed that the standing genetic variation in the gut microbiota helps to bypass this fitness-resistance trade-off by supporting the coexistence of subpopulations carrying either low-cost/low-resistance or high-cost/high-resistance PVs. The stability of this coexistence will be influenced by the frequency of the antibiotic treatment, which is usually high in hospitalised patients. Indeed, because of the sampling and isolation protocol used in R-GNOSIS (isolation of one clone per species and time point), the role of preexisting genetic diversity in AMR evolution is probably vastly underestimated in our analysis. Second, we previously reported that pOXA-48 conjugation is pervasive in hospitalised patients and leads to long-term plasmid carriage in their gut²³, promoting plasmid maintenance in the bacterial community through source-sink dynamics^13,30. In that study, we described an in vivo pOXA-48 conjugation event in patient HKH involving the same E. coli clone in which we have now described subsequent plasmid evolution. The plasmid dynamics described in this patient perfectly exemplify the ability of pOXA-48 to spread rapidly and to evolve in the gut microbiota of hospitalised patients. These results highlight the need to consider AMR ecology and evolution in order to develop more rational strategies to counteract AMR in complex bacterial communities, such as the gut microbiota.

Our study also highlights the need to take into account two previously neglected issues in the study of AMR evolution. The first of these is the importance of analysing AMR evolution directly in the wild-type, clinically relevant bacterial strains. Our results showed that although the laboratory E. coli strain J53 provides reasonably good qualitative predictions of PVs effects, plasmid-associated fitness costs tend to be much higher than in the wild-type bacteria. This discrepancy could lead to erroneous predictions about the survival of AMR strains in the gut of patients. The second issue is the importance of considering the preexisting genetic diversity in bacterial communities when assessing the potential for AMR evolution. This is actually a key limitation of our study, because the low sampling depth of our collection strongly reduced our ability to estimate gut microbiota heterogeneity, impacting the resolution of our analysis. Despite this limitation, our results from patient JWC, together with findings from other recent studies^24,31, show that genetic diversity fuels AMR evolution. More generally, our results highlight once again the previously reported relevance of standing genetic diversity in the within-patient evolution of bacterial infections^32–34 Most research to date on within-patient evolution of AMR (including this study) has failed properly to screen for community diversity. In future studies, addressing these issues will produce more accurate predictions and help to develop better intervention strategies against AMR in clinical settings.

Methods

Bacterial strains, and culture conditions

Samples included in this study were recovered as part of a study approved by the Hospital Universitario Ramon y Cajal Ethics Committee (Reference 251/13), which waived the need for informed consent from patients on the basis that the study was assessing ward-level effects and it was of minimal risk. For patient’s anonymization, a previously-assigned randomized three-letter code was used²³. For this study, we used the previously characterised R-GNOSIS collection: 28,089 rectal samples from 9,275 patients were collected from hospitalized patients admitted in the Hospital Universitario Ramon y Cajal (Madrid, Spain) from March 2014 to July 2016, as part of a surveillance-screening for detecting ESBL/carbapenemase-carriers (R-GNOSIS-FP7-HEALTH-F3-2011-282512, https://www.r-gnosis.eu/). Rectal samples were obtained from patients within 72 hours of ward admission; weekly additional samples were recovered in patients hospitalized every 2-7 days; and a final sample at discharge was obtained in those patients with a hospital stay ≥3 days. For this study only pOXA-48-carrying enterobacteria were included. For each sampling timepoint a single isolate per enterobacterial species carrying pOXA-48-like-plasmids was stored. To increase the resolution of the PV analysis, we complemented to the R-GNOSIS collection with all the additional pOXA-48-carrying enterobacteria isolated from patients in our hospital since the plasmid was first reported in 2012 and till the end of the study period (n=70). More information on the patients colonized by pOXA-48-carrying enterobacteria in this period can be found in ^23,25. See Supplementary Data 4 for information on synthetic plasmids and primers. All experiments were performed in Lennox lysogeny broth (LB) which was -when indicated- supplemented with 15 g/L agar (CONDA, Spain). Mueller Hinton II broth (Oxoid) was used for IC90 determination and results were comparable with those obtained in LB. Amoxicillin+clavulanic acid (Sandoz, Spain), meropenem (Aurovitas, Spain), kanamycin, ertapenem, chloramphenicol, apramycin, streptomycin, sodium azide and carbenicillin (Merck, Spain) were used in this study.

Plasmid construction

pBGC²⁴ was used to construct pBGA by exchanging the cat gene (chloramphenicol resistance) with aac(3)-IV (apramycin resistance) from pMDIAI³⁵ by Gibson assembly (New England Biolabs, UK). pLC10-Apra was constructed by exchanging the aph(3’)-Ia gene (Kanamycin resistance) with the aac(3)-IV gene, by Gibson assembly. Plasmids pLC10-Kan/pLC10-Apra carry the Streptococcus pyogenes cas9 gene under the control of a Ptet promoter inducible by anhydrotetracycline (derived from pWJ153³⁶), cloned on a thermosensitive pSC101 plasmid backbone with a guide RNA under the control of a Ptrc promoter derived from pCas³⁷ (Addgene plasmid #62225). Single guide RNA (sgRNA) targeting pOXA-48 pemK gene (Fig.1A) was introduced into pLC10-Kan by golden gate assembly³⁸ (New England Biolabs, UK). pLC10-Apra was constructed by exchanging the aph(3’)-Ia gene (Kanamycin resistance) with the aac(3)-IV gene, by Gibson assembly.

gDNA extraction, short- and long-read sequencing

Genomic DNA was extracted using the Wizard genomic DNA purification kit (Promega). Short-read sequencing data from wild-type strains was obtained from ²³ (BioProject PRJNA626430). Additionally, E. coli J53 transconjugants/transformants and the wild-type strains involved in within-patient pOXA-48 evolution (K163, K165, C288, C289, K153 and K229) were sequenced in the Microbial Genome Sequencing Center (MIGS, USA) using NextSeq 2000 platform (coverage>100x). Long-read sequencing (MinION) was performed in MIGS for the wild-type strains involved in within-patient pOXA-48 (coverage>100x). Sequencing data are available under BioProject PRJNA838107. Short-reads from MiGS were trimmed with Trim Galore v0.6.4 (https://github.com/FelixKrueger/TrimGalore), using a quality threshold of 20 and removing adapters and reads <50 bp. Filtlong v0.2.1 (https://github.com/rrwick/Filtlong) was used for filtering long-reads.

Assembly and analysis of pOXA-48 variants

R-GNOSIS genomes were assembled as in²³. pOXA-48_K8 (MT441554) was used as reference in variant calling using Snippy v4.6.0 and plasmids sharing 72% of the pOXA-48 core-genome were selected (n=224). Then, nucleotide variants in 48,500-48,853 and 14,883-16,638 zones were discarded. Mutations in 48,500-48,853 were discarded as they were identified by Sanger sequencing (Macrogen, Spain) as false positives during assembly (Illumina data). This zone contains highly repeated nucleotides which Illumina cannot resolve properly. In 14,883-16,638 pOXA-48 contained a group-II intron (ltrA). ltrA sequence was blasted (BLASTn³⁹ v2.11.0) against the assemblies of each strain to confirm its presence/absence in each PVs. Identity differences in ltrA sequence were not considered for PVs. Insertions were not detected in hybrid assemblies and were not considered for strains sequenced just with Illumina technology. However, we could manually detect a bla_CTX-M-15 gene insertion in position 7,018 in PV-D, by comparing assemblies and in vitro validating by PCR. Deletions between PVs were in silico detected with BRIG v0.95⁴⁰ and validated by PCR amplification. We defined PVs as pOXA-48-like plasmids isolated in R-GNOSIS that share at least a 72% core-genome with pOXA-48_K8 but presented SNPs and/or indels when compared to it.

Introducing pOXA-48 variants into bacterial isolates

A subset of 14 PVs was selected for further investigation based on the following criteria: i) PVs carrying non-synonymous mutations/deletions covering a wide representation of different genes and functions and avoiding PVs with redundant mutations in the same genes, ii) PVs carrying insertions and large rearrangements and iii) PVs with intergenic mutations near to housekeeping plasmid genes, such as genes involved in replication, conjugation or partition. Wild-type strains (donors) and E. coli J53 (recipient) were streaked from freezer stocks onto solid LB agar with antibiotic selection: ertapenem (0.5 mg/L) and sodium azide (100 mg/L), respectively and incubated overnight at 37°C. Several donor colonies and one recipient colony were independently inoculated in 2 mL of LB in 15-mL tubes and cultured for 6 hours (37°C and 250 rpm, Thermo Scientific™ MaxQ™ 8000). Cultures were centrifuged (15 minutes, 1,500 g) and cells were mixed in 1:2 proportion (donor:recipient) and spotted onto solid LB medium overnight at 37°C. Transconjugants were selected by streaking the mix on LB with ertapenem and sodium azide. The presence of PVs in bacteria was confirmed by Illumina sequencing (MIGS). Additionally, each PVs was validated by PCR amplification and Sanger sequencing. For PV-A, which does not confer AMR, donors and recipients were mixed in 10:1 proportion and transconjugants were selected with sodium azide for E. coli J53 or Streptomycin 100 mg/L for the wild-type strain. The presence of the plasmid was confirmed by PCR screening multiple colonies. PV-B was isolated with the NucleoBond Xtra Midi Plus kit (MACHEREY-NAGEL, USA), and introduced into bacteria by electroporation as in⁴¹. Transformants were selected in LB agar with amoxicillin 200 mg/L + clavulanic acid 40mg/L.

Assembly and analysis of E. coli J53 carrying different PVs

Genomes were assembled using SPAdes⁴² v3.15.2. Assembly quality was assessed with Quast ⁴³ v5.0.2. All assemblies reached a size of 4.6-4.8 Mb and contigs >500 bp count was under 110. Prokka⁴⁴ v1.14.6 was used to annotate genomes. Snippy v4.6.0 (https://github.com/tseemann/snippy) was used to identify variants in the E. coli J53 genome by mapping Illumina reads back to its assembly. Variants in the J53 strains carrying PVs were called with Snippy and breseq⁴⁵ v0.35.6 using the annotated E. coli J53 genome as reference. Variants matching in J53 were discarded as assembly errors. From breseq output only predicted mutations and unassigned missing coverage (MC) were analysed because of Illumina data limitations. Snippy was used in a reverse approach, mapping the reads of E. coli J53 against the assemblies of the PVs-carriers. Unidentified mutations not identified in both comparisons and by both software were discarded. For pOXA-48 analysis Snippy and breseq were run using as reference pOXA-48_K8 (MT441554). Only mutations called by both programs, as well as MC and JC from breseq, were considered. The sequence of the ltrA gene was blasted (BLASTn³⁹ v2.11.0) against the assemblies of J53 and the PVs carriers. The contig containing ltrA had similar length in all assemblies and different coverage than of chromosomal contigs, indicating that the ltrA gene did not move into the chromosome of J53. Plasmid replicons were detected with ABRicate v1.0.1 (https://github.com/tseemann/abricate) using the plasmidfinder database⁴⁶. Resfinder database⁴⁷ and ABRicate were used to discard the presence of other resistance genes (Supplementary Data 2).

Relative fitness determination by competition assays

Competition assays were performed by using GFP-tagged strains to distinguish between populations with flow cytometry (CytoFLEX Platform Beckman Coulter Life Sciences, USA). Parameters were: 50 μl min−1 flow rate, 22 μm core size, and 10,000 events per well. Competitions were performed by competing each genotype against a GFP-tagged strain. In E. coli J53, each genotype had 6 replicates and the common competitor was the plasmid-free strain with pBGC²⁴. In clinical strains, competitions were performed by competing each genotype (plasmid-free, and the same strain carrying different PVs against a common competitor). Each genotype was obtained from independent conjugation/transformation events, resulting in 3 independent replicates measured 6 times each (n=18 for each genotype). The common competitor for clinical strains were HKH*/PV-K + pBGA, JWC*/PV-A + pBGC and WDV*/PV-B + pBGA for each case. The common competitors carried PVs to avoid conjugative transfer during competition assays through plasmid exclusion mechanisms⁴⁸. Note that pBGA only differs from pBGC in the AMR gene (apramycin & chloramphenicol resistance respectively). These plasmids contain a gfp gene which is under the control of the P_BAD promoter, so GFP production is controlled by the presence of L-Arabinose. Pre-cultures were incubated overnight in LB in 96-well plates at 250 rpm (Thermo Scientific™ MaxQ™ 8000) 37°C, then mixed 1:1 and diluted 400-fold in 200 μl of fresh LB in 96-well plates (Thermo Scientific, Denmark), and incubated during 24 in the same conditions. The initial populations were mixed (1:1) followed by diluting 400-fold in 200 μl of NaCl 0.9% with L-arabinose 0.5 % (Sigma, Spain) and incubated at 37 °C at 250 rpm during 1.5 hours to induce GFP expression. After 24 hours, final proportions were determined as described above. The fitness of each strain relative to the GFP-tagged one was determined using equation (1):

$$w=\frac{\ln \left(\frac{Nf}{Ni}\right)}{\ln \left(\frac{Nfgfp}{Nigfp}\right)}$$ (1)

Where w is the relative fitness of each strain carrying a determined pOXA-48 variant compared to the GFP-tagged competitor. Ni and Nf are the number of cells of gfp-free clones at the beginning (Ni) and end (Nf) of the competition. Ni gfp and Nf gfp are the number of cells of the common GFP-tagged competitor at the beginning and end of the competition respectively. We discarded PVs loss during the competition by growing individually plasmid-carrying bacteria on LB agar and plasmid-selective antibiotics and counting colony forming units per mL at the beginning and end of the assay (PV-A was tested by PCR). Relative fitness (w) was normalised using the w from the common competitors in each case. An underrepresentation of plasmid costs during the competition assay in E. coli J53 due to conjugative transfer was also discarded by comparing growth curve data (using area under the growth curve, AUC) with relative fitness (Extended Data Fig. 2B).

Growth curves

Growth curves were performed as in⁴⁹. Briefly, strains were streaked from freezer stocks onto solid LB-agar and incubated overnight at 37°C. The next day single colonies were grown in 2 mL of LB and incubated overnight at 37°C with continuous shaking (250 rpm, Thermo Scientific™ MaxQ™ 8000). Six overnight cultures were diluted 1:1,000 into fresh LB in flat-bottom 96-well plates (Thermo Scientific, Denmark), which were incubated during 24 hours at 37 °C 250 rpm. Optical densities (OD600) were measured every 10 minutes during the incubation in a plate reader (Synergy HTX Multi-Mode Reader, BioTek Instruments, USA). The area under the growth curve (AUC) was determined by using the growthrates v0.8.2 & flux v0.3-0 packages in Rstudio 2021.09.2+382. When determining plasmid-variants cost in E. coli J53, normalised AUC was calculated by dividing the AUC of each pOXA-48-carrying isolate by the average value of the AUC of the pOXA-48-free isolate from each plate.

Antimicrobial susceptibility testing

Bacterial AMR profile was determined by (i) LB-growth curves in the presence of different antibiotics and (ii) calculation of inhibitory concentration 90 (IC90) which corresponds to the antibiotic concentration inhibiting 90% of the bacterial growth in the absence of antibiotics. For (i) we used the protocol described above and for (ii) strains were streaked from freezer stocks onto solid MH-agar medium and incubated overnight (37°C). Then, single colonies of bacterial cells (n=5 or 10) were inoculated in parallel in liquid MH starter cultures and incubated at 37°C for 24 hours at 250 rpm. Later, each culture was diluted 1:1,000 in MH medium (~10⁶ cfu) and 200 μl of the final solution were added to a flat-bottom 96-well plate (Thermo Scientific, Denmark) containing the appropriate antibiotic concentration. Antibiotics tested were ertapenem, meropenem and amoxicillin+clavulanic acid. IC90 values were measured after 24 hours of incubation (37°C). Optical density at 600 nm (OD600) was determined in a Synergy HTX (BioTek Instruments, USA) plate reader after 30 seconds of orbital shaking. MH containing each antibiotic concentration was used as blank.

Determining plasmid transfer rate in E. coli J53

PVs transfer rate was evaluated using E. coli J53 as donor and E. coli J53/pBGC (a non-mobilizable and chloramphenicol-resistant plasmid) as recipient. Donors and recipients were streaked in selective agar (ertapenem 0.5mg/L or chloramphenicol 50mg/L, respectively). After an overnight incubation at 37°C, colonies of each donor and the recipient strain were independently inoculated in 2 mL of LB in 15-mL culture tubes and incubated overnight at 37 °C and 250 rpm (Thermo Scientific™ MaxQ™ 8000). Then, 100 μl of donor and recipient were mixed in a 1:1 proportion and incubated on a LB agar plate at 37°C for 2 hours. Subsequently, serial dilutions of each mix were prepared in sterile NaCl 0.9% and plated on selective media for each genotype (carbenicillin 100 mg/L^-1, Chloramphenicol 50mg/L and both antibiotics together). Conjugation rates were determined using the end-point method for solid surfaces as in²³.

Plasmid copy number determination by quantitative PCR (qPCR)

Each genotype was streaked in LB agar and incubated overnight at 37°C. The next day 2 independent colonies were resuspended in 800 μl in sterile water (Fisher Scientific, Spain) and boiled for 10 minutes (95°C). Each sample was centrifuged to spin down cellular debris. Then, 3 independent reactions per colony were performed in triplicate, with 1 μl of the supernatant as DNA template and using with the NZYSupreme qPCR Green Master Mix (2x), ROX plus kit (NZYtech, Portugal) and the 7500 Real Time PCR System (Applied Biosystems, USA). Targeted plasmid and chromosome genes were bla_OXA-48 (amplicon size 200 pb; efficiency 97.35-98.09%, r²=0.996-0.986) & dnaE (chromosomal gene with one copy, amplicon size 200 bp; efficiency 98.44-100.64%, r²= 0.989-0.996) respectively. The efficiency was calculated using serial ¼ dilutions of K8 strain (PV-I) and J53/PV-I as in⁵⁰. The amplification conditions were: 5 minutes denaturation (95°C) followed by 30 cycles of 15 seconds denaturation, 30 seconds annealing (55°C) and 30 seconds extension (60°C). The relative plasmid copy number was calculated using equation (2):

$$PCN=\frac{{(1+Ec)}^{CTC}}{{(1+Ep)}^{CTp}}$$ (2)

where PCN is the plasmid copy number per chromosome, Ec and Ep are the efficiencies of the chromosomal and plasmid reactions (relative to 1), and CTc and CTp are the threshold cycles for chromosomal and plasmid reactions.

Curing and reintroducing PVs into the clinical isolates

Two different plasmid versions carrying CRISPR/Cas9 machinery were used in this project: pLC10-Kan (kanamycin resistant) and pLC10-Apra (apramycin resistant). pLC10-Kan was used in JWC* and HKH* and pLC10-Apra in WDV*. First pOXA-48 carrying strains were made competent following the protocol described in ⁴⁰. Then each pLC10 was introduced in the cells by electroporation using 0.1 cm cuvettes and 1.8 kV pulse (MicroPulser Electroporator, Biorad Spain). Transformants were selected on LB agar plates with kanamycin 250-512mg/L or apramycin 30mg/L for each case. Transformants were verified by PCR (Supplementary Data 4). Then, CRISPR/Cas9 machinery was induced by resuspending several transformant colonies in 500 μl of LB with kanamycin or apramycin, 0.2 mg/L anhydrotetracycline (aTc), to activate Cas9 expression, and IPTG 0.08 mM to enhance sgRNA expression. Then, suspensions were incubated for 2 hours at 30°C with agitation (250 rpm, Thermo Scientific™ MaxQ™ 8000) and was streaked and incubated overnight at 37°C on LB agar to cure pLC10. Note that pLC10 oriC is based on pSC101 and codes for a thermosensitive replication protein. The next day single colonies were streaked parallelly in LB agar and LB agar supplemented with ertapenem 0.5 mg/L, kanamycin or apramycin. Only colonies that were sensitive to both antibiotics were recovered and sequenced by Illumina (MIGS, USA). Then different plasmid variants were re-introduced in triplicate (in parallel) to plasmid-free cells as described above.

Analysis of strains involved in within-patient AMR evolution

4 potential cases of within-patient pOXA-48 evolution were identified: HKH, JWC, WDV and HAX. HAX was discarded because the PVs differed from each other just by a synonymous SNP (Supplementary Data 1). Unicycler⁵¹ v0.4.9 with default parameters was used to obtain hybrid-assemblies from K153, K229, C288, C289, K163 and K165 strains. Long-reads were also assembled with Flye⁵² v2.9 and circularization was confirmed in Bandage⁵³ v0.8.1. Medaka v1.4.3 (https://github.com/nanoporetech/medaka) was used to obtain consensus sequences. Several rounds of Pilon⁵⁴ v1.24 were performed mapping the trimmed Illumina reads. Contigs were rotated with Circlator fixstart⁵⁵ v1.5.5. Long-read assembly quality was controlled in IGV⁵⁶ v2.11.1. PVs assemblies were confirmed by mapping short- and long-reads and by aligning the assemblies to the reference pOXA-48_K8 (MT441554) with BWA-MEM⁵⁷ v0.7.17 and minimap2⁵⁸ v2.21. Alignments were visualized in IGV. Closed assemblies were annotated with PGAP⁵⁹ v2021-07-01.build5508. Breseq v0.36.0 was used to identify SNPs and structural variants. To discard false-positive calls different combinations of breseq runs were performed. For K164, K165-2 cured, K165-1, K165-3, K165-4, K165-5, K165-6, K165-7 and K166 (JWC), K151, K152, K229 cured (WDV) and C289 cured (HKH), the trimmed reads were mapped to the closed strains from their respective patients. ABRicate v1.0.1 with the plasmidfinder and resfinder databases was used to confirm clonality and isogeneity between within-patient evolved and cured strains. Further details on workflow and analysis criteria are provided in https://github.com/LaboraTORIbio/within_patient_evolution.

Construction of phylogenetic trees

Snippy v4.6.0 was used to find SNPs between all E. coli ST10 & ST744 (reference C288), K. pneumoniae ST11 (reference K153) and K. pneumoniae ST307 (reference K163) from the R-GNOSIS collection. Strain K25 (ST11) was removed from the analysis because the fastq files were truncated. Snippy-core (https://github.com/tseemann/snippy) was used to find the core genome. Strain K78 (ST11) was removed for diverging too much from the rest of the strains. Gubbins⁶⁰ v3.1.4 was used to remove recombinant regions and SNPs were extracted with snp-sites⁶¹ v2.5.1. Maximum-likelihood trees were constructed with IQ-TREE⁶² v1.6.12 from the extracted alignments with best evolutionary model detection and an ultrafast bootstrap of 1000 optimized by hill-climbing nearest neighbour interchange (NNI) on the corresponding bootstrap alignment. Trees were visualized and edited in iTOL⁶³ and Inkscape v0.17.

Statistical analyses

All statistical analyses were performed in Rstudio 2021.09.2+382 (R v4.1.1 2021-08-10) with packages rstatix v0.7.0, tidyverse v1.3.1 and car v3.0-12. To test homoscedasticity and the normality of data for each dataset, Shapiro-Wilk test, Levene's Test and Bartlett’s test were performed. Then according to each data structure parametric and nonparametric tests were performed (see main manuscript for each test). We performed a generalized additive model (GAM) using the R package mgcv v1.8-40 with relative fitness as dependent variable and plasmid copy number, antimicrobial resistance and conjugation rates as explanatory variables. A non-linear fitting for the independent variables was used in the model.

Extended Data

Extended Data Fig. 1.

Extended Data Fig. 2.

Extended Data Fig. 3.

Extended Data Fig. 4.

Extended Data Fig. 5.

Data availability

The sequence data supporting the findings of this study are available in the National Center for Biotechnology Information Database with the accession code PRJNA838107 (https://www.ncbi.nlm.nih.gov/bioproject/838107). The raw data obtained in this study are available on Supplementary Data 5. The remaining R-GNOSIS sequences can be found in ²³.

Code availability

The code generated during the study can be found in GitHub (https://github.com/LaboraTORIbio/within_patient_evolution).