Antibiotic heteroresistance generated by multi-copy plasmids

Heteroresistance – in which a clonal bacterial population contains a cell subpopulation with higher resistance to antibiotics than the main population – is a growing clinical problem that complicates susceptibility determination and threatens therapeutic success. Despite the high prevalence of heteroresistance in clinical settings, the underlying genetic mechanisms that stably maintain heterogeneous bacterial populations are poorly understood. Using fluorescence microscopy, single-cell microfluidics, and quantitative image analysis, we show that random replication and segregation of multicopy plasmids produce populations of bacterium Escherichia coli MG1655 in which cells with low-and high-plasmid copy numbers stably co-exist. By combining stochastic simulations of a computational model with high-throughput single-cell measurements of blaTEM-1 expression, we show that copy number variability confers the bacterial population with transient resistance to a lethal concentration of a β -lactam antibiotic. Moreover, this surviving, high plasmid copy minority is capable of regenerating a heterogeneous bacterial population with low and high plasmid copy numbers through segregational instability, rapidly alleviating the fitness burden of carrying large numbers of plasmids. Our results provide further support for the tenet that plasmids are more than simple vehicles for horizontal transmission of genetic information between cells, as they can also drive bacterial adaptation in dynamic environments by providing a platform for rapid amplification and attenuation of gene copy number that can accelerate the rate of resistance adaptation and can lead to treatment failure.

Heteroresistance -in which a clonal bacterial population contains a cell subpopulation with higher resistance to antibiotics than the main population -is a growing clinical problem that complicates susceptibility determination and threatens therapeutic success. Despite the high prevalence of heteroresistance in clinical settings, the underlying genetic mechanisms that stably maintain heterogeneous bacterial populations are poorly understood. Using fluorescence microscopy, single-cell microfluidics, and quantitative image analysis, we show that random replication and segregation of multicopy plasmids produce populations of bacterium Escherichia coli MG1655 in which cells with low-and high-plasmid copy numbers stably co-exist. By combining stochastic simulations of a computational model with high-throughput single-cell measurements of bla TEM-1 expression, we show that copy number variability confers the bacterial population with transient resistance to a lethal concentration of a β -lactam antibiotic. Moreover, this surviving, high plasmid copy minority is capable of regenerating a heterogeneous bacterial population with low and high plasmid copy numbers through segregational instability, rapidly alleviating the fitness burden of carrying large numbers of plasmids. Our results provide further support for the tenet that plasmids are more than simple vehicles for horizontal transmission of genetic information between cells, as they can also drive bacterial adaptation in dynamic environments by providing a platform for rapid amplification and attenuation of gene copy number that can accelerate the rate of resistance adaptation and can lead to treatment failure.

37
The evolution and spread of antimicrobial resistance in clinical pathogens represent a major public health problem that threatens to become a global crisis. 1 In general, drug resistance is considered to be the consequence of stable genetic mutations or the acquisition of antibiotic resistance genes through horizontal gene transfer. 2 However, treatment failure can also result from the presence of subpopulations of bacterial cells with higher levels of resistance than those of the rest of the population. 3 This phenomenon is known as heteroresistance 4,5 and has been identified in diverse bacterial species and in a wide range of antimicrobial classes. [6][7][8] In the clinic, heteroresistance due to spontaneous tandem gene amplification has been proposed as a plausible cause of treatment failure, 24 with the incidences likely to be underestimated due to the intrinsic limitations of standard microbiology assays. 25 A recent large-scale analysis of heteroresistant clinical isolates found a high incidence of genomic amplifications that increased resistance to multiple antibiotics. 26 Interestingly, whole-genome sequencing revealed that, while some duplications occurred in large chromosomal regions containing known drug resistance genes, a considerable fraction of sequence amplifications were found in plasmids.
Plasmids are DNA molecules that replicate independently of the chromosome and play an essential role in the dissemination of resistance genes among clinically important pathogens. 27 Crucially, plasmids can be present in multiple copies per cell, from a few copies to dozens for high-copy plasmids. Although some plasmids can be transferred horizontally, thus spreading resistance genes between bacterial hosts, a large fraction of plasmids are non-conjugative and are carried in multiple copies per cell. 28 A recent clinical study showed that a large fraction of pathogenic Escherichia coli isolates carry small ColE1 plasmids. 29 The number of plasmids carried by each cell is a key driver of virulence 30 and horizontal gene transfer. 31 Furthermore, cells within a biofilm contain high plasmid copy numbers and therefore have elevated transcription of antibiotic resistance genes. 32 For multicopy plasmids lacking active partitioning or postsegregational killing mechanisms, 33 segregation occurs randomly upon division, with the probability of a plasmid being inherited to a given cell following a binomial distribution. [34][35][36] The interaction between replication and segregation, and the complex population dynamics this produces 37, 38 is known to enhance bacterial adaptation to novel environmental conditions, 39 as well as to determine the repertoire of genes carried in plasmids 40 and their stability in the absence of selection. 41, 42 Moreover, recent studies have shown that multicopy plasmids can accelerate bacterial adaptation, 43 for instance by promoting intracellular genetic diversity 44 and increasing the probability of the appearance of beneficial mutations and subsequently amplifying mutant gene expression. 45 In addition to amplifying gene dosage, an increase in copy number is also associated with a decrease in the probability of plasmid loss and with a higher metabolic burden. 46 A consequence of this tradeoff is that plasmid replication is subject to two conflicting levels of selection: 35, 47, 48 plasmids that overreplicate have a higher chance of overcoming segregational loss and becoming fixed in descendant cells, but cells with more plasmid copies have a lower probability of becoming fixed in the population. As a result, plasmid control is a tightly regulated process 49 that depends on the host's genetic 50 and physiological state, 51 as well as on the extracellular environmental conditions. 52, 53 For high-copy plasmids, however, replicative noise emerges as intracellular selection favors overreplication, thereby relieving intracellular selection for precise copy number control. 35 100 To investigate the distribution of plasmids in bacterial populations, we used an experimental model system consisting of E. coli MG1655 carrying pBGT, a ColE1-like plasmid containing a GFP fluorescent marker (eGFPmut2) and bla TEM-1 , a gene that encodes a TEM-1 β -lactamase, which inactivates βlactam antibiotics by hydrolyzing the β -lactam ring. 54 β -lactam resistance genes are generally located on plasmids and, in particular, TEM-1 has a plasmid origin, with more than two-hundred TEM β -lactamase variants descending from this allele recorded. 55 We denote the strain carrying this wellcharacterized, 45, 56 non-conjugative, and multicopy plasmid as MG/pBGT (average copy number=19.12, s.d.= 1.53; Figure 1A-C). 45 As a control, we used a fluorescently tagged strain carrying a chromosomally encoded bla TEM-1 , which we term MG:GT. Moreover, to explore the association between PCN and fluorescence, we also used strains obtained in a previous experimental evolution study, 45 with mutations in the origin of replication (Table S1) that result in a high mean PCN, with correspondingly high fluorescence intensity and elevated drug resistance compared with MG/pBGT.

Environmental modulation of PCN distributions in bacterial populations
In a recent study, direct, fluorescent-reporter-based measurement of PCN, promoter activity, and protein abundance at single-cell resolution revealed a positive correlation between PCN and protein expression. 57 In our experimental system, we similarly observed a correlation between PCN measured by qPCR 39 and fluorescence intensity quantified using a fluorescence spectrophotometer (R 2 = 0.9387; Figure 1D). To validate the correlation between PCN and GFP in our system, we sorted the plasmidbearing cell population according to GFP intensity into clusters with low, medium, and high fluorescence and confirmed the positive correlation between fluorescence and mean PCN estimated by qPCR (R 2 = 0.879; Figure S1). To measure the effect of the strength of antibiotic selection pressure on the distribution of PCN, we exposed a population of MG/pBGT cells to a range of ampicillin (AMP) concentrations, and used flow cytometry to measure GFP abundance in single cells. We found that the mean GFP abundance increased with the strength of selection ( Figure 1E), and that the coefficient of variation for the PCN distribution decreased as a function of drug concentration (R 2 = 0.593, p-value< 0.01; Figure S3). When the same experiment was repeated with MG:GT cells, mean fluorescence and its coefficient of variation remained constant accross all AMP concentrations (R 2 = 0.052, p-value> 0.5; Supplementary Figure S2). (in blue) and eGFPmut2 (in green). The reading frames for genes are represented with arrows, with arrowheads indicating the direction of transcription. B) DIC microscopy image of a plasmid-bearing population (MG/pBGT). C) Fluorescence microscopy image of this population shows high levels of GFP heterogeneity between cells. D) Mean fluorescence and mean plasmid copy number are positively correlated in bacterial populations. The circle's diameter is proportional to each strain's drug resistance level. E) Fluorescence distributions of MG/pBGT exposed to a range of ampicillin concentrations. 139 The increase in fluorescence that we observed in response to AMP could be explained by a uniform increase in resistance levels in all cells in the population (e.g. by increasing the rate of plasmid replication), or by cell-to-cell heterogeneity in resistance levels (e.g. pre-existing copy number variability in the population). Because population-level experiments cannot be used to contrast both hypothesis, we measured the response of individual MG/pBGT cells to AMP exposure using a microfluidic chemostat and fluorescent microscopy (Methods). Using this set-up, we followed the life history of individual plasmid-bearing cells exposed to an antibiotic ramp of linearly increasing AMP concentration. In these experiments, the longest surviving MG/pBGT cells were those with high fluorescence before antibiotic exposure (Figure 2A and Movie S1). Crucially, the fluorescence of individual cells remained constant throughout the antibiotic ramp ( Figure 2B), suggesting that the population-level increase in mean GFP is a consequence of the antibiotic killing low PCN cells, and not the result of individual cells upregulating plasmid replication or bla TEM-1 expression (see also Figure S4).

PCN variability enhances the survival of bacterial populations exposed to fluctuating selection
To determine if between-cell differences in drug resistance produced heteroresistance at a populationlevel, we examined the response to AMP of 88 clonal populations of plasmid-bearing strains with different mean PCNs (strains MG/pBGT, MG/G54U and MG/G55U, with 19, 44, and 88 plasmid copies, respectively). When the cultures reached exponential growth, ∼ 1% of each population was transferred to an environment containing replenished media and a lethal AMP concentration ( Figure  3A). After 30 minutes, a sample was transferred back to drug-free medium. This sampling process was repeated every 30 minutes and, for each duration of drug exposure, we counted the number of replicates showing growth after 24h. Relative to MG:GT, all plasmid-bearing populations exhibited increased survival of fluctuating selection ( Figure 3B; log-rank test, p-value< 0.005). For instance, the probability of survival after 90 minutes of AMP exposure was > 50% for all plasmid-bearing strains, whereas < 5% of the MG:GT replicate populations survived. It should be noted that the lethal drug concentration was determined independently for each strain (see Table S1 for MICs used). For each strain, we estimated the duration of drug exposure such that the probability of survival was 50% (60 min for MG:GT at 2 mg/mL AMP, and 80 min at 32 mg/mL AMP for MG/pBGT; dotted lines in Figure 3B). We refer to exposure to this concentration and duration as a semi-lethal pulse.   To confirm that the increased tolerance to a semi-lethal pulse presented by plasmid-bearing strains was not a consequence of a decrease in growth rate associated with the metabolic burden inherent to carrying plasmids (rather than selection of a subpopulation with more copies of bla TEM-1 ), we performed a survival assay for MG/pBGT in the presence of 256 µg/L of the β -lactamase inhibitor sulbactam. As expected, fluorescence remained constant independently of the ampicillin concentration, and only one out of eight populations survived exposure to 2 mg/mL of AMP ( Figure S5).

Quantifying heterogeneity in bla TEM-1 expression and survival after a semi-lethal pulse
In a microfluidic experiment, MG/pBGT and MG:GT populations were exposed separately to a semilethal pulse of AMP, with the critical dose and duration of the antibiotic pulse determined independently for each strain ( Figure 4A). We acquired time-series of the fluorescent intensity of individual cells, recorded division events, and estimated the duplication rates of 5, 810 lineages for MG/pBGT and 1, 077 MG:GT lineages, respectively obtained from 46 and 8 separate microfluidic chambers (see Supplementary Movies 2 and 3 for sample time-lapse movies). The criteria for including a single-cell lineage in the analysis was that they were observed for a period spanning the antibiotic pulse. AMP-induced cell lysis was estimated by staining the medium with rhodamine and measuring the accumulation of fluorescent dye. After a recovery period in drug-free medium, cells were classified according to whether they died or survived the semi-lethal pulse. As the antibiotic concentration and duration of treatment were determined separately for each strain, we expected the semi-lethal pulse to kill approximately half the population. In line with this prediction, only 46.7% of MG:GT cells and 44.2% of MG/pBGT cells survived the antibiotic pulse ( Figure 4B). A retrospective analysis of surviving and non-surviving cells revealed that surviving cells had an elevated duplication rate, measured as the time elapsed between cell division events (87.09 and 106.7 minutes, respectively; Figure S6; p-value< 0.005). Similarly, surviving cells had a higher rate of elongation (changes in cell length between consecutive frames) than cells that were killed ( Figure S7; p-value< 0.005). These results suggest that an enhanced probability of survival is a consequence not of reduced metabolic activity, but of heterogeneity in bla TEM-1 expression. Changes in the fraction of surviving cells as a function of GFP expression before drug exposure are shown in Figure 4C. As expected, cells with very high GFP expression had a high probability of survival (54% survival for the top quartile), whereas the mean survival rate for cells in the bottom quartile was below 34%. Interestingly, survival probability was not a monotonously increasing function of GFP intensity, since high survival rates were also observed in cells with intermediate GFP expression ( Figure  4D).

Plasmid-driven phenotypic noise produces a heterogeneous stress response
To investigate if this bimodality in the survival distribution is a consequence of a heterogeneous stress response triggered by a subpopulation of cells, we exposed a MG:GT population to a semi-lethal pulse of AMP and recapitulated the life history of the surviving cells ( Figure 5A-B). Note that shortly after being exposed to the antibiotic, some cells ceased dividing but continued to grow, thus producing filaments (see also Figure 4E). Conditional filamentation can be triggered by multiple molecular mechanisms, 58 including a general stress response -the SOS regulatory network -that regulates the expression of over 50 genes involved in DNA repair, DNA damage tolerance, and the induction of a DNA damage checkpoint that transiently suppresses cell division. 59 In particular, the SOS response can be triggered by the binding of β -lactamase molecules to penicillinbinding protein 3 (PBP3). Lactamase-bound PBP3 acts through DpiBA, a two-component signal transduction system 60 that induces sulA, which in turn inhibits septation by blocking FtsZ polymerization. As a result, cell division is suppressed and bacterial filaments are produced. 61, 62 Crucially, once the stress is removed, filamented cells reorganize the FtsZ ring, divide, and resume normal growth. 63 Furthermore, consistent with previous studies, 65 the temporal expression of genes in the SOS system appeared to be tightly regulated, with 61.4% of cells in the MG:GT population responding synchronously to the antibiotic input and producing filaments (we define a filamented cell as a cell with more than two standard deviations from the mean length of the population before drug exposure). In contrast, the plasmid-bearing population produced a very heterogeneous response, with only 17.1% of cells producing filaments ( Figure 5C-D). This was expected, as we have established that variability in PCN maintains a subpopulation of cells that overproduce β -lactamase and hence avoid triggering the stress response by maintaining a low periplasmic AMP concentration. Conversely, cells with low PCN are killed by the antibiotic before they can trigger the SOS response. Histograms of GFP fluorescence in cells of each subpopulation before the introduction of AMP was introduced into the microfluidic device are shown in Figure S8. As expected, the MG:GT population exhibited low variance, with no significant differences in mean GFP intensity detected between subpopulations. In contrast, the plasmid-bearing population exhibited a GFP intensity distribution with high variance. We classified each cell according to whether it was killed or survived drug exposure and according to whether or not the stress response was triggered. Surviving cells in the MG/pBGT population, either had high fluorescence intensity and did not trigger the SOS response (a consequence of increased β -lactamase synthesis), or had intermediate GFP fluorescence and survived antibiotic exposure by elongating and delaying cell division. We also performed an exploratory data analysis, which showed that while PCN (measured by proxy through GFP intensity) is important for cell survival, so is cell length at the moment of the environmental perturbation (see PCA plot in Figure S9). This analysis confirmed that cells with increased survival are small cells with high GFP fluorescence, or cells that were already filamented when exposed to AMP ( Figure S10). Our data suggests that plasmid-driven phenotypic noise produces random conditional filamentation, thus enabling the population to adapt to a rapid increase in drug concentration.

High levels of antibiotic resistance are unstable in the absence of selection
In our microfluidics data, the mean time elapsed between cell duplication events was significantly different between the two strains (36.6 minutes for MG:GT and 88.2 minutes for MG/pBGT; p-value< 0.001; Figure S11). Similarly, at the population-level, a comparison of growth rate in strains with different PCNs with respect to plasmid-free cells revealed a negative correlation between growth rate and mean PCN in the absence of selection for plasmid-encoded genes (R 2 =0.562; Figure S12). The cost associated with bearing plasmids is well-documented, [66][67][68] particularly for ColE1-like plasmids, 69,70 and has been reported for multiple plasmid-host associations in a wide range of bacterial species. 45, 71-73 The burden associated with plasmid carriage is highly variable and depends on the interaction between plasmids and their bacterial hosts. 50 This fitness cost can be ameliorated through mutations in genes located either on the chromosome or the plasmid. [74][75][76][77] In addition to these compensatory mutations, another strategy to ameliorate the burden of carrying high-copy plasmids is to reduce the number of plasmids carried per cell. For instance, a previous experimental evolution study reported that mutations near the origin of replication generated a 10-fold amplification in mean PCN, but at a very high fitness cost that resulted in high levels of antibiotic resistance being unstable in the population once the antibiotic was removed. 41 To assess how rapidly PCN amplification is reversed once the antibiotic is withdrawn, we performed a three-season serial dilution experiment in which a MG/pBGT population was exposed to fluctuating selection (season 1, drug-free; season 2, 32 mg/ml AMP; season 3, drug-free). The GFP fluorescence distribution was recorded at the end of each season ( Figure 6A). In the presence of AMP, the GFP fluorescence distribution shifted to high expression but rapidly returned to the original fluorescence distribution once the antibiotic was removed. This effect was also observed with high-copy plasmids ( Figure S13). Repeat runs of the experiment with different drug concentrations revealed that mean GFP fluorescence of the MG/pBGT population increased proportionally to the strength of selection, and the shift towards higher copy number cells was rapidly reversed after removing the antibiotic ( Figure 6C). In contrast, the GFP intensity distribution in MG:GT cultures was the same independently of the presence of antibiotic in the medium ( Figure 6B). . Note that the antibiotic shifts the GFP distribution to the right (green area) and is later restored when the antibiotic is removed. B) GFP histogram for MG:GT reveals that GFP distributions coincide independently of the environmental drug concentration. C) Increase in mean fluorescence in the presence of antibiotics is correlated with drug dose (darker red, higher drug concentrations). Once the antibiotic is removed, mean GFP intensity is restored to pre-exposure levels. The black line shows that fluorescent intensity for MG:GT remains constant during the experiment. 325 To further explore the interaction between the stochastic plasmid dynamics and the strength of selection for plasmid-encoded genes, we used a multi-level computational model that incorporates intracellular plasmid dynamics into an ecological framework (Methods). Briefly, the agent-based model explicitly simulates key cellular processes: cell duplication, resource-dependent growth, antimicrobial-induced death, and random plasmid replication and segregation. Propensities of each process are determined from the concentrations of a limiting resource and a bactericidal antibiotic present in a well-mixed environment.  Figure 7A shows numerical realizations of the agent-based model simulating an exponentially-growing population of cells descended from a parental plasmid-bearing cell. We considered the number of plasmids carried by each cell as a time-dependent variable subject to two main sources of noise: (1) imperfect PCN control, 78 with plasmid replication occurring in discrete events distributed stochastically over time, and (2) plasmid segregation occurring randomly between daughter cells upon division. A consequence of this stochastic plasmid dynamics is that PCN in any individual cell is highly variable over time and, as the culture is no longer synchronous after a few cell duplications, plasmid-bearing populations exhibit high levels of copy number heterogeneity. This heterogeneity results in a PCN distribution with large variance ( Figure 7B). Based on our experimental data and previous reports, 79, 80 we assume a linear relationship between PCN and gene dosage. Therefore the probability of an individual cell dying upon exposure to a given antibiotic concentration can be estimated from the number of plasmid copies it carries and the degree of resistance conferred by each plasmid-encoded gene. For instance, if we assume that every cell in the population is equally sensitive to the antibiotic (i.e. a population with low-variance PCN distribution), then we find a drug concentration that kills all cells simultaneously (a dose referred to in the clinical literature as the minimum inhibitory concentration, MIC). Hence the survival probability function of such a homogeneous population is a stepwise function that switches from 1 to 0 at this critical drug concentration (black dotted line in Figure 7C). However, when we consider a heterogeneous population characterized by a PCN distribution with large variance then, by definition, the population contains cells with fewer or more gene copies than the expected value (green lines in Figure 7B). This implies that the survival probability of heterogeneous populations is lower than that predicted for a homogeneous population at sub-MIC concentrations and higher than the predicted value in high-drug environments ( Figure 7C). Indeed, temporal changes in PCN can result in cells with differing degrees of drug susceptibility; as a result, when antibiotics are introduced into the system, only that fraction of cells that had overreplicated the plasmid was able to survive drug exposure ( Figure 7D). In our computational experiments, exposure to antibiotics reduced total bacterial density, but, as cells with low levels of resistance are cleared first from the population, the PCN distribution shifts towards higher values (red lines in Figure 7E). Furthermore, the computational model predicts that the intensity of drug-induced PCN amplification in the population is proportional to the strength of the selective pressure ( Figure 7F), with selection for high-copy plasmid cells occurring even at sub-lethal drug concentrations (resulting from killing cells with fewer plasmids than the mean PCN).

Stochastic plasmid dynamics promotes heteroresistance in a computational model
Moreover, once the antibiotic was withdrawn, cells that survived continued to grow and divide, therefore randomly replicating and segregating plasmids (see Figure S14). A consequence of this stochastic plasmid dynamics is that cells with low PCN are readily produced through segregational drift. These low-copy cells are at a competitive advantage relative to high-copy sub-populations, and consequently the mean PCN of the population returns to the level observed prior to antibiotic exposure. Similarly with the experimental data, repetition of the computational experiment for different selection strengths revealed that the degree of PCN amplification appears to be correlated not only with the strength of selection, but also with its rate of decay once the antibiotic is removed.

395
The evolution of antimicrobial resistance in response to the industrialized consumption of antibiotics, specifically those of the β -lactam class, is one of the most serious health threats societies face today. 81 In clinical isolates, heteroresistance can be the result of unstable genomic amplifications, 26 and has been shown to be the first stage in the progression to β -lactam resistance. 82 Taken together, our data show that cell-to-cell differences in PCN in a clonal population can produce heterogeneity in drug susceptibility in the population, thus enabling plasmid-bearing populations to implement a nonresponsive adaptive strategy that increases their survival in a context of fluctuating selection pressures. Microfluidics uniquely enabled us to connect the plasmid copy number of individual cell lineages (GFP fluorescence) to their phenotypic variability (survival, elongation, or death) under antibiotic pressure and to examine their fate after the antibiotic was removed. Single-cell traces also allowed us to compare our experiments to dynamic computational models. Moreover, the combination of high-throughput fluorescence measurements with single-cell and populationlevel susceptibility assays enabled us to show that PCN distribution is modulated by the strength of selection for plasmid-encoded genes, rapidly increasing the mean resistance of the population during selective conditions. Our analysis focused on non-conjugative, multi-copy plasmids that are usually carried at around 10-30 copies per cell; however, plasmid-driven phenotypic noise is not exclusive to high-copy plasmids. 79 A recent study showed that a conjugative, low PCN populations (1-8 copies per cell) also exhibited large copy number heterogeneity that resulted in noisy expression of plasmidencoded genes. 83 We also found that PCN heterogeneity promoted variability in the SOS system, a stress response mechanism that is known to increase resistance to heavy metals 84, 85 and antimicrobial substances. 86, 87 This stress response is also known to increase genetic variation 88 by promoting bacterial mutagenesis 89, 90 and enabling the horizontal transmission of virulence factors 91 and antibiotic resistance genes. 92 The SOS system also produces bacterial filaments, which have been shown to be an adaptive trait with many benefits, 58 including the promotion of tissue colonization 93 and increased tolerance to cell wall damage produced by the antibiotics used in this study. 64 Our study, combined with previous reports, shows that having a phenotypically diverse population is an effective adaptive strategy to survive fluctuating environmental conditions. 95-97 Transitions between phenotypic states can result from promoter noise; 98 asymmetry in the cell division process; 99 or stochastic fluctuations in the concentrations of proteins, mRNAs, and other macromolecules present at low-copy numbers in the cell. 100-102 We proposed that the stochastic nature of plasmid replication and segregation also produces heterogeneous populations, in which a minority of cells that carried more copies of a plasmid encoding the antibiotic resistance gene bla TEM-1 survived exposure to a lethal AMP concentration.
Upon removal of the antibiotic from the environment, surviving cells continued growing and dividing, therefore replicating and segregating plasmids. As a result, low PCN cells with increased competitive fitness relative to the highly-tolerant subpopulation emerged, therefore restoring drug susceptibility and compensating for the cost imposed by bearing multiple plasmid copies. Altogether, these results indicate that multicopy plasmids provide a platform for implementing a reversible phenotypic tolerance mechanism that rapidly compensates for the burden of carrying multiple plasmid copies. Furthermore, transient amplification of selective genes encoded in multicopy plasmids may not be exclusive to bla TEM-1 , as similar effects would be achieved by antimicrobial resistance genes encoding efflux proteins or other drug-modifying enzymes. 8, 103-105 Other systems where gene dosage is relevant and that scale with gene copy number may also use multi-copy plasmids as platforms for fine-tuning gene activity. 106, 107 A recent study showed that precise control of gene expression in genetic engineering and synthetic biology can be achieved by tuning PCN in individual cells. 80 This provides a promising tool for the optimization of synthetic circuits, but also represents a novel approach that can be used for the design of rational treatment strategies that are effective at suppressing heteroresistant populations.

Materials and Methods
Bacterial strains and culture conditions 462 In this study, we used Escherichia coli K12 MG1655 bearing a ColE1-like (p15a) plasmid, pBGT, encoding for the β -lactamase resistance gene bla TEM-1 that confers resistance to ampicillin, an eGFPmut2 gene under an arabinose inducible promoter, and the araC repressor. Mean PCN=19.12, s.d.= 1.53. 45 As a control, a strain E. coli K12 MG1655 was used, carrying the pBADg f p2, araC, and the bla TEM-1 integrated into the chromosome through the λ -phage. Strains bearing plasmid variants G54U and G55U contained a point mutation in the origin of replication: G to U changes at positions 54 and 55 of the RNAI placed in the loop of the central hairpin and affect the RNAI-RNAII kissing complex that controls plasmid replication and PCN. All experiments were conducted in Lysogeny Broth-Lenox (LB) (Sigma-L3022) supplemented with arabinose (0.5% w/v) and appropriate ampicillin concentrations were supplemented as indicated in each experiment. Arabinose stocks solutions were prepared at 20% w/v by diluting 2 g of arabinose (Sigma-A91906) in 10 ml DD water sterilized by 0.22 µm filtration. AMP stock solutions (100 mg/ml) were prepared by diluting ampicillin (Sigma-A0166) directly in 0.5% w/v arabinose LB. Antibiotic susceptibility determination 476 The minimum inhibitory concentration (MIC) of different strains was calculated using dose-response curves performed in 200 µL of liquid media. 96-well plates (Corning CLS3370) supplemented with LB (0.5% w/v arabinose) and a logarithmically-separated range of drug concentrations were used. Antibiotic plates were inoculated from a master plate using a 96-pin microplate replicator (Boekel 140500). Inoculation plates were prepared by adding 200 µL of overnight culture into each well and incubating at 37 • C with 200 rpm shaking. Optical density measurements were performed using a BioTek ELx808 Absorbance Microplate Reader at 630 nm. MIC was determined when the reader was unable to detect bacterial growth (2, 32, 43, and 46 mg/mL for strains MG:GT, pBGT, G54U, and G55U, respectively). Plasmid copy number determination 486 PCN per chromosome was determined using quantitative polymerase chain reaction (qPCR) with a CFX96 Touch Real-Time PCR Detection System. Specific primers were used for the E. coli's dxs monocopy gene as chromosomal reference (dxs-F CGAGAAACTGGCGATCCTTA, dxs-R CTTCAT-CAAGCGGTTTCACA) and primers for the bla TEM-1 plasmid-encoded gene (Tem-F: ACATTTC-CGTGTCGCCCTT, Tem-R: CACTCGTGCACCCAACTGA) both with amplicon sizes 100 bp as previously described. 45 In short, samples were prepared following a previously published protocol: 108 100 µl culture samples were centrifuged at 16,000 g for 60", the supernatant was removed, and the pellet was resuspended in an equal volume of MilliQ water. Then, samples were boiled at 95 • C for 10' using a thermoblock and stored at -20 • C for later use. Primers were diluted in TE buffer at 10 µM and stored a -20 • C. Primers' final concentration was 300 nM. qPCR reactions were performed using SYBR Select Master Mix (Applied Biosystems -4472908) in 96-well flat-bottom polystyrene microplates

515
To determine competitive fitness in the absence of antibiotics, each strain was cultured in a 96-well plate with LB supplemented with arabinose 0.5% w/v. A Synergy H1 microplate reader was used to obtain the growth kinetics of each strain by inoculating a 96-well plate with an overnight culture of each strain and growing at 37 • C for 24 hours, reading every 20 minutes, after 30 seconds of shaking. Maximum growth rate estimates were obtained by fitting the mean optical density of N=8 using the R package GrowthRates using non-parametric smoothing splines fit. 110

522
Strains of MG::GT and MG/pBGT were exposed to a three-season serial transfer experiment using 96-well plates (8 replicates per strain). An initial inoculation plate was made by putting 200 ml of overnight culture per well. Season 1 (LB) was inoculated from an inoculation plate using a microplate pin replicator. Season 2 (LB-AMP) was inoculated from season 1 after 12 hours of growth. We used the following ampicillin gradient: 0, 1/128, 1/64, 1/32, 1/16, 1/8, 1/4, 1/2, 1, and 2 MIC units. In season 3, cultures were transferred to a new LB plate after 12 hours of growth, allowing bacteria to grow for another 12 hours. Plates were sealed using an X-Pierce film (Sigma Z722529) perforating every well to avoid condensation and grown at 37 • C inside a BioTek ELx808 Absorbance Microplate Reader. Measurements were taken every 20 minutes, after 30 seconds of linear shaking at 567 cpm (3 mm).
At the end of each season, end-point fluorescence intensity was measured using a BioTek Synergy H1 using OD (630nm) and eGFP (479.520nm). Plates were then stored at 4 • C before imaging flow cytometry was performed the following day. A complete independent four-replicate experiment was performed for each strain. DNA samples were extracted at the end of each season to quantify PCN.

Population-level survival assay
Strains were grown in an LB+Amp media in a 96-well plate under a concentration of AMP determined based on the MIC of each strain. For each LB+AMP plate, we considered 88 populations growing in antibiotics and 8 without antibiotics as controls. Inoculated plates were incubated in a BioTek ELx808 absorbance microplate reader at 30 • C, with optical density measurements (630nm) obtained every 30 min, after 1 min of shaking. After each read, plates were taken out, and a plate sample was taken with a microplate replicator to inoculate a new LB plate. Samples were taken every 30 min, from 0 to 8 hours, then at 18 and 24 hours. New plates were grown in a static incubator at 30 • C for 24 hours. Growth was measured using OD (630nm) and eGFP (479,520 nm) in a Synergy H1 microplate reader after 5 min shaking. An additional experiment was performed for the MG:GT and MG/pBGT strains sampling every 2 hours from 0 to 12 hours and a final sampling at 24 hours. For the β -lactamase inhibition assay, sulbactam (Sigma-S9701) was used. First, the ampicillin concentration was fixed to be that of the MIC of MG:GT (2 mg/ml). Then, a sulbactam dose-response experiment with MG/pBGT was performed and found that the minimum sulbactam concentration achieved that complete growth suppression was 256 µg/l. Critical AMP and sulbactam concentrations were used to performed a population-level survival assay consisting on exposing 8 replicate populations to fluctuating selection: LB− → LB+AMP+sulbactam − → LB. Samples of four replicates were used for flow cytometry, and the remaining four replicate samples were used for PCN quantification. Single-cell microfluidics 555 A microfluidic device built-in PDMS (polydimethylsiloxane; Sylgard 04019862) from molds manufactured by Micro resist technology GmbH using soft photolithography (SU-8 2000.5) was used for this study. In particular, a micro-chemostat that contains two media inputs and 48 rectangular chambers (40x50x0.95µm 3 ). 111 Each confinement chamber traps approximately 1, 000 cells in the same focal plane, enabling us to use time-lapse microscopy to follow thousands of individual cells in time. Chips were fabricated by pouring PDMS into the mold before baking it for 2 hours at 65 • C. Solid chip prints were cut, punched, and bound to a glass coverslip using a plasma cleaner machine (Harrick Plasma -PDC-001) at full power for 1 min and 15 sec. Then we baked them again overnight at 45 • C to ensure binding. Moreover, for each strain, MG/pBGT and MG:GT200, a 1 l titration flask was inoculated with 200µl of an overnight culture (LB at 30 • C and 200 rpm) when the culture reached 0.2-0.3 OD630; it was split into 4 falcon tubes and centrifuged for 5 min at 7, 000 rpm. Supernatant was disposed of, and cells were resuspended by serial transfers into 5 ml of fresh media supplemented with arabinose 0.5% w/v. This dense culture was used to inoculate the microfluidic device. Data acquisitions started 5 hrs after the device chambers were filled and cells were growing exponentially. After 60 minutes of growth, we switched the environment from LB to LB+AMP. Drug concentration was determined independently for each strain (2mg/ml and 8mg/ml for MG:GT and MG/pBGT, respectively). Media and antibiotics were introduced into the microfluidic device using a bespoke dynamic pressure control system based on vertical linear actuators (adapted from 112 ). The duration 20/38 of drug exposure was determined based on the time elapsed before the probability of survival of the population exposed to the MIC is below 50% (a semi-lethal pulse; an exposure of 2mg/ml for 60 min for MG:GT, and of 8mg/ml for 80 min for MG/pBGT). At the end of the period of drug exposure, the population was transferred to a drug-free environment and grown for 120 min for MG:GT, and 100 min for MG/pBGT. Growth media was supplemented with arabinose at 0.5% and Tween20 (Sigma-P2287) at 0.075%, and filtered with .22µm filters. Experiments were conducted at 30 • C, and the ampicillin media was stained by adding 5 µl and 3 µl of a fluorescent dye (rhodamine, Sigma S1402) in 100ml of media used to grow MG:GT and MG/pBGT cells, respectively. This red fluorescent dye allowed us to calibrate media inputs inside the microfluidic device and also worked as a dead-cell marker. Rhodamine stock solution was prepared, diluting the powder in ethanol, and stored at 4 • C. Microscopy was performed in a Nikon Ti-E inverted microscope equipped with Nikon's Perfect Focus System and a motorized stage. Temperature control is achieved with a Lexan Enclosure Unit with Okotouch. The microscope was controlled with NIS-Elements 4.20 AR software. Image acquisition was taken with a 100x Plan APO objective without analog gain and with the field and aperture diaphragms as closed as possible to avoid photobleaching. DIC channel captures were made with a 9v DIA-lamp intensity, red channel (excitation from 540 to 580nm, emission from 600 to 660nm filter), green channel (excitation from 455 to 485nm, emission from 500 to 545nm). Exposure times were 200ms, 200ms, and 600ms for DIC, green and red channels, respectively. Multi-channel, multi-position images were obtained every 10 minutes in the following order: Red, Green, Lamp-ON, DIC, Lamp-OFF. We added the Lamp-ON optical configuration to allow the bright-light lamp to be fully powered before acquiring the DIC image, while the Lamp-OFF configuration was added to make sure that the lamp was completely off before capturing the next position. Microscopy time-lapse images were analyzed using µJ, an ImageJ-Python-Napari image analysis pipeline that implements Deep Learning for image segmentation. In short, the pipeline uses ImageJ macros to arrange and manipulate microscopy images. Image segmentation was performed using DeepCell. 113 Binary masks were corrected manually using bespoke ImageJ macros. Cell tracking was performed using a nearest-neighbor weighted algorithm coded in Python. Cell-tracking was corrected manually using a custom cell viewer coded in Napari. 114 Lineage reconstruction was performed in Python, obtaining thousands of single-cell time-series of fluorescent intensity and cell length, as well as time-resolved population-level statistics, including the probability of survival to the antibiotic shock and the distribution of fluorescent intensities. Our cell viewer also allows easy lineage data visualization and plotting. Code used to analyze images is available in a public repository: https://github.com/ccgesb-lab/uJ/

Supplementary material
Movie S1: MG/pBGT exposed to an antibiotic ramp. Movie S2: MG/pBGT exposed to a semi-lethal pulse of AMP. Movie S3: MG:GT exposed to a semi-lethal pulse of AMP. For the GFP fluorescent intensity, the chromosomal strain exhibits a stable expression over time, while the plasmid-bearing strain shows a decline in GFP observed for the population that did not survive. C) GFP intensity as a function of time. Note that MG:GT cells that did not filament continued to grow past the filamentation threshold (horizontal dotted line) even after the antibiotic is withdrawn. Eventually mean length of the population reduces as filamented cells resolve and continue growing normally. For the plasmid-bearing strain, note that filamented cells that were killed exhibited a larger cell length than surviving cells when the antibiotic was introduced into the environment.  Figure S5. Survival assay with AMP and a β -lactamase inhibitor. A) Growth curves obtained for MG/pBGT populations exposed to 2 mg/ml of AMP and a range of sulbactam concentrations (low sulbactam doses in light blue, high concentrations in purple). B) Final optical density as a function of sulbactam concentration. We consider that bacterial growth is completely suppressed at concentrations higher than 256µg/l of sulbactam. C) Optical density (OD 600 ) measured after 12 hours of growth in a 3-season survival assay (season 1: LB; season 2: LB + sulbactam (256µg/l) + AMP (2 mg/ml); season 3: LB). Gray lines represent different replicates (N = 8), with the mean OD 600 represented with a black line. Of note, only one replicate exhibited growth after the recovery period. D) Normalized fluorescence intensity of populations exposed to a 3-season serial dilution experiment. Note that supplementing the media with sulbactam reduced the relative fluorescent intensity exhibited by the population during the drug exposure period, in contrast to previous experiments performed in the absence of sulbactam, where we observed an increase in fluorescence during AMP exposure.  Figure S8. Histograms of fluorescent intensity for classified cells. A) Cells in MG:GT exhibit a fluorescent distribution with low variance and with no significant differences in mean GFP between cells that produced filaments and were killed (red) or survived (green), as well as for cells that did not produce filaments and died (blue), and those that survived drug exposure (purple). B) GFP distributions of the plasmid-bearing population exhibit large variance. Cells that survived showed increased mean fluorescence relative to cells that were killed. For surviving cells, mean GFP was significantly lower for cells that did not produce filaments with respect to cells that triggered the SOS response system.  Growth rate (h 1 ) Figure S12. Fitness cost estimated at a population-level. Growth rates for different strains obtained by fitting a growth curve using non-parametric smoothing splines. As expected, there is a negative correlation between fitness in drug-free environments and the number of plasmids carried by each cell. Growth rates ANOVA p-value is 2.91e-07 indicating significant differences. A follow-up Tukey's Honest Significant Differences analysis yields the following strain pairs with p-value < 0.05 : WT-pBGT, pBGT-MGGT, pBGT-G54U.  Figure S13. Rapid gene amplification is unstable in high-copy plasmids. A) Fold increase in PCN (relative to season 1) for strain MG/G54U in a three-season serial dilution experiment (black line represents the mean over N = 4 replicates, in grey). During the second season, a sub-lethal concentration of AMP is deployed, selecting for high-copy plasmid cells, therefore increasing five-fold the mean PCN in the population. In the third season, the antibiotic is removed and the mean GFP fluorescence intensity decrease to the levels exhibited prior to drug exposure. B) Mean GFP intensity for MG/G55U also shows a rapid increase in fluorescence during drug exposure and a rapid decline once the drug is removed.