OUP user menu

The importance of being persistent: heterogeneity of bacterial populations under antibiotic stress

Orit Gefen, Nathalie Q. Balaban
DOI: http://dx.doi.org/10.1111/j.1574-6976.2008.00156.x 704-717 First published online: 1 July 2009


While the DNA sequence is largely responsible for transmitting phenotypic traits over evolutionary time, organisms are also considerably affected by phenotypic variations that persist for more than one generation, with no direct change in the organisms' DNA sequence. In contrast to genetic variation, which is passed on over many generations, the phenotypic variation generated by nongenetic mechanisms is difficult to study due to the inherently limited life time of states that are not encoded in the DNA sequence, but makes it possible for the ‘memory’ of past environments to influence future organisms. One striking example of phenotypic variation is the phenomenon of bacterial persistence, whereby genetically identical bacterial populations respond heterogeneously to antibiotic treatment. Our aim is to review several experimental and theoretical approaches to the study of persistence. We define persistence as a characteristic of a heterogeneous bacterial population that is taken as a generic example through which we illustrate the approach and study the dynamics of population variability. The clinical and evolutionary implications of persistence are discussed in light of the mathematical description. This approach should be of relevance to the study of other phenomena in which nongenetic variability is involved, such as cellular differentiation or the response of cancer cells to treatment.

  • nongenetic inheritance
  • antibiotics
  • single cells
  • microfluidics
  • resistance
  • tolerance


The phenomenon of bacterial persistence was observed as early as in 1942, when Staphyloccocal infections were seen to recur, even following extensive treatments with high doses of penicillin (Hobby et al., 1942). In 1944, Bigger realized that, contrary to the belief prevalent at the time, penicillin is not bacteriostatic, but rather kills staphylococci. He also realized that the complete sterilization of cultures treated with penicillin is not easily achieved (Bigger et al., 1944a, b). He attributed the failure in sterilization to a minute fraction of the population, consisting of bacteria, that he named ‘persisters,’ which he evaluated at about one per million. Unlike resistance, which is genetically acquired and passed on to subsequent generations, persistence is a transient phenotype typically observed when a population of cells is exposed to antibiotics and the number of survivors is monitored. An example of the resulting killing curve is shown in Fig. 1a. The initial steep decrease in survival, characterized by a fast death rate, is followed by a much slower decrease, which reveals the existence of persister cells. These cells have not genetically acquired resistance; they regrow a population that is still sensitive to the antibiotic treatment. The exposure to the same antibiotic of the population regrown from these survivors results in a killing curve identical to that of the original bacterial population (Fig. 1b, solid line). For many years, this phenomenon was overlooked, mainly because in most bacterial infections, persisters are eventually eliminated by the immune system. However, it became increasingly obvious that persistence is the main problem in diseases such as tuberculosis, where the immune system proves ineffective and a single surviving bacterium can start an infection (Stewart et al., 2003).


Persistence in batch cultures. (a) Killing curve of Escherichia coli high-persistence strain hipA7 under ampicillin. At t=0, bacteria are exposed to ampicillin and the fraction of survivors is monitored. The steep initial decrease is followed by a slower rate, revealing persister bacteria. The solid line represents a fit to the experimental data using the mathematical model presented in Eqn. (2) [adapted from Balaban et al. (2004)]. (b) Schematic killing curves of different responses to antibiotics: tolerance (dashed–dotted line), a sensitive population that initially contains resistant mutants (dashed line) and persistence (solid line). Here the first antibiotic treatment (t=0 h) is followed by a period of regrowth without antibiotics and a second exposure to antibiotics (t=12 h). (c) Schematic killing curves of responses of uniform populations to antibiotics with different killing rates, k: sensitive wt (solid line), tolerance with lower k (dashed line) and resistance with positive growth rate μ (dashed-dotted line). (d) Schematic representation of the stochastic switching between persisters, p, and normally growing cells, n. In Type I persistence, persisters spontaneously switch to normally growing cells, but are formed only after a stress signal, whereas in Type II persistence, the switching occurs spontaneously in both directions.

Despite having been first observed more than 60 years ago, bacterial persistence remains a puzzle. Several hypotheses have been put forward in order to explain the heterogeneous response to antibiotics in apparently uniform populations (Lewis, 2000). Already in 1944, Bigger had suggested that persisters are in a dormant state (Bigger, 1944a, b). It has also been proposed, for instance, that persistent bacteria are those that find themselves in some ‘protected’ part of the cell cycle at the time of exposure to antibiotics. Persistent bacteria have also been viewed as being those that are able to adapt rapidly to antibiotic stress, for instance by switching on/off genes linked to general stress responses. Alternatively, it has been suggested that persisters are defective in a programmed cell death module that is triggered by the antibiotics (Lewis, 2000; Sat et al., 2001).

Apart from being of considerable clinical relevance, the study of persistence has revealed the inherent nongenetic heterogeneity of bacterial populations, which might constitute a general adaptation to variable environments.

Experimental analysis of bacterial persistence

The study of persistence requires the measurement of survival of the bacterial population over time

As discussed in Mathematical analysis (see also Fig. 1b), survival assays that evaluate the survival after only one specific time interval of antibiotic treatment cannot unequivocally measure persistence, and continuous measurements of survival over time are needed. Several methods have been devised for measuring killing curves in batch cultures; among them the serial dilutions plating on solid medium, the most probable number (Hurley & Roscoe, 1983), viable stains and luciferase (Virta et al., 1998). The extremely low number of persisters makes the latter methods inadequate. Most studies have used serial dilutions plating to determine the extent of persistence, which manifests itself as a decreasing death rate at the longer exposure times. Batch cultures have been used to study the dependence of persistence on the type of antibiotic or killing agent used (Moyed & Bertrand, 1983; Wolfson et al., 1990), on temperature (Scherrer & Moyed, 1988) and on the bacteria's growth phase (Korch et al., 2003). Two types of persister bacteria were observed (Balaban et al., 2004): (1) Type I persisters, which appear only after the culture has reached the stationary phase (Korch et al., 2003), and (2) Type II persisters, which are continuously generated during the exponential growth of the population. The difference between the two types of persisters is illustrated in Fig. 1d and will be further discussed below. The study of persistence was advanced with the isolation of mutant strains with levels of persistence that can be observed under the microscope, and with the development of various genetic analysis techniques, as described below.

Genetic analysis of persistence-related genes

In order to increase the number of persisters and to identify putative genes involved in persistence, Moyed and colleagues (Moyed & Bertrand, 1983) developed a screen to identify mutants of Escherichia coli K-12 that exhibit increased persistence to ampicillin treatment, an antibiotic that targets the cell wall assembly. They isolated high-persistence mutants with a persistence level 1000-fold higher than its parental strain, and identified a locus named hip for high persistence. The location of the mutations was later mapped to minutes 33.8 on the chromosome (Moyed & Broderick, 1986), in a two-gene operon, named hipBA (Black et al., 1991). This module was found to follow the main characteristics of toxin–antitoxin modules. Overexpression of HipA causes growth arrest (Falla & Chopra, 1998; Correia et al., 2006) and is counteracted by HipB (Korch & Hill, 2006). Hill et al. sequenced the hipBA locus of the high-persistence mutant, hipA7, and identified two point mutations (Korch et al., 2003). Despite thorough analysis of the hipBA module, the key feature of persistence – namely the coexistence of growth arrested and normally growing cells – has not yet been explained. An additional locus, hipQ, was identified in a screen for persistence to quinolones, which inhibit DNA replication (Wolfson et al., 1989). The hipQ locus was mapped to minutes 2 on the chromosome, but its precise location and the relevant genes are still unknown. It was found that hipA7 and hipQ mutant strains exhibit increased persistence under inhibitors to both DNA replication and cell wall synthesis (Wolfson et al., 1990). Despite their apparent similarity, these loci were later found to generate different types of persistence (see Mathematical analysis). High Type I persistence was observed in the hipA7 mutant and high Type II persistence in the hipQ mutant. Wild-type (wt) E. coli seems to generate both types of persister bacteria (Balaban et al., 2004).

Recent genetic analyses, using a library of overexpression gene constructs or transposon mutagenesis, have identified several genes –glpD, plsB (Spoering et al., 2006) and phoU (Li & Zhang, 2007) – that are involved either in the formation or the maintenance of persistence, but the role they play in the persistence phenomenon is still unclear (Lamarche et al., 2008). According to the protocols used in these works, where cultures were inoculated with a large inoculum from the stationary phase, it is likely that the candidate genes identified are involved in Type I persistence. It would be interesting to apply the same genetic screens and identify genes involved specifically in Type II persistence. This would require avoiding the formation of Type I persistence by serial subcultures. In effect, because Type I persisters require passage through the stationary phase for their formation, keeping a culture constantly at exponential growth in serial subcultures can reduce their fraction in the population to zero (Balaban et al., 2004; Keren et al., 2004).

The picture that emerges so far from genetic studies is not pointing clearly at one single mechanism responsible for persister formation. Rather, it seems likely that the activation of different stress modules results in growth arrest and that various different genetic pathways converge toward persistence (Dhar & McKinney, 2007).

Single-cell observations in microfluidic devices: persisters have a reduced growth rate before the antibiotic treatment

Although batch culture measurements made possible the discovery and initial characterization of persistence, the discrimination of the various putative mechanisms behind the persistence phenomenon required direct observation under the microscope. The high level of persistence in previously identified hip mutant strains facilitated the direct examination of persisters under the microscope. The goal was to monitor persister cells before, during and after being subjected to antibiotic treatment. However, the task of keeping single bacteria under the microscope for many hours under variable conditions required new techniques. Microfluidic devices were designed and fabricated, with soft lithography being used to pattern layers of a transparent polymer (McDonald et al., 2000; Quake & Scherer, 2000). These devices made it possible for the growth of individual bacteria to be recorded under various growth conditions, while trapped between the patterned layers (Balaban et al., 2004). The bacteria's response to antibiotic treatment could then be monitored and the rare survivors could be analyzed. A typical experiment is presented in Fig. 2a–f: bacteria are first introduced in the device and grown under normal growth conditions; after a few divisions, antibiotic is added and results in the lysis of most of the bacteria; a few hours later, the antibiotic is removed and the regrowth of persister cells is monitored; the antibiotic treatment is maintained until the fast exponential killing ends (see Fig. 1a). Remarkably, even before the antibiotic treatment, all the observed persisters were clearly distinguishable from normal cells by their reduced growth rate. This observation was sufficient to explain their low sensitivity to the antibiotic, as cells with reduced growth rates are known to be less sensitive to many antibiotics that target the cell replication machinery. The quantitative characterization of the dependence of growth rate on the killing rate was demonstrated when E. coli populations growing at various growth rates in glycerol-limited chemostats were shown to be killed by β-lactams at a rate that is directly proportional to the growth rate (Tuomanen et al., 1986).


Persistence at the level of single cells. (a–f) Single-cell observations of persister cells in a microfluidic device, where bacteria are trapped in microlines. Initially (a–c), the bacteria grow in fresh medium, forming linear microcolonies. Then (d), an antibiotic is added to the medium, killing most of the bacteria. When the antibiotic is removed (e, f), a single persister bacterium that switches back to normal growth is observed. This bacterium was not growing already before and during the antibiotic treatment, and was thus protected from antibiotic killing [adapted from Balaban et al. (2004)]. (g, h) Schematics of possible induction curves for normally growing (black) and persister (red) cells. In (g), the induction strength is lower in persister bacteria, but the time dependence of the response is similar, whereas in (h), the induction strength is identical for persisters and for normal bacteria, but the response is slower for persister bacteria. (i) Experimental determination of the induction curves for persisters (red) and normally growing cells (black). The initial response is identical. Only after 1.5 h do persisters stop responding to the induction signal [adapted from Gefen et al. (2008)]. (j) Current model for the formation of Type I persisters after passage through the stationary phase.

Persistence in the hip strains of E. coli is therefore linked to the inherent heterogeneity of growth rates in the bacterial population, rather than to the interaction between a specific antibiotic treatment and the bacteria. The population contains subpopulations that differ in their growth rate: fast and slow growers. In effect, the role played by the antibiotic is mainly to reveal the heterogeneity of growth rates, namely to eliminate the fast-growing cells that normally mask the presence of the slow growers.

While our understanding of bacterial persistence was advanced by single-cell studies, the molecular mechanism responsible for the generation of subpopulations of bacteria with reduced growth rates is still far from being understood.

Gene expression profiles of persister bacteria

In order to shed light on the genes involved in persisters' formation, gene expression profiling of persister bacteria was undertaken. Persister populations were initially enriched by ampicillin treatment. Genes whose expression increased with enrichment were classified as specific persister genes (Keren et al., 2004). These genes included the SOS stress response genes, in particular sulA, which inhibits cell division, as well as several toxin–antitoxin modules. One possibility would be that persister cells are spontaneously preinduced for regulatory networks that arrest replication and growth (Debbia et al., 2001). However, the discrimination between the response to the ampicillin treatment and the expression of specific persister genes proved difficult in the gene expression assay. For example, low doses of ampicillin have been shown to trigger the SOS response (Miller et al., 2004), which makes it difficult to attribute the elevated SOS gene expression level to persistence per se. An improved enrichment technique was devised, whereby persisters were sorted by fluorescent activated cell sorting (FACS) according to their fluorescence level (Shah et al., 2006). Here, the assumption, based on single-cell observations, was that persisters do not grow shortly after the stationary phase and therefore show a low level of a short-lived variant of green fluorescent protein (GFP), whereas normal cells continuously produce new GFP proteins. The procedure therefore selects for Type I persisters. Weakly fluorescent cells were sorted by FACS and their gene expression levels were compared with those of the brighter cells taken either from exponential- or stationary-phase cultures. The results show that persisters of the wt E. coli differ from both stationary-phase cells and exponentially growing cells, that they downregulate chemotaxis genes and overexpress several toxin–antitoxin genes. These studies suggest that persisters might be the result of a specific differentiation program and not the mere byproduct of a defective growth cycle.

Single-cell gene induction reveals the Achilles heel of persister bacteria

Persister cells exhibit a reduced growth rate already before antibiotic treatments. For Type I persistence, generated upon starvation, persisters were believed to reach their shutdown state at the stationary phase and to remain dormant even when transferred to fresh medium. The difficulty of characterizing these nongrowing bacteria was evidenced by the subtlety in properly isolating the subpopulation. Another approach that did not require the sorting out of persister bacteria consisted of introducing reporter plasmids with simple synthetic genetic circuits into the bacterial population, and then measuring the single-cell gene expression dynamics of persister bacteria (Gefen et al., 2008). Well-characterized inducible promoters were fused to fluorescent protein genes and induced under the microscope, so that fluorescence increase of persisters and normally growing cells could be compared. In normally growing cells, the fluorescence increase reached a steady state on a time scale comparable to the division time. Two main parameters could be extracted from the shape of the fluorescence increase: the time to steady state and the strength of the induction (Alon et al., 2007). Based on the previous understanding that persister bacteria become dormant at the stationary phase and remain that way despite being exposed to fresh medium, persister cells were expected to respond to the induction signal either more weakly (Fig. 2g) or more slowly (Fig. 2h), when compared with normally growing bacteria. However, when the induction process was monitored in a hipA7 stationary-phase population containing both normal and persister bacteria, the surprising observation was that persister bacteria responded to the induction as quickly and as strongly as normal cells during the first 1.5 h following their transfer to fresh medium. Only thereafter did they stop responding to the induction signal and arrested protein production (Fig. 2i). This suggested the existence of a checkpoint, at the end of a 1.5-h time window, when bacteria fully differentiate into dormant persister cells (Fig. 2j). During this window of opportunity, following exposure to fresh nutrients, persistence to antibiotic treatment was found to decrease by an order of magnitude. It is interesting to note that a related phenomenon was reported for Staphylococcus aureus, after dilution in large volumes of broth (Bigger, 1944a, b). The technique of single-cell gene induction could serve as a general characterization tool of nongrowing cells such as those found in biofilms (Lewis, 2005) or in the viable but not culturable state (Oliver, 2005).

Mathematical analysis

The mathematical description of persistence presented in this section starts with a mathematical definition of persistence and follows with a simple model of phenotypic switching. The model aims at pointing out the relevant parameters, and at ways to determine those parameters experimentally. We start with a model of a population variability that arises through the switching between two very different growth phenotypes. We discuss at the end how this model can be easily extended to treat any number of discrete states, and therefore, also population variability that originates from a continuous phenotype.

The mathematical analysis, which might seem a mere mathematical interlude, in effect provides quantitative predictive power. For example, it predicts that a cyclic exposure to antibiotics, as proposed by (Bigger, 1944a, b), would not be effective at fighting persistence, while suggesting alternative strategies. Furthermore, as shown below, the quantitative measurements of the key parameters that characterize persistence make possible the discrimination between seemingly similar persistence phenotypes. Finally, the mathematical analysis of persistence allows making a quantitative prediction on evolutionary processes that would promote the emergence of persistence, as well as evaluating its ecological implications.

Persistence is an attribute of a heterogenous population

The mathematical analysis of persistence seeks to define the phenomenon as precisely as possible, so as to discriminate between different phenomena that also result in an increased survival of bacteria under antibiotic treatments (Levin & Rozen, 2006). We will start with the simple equation that describes the death of a uniform population under antibiotic treatment. Assuming an interaction between bacteria and antibiotics that does not change with time and a homogeneous bacterial population, one may assign a single parameter, the killing rate k, that represents the probability for bacteria to be killed, per unit time. The dependence on time of the total population, N(t), is then described by Embedded Image 1

The solution of Eqn. (1) is an exponential decay of the population with time: Embedded Image. A single killing rate k can be used to characterize uniform populations. A population that is killed at a much slower rate than the sensitive wt population is defined as tolerant namely: kwt>ktolerant≥0; a population that grows in the presence of the antibiotic is resistant, and the killing rate will be replaced by a growth rate μ, which characterizes the increase of the population with time: Embedded Image (Fig. 1c).

A persistent population cannot be characterized by a single killing rate, but with at least two different rates, each of which characterizes different subpopulations (Fig. 1a). Here, persistence is defined as the heterogeneous killing of the bacterial population by antibiotics, namely the coexistence of cells that are rapidly killed and of cells that are killed very slowly. The cells that are killed rapidly have the same killing rate as the wt, while the cells that are killed very slowly, namely persisters, are in fact ‘tolerant’ to the antibiotic treatment.

Different phenomena involving reduced killing by antibiotics have often been mixed up in the literature. The confusion arises mainly because persistence, unlike resistance or tolerance, is an attribute of a heterogeneous population of cells and cannot characterize a single cell. Several studies have reported persistence by measuring the survival of bacterial populations to antibiotics at only one time point after treatment. In Fig. 1b, it can be seen that different phenomena might lead to similar survival if the population's survival is measured only at a single time point, as in this example after 7.5 h of antibiotic treatment (Vazquez-Laslop et al., 2006). Therefore, survival per se cannot be regarded as the fingerprint of persistence. Rather, it is the heterogeneity of killing rates that needs to be assessed for demonstrating persistence.

A simple model of persistence

As discussed above, a population exhibiting the persistence phenomenon may be described as consisting of at least two subpopulations, each characterized by a different killing rate: a ‘normal’ (n) population of bacteria that is killed rapidly by antibiotics at rate kn, and a tolerant, ‘persister’ (p), subpopulation of bacteria that dies at a much slower rate kpkn. These different killing rates are directly linked to the subpopulations' different growth rates, namely, μp≪μn, where μp and μn are the growth rates of persister and normal bacteria, respectively. A distinct persistent state can be defined when the growth rate of persisters is reduced far beyond the natural variation of growth rates expected in the population around its mean. Typically, for several bacterial populations, the coefficient of variation of division times has been determined to be around 0.3 (Powell et al., 1955). The number of persister bacteria generated in a time interval Δt is proportional to the rate at which normal cells switch to persister bacteria, a, and to the total number of normal bacteria at that time, n(t). Therefore, the total persister population p(t) increases by a·n(t)·Δt while the normal population, n(t), decreases by the same amount. Similarly, the switching of persisters back to normal cells with rate b increases the number of normal cells by b·p(t)·Δt and decreases the number of persisters by the same amount. Also taking into account the growth of each subpopulation leads to: Embedded Image (2,3)

Here, normal and persister bacteria grow without antibiotics at growth rates μn and μp, respectively. Under antibiotic treatment, the same equations apply, but the growth rates μn and μp need to be replaced by −kn and −kp, respectively, for the killing rates.

Different persistence types

The above model has led to the characterization of persistent populations in terms of the two rate constants a and b. When these rate constants were measured for different E. coli strains, two different persisters formation processes were observed, and two types of persisters were defined (Balaban et al., 2004) (see Fig. 1d):

  1. ‘Type I persisters’–a=0; b≠0 – formed in response to an external trigger only; for example, starvation, followed by full differentiation (Gefen et al., 2008).

  2. ‘Type II persisters’–a≠0; b≠0 – formed stochastically and continuously during the population growth.

The mathematical characterization of persistence facilitates the discrimination between intrinsically different phenomena, which previously appeared to be very similar experimentally. For example, the killing curves of the high-persistence mutants in Fig. 3a and b are difficult to tell apart, until the quantitative measurement of their switching rates reveals different mechanisms: hipA7 persisters are of Type I, whereas hipQ persisters belong to Type II. Typically, wt bacterial populations contain a small fraction of both types of persister bacteria. Whereas the number of Type I persisters slightly decreases during the populations' exponential growth and suddenly increases at the onset of the stationary phase (Fig. 3c), the number of Type II persisters typically increases with growth (Fig. 3d).


Experimental determination of the switching rates. (a) and (b) Killing curves under antibiotic treatments of hipA mutant (Type I) and hipQ (Type II) strains, respectively [adapted with permission from [Wolfson et al. (1990)]. (c) Schematic growth of Type I persisters (red) vs. normal cells (black). The persisters' number decreases slowly during population growth and increases at the beginning of the stationary phase. (d) Schematic growth of Type II persisters (red) vs. normal cells (black). The persisters' number increases with growth, typically remaining at a constant fraction due to stochastic switching of normal bacteria to persisters. (e) Regrowth of the total population (n+p) after an extensive antibiotic treatment. Type II persisters cells grow slowly and constitute the majority of the population until a switching event occurs and generates normal, fast-growing cells. The solid line represents a fit according to the solution of Eqns (2) and (3) [adapted from Balaban et al. (2004)].

Passage through the stationary phase has been shown to be required for Type I persistence in the hipA7 mutant, although the full differentiation is attained only during regrowth. The stationary phase is a complex process in which several parameters change (Kolter et al., 1993; Nystrom et al., 2004), among them, nutrients, pH, cell density, etc. Starvation for different nutrients might lead to different states (Peterson et al., 2005). Type I persisters in the hipA7 strain have been linked to amino acid limitation and ppGpp levels (Korch et al., 2003), but other starvation conditions as well as quorum-sensing factors should be investigated. Type I persisters keep the ‘memory’ of dormancy on a time-scale that is much longer than the typical division time, but is not passed on to their descendants. Once these Type I persisters divide, they become normal cells and memory is lost. In contrast, the Type II persisters observed in the hipQ strain have a reduced growth rate that is passed on over several divisions before switching back to normal cells, thus leading to nongenetic inheritance.

Equations (2) and (3) rely on simplifying assumptions on the interaction between bacterial populations and antibiotics, but nevertheless capture the essential features of the populations' dynamics. Furthermore, they can be analytically solved (see Supporting Information, Appendix S1), and the time constants that dominate the dynamics can be extracted from experiments, as discussed below.

Other models have included additional factors such as the dependence of persistence on the antibiotic concentration, the ability of bacteria to mutate and become resistant (Levin & Rozen, 2006) or the specific mechanism responsible for the switching (Klapper et al., 2007). This latter work proposes bacterial aging as the differentiation mechanism between persisters and normal cells, and might explain persistence in wt populations. These mathematical descriptions enable identifying other relevant parameters governing the population dynamics and will require further experimental data both in vitro and in vivo to tell apart.

Experimental determination of the parameters

Once the mechanism of persistence is demonstrated at the level of single cells, measurements of the parameters can be made in batch cultures. Figure 3e plots the growth of the total population (N=n+p) in a strain with elevated Type II persistence after an extensive antibiotic treatment that killed all normal cells. The initial conditions are therefore n(t0)=0, p(t0)=N(t0).

The behavior of the total population is initially dominated by the slow regrowth of the persisters. Then, the switching of persisters to normal cells results in a fast growing population. The two-parameter fitting procedure using the general solution (Appendix S1) allows the measurement of b and μp. At the same time, the growth of the persister cells can be monitored and fitted with the same parameters. By departing from different initial conditions for Eqn. (A1) (Appendix S1) [n(t0)=N, p(t0)=0] and monitoring the rate of appearance of persisters, a can be determined.

The determination of the parameters from batch culture avoids the inherent stochasticity apparent in the single-cell behavior. However, it is imperative to compare the parameters extracted from batch cultures with those evaluated from the single-cell experiments. Single-cell experiments are also essential to prove that persisters switch back to normal fast-growing cells. Otherwise, the curve in Fig. 3e might be the result of very rare survivors of the normal population managing to outgrow the slow-growing population of persisters.

The persistence observed in the high-persistent mutant isolated by Moyed can be fully described by this two-state model. This is already apparent in the killing curve of that strain (Fig. 1a) that is well fitted with the sum of two exponential decays – a fast killing of the normal bacteria, followed by the much slower killing of the persister bacteria. However, in general, population variability might be due to more than two subpopulations and even due to a continuum of phenotypes. The resulting killing curves would then depart from the initial exponential decay in a smooth transition between the different states. For example, the killing curve of the wt K-12 can be fitted with three subpopulations, each characterized by a different killing rate (Balaban et al., 2004). The model described above can then be extended to three subpopulations that can each switch between the different phenotypes. In the general case, the model can describe switching between M different phenotypic states, and will consist of M differential equations similar to Eqns (2) and (3), whose parameters, Mij, represent the switching rate from phenotype i to phenotype j. A continuum of states can also be described by the same formalism by increasing the number of states. In practice, experimental determination of switching rates for more than a few subpopulations can be difficult, because of the limited dynamic range of measurements. Experimentally, characteristic switching times can be measured by the time it takes for a certain fraction of the distribution to span the whole distribution. For example, the high- and low-fluorescence fractions of a continuous distribution of GFP in a population have been sorted out by FACS and regrown. The number of generations needed to recover the whole span of GFP fluorescence in the population was measured and found to be highly asymmetric (Brenner et al., 2006).

Clinical implications

Most of the studies cited above have been performed on the E. coli K-12 non pathogenic strains, but persistence is not restricted to laboratory strains. It is a ubiquitous feature of microbial populations and has also been observed in many pathogens, including pathogenic E. coli (Marcusson et al., 2005), Mycobacterium tuberculosis (Wallis et al., 1999), S. aureus (Bigger et al., 1944a, b) and Pseudomonas aeruginosa (Keren et al., 2004). A survival higher than expected at long exposure times has also been reported for Candida albicans in both in vitro and in vivo studies (LaFleur et al., 2006; Pacini et al., 2007).

If persister bacteria are an inherent part of all microbial populations, the question arises as to why the phenomenon was overlooked for so many years. As it happens, the survival to antibiotic treatments of a small population of pathogens is not always problematic. Many antibiotics are bacteriostatic – that is, they prevent the growth of bacteria without killing them. Once bacteriostatic antibiotics are washed away, high survival of bacteria will be observed in vitro, even though, in vivo, they might be as effective in fighting infections as bacteriocidal antibiotics. In most infections, our immune system is able to cope with surviving bacteria, provided they do not grow too fast, but persistence will turn out to be a major problem in all cases where bacteria manage to evade our immune system. These instances have been described in (Nataro et al., 2000) and span 3 major mechanisms: (1) in immunosuppressed hosts, (2) in pathogens that have adapted to our immune system and (3) in niches within the body that are less accessible to the immune system. The first mechanism affects specific hosts, the second depends on the pathogen strain and the last can be found in almost any host–pathogen combination.

In immunosuppressed hosts, the persister bacteria might be the main reason for lethality. These include HIV and cystic fibrosis patients who mainly die from high-persistence pathogens: M. tuberculosis or Mycobacterium avium and P. aeruginosa, respectively (Nataro et al., 2000).

Several pathogens are so well adapted to our immune system that they are able to live intracellularly in the immune system cells. The minimum lethal dose of pathogens that kills 50% of non-immunocompromised hosts can be as low as 1. The survival of even a few persister pathogens in the host may have fateful consequences so that it is essential to find antibiotic treatments that are effective against them. However, when antibiotics are tested in the laboratory against specific infections, the assay usually performs on the population as a whole. In order to illustrate why assays that test the effect of the antibiotics specifically on persisters are needed, imagine a population displaying persistence, with a fraction of 1/1000 persisters: an antibiotic that, within a few hours, will kill 98% of the persister bacteria and 98% of normal bacteria will not be chosen over another antibiotic that kills 99% of normal bacteria and 0% of persisters.

Finally, even pathogens that do not usually evade our immune system may, under certain circumstances, find themselves in specific niches that protect them from either detection or killing. A striking example of the importance of the physical surroundings of the infection was shown 50 years ago in experiments carried out in human volunteers infected with different amounts of S. aureus (Elek & Conen, 1957). The purpose was to determine the minimal amount of bacteria needed to establish a significant pus-forming subcutaneous infection. The number of bacteria needed in the injection was above a few millions. However, when bacteria were injected together with a sterile foreign body, such as a silk suture, an infection developed with as few as 100 bacteria, and was difficult to cure with antibiotics. The reason for the much higher infectivity of pathogens in the presence of a foreign body remained a puzzle for many years. We now know that foreign bodies often promote the development of biofilms (Fux et al., 2005). When biofilms were shown to be linked to the failure of antibiotic treatments, it was proposed that a reduced diffusion of antibiotics might be the cause (Hoyle et al., 1992). Today, it is understood that biofilms promote the survival of bacteria by two main mechanisms: by creating a niche where bacteria can evade the immune system (Costerton et al., 2001) and by triggering dormancy, thus generating Type I persistence (Lewis, 2007).

The mathematical model of bacterial persistence suggests several lines of action toward a reduction of persistence. First, the killing rate of persisters could be increased. To this end, specific tests for antibiotics efficiency against persisters should be adopted. Therapies combining antibiotics that are effective against persisters with conventional antibiotics that kill normal bacteria should be tested for synergistic effects. An attractive strategy would be to reduce persistence by increasing b, the switching rate from persisters to normal cells. Growth factors, quorum-sensing-related factors or resuscitation-promoting factors (Mukamolova et al., 2002a,b; Keep et al., 2006) might be promising candidates as agents that increase b. For example, it was found that exposing Type I persisters to a fresh growth medium together with the antibiotic treatment reduces persistence by 90% (Gefen et al., 2008). Finally, preventing persister formation by decreasing a in the case of Type II persistence, or by preventing the detection of the triggering event in the case of Type I persistence, might provide another line of action against persistence.

Persistence to other stresses

Several stresses inflicted upon different microorganisms have been shown to result in biphasic killing curves similar to the solid line in Fig. 1b that cannot be explained by the appearance of resistant mutants. It has been observed for many different antimicrobials–microorganisms interactions, as described above, as well as for acid stress in E. coli (Booth, 2002) and yeast (Steels et al., 2000; Rando & Verstrepen, 2007), phages (Fischer et al., 2004; Pearl et al., 2008) and metals at toxic concentrations (Harrison et al., 2007). The phenomenon seems to be ubiquitous, and it is our hope that the recent interest in nongenetic variability will lead to more systematic studies of the survival fraction under the different stresses.

The mathematical description developed for bacterial persistence to antibiotics should also describe the population dynamics under those stresses, if the main reason for the biphasic killing in those stresses is the pre-existing bimodality of growth rates in the population. The experimental approach used to characterize persistence to antibiotics, using single-cell observations in microfluidic devices, should enable revealing the mechanisms behind persistence to these stresses, and determine whether the heterogeneity is pre-existing or rather stems from an adaptive response to the stress.

Pre-existing slow growth or dormancy protects bacteria against various stresses, as pointed out by Booth, (2002). High-persistence mutants exhibit higher survival fractions to different classes of antibiotics as well as to heat shock, thyminless death, nalixidic acid (Scherrer & Moyed, 1988) and prophage induction (Pearl et al., 2008). However, persister cells seem to be killed as effectively as normally growing cells by antibiotics such as kanamycin, and by stresses such as UV radiation (Scherrer & Moyed, 1988) or lytic phage infections (Pearl et al., 2008). What determines whether a killing agent will be effective against persister bacteria? An insight into the kind of stresses that might kill persistent bacteria comes from studies of the interaction between persister bacteria and the λ phage; whereas persister bacteria survive killing by prophage induction, namely by phages that have integrated into the bacterial chromosome, they do not survive infections by lytic phages. Single-cell observations of the timing of killing by lytic phages have shown that persister bacteria are not directly killed by the phage and that it is only when persisters switch to normal growth that the lytic process proceeds and lysis occurs. Experiments, in which bacteria infected with lytic phages were observed under the microscope, have revealed that the expression of phage genes starts within minutes in normal cells, whereas persister cells were killed only after many hours, on a time scale that is solely determined by the rate of switching from persisters to normally growing cells. Infecting phages are able to attach to the bacteria in spite of extensive washing out, and to be able to wait for as long as a day for the switch to occur (Pearl et al., 2008). Antibiotics, however, act in a different manner: once the treatment is stopped and the antibiotic is washed away, the persisters that switch back to normal growth are out of danger. Treatments that will effectively kill persister bacteria will be those that inflict an irreversible damage, or would remain active inside the dormant bacterium until it wakes up. Such drugs would need to be irreversibly altered upon entry into the bacterium in such a way that prevents their diffusion out, once the drug is removed from the environment.

Ecological and evolutionary implications: persistence as a survival strategy in fluctuating environments

As defined above, persistence is a characteristic of a heterogeneous population of cells; it may therefore present an advantage at the level of the population and not at the level of single cells. What are the conditions under which phenotypic variability is advantageous over the regular adaptive responses to stress (Jablonka & Lamb, 1995)? Lachman and colleagues analyzed the advantage of switching between two different phenotypes, when a population of organisms is subjected to an environment that cycles between two states, each advantageous to a different phenotype (Lachmann & Jablonka, 1996) (Fig. 4a). They found that a cyclic environment would strongly select for a phenotypic switch of the same frequency. In the case of persistence, the two environments would be one with and the other without antibiotics, which would select for persisters or normally growing bacteria, respectively. Using the mathematical analysis presented in Mathematical analysis, Kussell et al. (2005) extended the results of Lachmann & Jablonka to environmental fluctuations that might each have a different time scale and mapped the phase-space region of parameters that would be advantageous for the hipQ strain over the wt strain (Fig. 4b). The analysis on evolutionary time scales shows that the two main parameters that evolution would select for are the two switching rates a and b (see Mathematical analysis), whereas the death and growth rates play a lesser role. Their conclusions suggest that persistence might have evolved as a bet-hedging strategy to cope with a fluctuating stressful environment. In a recent competition experiment between a high- and a low-persistence strain under cyclic prophage induction conditions, persistence was shown to provide a strong selective advantage (Pearl et al., 2008). Therefore, the bet-hedging strategy of persistence may be a general nonadaptive stress response that could have evolved not only under antibiotic stress (Gardner et al., 2007) but also under the ubiquitous stress imposed on bacteria by prophages in the wild. Further analysis of the costs and benefits of evolving a persistence strategy vs. evolving an adaptive response has been carried out by (Kussell & Leibler, 2005). The results may be generalized for many different biological systems where stochastic switching occurs (Dubnau & Losick, 2006; Smits et al., 2006), and have recently been experimentally demonstrated even when the selective pressure is not as high as under antibiotic treatments (Acar et al., 2008). Further analyses of the advantage of stochastic switching have shown surprising analogies between systems as far apart as bacterial competence and photoreceptor differentiation in the Drosophila retina (Losick & Desplan, 2008).


Evolutionary and ecological implications of persistence. (a) Schematic representation of a bacterial population with two phenotypic variants under a fluctuating environment. The environment fluctuates between two states (red and gray), each advantageous to a different phenotype. In the red environment, red bacteria outcompete gray bacteria, whereas in the gray environment, the situation is reversed. (b) Phase-space representation of the result of a competition experiment between strains with different switching rates in a fluctuating environment [adapted from Kussell et al. (2005)].


Renewed interest in persistence has advanced our understanding of the phenomenon from various different angles. Genetic analyses have identified several modules involved in the persistence phenomenon, single-cell analyses have demonstrated the link to growth rate variability, and mathematical models have led to quantitative predictions of population dynamics. Clinical research on pathogens that have long been recognized as being involved in persistent infections such as tuberculosis has provided an insight from a different perspective. The challenge remains to combine the different approaches toward a comprehensive understanding of the mechanisms behind persistence, so that general strategies against persistence can be devised. For example, it would be interesting to understand whether the trigger of Type I persistence at the stationary phase depends on starvation conditions, on cell density and, in vivo, on the immune response. At the level of gene regulation, the goal remains to understand the type of processes that can lead to the observed differentiation of the population into distinct subpopulations (Kaern et al., 2005). Theoretical models that have been successful at describing other nongenetic mechanisms (Smits et al., 2006; Suel et al., 2006; Maamar et al., 2007) might be relevant for persistence. However, experimental data that would enable to discriminate between different models of stochastic switching (Booth et al., 2002), or of sensitivity to initial conditions (Lu et al., 2007), are still lacking. Beyond its relevance for bacterial infections, the heterogeneity of growth rates inherent to bacterial populations might have wide implications, for example, for the stability of bacteria and phages in the environment (Pearl et al., 2008) or the evolution of resistance (Levin & Rozen, 2006). The phenotypic variability that becomes apparent after extensive antibiotic treatments might only be the tip of the iceberg of the general phenotypic variability that occurs in populations, which should force us to rethink the way in which populations of cells respond to the environment, adapt and evolve.

Supporting Information

Additional Supporting Information may be found in the online version of this article.

Appendix S1: Mathematical description of persistence: General equations describing the dynamics of two sub-populations, here denoted as “normal” (n) and “persister” (p) cells. Each sub-population is characterized by a growth/death rate, μn and μp respectively and by rates of switching from one population to the other (a,b). Type I persisters are generated by a trigger event during stationary phase, while type II persisters are continuously generated during growth.

Please note: Wiley-Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.


We would like to thank Charlotte Balaban, Doron Azulay and the members of the lab for comments and suggestions on the manuscript, and Collin Block, Ofer Biham and Stanislas Leibler for illuminating discussions. N.Q.B. acknowledges funding from the Human Frontier Science Program and the Horowitz Foundation. O.G. is supported by the Israel Ministry of Science.


  • Editor: David Gutnick


View Abstract