We measured the level of inhibition by anticompetitor toxins by spotting a dilution series of a cell free extract collected from 48 h old P. aeruginosa
PA01 or PA14 culture onto a lawn of one of 55 different clinical isolates growing on a solid surface. The natural isolates differ in their genetic distance to the producing strain; genetic distance is quantified using full genome digests. The lowest concentration of cell free extract that gave rise to inhibition of the clinical isolate was used to calculate the inhibition score (Figure ) [14
]. Figure depicts the level of inhibition by both PA01 and PA14 as a function of genetic distance of toxin producing strain to the clinical isolates.
Figure 1 Inhibition assay. Lawn of a Pseudomonas aeruginosa natural isolate growing on the surface of an agar plate. Spots of pyocin containing cell free extract from a laboratory strain of P. aeruginosa PA01 were applied on the lawn at different dilutions. The (more ...)
Figure 2 Inhibition by toxin containing cell free extract. Inhibition of clinical isolates by toxins in cell free extract collected from laboratory strains PA01 and PA14 as a function of genetic distance (Jaccard similarity) between toxin producer and clinical (more ...)
Our results lend strong support to the idea that toxins are most effective when active against genotypes of intermediate genetic distance relative to the focal strain. The relationship between inhibition and genetic distance is unimodal, peaking at intermediate genetic distance for both toxin producers PA01 and PA14. This result is confirmed more formally by noting that a quadratic model with an internal maximum is a better descriptor of the data than a linear model (Table ; in the linear regressions, the linear term is not significant), by the lower AIC (Aikake's Information Criterion) values for the quadratic models than the linear models (Table ) and by an F-ratio test asking if adding the quadratic term provides a significantly better fit than the linear model (PA01, F1,48
= 5.96, P
= 0.018; PA14, F1,42
= 17.56, P
= 0.00014). We also tested for the existence of an internal maximum in the data using a Mitchell-Olds and Shaw (MOS) test (as implemented in the R package vegan) following Mittelbach et al. (2001) [33
]. This approach tests the null hypothesis that a quadratic function, fitted to the data, has no stationary point (either a maximum or minimum) within the range provided. Our results reject this null hypothesis for both PA01 and PA14 at the P <
0.1 level (PA01: P
= 0.072; PA14: P
= 0.0006), the same criterion used in Mittelbach et al. (2001) [33
]. Since the sign of the quadratic coefficient for both producer strains was negative, the results of the MOS test indicate the presence of a statistically significant 'hump' within the range of genetic distances examined. The evidence for an internal hump is somewhat weaker for PA01 than PA14 but we note that our test is conservative, as we have not included data on the effectiveness of either strain at inhibiting itself. As both of these values are zero (see Methods), including these values would produce a much more pronounced hump.
Linear and quadratic regressions of inhibition of clinical isolates by sterile (non heat treated) cell free extract of PA01 and PA14 cultures as function of genetic distance (Figure 2)
To verify that genetic distance correlates with resource use, we measured the metabolic similarity of toxin producing strains to the clinical isolates using Biolog plates (see Methods). Metabolic profiles become more divergent with increasing genetic distance, as expected, reflected in the significantly negative linear relationship observed between Jaccard distance and metabolic correlation between pairs of strains (PA01: slope ± standard error = -0.493 ± 0.213; multiple R2
= 0.098, t,49
= -2.312, P
= 0.025; PA14: slope ± standard error = -0.644 ± 0.208, multiple R2
= 0.164, t49
= -3.104, P = 0.0032). These results lend support to the idea that genetic distance is linked to ecological divergence. It is further notable that inhibition score peaked at intermediate metabolic similarities for both PA01 and PA14 but was statistically significant only for PA14 (see Additional file 1
: Table S1 and Additional file 2
: Figure S1; F-ratio test on the fitting of the quadratic term, PA01: F1,48
= 0.176, P
= 0.68; PA14: F1,42
= 7.00, P
It is not immediately obvious why we detected a significant quadratic relationship between inhibition score and metabolic similarity in one strain but not the other. One possibility is that the Biolog plates we used here, which provide profiles on carbon substrate metabolism, represent one of many possible dimensions along which ecological divergence can proceed. Under this interpretation, metabolic divergence in carbon substrate use may reflect a correlated response to divergence in other ecologically important factors such as ability to grow as a biofilm, resistance to various stressors like pH, temperature, or salinity, and possibly even predation, that we did not measure. Thus we would still expect to see some relationship between metabolic similarity and genetic distance, as we did for PA01, even if this is not the sole target of ecological divergence. There are any number of other differences between PA01 and PA14 that could be responsible for this difference. PA14 has a slightly larger genome than PA01 (6.5 Mbp and 6.3 Mbp, respectively) and contains a number of unique 'pathogenicity islands' that are thought to be associated with a generally increased level of virulence in most hosts [34
]. It also is thought to produce only R- and F-type pyocins, whereas PA01 produces all three types (R, F, and S) [4
]. It is notable that S-pyocins differ from both R- and F-pyocins in that they are oligopeptides whereas R- and F-pyocins are both phage-like structures. Why or how the differences in genome content, size, or pyocin identity affects the relationship between inhibition score and metabolic similarity remains an open question, however.
What agents are responsible for killing in our experiments? Bacteriophage were clearly not responsible. If bacteriophage were causing the inhibition of clinical isolates, they would be able to amplify themselves in an exponential culture of the same clinical isolate. This was not the case (see Methods). Three lines of evidence suggest, rather, that toxic compounds such as pyocins or exotoxins excreted by PA01 and PA14 are the main killing agent. The first is that PA01 and PA14 are not killed by their own supernatant. Such a result is consistent with the idea that the toxins are pyocins, as pyocin production involves specific immunity genes that confer resistance by preventing lysis in non-producing kin [4
], although it does not rule out the possibility that other toxins with similar immunity properties are also involved. If killing were associated with a non-specific toxic compound such as some waste product, we would have expected both producer strains to be susceptible to killing and killing would most likely also not depend on genetic or metabolic similarity. Second, repeating the inhibition assay with heat-treated supernatant eliminates killing (Figure ; both linear and quadratic regressions are non-significant), providing strong support for the idea that the killing compounds are proteins. Third, and most interestingly, inhibition by PA01 is stronger, on average, than that by PA14 (mean log inhibition score for PA01 = 1.51; mean log inhibition score for PA14 = 0.95; t-test, t93
= 6.05, P
< 0.0001), a result that is likely due to the fact that PA01 produces a larger array of pyocins than PA14, including S-type pyocins [4
Figure 3 Inhibition by heat treated cell free extract. Inhibition of clinical isolates by heat treated cell free extract collected from laboratory strains PA01 and PA14 as a function of genetic distance (Jaccard similarity). No regression gave a significant fit (more ...)
Although these three lines of evidence point suggestively to pyocins as being the main killing agent, we have not conducted an explicit test of this hypothesis by, for example, repeating our assays with pyocin knock-out strains. Although it may be possible to conduct such a test by focusing on the prtR/N regulator, which is thought to be a global regulator of known pyocins [4
], it is not clear that such a test would be conclusive since a number of the pyocins in both PA01 and PA14 have yet to be isolated [18
] and there may exist other exotoxins that behave in similar ways to pyocins. Note also that knowing the mechanism of killing, while of obvious interest, is in many ways of secondary importance to the observation that the effectiveness of killing depends in a regular way on genetic distance, at least in the strains we have studied here.
Our main result is that the strength of antagonistic interactions peak at intermediate genetic distance. This pattern is strikingly similar to that expected from theoretical [37
] and experimental [38
] kin selection models for selection using mixed populations of two strains at various ratios to adjust relatedness and considering one bacteriocin and one immunity protein. These models have emphasized how the cost of bacteriocin production is affected by the social environment: bacteriocin production is not favored when producers are both common, because the majority of competitors are kin and so immune to the bacteriocin, and rare, because there are now too few kin to enjoy the benefits of the extra resources. This is clearly not an appropriate interpretation of our results because we did not manipulate the frequency of producers and non-producers in our experimental system to adjust relatedness, as Inglis et al. [38
] have done using degree of kinship as a measure of relatedness.
Rather, our results provide some evidence consistent with the idea that ecological divergence may be important in mediating social interactions. It is notable that the explanation for the ineffectiveness of toxins at inhibiting closely related genotypes (i.e. short genetic distance) in our experiment is likely similar to that in kin selection models: they share a degree of immunity to each other's toxins. However, the ineffectiveness of toxins against distantly related genotypes in our system is probably not directly tied to kin selection. Because increasing genetic divergence is accompanied by reduced overlap in resource use, distantly related genotypes are unlikely to compete for similar resources and so the resources liberated through antagonism are therefore unlikely to benefit the producer [8
]. The production of antagonistic traits such as bacteriocins in this situation is therefore likely to be costly and so selection should lead to decreased levels of antagonism. Our observation of decreased antagonism among distantly related strains, at least for PA14, is consistent with this interpretation.
This interpretation is in accordance with the idea that the evolution of specific antagonism against conspecifics, as is often observed for bacteriocins, could evolve as a direct response to intraspecific resource competition and is consistent with models describing how fitness varies as a function of the strength of resource competition between types: the strength of resource competition, and so the loss of fitness, is highest when resource overlap is intermediate [41
]. To the extent that the opportunity or intensity of resource competition is enhanced through the physical proximity of co-occurring strains in a given habitat [42
], such as a CF lung, then this may further promote the evolution of antagonistic interactions such as those mediated by bacteriocins. It remains to be seen whether our results are specific to the strains we used in this study or whether they apply more broadly to non-CF strains of P. aeruginosa
or other species. This will be an important avenue for future research.
It is not possible with our data to distinguish the specific mechanisms causing variation in toxin susceptibility. If bacteriocins are indeed responsible for killing, then one possibility is that selection targets the total amount of bacteriocin production or the efficiency with which bacteriocins inhibit or kill their victims. It is also possible that the target of selection is the number of receptor sites for bacteriocins in target strains. Deciding among these alternatives requires follow-up experiments that focus on finding the evolutionary origin of bacteriocins using direct competition experiments of producer stains and several target strains to ask under what conditions and by what mechanism bacteriocins aid producer populations to invade populations of sensitive strains [43
]. These experiments would however be very elaborate since the effect of many complicating factors such as frequency dependence, cross-feeding, the viscosity of the environment and exact costs of producing bacteriocins would have to be determined for the interaction of the producer and each target strain. It is even possible that the high specificity of bacteriocins results from their having evolved initially as a by-product of selection for fertility-recognition systems such as conjugation that were later co-opted for use as bacteriocidal agents [49
]. Investigating the relationship between bacteriocin diversity and conjugation frequency or recombination could help shed some light on this issue.
Our results have important implications for understanding of the dynamics of infection in clinical settings. We have firmly established that toxic compounds with high specificity mediate bacterial interactions as antagonistic agents, for instance in structuring pathogen populations in patients with a mixed P. aeruginosa
]. Social evolution theory predicts that selection for antagonism among pathogenic strains should be accompanied by reduced virulence to the host. The consequences of P. aeruginosa
infection on patient morbidity and mortality may therefore depend to some extent on the particular strains present. Furthermore, it is notable that recent research on CF patients from Ontario suggests that 25% of Ontario patients who are infected with P. aeruginosa
are infected with one of two predominant epidemic strains. It may be that the predominance of these epidemic strains is due to the production of specific antagonistic agents such as pyocins [13
]. This is an intriguing hypothesis that awaits further testing. As a start, we have confirmed that three of our clinical isolates produce toxic substances that kill or inhibit other clinical isolates (data not shown). Thus the antagonistic interactions we have studied here do happen among clinical isolates and are not just the consequence of using strains PA01 and PA14 as producers in our study [13
Understanding the way toxins such as pyocins kill P. aeruginosa
strains, and how this is modulated by genetic relatedness, may also provide insight into the development of novel therapeutic interventions, for example by evolving pyocins specifically against strains that predominate in infections. They can thus be considered designer drugs [7
] and will be a much more direct agent to treatment of the disease than the current practice of using broad spectrum antibiotics against which wide spread resistance exists [46
]. Interestingly, pyocins are not new in a clinical setting: it has been shown that pyocins slow down the development of several forms of cancer in mammalian cells [47
]. Also, membrane vesicles produced by P. aeruginosa
have been suggested as novel therapeutic agents [23
]. However they may be even more effective when used in a targeted way against known infections. The similarity between strains can then be used as a predictor of the intensity of the antagonistic interaction and thus the effectiveness of the pyocin.