Abstract
Bacterial resistance to drugs is a growing problem, and one that is inspiring a search for new classes of anti-bacterial compounds. This has generated interest in antimicrobial peptides (AMPs), which are key components of innate immune defences. In contrast to conventional antibiotics, however, little is known about how bacteria respond to AMPs, and whether they modify their phenotypic responses based on their prior experiences. Here, we explore whether prior exposure to sublethal doses of AMPs increases bacterial survival and abets the evolution of resistance. We show that Escherichia coli cells primed by sublethal doses of AMPs develop tolerance and generate more persister cells. Priming with the AMPs melittin and pexiganan leads to bacterial production of curli and colanic acid, respectively. Based on the phenotypic data we developed a population dynamic model to show how priming increases persistence and tolerance. The model predicts that priming delays the clearance of infections and fuels the evolution of genetic resistance. Since AMPs are immune effectors our results suggest that the optimal strategy to reduce problems caused by tolerant or persistent cells requires both (a) high concentrations of and (b) fast and long-lasting expression of AMPs. We anticipate that the effects discussed here will apply to many AMPs as well as other drugs that target the cell surface. Our findings also offer a new understanding of phenotypic drug resistance and could lead to measures that slow the evolution of resistance while improving the treatment of persistent infections.
Main Text
Sublethal concentrations of antibiotics increase bacterial tolerance and persistence and therefore contribute to antibiotic resistance (1–4). Low, sublethal levels of antibiotics are common: whether in the environment or during medical application of antibiotics where the pharmacodynamics start at zero. While sublethal concentrations, tolerance and persistence have been of great interest for antibiotics (1–4), little is known about antimicrobial peptides (AMPs) which are novel drug candidates (5) but importantly also key effectors of innate immune defences (6). After infection, the induction of AMPs results in sublethal concentrations before killing concentrations are reached. But AMPs differ significantly from antibiotics in their pharmacodynamics (7) and they kill cells within minutes (8) rather than hours (9). Here we explore whether prior exposure to sublethal doses of AMPs increase bacterial survival and the risk of resistance, as is the case for antibiotics.
We find that a sublethal dose of certain antimicrobial peptides can induce increased tolerance and/or persistence (Fig. S1) in bacteria and hence prime (10) them for the exposure to a subsequent lethal dose. We identify the underlying molecular mechanisms and capture the population dynamics by adapting a classic mathematical model of persistence (11). With computational simulations, we then illustrate that increasing tolerance and persistence has a positive effect on bacterial survival and the emergence of resistance.
First, we studied the effect of priming E. coli K-12 using two antimicrobial peptides, melittin from the honeybee, and pexiganan, the first eukaryotic AMP used as a drug (6). We primed using a fraction of the minimal inhibitory concentration (0.1xMIC, table S1) and then monitored bacterial survival over time under lethal concentrations of the respective AMPs (10 x MIC, table S1). We found that the priming treatment resulted in much higher E. coli survival (Fig. 1 A, B).
The decline of the time kill curves is biphasic suggesting two subpopulations (11–14). We excluded that deviations from monophasic decline arise because of decreasing antimicrobial concentrations over time (Fig. S2) and fitted the time-kill curves to a biphasic linear function. For both AMPs, bacterial populations declined faster during the first than the second phase (Fig. 1, Table S2, S3). Tolerance, the decline of bacterial populations in the first phase (4), was significantly higher in primed than in naïve bacteria for both AMPs. Primed bacteria showed higher survival in the second phase, indicating higher numbers of persisters (4). The change in population size in the second phase, however, was not significantly different between primed and naïve populations, indicating that the dynamics of stochastically switching into and out of the persister state in the second phase are not influenced (Fig. 1). In short, priming with AMPs allow bacteria to survive better by increasing both bacterial tolerance and persistence.
To understand how priming leads to tolerance and persistence, we used RNAseq of cells exposed to priming concentrations of AMPs (Tables S4, 5). Melitin induced up-regulation of curli fimbriae and pexiganan induced the colanic acid synthesis (Fig. 2, S3). Curli is an important virulence factor (15) and component of extra-cellular matrix and protects against AMPs (16). Likewise, colanic acid capsules protect against AMPs and antibiotics (17). Both AMPs also induced significant overlap in gene expression related to osmotic shock (Fig. 2, S3). The removal of an essential gene for colonic acid production completely abolished the priming effect by pexiganan. A curli-deficient mutant showed a decrease of the priming effect induced by melittin (Fig. S4).
By phase contrast imaging, we observed the formation of a characteristic colonic acid capsule in pexiganan-primed but not in naïve cells (Fig. S5, S6) and the priming response was homogenous. For melittin primed bacteria curli induction was documented with a specific curli binding chemical (Fig. S6, S7) and this response was homogenous (Fig. S8).
Afrter lethal exposure to melittin and pexiganan we observed a high degree of heterogeneity regarding the killing rate of individual cells in primed bacteria but not in controls (Fig. S8). Primed cells also aggregate with stronger effect in pexiganan-treated cells (Fig. S8).
A decrease of intracellular ATP increases persistence under the antibiotic ciprofloxacin (18). Because ATP leakage is a hallmark of AMP-treated bacteria (19), we exposed melitin and pexiganan primed populations and controls to ciprofloxacin (18). This resulted in a highly significant increase in the number of persisters (Fig. S9). The level of leaked ATP in the culture supernatant was significantly higher in primed bacteria for both AMPs (Fig. S9). The pre-treatment with melittin or pexiganan does not change the MIC of E. coli to melittin, pexiganan or ciprofloxacin consistent with the definition of persisters (4) (Table S5).
To quantify the influence of priming on bacterial tolerance and persistence, phenomena inherently linked to the growth dynamics and subpopulation structure, we developed a population dynamics model that captures tolerance and persistence. We build on a two-state population model previously developed (11) to describe bacterial antibiotic persistence (Fig. S10). This model assumes that bacteria exist in two phenotypic states, normal cells (N) and persisters (P). The two subpopulations N and P differ in their susceptibility to AMPs, a difference that is implemented as differing net growth rates rN and rP, respectively. Bacteria switch from subpopulation N to P with the rate sN and back with the rate sP. We quantified the model parameters by fitting the analytical solution of the model equations to the time-kill datasets of melittin and pexiganan (Fig. 3A,B and Table S6). In a first fit with four free parameters (rN, sN, rP, and sP,), the parameter rp was not significantly different from 0 (Fig. S11). In the fit with three free parameters, rN, sN, and sP, priming affected the parameter rN and sP(Fig. S12): rN increased due to priming, which translates into increased tolerance (see suppl) and sPdecreased. Together, an increase in rN and decrease in sPresult in higher persistence (Fig. 3C, D).
To assess the influence of priming on possible treatment success and resistance evolution, we extended a previously developed predictive framework (7) by a persistent subpopulation (Fig. S10). We then explored the effects of priming on tolerance and persistence, individually and in combination, which would be challenging empirically. Using our predictive framework, we investigated the effect of priming on survival based on a zero-order pharmacokinetic profile (7, 20) (Fig. S13) and with parameterized pharmacodynamics functions (Fig. S14).
We found that survival of the population was highly dependent on tolerance (Fig. 4A, B, S15). The presence of persistent cells alone only marginally increased time until clearance. When we implemented a decrease in switching rate sP, the time until clearance of the bacterial population is extended at high treatment intensities (high concentrations Amax). The results do not qualitatively change for larger pharmacokinetic decay rates (k) typical for AMPs (Fig. S16). Taken together, an increase in tolerance alone resulted in higher survival independent of persister cells. An increase in persister cells further increased survival at high antimicrobial concentrations.
Next, we assessed how priming affects resistance emergence. Generally, resistance evolution depends on the population size, mutation rate and the replication rate. While priming increases bacterial survival, hence the population size, it also increases the number of persisters that do not replicate and hence are a poor source of resistant mutations (11) (Fig 1, Fig. S17). Our simulations revealed that the beneficial effects of priming on survival due to increased persistence did not translate into an increased probability of resistance emergence and establishment (Fig. 4B, S15, S16). The probability of resistance emergence was mainly influenced by the effect of priming on tolerance.
We find that sublethal dosing of the AMPs melittin and pexiganan primes bacterial cells to increase both, tolerance and persistence. This differs from antibiotics, which usually increase either tolerance or persistence (4). The molecular basis of the induction of AMP tolerance and persistence relies on modifications of bacterial envelopes involving curli or colonic acid respectively. Interestingly, the activation of both pathways shows different dynamics in biofilm formation (21). Because curli and colanic acid are important component of the biofilm matrix, triggering their expression by sublethal levels of AMPs could potentially catalyse the biofilm formation. Within a host, if the immune system fails to clear the pathogens, the AMP-priming effect may thereby favour the transition from acute to chronic bacterial infections, where biofilms prevail.
Sublethal concentrations of antimicrobials are common and cannot only directly select for bona fide resistance (22) but can also generate phenotypic resistance indirectly (3). Priming by AMPs likely plays a role in infection vectors: in the flea gut the PhoQ-PhoP system is induced in Yersinia pestis by AMPs leading to biofilm formation that enhances transmission to the final host (23). It is not clear as yet if phenotypic AMP-resistance will facilitate opportunistic infections in a way similar to genetic AMP-resistance, as has been shown for S. aureus (24), but in the light of our results it seems likely. Our combined theoretical and empirical results suggest that in hosts the optimal strategy of AMP-expression requires three components: (i) high concentrations to clear bacteria quickly (ii) fast up-regulation to avoid priming and (iii) and long up-regulation to clear all of the targeted bacteria. We anticipate that these requirements also hold for the medical application of AMPs.
Methods
Bacteria and growth conditions
The E. coli MG1655 was used as bacterial models for all experiments. All cultures related to antimicrobial tests were carried out in Mueller-Hinton I Broth (Sigma). For genetic manipulations, the strain and its derivatives were routinely cultured in Lysogeny Broth (LB medium) or SOB, supplemented with antibiotics when appropriate. Other constructed and strains used in these study are listed in table X of this section.
Minimal inhibitory concentration (MIC)
MICs were determined according to CLSI recommendations by a microdilution method in MHB with some modifications. Inoculum size that was adjusted to 2×107 CFU/ml from a regrowth of overnight cultures to be consistent with the downstream experiments. The MIC was defined as the antimicrobial concentration that inhibited growth after 24 h of incubation in liquid MH medium at 37°C. Polypropylene non-binding plates (Th. Geyer, Germany) were used for all experiments. The MIC was considered as the antimicrobial concentration that inhibited growth after 24 hours of incubation in liquid medium at 37°C.
Priming experiments
Starting from 1×108 CFU/ml, where bacteria were exposed (stimulus) to 1/10 MIC of melittin or pexiganan during 30 minutes at 37°C with soft shaking. The tubes were centrifuged at 4000 g for 10 minutes and allow to recover during 60 minutes. The cells were challenged (trigger of priming response) with a concentration equivalent to an 10X MIC. The cultures were diluted and plated to determine cell viability. Five replicas per culture were used, and every experiment was repeated twice. Non treated cells were used as a control.
Persister Antibiotic survival assay
Bacterial cultures were inoculated at 1:100 from a 16-hour overnight culture into MHB medium. Cell were grown during 2 h to reach approximately to 2 × 108 CFU/ml). The cultures were treated with priming concentrations (1/10 MIC) of melittin and pexiganan during 30 minutes. Non-treated cultures were used as control. All cultures were washed and centrifuged twice to remove the treatment. The supernatants were used to determine ATP concentration using a Molecular Probes ATP Determination Kit (Thermo Fisher Scientific, Germany). Bacteria were resuspended in equal volume of fresh medium and an aliquot from each culture was taken to determine the number of bacteria at t=0 by diluting and plating in MHB agar. Following, ciprofloxacin was added for a final concentration of 2 µg/ml to treated tubes and to non-treated AMPs control. The cultures were incubated during four hours. Then, bacteria were washed twice with 0.9 % NaCl and plated on MHB agar to determine the counts of survival fractions. The percent survival was calculated as the ration between described previously 1. Briefly final CFU/CFU at 0 h) × 100. The results are presented as the average from 5 independent replicas.
Construction and verification of deletion mutants
We inactivated the major curli subunit protein gene csgA and the colonic acid precursor gene wza. Although both pathways involve many genes, the removal of these two components impair the production of both substances respectively. These mutants were generated in E. coli K-12 strain MG1655 following the methodology described elsewhere 2 with some modifications because we used as template the genomic DNA of the Keio collection 3. Briefly, we extracted genomic DNA from the mutants csgA::Kn and wza::Kn of the Keio collection (E. coli BW25113) and amplified by PRC the flanking regions of the kanamycin resistance cassette disrupting both genes and including an appropriate homology sequence. For the csgA mutant we used the primers 5’-GATGCCAGTATTTCGCAAGGTG-3’ and 5’-GGTTATCTGACTGGAAAGTGCC-3’ while primers 5’-TAGCGTGTCTGGATGCCTG-3’ and 5’-CCACTTTCAGCTCCGGGT-3’ were used for wza. The PCR products were purified and electroporated in the E. coli MG1655 carrying a red recombinase helper plasmid, pKD46. The strain was grown in 10 ml SOB medium with ampicillin (100 µg/ml) and L-arabinose at 30°C to an OD600 of ∼0.5 and then made electrocompetent by washing and centrifuge them with a cold solution of glycerol 10%. Competent cells in 80 µl aliquots were electroporated with 200 ng of PCR product. Cells were added immediately 0.9 ml of SOC, incubated during 2 h at 37°C, and then 100 µl aliquots spread onto LB agar with kanamycin (30 µg/ml). The correct inactivation of genes was verified by PCR. The antibiotic resistant cassette (Kn) was removed for both mutants using the flippase plasmid pCP20.
Transcriptome sequencing
The transcriptome sequencing of primed cells was determined on samples treated identically as described for the priming experiments. Total RNA from 108 cell per sample was isolated using the RNAeasy Isolation kit (Qiagen, Germany). Traces of genomic DNA were removed from 10 µg of RNA by digestion in a total volume of 500 μl containing 20 units of TURBO DNase, incubated for 30 minutes at 37°C, immediately followed by RNeasy (Qiagen) clean-up and elution in 30 μl of RNase-free water. Following DNase treatment, RNA integrity was assessed using Agilent RNA 6000 Nano kit and 2100 Bioanalyzer instrument (both from Agilent Technologies). Total RNA was depleted from ribosomal RNA using the Ribo-Zero Depletion Kit for Gram-negative bacteria (Illumina, USA). Libraries were prepared using a TruSeq Stranded Total RNA library preparation kit (Illumina, USA) and were sequenced on a MiSeq platform.
Transcript abundances were derived from pseudo-alignment of reads to the cDNA sequences from the ASM584v2 assembly of Escherichia coli MG1655 (ENA accession GCA_000005845.2) using Salmon version 0.7.2 with default parameters 4. Differential gene expression was analyzed using the R package DESeq2 5 in conjunction with tximport6. Pairwise contrasts were performed between the control and each AMP treatment with empirical bayesian shrinkage of both dispersion parameters and fold-change estimation. We defined genes as being significantly differentially expressed when the absolute fold-change in expression was greater than 2, at an FDR-adjusted p-value of less than 0.05. The variance-stabilizing transformation was used to remove the dependence of the variance on the mean and to transform data to the log2 scale prior to ordination using principal component analysis. Quality of RNAseq data were contrasted by Euclidian distance and symmetry of data reads distribution (Fig. S18).
Observation at single cell level
To observe cell reaction at single cell level during priming experiments, we used an ad hoc microfluidic device developed for this project. It consisted in a main channel for bacterial inoculation and medium perfusion and with several lateral compartments which dimensions are around 1.5 µm height (ensuring all bacteria are kept in focus) and square 200 µm width corresponding to a field size of the used microscope at 1000x magnification (Fig. SX8). The chip was designed in Autocad (Fig. S19). We started the replication of our microfluidic chips from a custom made (Sigatec SA) silicon (SiO) master. This silicon master itself was first replicated in Smooth-Cast 310 (Bentley advanced material). Soft lithography produced the chips in PDMS (Sylgard Silicone Elastomer Base and Curing Agent mixed in 10:1 ratio). The PDMS chips were cured overnight at 75°C in an incubator. We punched an inlet and outlet hole for the laminar flow in each chip using a biopsy puncher of 0.5 mm (outer diameter). The chips were bonded to a glass cover slide (24×60 mm) after a 30-second air plasma treatment (PDC-002, Harrick Plasma). Before use, the assembled chip was treated for 15 seconds in air plasma and immediately injected it with filtered MHB medium for passivation. We left the activated chip to incubate for a least 1 hour before loading the bacteria. The devices were loaded to full capacity with a bacterial suspension containing nearly 2×108 CFU/ml (exponentially growing bacteria, 0.5 OD600). Cell suspension was injected into the main channel of the chip using a blunt-end 23G needle attached to a 1ml syringe. We centrifuged the loaded chip at 200 x g for 10 min using in-house adapters, checked the loading.
After we loaded the bacteria cells in the dead end side channels, we connected the chip to a syringe pump (AL-6000, WPI, Germany) and placed the chip under an inverted microscope. A continuous laminar flow (100µl/h) of MHB through the central channel throughout the experiment (SM Fig. 1). For the life cell imaging, after infusion with priming or triggering concentration of AMPs, we injected MHB supplemented with bacterial Live/Dead stain kit solutions (Thermo Scientific, Germany) for a final concentration of 0.1 µl/ml of MHB. We took pictures of at least 20 fields per treatment. Fluorescent images were taken of each field of view with simultaneous acquisition in red and green fluorescent channels during a time interval of no more than 2 minutes per treatment with a Nikon Ti-2 inverted microscope (Nikon, Japan). Cells were observed with the 100× objective and controlled by Nis Element AR software. The chip holder is temperature controlled at 37°C.
Determination of melittin-induced curli
The production of curli due to melittin treatment was determined by using the fluorescent dye EctracerTM680 (Ebba Biotech, Sweden) that stain extracellular curli. EctracerTM680 was used according to the instructions of manufacturers. Bacterial cultures were treated on chip with priming and triggering concentrations of melittin as described for single cell observation section omitting the live/dead staining. After priming and triggering, the channels were perfused MHB supplemented with EctracerTM680 in a proportion of 1/1000 related to the medium. Cells were observed with the red channel fluorescence for Cy5 dye using a Nikon Eclipse Ti2 inverted optical microscope using the 100X oil objective. Two independent samples were prepared for each group (primed and naïve cells).
SEM of E. coli treated antimicrobial peptides
Approximately 2×107 CFU/ml E. coli MG1655 were treated with 1/10 MIC of pexiganan and melittin during 30 minutes. The cultures were concentrated 10 times by a quick centrifugation step of 1 minute at 8 000 g and resuspended in 1/10 of its own supernatant. and resuspension 10 µl drops were placed on a circular glass cover slip (1.5 cm of diameter). The drops were fixed with osmium tetroxide vapors during one minute and allow to dry in a laminar flow cabinet. The cover glasses were mounted on aluminum stubs using double-sided adhesive tape and coated with gold in a sputter coater (SCD-040; Balzers, Union, Liechtenstein). The specimens were examined with a FEI Quanta 200 SEM (FEI Co., Hillsboro, OR) operating at an accelerating voltage of 15 kV under high vacuum mode at different magnifications. At least 5 fields from two independent replicas.
Statistical analysis
To analyze the priming data, we first tested if the dynamics of the depicted in the time-kill curves are biphasic.
We fitted the function
to the time-kill data of each AMP and for primed and naïve populations individually using RSS.
Here, tkink is the time point, at which the population dynamics switch from the first phase to the second phase and m1 and m2 are the slopes of the first and the second decline, respectively.
Note that m1 is a direct measure of tolerance. The standard error (SE) was calculated as Here, 𝜃 denotes to the estimated parameter values of m1, m2, and tkinkand 𝐼(𝜃) is the expected Fisher information. The 95% CI interval was calculated as 𝜃 ± 1.96 ∗ SE(𝜃).
Population models
To describe bacterial population dynamics, we used the two-state model by Balaban et al. 7 (Fig. S8): In this model, the population B(t) consists of 2 subpopulations N(t) and P(t), with B(t) = N(t)+P(t). The rate of change of the population is determined by the net growth rate of N and P, rn and rp, and the switching rate from N to P, sn, and the switching rate from P to N, sp. The analytical solution of this ODE system 7,8 is With The model was fitted by minimizing the RSS, similar to the above. For the starting conditions (N(t=0), P(t=0), we assumed that the ratio of N/P was constant over time when the exposure to lethal concentrations of AMPs started. N(t=0) and P(t=0) were therefore calculated using the eigenvector that corresponds to the largest eigenvalue of a system without antimicrobials. Here, we assumed that the parameter rN is equal the net growth rate in absence of antimicrobials, 𝑟M = 𝜓max. The parameter 𝜓max was estimated based on the time-kill curve of bacterial population that grow in absence of antimicrobials (see below). The eigenvector contains information about the ratio of N and P for . Resulting, and was estimated from the data. Confidence intervals were calculated as described above. In a pre-analysis, we used 4 free model parameters that were fitted: rN, rP, sN and sP. The parameter rP was not significant from (Fig. S4). Therefore, we set the parameter rPto 0 and fitted the remaining 3 parameters to the data (table S6).
Tolerance and persistence in terms of model parameters
The measure of tolerance is the slope m1. Komarova and Wodarz 8 showed that the slope can directly be linked to the population model parameters. In our notation, Note that the first phase is mainly influenced by rN (Fig. S9), therefore, 𝑚$ ≈ 𝑟M. Persistent cell numbers at time t were calculated with the analytic solution:
Pharmacokinetic and pharmacodynamic function
We used the pharmacokinetic function , with and 𝑛 = 0,1,2 … described previously elsewhere 9. In our simulations, we fixed the decay parameter k and varied the drug input Amax. To describe the effect of the AMPs on the bacterial population, we used the pharmacodynamic (PD) function 𝜓(𝐴) 9,10, with The parameter 𝜓max describes the net growth rate in absence of antimicrobials (𝜓max =𝜓(𝐴 = 0)). The antimicrobial effect e(A) is dependent on the antimicrobial concentration and is the defined with 𝜓max, 𝜓min, the net growth rate in presence of large amounts of antimicrobials (𝜓(𝐴 → ∞)), with the MIC, the antimicrobial concentration that results in no growth (𝜓(𝐴 = 𝑀𝐼𝐶) = 0) and with 𝜅, which determines the steepness of the PD curve.
The PD function was fitted to the time-kill curves (Fig. S6), as described by Regoes et al. 9. In short, we used log-linear regressions of the time kill curves within the time-points 0h and 1h to estimate the change of the bacterial population over time, i.e. the slopes of the log-linear regression. We fixed the parameter 𝜓max and fitted the 3 remaining parameters of the PD function with the Markov-Chain-Monte-Carlo method.
Stochastic simulations
To simulate resistance evolution with stochastic simulations, we expanded a previously developed framework for bacterial population dynamics 11. The framework models bacterial population dynamics exposed to changing levels of antimicrobials and allows for resistance evolution. In the simulations, the change in population size of a sensitive strain S, with S = N+P, and of a resistant strain R were described with the following ODE system: Here, the replication rate is assumed to be equal to the maximum net growth rate 𝜓max. The effect of the antimicrobial 𝑒N(𝐴) is explained above. Note that we assume that bacteria in class P do not grow and are not affected by antimicrobials. We also assumed that the switching rates are constant. To describe the effect of an antimicrobial on the strain R, 𝑒R(𝐴), we use the same parameter set than with 𝑒 (𝐴), except for the . In the simulations, we differentiated between the following cases: (i) naïve bacteria, homogeneous population (no persistence), (ii) naïve bacteria, heterogeneous population, (iii) primed bacteria (increase in rN), homogeneous population, (iv) primed bacteria (increase in rN), heterogeneous population (v) primed bacteria (increase in rN and decrease in sP). The stochastic simulations were run 1000 times for each antimicrobial, for each case, and for a variety of input antimicrobial concentrations Amax. All parameter values are listed in table S7. The intensity is the input antimicrobial concentrations Amax. The simulations were run for t=7d. Time until clearance and probability of resistance evolution were calculated as the mean of the value over the 1000 simulations.
Implementation
Statistical testing, simulations and plots were done in R version 3.3.2 12, using Rstudio version 1.0.143 13. We used the following R-packages: (i) for plotting: sfsmisc 14, (ii) for fitting the PD function: rjags(iii) for stochastic simulations: adaptivetau 15. We used Mathematica version 11.0 16 to determine the analytical solutions of the population models.
Acknowledgments
We are grateful to Sophie Armitage and Olivia Judson for comments on the manuscript.