Cooperation and Cheating through a Secreted Aminopeptidase in the Pseudomonas aeruginosa RpoS Response.

Bacterial stress responses are generally considered protective measures taken by individual cells. Enabled by an experimental evolution approach, we describe a contrasting property, collective nutrient acquisition, in the RpoS-dependent stress response of the opportunistic human pathogen P. aeruginosa. Specifically, we identify the secreted P. aeruginosa aminopeptidase (PaAP) as an essential RpoS-controlled function in extracellular proteolysis. As a secreted “public good,” PaAP permits cheating by rpoS mutants that save the metabolic costs of expressing RpoS-controlled genes dispensable under the given growth conditions. Proteolytic enzymes are important virulence factors in P. aeruginosa pathogenesis and constitute a potential target for antimicrobial therapy. More broadly, our work contributes to recent findings in higher organisms that stress affects not only individual fitness and competitiveness but also cooperative behavior.

study by Sandoz et al. and 21 isolates from the more recent study by Wilder et al. (two replicate lines each [20,23]). In vitro evolution experiments in both studies were conducted in the same laboratory under identical culture conditions. We grouped the isolates into three classes based on two QS-dependent phenotypes, skim milk proteolysis and growth on adenosine as the sole C source (Table 1). Adenosine utilization offers an additional, easily scorable phenotype that directly depends on a functional LasR (28). Our classification confirmed previous results (20,23). Class I isolates are protease and adenosine positive, class II isolates are protease positive but adenosine negative, and class III isolates are both protease and adenosine negative. Representatives of these three classes were observed in all four replicate experiments. We had previously interpreted class I isolates as QS-proficient "cooperators" with intact LasR, class II isolates as moderately QS-deficient "hybrids" with partially functional LasR, and class III isolates as fully QS-deficient "cheaters" with nonfunctional LasR (20,22,23). We report below that the interpretation of class I isolates as cooperators requires revision.
Sequencing was performed on an Illumina HiSeq 3000 platform with 150-bp pairedend reads. After low-quality reads were trimmed and sequences were aligned to the parent PAO1 strain as a reference, mutations were called and assessed (Table 1; see also Fig. S1 in the supplemental material). Sequencing revealed psdR mutations in all but five isolates. Upon closer examination, we found that a contig gap in the assembled genome fell in the middle of the psdR gene in four of the five cases. Due to this and the discovery of mutations at this locus in all of the other isolates, we believe that each isolate with a contig gap does indeed contain a mutation in this gene. This interpretation is consistent with our previous work, which characterized widespread loss of PsdR as an important adaptation to the proteolytic growth environment (22). In the class III genomes, we found a mutation in the lasR coding region in five of six isolates. In the remaining isolate, there was an insertion in the Ϫ15 position from the lasR translation start site that presumably disrupts translation by hampering ribosome binding. Class II genomes had mutations in lasR and psdR, in addition to other loci that were not consistent throughout the group. The abundance of lasR mutations in these two classes and in all four replicate experiments also confirms our previous work (20,22,23). Importantly, in the class I genomes we found a mutation in rpoS in 6 of 14 isolates (43%) in three of four replicate experiments. An rpoS mutation was identified in only one class II isolate and one class III isolate. This was the only novel mutation consistent within a group, making it a compelling candidate for further study.
The P. aeruginosa rpoS gene encodes an alternative sigma factor of 334 amino acids in length. The various nonsynonymous rpoS mutations result in an amino acid substitution (rpoS1, rpoS3, and rpoS4), a truncation (rpoS2), a frameshift (rpoS5 and rpoS6), or a lost start codon (rpoS7) ( Table 2). We also mapped each RpoS mutation onto a three-dimensional protein structure (Fig. 1), indicating no particular domain preference or clustering. rpoS1, rpoS2, and rpoS4 harbor mutations in a DNA-binding region that presumably disrupt promoter recognition and transcription initiation (29). rpoS5, rpoS6, and rpoS7 encode either a nontranslated or an aberrant protein. As such, we presume that each mutation results in at least partial loss of function.
To investigate the nature of the fitness costs and benefits conferred by the loss of rpoS, we chose rpoS5 for further analysis. An early frameshift at codon 20 very likely results in a complete loss-of-function protein, thereby simplifying interpretations of mutant phenotypes (Table 2). We cloned the rpoS5 allele and transferred it into the PAO1 WT to analyze the effects of this mutation. For comparison, we also utilized a previously constructed rpoS loss-of-function mutant in which the rpoS gene was disrupted by the insertion of a gentamicin resistance (Gm r ) cassette (11).
RpoS mutant frequency during in vitro evolution. First, we considered the sequence of major mutational events during in vitro evolution. We know from previous work that psdR mutations rise to fixation by day 4 of subculturing and that lasR mutations comprise, on average, about a quarter of the population at day 12 (22). To evaluate how early rpoS mutations emerge, we sequenced the rpoS gene of 20 class I   (23). None of the isolates harbored a mutation, indicating that rpoS mutations are selected gradually from within an established psdR mutant population. To more precisely quantify the frequency of rpoS mutants over time, we devised an assay suitable for large-scale screening of rpoS-deficient phenotypes. To this end, we used Biolog phenotype arrays to identify a specific growth condition that would enable differentiation of the WT strain from an rpoS mutant. Among 384 different C and nitrogen (N) sources, only one produced a strong phenotype (see Data Set S1 in the supplemental material). The WT strain grew on the compound N-acetyl-L-glutamic acid (NAG) as either a C or N source, whereas the rpoS5 mutant did not. The same growth differences manifested in our M9-based minimal medium ( Fig. 2A). The molecular basis of the NAG phenotype is not yet known. One pathway for NAG utilization appears to involve periplasmic conversion to glutamate by a deacetylase (30). However, the corresponding gene has not been identified and its identity is not apparent from our list of RpoS-induced genes (11).
Our screening assay for the identification of evolved rpoS mutant isolates was based on a solid minimal medium with NAG as the sole N source. The defined rpoS5 and rpoS::Gm r mutations conferred no growth on this medium, whereas all other major mutations previously described in our in vitro evolution system did (lasR, psdR, pqsR; Fig. 2B). We initially screened all of the sequenced isolates listed in Table 1 and qualitatively evaluated their growth on solid medium. We found a good correlation between the presence of an rpoS mutation and a growth defect on NAG. Of the eight sequenced isolates with an rpoS mutation, seven showed no to marginal growth  ( Table 2). Only one (isolate TR02 with rpoS1) showed WT-like growth. The resulting mutation-a single amino acid substitution near the C terminus of the proteinapparently has no major impact on RpoS function. To obtain additional quantitative information about the degree of NAG deficiency, we grew all eight evolved mutants in NAG-liquid culture and determined cell densities ( Table 2). In this format, all of the rpoS mutant isolates, including TR02, showed much reduced growth compared with the WT. However, consistent with the plate assay, TR02 grew to a significantly higher density than the other isolates. In contrast, all isolates with a WT rpoS allele showed robust growth on solid NAG medium that was indistinguishable from the level seen with the WT parent (Table 1). We conclude that our plate assay reliably identifies NAG-deficient isolates as rpoS mutants and is suitable for large-scale screening, while recognizing the limitation that a small fraction of partial loss-of-function rpoS mutants may escape detection.
We screened all available isolates from the Wilder et al. study (two replicate lineages, with 92 isolates each from days 4, 8, 12, 16, and 20). rpoS mutants were detected as early as day 12 and increased in frequency on days 16 and 20 (Fig. 2C). With seven to eight generations per 24-h cycle during experimental evolution (22), this increase in frequency amounts to an average growth rate of the rpoS mutant subpopulation that is 7% to 11% higher than that of the remaining population. Combining the rpoS data with the available information on lasR-deficient phenotypes (20, 23), we found that rpoS mutants lagged behind lasR mutants in frequency and that very few isolates were both rpoS and lasR deficient. Thus, mutations in rpoS arose largely independently of those in lasR and cooccurred in only two isolates (TR02 and TR16). In TR02, the lasR mutation appears to have emerged prior to rpoS, because another isolate from the same replicate experiment (TR09) was found to harbor the same lasR allele (A228V) with no rpoS mutation.
RpoS mutants as social cheaters. To elucidate the phenotypic consequences of rpoS deficiency, we first evaluated the growth of individual strains in casein medium. We considered the effect of the rpoS5 allele in the WT and in the psdR1 mutant backgrounds. The psdR1 mutation results in the deletion of six amino acids close to the carboxyl terminus conferring a complete loss of function (22). We compared the growth of the rpoS5 mutant and the rpoS5 psdR1 double mutant to that of the WT, the psdR1 mutant, and the lasR deletion mutant (Fig. 3A). We also included the previously constructed rpoS::Gm r insertion mutant (31 confirming the notion that the evolved rpoS allele confers complete loss of function. Whereas the psdR1 mutant grew faster than the WT, the rpoS5 mutant showed a substantial growth defect similar to that seen with the lasR mutant. Its highest density was reached within 6 h and plateaued between 6 and 24 h. Even the psdR1 rpoS5 mutant followed this same pattern, indicating that a nonfunctional RpoS is detrimental to proteolytic growth in pure culture. Of note, initial low-density growth of the protease-deficient lasR mutant is routinely observed in this medium (20,22,23), presumably because of low-molecular-weight impurities in the commercial casein preparation. We also expressed the growth of all strains as absolute-fitness data (the average growth rate within a 24-h period; Fig. 3B) and assigned statistically significant differences to the patterns observed in Fig. 3A. Even though rpoS mutants showed low absolute fitness, they were nevertheless prevalent in the population during experimental evolution (Fig. 2C), suggesting a high relative fitness in the presence of other rpoS-positive cells. We therefore tested the relative fitness of the rpoS::Gm r mutant and the psdR1 rpoS::Gm r mutant in coculture with either the WT or the psdR1 mutant (Fig. 3C). Relative fitness is expressed as the ratio of average growth rates during the cultivation period. To distinguish strains in  (23) and are consistent with the notation used in Table 1. Frequencies of isolates with no or marginal growth on NAG as N source, on adenosine as C source, or on both media are shown.
coculture, we employed rpoS mutant strains harboring the gentamicin resistance cassette for subsequent antibiotic selection. We were particularly interested in the ability of the mutants to invade a resident population when initially rare. Inoculated at 1% initial frequency, the rpoS::Gm r and psdR1 rpoS::Gm r mutants were enriched in WT and psdR1 mutant cocultures, respectively. Likewise, the lasR and the psdR1 lasR mutants were enriched in their corresponding cocultures, confirming previous results (22). In all cases, the relative fitness of the rare strain is above 1. The relative fitness differences between strain pairs qualitatively match the sequence in which psdR, lasR, and rpoS mutations emerged during in vitro evolution (i.e., w psdR versus WT Ͼ w psdR1 lasR versus psdR Ͼ w psdR1 rpoS versus psdR ). In sum, we found that rpoS mutants have low absolute fitness in isolation but high relative fitness in mixed culture. The combination of low absolute fitness with high relative fitness indicates that rpoS mutants, like lasR mutants, are social cheaters that take advantage of the secreted public goods supplied by others without investing in costly production themselves. This obligate, parasitic lifestyle depends on the presence of cooperators, consistent with the observation that rpoS mutants comprise only a subset of the evolved population (Table 1 and Fig. 2C).
No apparent RpoS-dependent effect on proteolysis. Evidence thus far indicates that rpoS mutants are cheaters that may utilize public goods produced by the WT, most obviously secreted enzymes involved in casein proteolysis. We therefore tested the proteolytic capabilities of all our strains to identify differences (Fig. 4). To accomplish this, we measured the level of proteolysis on skim milk agar when incubated with cell-free culture supernatant. The supernatant was obtained from stationary-phase LB liquid cultures. These growth conditions promote protease expression but do not also require proteolysis for growth (22). All culture supernatants, with the exception of the lasR mutant supernatant, produced a clearance zone indistinguishable from that of the WT. Thus, proteolytic ability as assessed with our plate assay does not explain the behavior of the rpoS mutant.
We note that this assay is similar but not identical to the original skim milk plate assay used to classify evolved phenotypes ( Table 1). The original skim milk assay qualitatively evaluates proteolytic activity through the zone of clearance surrounding a growing P. aeruginosa colony. The agar contains a small amount of LB medium, permitting some growth even of protease-deficient isolates. The supernatant assay shown in Fig. 4 is more quantitative because it strictly separates growth from proteolysis.
Functions responsible for the RpoS phenotype. To obtain additional insights into the mechanism underlying the rpoS cheater phenotype, we interrogated our previous microarray data of RpoS-dependent genes in P. aeruginosa PAO1 (11). Even though the RpoS regulon contains almost 800 genes, the information thus far obtained from our phenotypic characterization allows us to narrow our search. It is reasonable to assume that the cheater phenotype involves an RpoS-controlled public good, perhaps in the form of a secreted protease that escapes detection by the skim milk plate assay. This assay mainly detects the activity of the major casein endoprotease LasB elastase (32)(33)(34), but there are other extracellular proteases that might contribute to growth of P. aeruginosa in casein medium. Among the proteases secreted by P. aeruginosa, two are activated by RpoS (Table 3)  RpoS Response and Cheating ® pepB, and the endoprotease protease IV, encoded by piv (25,35). PaAP seems the more likely candidate. Its high level of induction by RpoS correlates with the strong casein growth phenotype. PaAP preferentially cleaves leucine but also cleaves other amino acids from the ends of peptides (25). This activity would not be required for caseinolytic activity visualized on skim milk plates but might still promote growth on casein by making individual amino acids available from the breakdown products of other proteases.
In order to examine the roles of PaAP and protease IV in proteolytic growth, we constructed pepB and piv mutants. The piv mutant grew as well in casein medium as the WT, ruling out its involvement in the rpoS mutant phenotype (Fig. 5A). The pepB mutant, in contrast, showed impaired growth in casein medium but showed no deficit in casein proteolysis on skim milk plates ( Fig. 5A and B). These phenotypes are identical to those of the rpoS mutant. To confirm that the pepB-encoded enzyme, PaAP, is the only factor contributing to the loss-of-function growth phenotype of the rpoS mutant, we conducted two different complementation assays. First, we added purified leucine aminopeptidase to our casein minimal medium and found that growth was restored in both the pepB and rpoS mutant strains (Fig. 5A). Second, we ectopically expressed pepB from an arabinose-inducible promoter and found that it fully complemented the growth defect of the rpoS and pepB mutants (Fig. 5C). The absolute fitness of all three genetically complemented strains was above that of the noncomplemented WT, presumably because pepB is produced at high levels as the result of immediate induction upon inoculation.
Next, we investigated relative fitness in WT cocultures. The pepB mutant had a relative fitness level significantly below that of the rpoS mutant (Fig. 5D), consistent with the idea that the cost savings accruing from loss of an individual cooperative gene is lower than the savings associated with loss of an entire regulon (36). The relative fitness level of the pepB mutant was even slightly below 1, as the pepB mutant grew less well than the WT in coculture. However, this value is above what would be expected if the pepB mutant did not benefit at all from the WT (0.89 Ϯ 0.04 versus 0.55 Ϯ 0.03, P value Ͻ 0.001 by two-sample t test; see Materials and Methods for details), suggesting that PaAP is indeed secreted and at least partially beneficial to nonproducing cells.
Considering the large size of the RpoS regulon, it is conceivable that some functions, particularly those repressed by RpoS, also provide a direct fitness benefit to the mutant during proteolytic growth. Interrogation of our previous transcriptome analysis identifies one candidate function, amino acid and dipeptide uptake (encoded by the mdp/dpp gene cluster) (11). The level of expression of these genes is about 2-fold-to-3-fold higher in the rpoS mutant than in the WT, providing a potential additional growth advantage to the cheater in mixed culture. Intriguingly, the same function is derepressed in psdR mutants, with about 50-fold-higher expression of mdpA and dppA in the mutants than in the WT (22). Given that psdR mutations arise before rpoS mutations during experimental evolution, it is questionable, however, whether the modest regulation by RpoS would have any additional effect. We tested these predictions with a dipeptide growth assay that measures the ability of strains to utilize the dipeptide Gly-Glu as the sole C source (22,24). As reported previously (22,24), the psdR mutant had a significant growth advantage compared with the WT (Fig. 6). The rpoS mutant also had a slight growth advantage compared with the WT, but the rpoS psdR mutant had no advantage compared with the psdR mutant, confirming our predictions. Because all evolved rpoS mutants harbor psdR mutations, the small effect of RpoS on mdp/dpp expression does not appear to contribute to their enrichment. A model that reconciles relative fitness differences between rpoS and pepB mutants. The RpoS control of PaAP explains why rpoS mutants grow well in WT coculture but not in isolation. What still needs some explaining is the difference in relative fitness between rpoS and pepB mutants. As indicated above, there likely is a substantial metabolic burden associated with expressing RpoS-dependent genes. Most genes are induced by RpoS, and the average level of induction is much higher than the average level of repression (11). This metabolic burden is relieved by loss of RpoS, but not by loss of PaAP alone, translating to lower fitness costs and a higher growth rate. However, the question remains why the relative fitness of the pepB mutant is even lower than that of the WT. A plausible explanation is that PaAP-producing cells have preferential access to the secreted enzyme, even in a well-mixed environment. For example, if PaAP were not fully secreted and remained partially associated with the cell RpoS Response and Cheating ® envelope, then digestion products might be directly captured by the producer before diffusing away. This property has indeed been reported for some microbial public goods (37,38). We conceptualized the interplay of the two factors, fitness costs on the one hand and capture efficiency on the other hand, in a simple mathematical model. We modeled the relative fitness of the cheater as the ratio of the cheater growth rate to the cooperator growth rate, the same metric as that used in our coculture assays. We expressed growth rates as a saturation function that is dependent on the concentration of the digestion products of PaAP as the relevant public good. Costs for the production of PaAP or for other pleiotropically linked traits reduce the growth rate of the cooperator, and capture efficiency reduces the growth rate of the cheater. As expected, the model shows that the cooperator does better when fitness costs are low and capture efficiency is high (Fig. 7). Conversely, the cheater does better when costs are high and capture efficiency is low. The difference in cost savings can explain a relative fitness level above 1 for the rpoS mutant and a relative fitness level below 1 for the pepB mutant.

DISCUSSION
In this study, we identified a social phenotype in the RpoS response of the opportunistic pathogen P. aeruginosa. RpoS was originally described for E. coli as a master regulator of a stress response to nutrient exhaustion in stationary phase (39). In P. aeruginosa, RpoS also controls a large number of genes in stationary phase, although its role in stress protection is less clear (11)(12)(13). We used a variety of methods to investigate rpoS mutations and the nature of their fitness benefit during proteolytic growth of P. aeruginosa, including a novel screening assay based on NAG as a strictly RpoS-dependent nutrient. In our in vitro evolution experiments, we detected exclusively nonsynonymous loss-of-function mutations in rpoS that continuously increased in frequency, indicating a positive selection process (Table 1 and Fig. 2C). We provided evidence that the selected phenotype is social and exhibits the hallmarks of cheating behavior: RpoS mutants had low absolute fitness in pure culture but high relative fitness in coculture with cooperating cells (Fig. 3). Cheating has been observed in several microbial systems, including QS, nutrient acquisition, and development (19,20,(40)(41)(42). It is considered a central problem in the evolution and maintenance of cooperative behavior, although mechanisms of cheater control have been described (21,43). The social phenotype of rpoS mutants can be explained by two properties: (i) the regulatory effect of RpoS on a single gene, pepB, that encodes the aminopeptidase PaAP (Fig. 5) and (ii) the large fitness cost associated with expressing genes in the global RpoS regulon ( Fig. 3 and 7). Regarding the first point, we found that PaAP is necessary for growth on protein as nutrient source and that its presence is sufficient to complement the growth defect of an rpoS mutant (Fig. 5). PaAP has been characterized as a Zn-dependent metallopeptidase that preferentially cleaves leucine residues from the N termini of peptides (25). In vitro, it has at least 10-fold-lower activity toward other residues (25). However, given the strong phenotype, we surmise that hydrolysis of residues other than leucine likely contributes to proteolytic growth in vivo. PaAP is a highly processed extracellular protease secreted by the Esx type II secretion pathway (44). It is found in soluble form in the extracellular milieu and as an abundant component of outer membrane vesicles that can affect the formation of biofilms (27,45). Consistent with its extracellular localization, we found that PaAP functions as a public good that is shared between WT and rpoS mutant cells.
Because the fitness of a targeted pepB mutant is lower than that of the WT in coculture, we propose that PaAP-producing cells have preferential access to the secreted enzyme or its products, even in a well-mixed environment. It is possible that PaAP is not fully secreted and remains partially associated with the cell envelope, such that digestion products are preferentially available to the producer. This property has been reported for the enzyme invertase in yeast (37) and has been suggested for the siderophore enterochelin in E. coli (38). The presumed localization of PaAP on the surface of P. aeruginosa outer membrane vesicles indeed renders the idea of a transient association with the cell envelope during the process of vesicle biogenesis plausible (46,47).
The second point explains the large relative fitness benefit associated with the rpoS mutant. While PaAP represents an individual cooperative trait, RpoS is a pleiotropic regulator that controls many other traits (11). There is a substantial metabolic burden associated with expression of hundreds of RpoS-dependent genes, and this burden is relieved by loss of RpoS but not by loss of PaAP alone. The same reasoning explains why mutations in another pleiotropic regulator, LasR, confer a large fitness benefit, whereas mutations in an individual LasR-controlled public good, elastase, do not (36). Consistent with this argument, in vitro evolution did not select for mutations in pepB or other protease genes. In turn, there is no pleiotropic cost associated with expression of other genes in the RpoS regulon. Under the specific conditions of cooperative growth, all but one of the genes controlled by RpoS are indeed dispensable. Instead, we provide evidence for a potentially beneficial function, peptide uptake and metabolism, that is enhanced in the absence of RpoS and that could add to the fitness advantage acquired from cheating. RpoS mutations are not reported for P. aeruginosa isolates from infections and natural environments, suggesting that the associated growth conditions are more varied, favoring a functional RpoS response in these contexts.
When we consider rpoS in the context of the other major mutations in our evolved isolates, the genotype-phenotype relationship is not always clear (Table 1). For example, we do not know why isolate TR02 with psdR, lasR, and rpoS mutations represents a class I phenotype whereas isolate TR09 with identical psdR and lasR mutations represents a class III phenotype. There may be as-yet-unknown interactions between the mutations, and there may be additional mutations that we did not identify, due to either sequencing gaps or limited sample size. What we can say is that the lasR A228V allele in isolates TR02 and TR09 retains partial function (22), explaining why they are not categorized as fully QS-deficient class II isolates. Of note, we did not identify mutations in the orphan QS regulator PqsR in any of our isolates. Our previous study revealed fluctuating frequencies of pqsR mutations that were as high as 60% on day 12 and as low as 6% on day 20 (23), suggesting that we may have missed pqsR mutations from day 20 simply due to limited sampling depth. Limited sampling likely also explains why we detected no rpoS mutations in one of the four replicates. The relative abundances of psdR, lasR, and rpoS mutations during in vitro evolution generally match their relative fitness differences in defined coculture (Fig. 2C and 3C) (22). Of course, there are additional factors that can influence mutant dynamics during in vitro evolution, such as differences in the abundances of preexisting mutations in the ancestral population, different propensities for loss-of-function mutations due to differing protein sizes and functions, and fitness contributions from other evolved mutations. We also found that the two major cheater lineages, rpoS and lasR, evolved largely independently of one another (Fig. 2C). This finding is consistent with the large overlap of the RpoS and LasR regulons (11), such that a second mutation would provide only a small additional reduction in fitness costs.
Our work highlights the power of experimental evolution, which complements targeted genetic analysis in the identification of functions in novel contexts. The literature on stress responses tends to emphasize nonsocial functions that protect the cell from harm. However, regulation of social functions makes sense if stress responses are interpreted as a response to ecological competition (5). Such a response is contingent on perception of nutrient limitation as interference competition from other microbes, and an adequate response to competition would be a counterattack through expression of cooperative secretions (e.g., toxins or antibiotics). While PaAP shares this cooperative property, there is no evidence that it also provides a competitive advantage by, for example, harming other cells or making digestion products available solely to P. aeruginosa. Beyond the microbial realm, recent research has shown complex, opposing effects of stress on cooperative behavior. Humans, for example, are inclined to cooperate under certain stressful situations, such as those requiring the immediate rescue of a person in danger (48). Female meerkats, in contrast, are less willing to cooperate with group members under conditions of stress imposed by dominant females (49).
Another important question relates to the general stability of cooperative populations during experimental evolution: can our comprehensive sequence analysis provide insights? We had previously observed that only one of five replicate populations collapsed after 20 days (approximately 200 generations) of experimental evolution, despite the rise of lasR cheater mutants (20,22,23). In this study, we identified yet another cheater type among the members of a phenotypic class previously considered to be "cooperators." Both RpoS and LasR cheaters are deficient in proteolysis, because pepB expression requires both LasR and RpoS. Thus, the associated cheater phenotypes are not complementary in the sense that the two cheat with respect to different traits (50). Nevertheless, it is possible that one cheater might be able to restrain the other. RpoS mutants overproduce the redox-active metabolite pyocyanin, and pyocyanin preferentially inhibits lasR mutants (51). In addition, as mentioned above, pervasive mutations in psdR provide a demonstrated nonsocial fitness advantage that defers population collapse. Some of the other novel mutations that we identified primarily in class II ("hybrid") isolates may work similarly. Class II isolates also harbor partial loss-of-function mutations in lasR that may "streamline" the QS regulon with respect to essential functions, thereby optimizing metabolic costs (22). Finally, the long-term trajectory of the in vitro-evolved populations beyond the 20-day growth period is not clear, and it is possible that they would all collapse eventually.
As mentioned, the expression of pepB is controlled not only by RpoS but also by the QS regulator LasR (11). Thus, pepB regulation integrates two factors, the density of self-cells and stress. The signaling mechanisms in QS are well understood, but the cues that trigger the RpoS response in P. aeruginosa are not. Nutrient starvation is likely an important trigger, perceived through the stringent response that directly senses the lack of amino acids (52,53). It is intriguing to speculate that the Lrp protein in P. aeruginosa, a homolog of the leucine-responsive protein in E. coli, is involved in sensing the products of leucine aminopeptidase and promoting RpoS-dependent transcription (54). In any case, it seems clear that the presence of a high density of self-cells ensures that peptidase secretion is effective, while nutrient stress provides information about the actual need for this enzyme. It is useful here to consider PaAP in the context of other proteases secreted by P. aeruginosa. Endoproteases and aminopeptidases likely work together during proteolytic growth (Fig. 8). Endoproteases such as LasB elastase break down peptide strands into smaller fragments (55,56). Elastase is activated by QS but not by RpoS (11,31). Thus, high cell density alone seems to be sufficient to anticipate a future need for this enzyme. PaAP is triggered by both cell density and starvation. PaAP frees individual amino acids from the peptide fragments that are taken up by the cell. Consistent with this scenario, proteolytic activation of PaAP seems to require elastase (25).
Taken together, we have provided evidence for a role of global stress responses beyond cellular protective functions. For the RpoS response in P. aeruginosa, one such role is the utilization of alternative nutrient sources available through cooperative secretions vulnerable to exploitation by cheaters.

MATERIALS AND METHODS
Strains and culture conditions. For all experiments, we used the P. aeruginosa PAO1 wild-type (WT) strain or mutant derivatives. For a complete list of strains, see Table 4. Unless otherwise specified, strains were grown in Lennox lysogeny broth (LB; also referred to as Luria-Bertani medium) buffered with 50 mM 3-(N-morpholino)-propanesulfonic acid (LB-MOPS; pH 7.0). All cultures were incubated at 37°C with activates expression of the extracellular endoprotease elastase (brown; encoded by lasB) that breaks casein peptides into smaller fragments (yellow chains; each circle represents one amino acid). Both RpoS and LasR activate expression of the extracellular aminopeptidase PaAP (orange; encoded by pepB) that cleaves individual amino acids from the ends of peptides. A potential association of PaAP with the cell envelope is indicated. Loss of LasR or RpoS regulation prevents expression of the respective extracellular enzymes, producing a cheater phenotype that benefits from the enzymes produced by cooperator cells. PsdR represses genes (dppA3 and mdpA) involved in the transport and intracellular processing of amino acids and dipeptides (green). Loss of PsdR leads to increased uptake and processing, representing a nonsocial adaptation. RpoS has a slight inhibitory effect on uptake and processing genes (dashed line), potentially further contributing to a nonsocial fitness increase in an rpoS mutant.
For whole-genome sequencing, isolates were chosen from in vitro evolution experiments according to QS-dependent phenotypes (20,22,23). The PAO rpoS5 and PAO psdR1 rpoS5 mutants were constructed using a two-step allelic exchange as described previously (57). The rpoS region was PCR amplified with the desired evolved frameshift mutation using primers 5=-NNNNNNAAGCTTAGGTCGTCG ATCGCAACGGTTC-3= and 5=-NNNNNNTCTAGACGTCACTCGACAGGCCATTCTTCTC-3= with flanking HindIII and XbaI restriction sites (underlined), respectively. The PCR fragment was cloned into XbaI-digested and HindIII-digested pEX18Tc for use in allelic exchange (57). The resulting construct was introduced into either the PAO1 WT or the previously constructed PAO psdR1 mutant by conjugation (57). The markerless ΔpepB mutant was constructed as follows. An in-frame deletion of residues 148 to 376 was generated by splicing by overlap extension PCR (SOE-PCR), using four primers. Fragment one was amplified with 5=-NNNNNNGAATTCGGCGGGAAGAATTTGGTGATG-3= (EcoRI site underlined) and 5=-CGACAGGTAGGTGA AATCCTTC-3=. Fragment two was amplified with 5=-GAAGGATTTCACCTACCTGTCGGGCAACTTCATCTATG ACGGC-3= (primer 2 complement underlined) and 5=-NNNNNNAAGCTTCTTCAACCTGTCGCCCAATC-3= (HindIII site underlined). The ΔpepB construct was then cloned into HindIII-digested and EcoRI-digested pEX18Gm for use in allelic exchange as described above. The PAO psdR1 rpoS::Gm r and PAO piv::Tc r mutants were constructed by transformation with chromosomal DNA as previously described (58). Briefly, 1.25 ml of recipient cells from an LB overnight culture was pelleted, washed, and resuspended in 100 l of 300 mM sucrose solution. Cells were electroporated after addition of 500 ng of purified chromosomal DNA from either strain PAO rpoS::Gm r or strain DH0001 (PAO piv::Tc r ) (14,33). The piv mutation was transferred from one PAO1 strain to another to ensure isogenic backgrounds. Transformants were selected on LB plates containing the appropriate antibiotic and were verified by PCR.
DNA preparation and sequencing. Chromosomal DNA was extracted from 30 isolates evolved in vitro and from the PAO1 parent strain using a Gentra Puregene yeast/bacteria kit from Qiagen. DNA was suspended in 100 l double-distilled water (ddH 2 O). Whole-genome sequencing was done using an Illumina HiSeq 3000 platform in the Center for Genome Research and Bioinformatics (CGRB) at Oregon State University, with 150-bp paired-end reads. Between 8,665,795 and 15,100,212 reads were obtained per sample (median, 12,497,204). Low-quality reads and primer sequences were trimmed using Trimmomatic (version 0.36) (59). Adapter sequences and 10 bases in leading or trailing positions were cut off, and reads that had a Phred score below 20 in a 4-base sliding window or were shorter than 40 bases were discarded (59). Reads were aligned, and genomes were assembled and manipulated with Bowtie2 (version 2.3.2) and SAMtools (version 1.3) (60, 61). Our PAO1 parent strain reads were aligned to  (63). Assembled genomes and mutations were visualized using Geneious (64). Targeted Sanger sequencing of rpoS was also conducted on DNA from 20 isolates from day 12 at the CGRB using the following primers flanking the gene: 5=-GCTTGAGTCGAACTCATGCAAG-3= and 5=-CGGCATTTATCTACTTAGGCTCA-3=. Absolute and relative fitness assays. To assess the growth and absolute fitness of individual strains, 4 ml of an LB-MOPS preculture was inoculated with a single colony from a freshly streaked plate and incubated for 18 h. A 4-ml volume of M9 minimal medium containing 1% (wt/vol) sodium caseinate as the sole C source was inoculated with a preculture aliquot to an optical density at 600 nm (OD 600 ) of 0.05 and incubated for 24 h. CFU counts per milliliter were determined by serial dilution and spot plating. For enzyme complementation, M9-casein cultures were supplemented with 0.05 U/ml of leucine aminopeptidase (from porcine kidney; Sigma-Aldrich CAS no. 9054-63-1). For genetic complementation, plasmid pSP401 harboring pepB under the control of the arabinose-inducible P BAD promoter (33) was introduced into the mutant and WT strains. Controls carried parent plasmid pJN105 (65). M9-casein cultures were supplemented with 0.2% (wt/vol) L-arabinose at the beginning of growth. Absolute fitness was calculated as average growth rate or Malthusian parameter (m), with m ϭ ln(N 1 /N 0 )/t, where N 1 is the final CFU count per milliliter after 24 h, N 0 is the initial CFU count per milliliter at time zero, and t is the culturing time in hours (66).
For cocultures, strain mixtures were formulated from individual LB-MOPS precultures according to OD 600 readings. To distinguish the strains in coculture, the rare strain carried a chromosomal antibiotic resistance marker (Gm r , Tc r , or Tp r ) ( Table 4). Tc r and Tp r markers each reside on a mini-Tn7 insertion at a neutral chromosomal site (67). They have no measurable effect on growth rate (23). A 600-l volume of M9-casein minimal medium was inoculated with the strain mixtures in a 96 deep-well block at a starting OD 600 of 0.02. Cocultures were incubated for 24 h. At 0 and 24 h, CFU counts per milliliter and strain ratios were determined by serial dilution and spot plating on LB plates with and without the respective antibiotic. Relative fitness values were calculated as the ratio (w) of the Malthusian parameters of the two competing strains (22,23) as follows: w cheater ϭ m cheater /m cooperator .
A hypothetical relative fitness value was calculated for the pepB mutant in WT coculture at 1% initial mutant frequency, under the assumption that there is no social interaction between the two strains. The values corresponding to the initial and final coculture densities (CFU counts per milliliter at 0 and 24 h) necessary for this calculation were determined as follows. The initial coculture densities of both strains were taken directly from the experiment in Fig. 5D. The final coculture densities were inferred from the respective single culture densities of the experiment in Fig. 5A. Specifically, the final coculture density of a pepB mutant that did not interact with the WT would be 1% of its final density in single culture, representing the low growth yield attainable with a proteolysis defect. The final coculture density of a WT strain that did not interact with the pepB mutant would essentially be as high as that in single culture, because the pepB mutant frequency remained negligible.
Proteolysis and adenosine growth assays. To confirm previously recorded QS-dependent phenotypes of each in vitro-evolved P. aeruginosa isolate, we conducted qualitative skim milk proteolysis and adenosine growth assays (20). A single freshly streaked colony from each isolate was patched onto skim milk plates (4% skim milk-quarter-strength LB) and adenosine agar plates (M9 minimal medium with 0.1% adenosine as the sole C source) and incubated at 37°C. After 18 h, skim milk proteolysis was evaluated as the zone of clearance surrounding each colony. After 48 h, growth was evaluated on adenosine plates. For both assays, the respective phenotypes were compared to those of a positive WT control and a negative lasR mutant control.
For a more quantitative assessment of casein proteolysis, we measured the proteolytic activity of culture supernatant on skim milk plates. The QS-controlled enzyme LasB elastase is the major casein protease in this assay (32,33). A liquid culture with 4 ml of LB-MOPS was inoculated with a single freshly streaked colony and grown for 24 h at 37°C. Cultures were centrifuged, and supernatant liquid was filtered (using a 0.2-m-pore-size filter). Wells cut out of skim milk agar plates with the back side of a Pasteur pipette were filled with 200 l of each culture supernatant. Agar plates were incubated for 24 h at 37°C. Each circular zone of proteolysis was measured from the well edge to the end of the zone.
Phenotype arrays and rpoS mutant screen. For Biolog screening, 96-well microtiter plates PM1, PM2A, and PM3B were used (Biolog Inc., Hayward, CA). PM1 and PM2A contained C sources, and PM2B contained N sources. The manufacturer's protocol for E. coli and other Gram-negative bacteria was followed. For inoculation, a suspension was made in inoculating fluid IF-0 from fresh LB plate cultures grown overnight. For PM3B, 20 mM sodium pyruvate was added as a C source. The starting OD 600 was 0.06 (85% transmittance). Plates were incubated for 24 h without shaking, and OD 600 was measured with a multifunction plate reader (Tecan M200).
To determine the NAG growth phenotype of defined and evolved strains, M9 minimal salts medium was formulated with NAG (0.5% [wt/vol]) either as the sole C or N source. When NAG was the C source, 0.1% (wt/vol) NH 4 Cl was used as the N source. When NAG was the N source, 0.5% (wt/vol) sodium pyruvate was used as the C source. For liquid cultures, the strains were inoculated from LB stationaryphase cultures to an OD 600 of 0.05. Cultures were grown for 18 h, and then OD 600 was measured. For plate cultures, strains were either patched individually with a toothpick or plated with a 96-well replicator from freshly grown LB source plates. Growth phenotypes were scored visually after 24 h of incubation.
Dipeptide growth assay. To evaluate the ability of each strain to grow with a dipeptide as the sole C source, growth assays were performed as described previously (22,24). Strains were precultured in M9 minimal medium with 0.5% (wt/vol) glucose for 18 h, washed in 1ϫ M9 salts, and inoculated at an OD 600 of 0.02 in 4 ml of M9 minimal medium containing 10 mM Gly-Glu dipeptide as the sole C source. Cultures were incubated at 37°C for up to 5 days, with OD 600 measurements taken every day.
Mathematical model. We modeled the relationship between cooperator and cheater strains by focusing on their growth rates, which directly correlate with the relative fitness values measured experimentally. A full mechanistic model based on ordinary differential equations would have many degrees of freedom due to unknown enzymatic and growth parameters. Assuming Monod kinetics, a generic growth rate constant is defined as ϭ max * P/(K s ϩ P), where max is the maximum growth rate, K s is the half-saturation constant, and P is the concentration of the growth substrate, which in our case is the product of the PaAP enzyme. We introduce the product capturing efficiency e and the relative cost c of rpoS-and pepB-dependent gene expression, respectively, as fractions between 0 and 1. Direct capture by the cooperator reduces the level of P available to the cheater by the factor (1 Ϫ e), and specific gene expression costs incurred by the cooperator reduce its growth rate by the factor (1 Ϫ c). The growth rates for cooperator and cheater then correspond to the following equations: μ cooperator ϭ ͑ 1 -c ͒ * μ max * P ⁄ ͑ K s ϩ P ͒ μ cheater ϭ μ max * ͑ 1 Ϫ e ͒ * P ⁄ ͓Ks ϩ ͑ 1 Ϫ e ͒ * P͔ For simplicity, we assume equal max and K s values for the two strains. The growth rate ratio of cheater to cooperator, which is equivalent to the ratio of Malthusian growth parameters or the relative cheater fitness parameter w, is then calculated as follows: We determined w for all combinations of c and e at a P/K s ratio of 1 and visualized the data as a heat map using Morpheus (https://software.broadinstitute.org/morpheus). Of note, costs incurred by the cooperator are equivalent to costs saved by the cheater. While the former reduces the cooperator growth rate by the factor (1 Ϫ c), the latter increases the cheater growth rate by the factor [1/(1 Ϫ c)]. Both result in the same w.
Statistical analysis. Statistical analysis was performed using GraphPad Prism version 7 (GraphPad Software, La Jolla, CA). Ordinary one-way analysis of variance (ANOVA) was used for Table 2, Fig. 3B and C, 4B, and 5A and B. Ordinary two-way ANOVA was used for Fig. 2A and 5C. Two-way repeated-measures ANOVA was used for Fig. 6. For all ANOVA variants, either Tukey's or Sidak's multiple-comparison test was used for pairwise comparisons (generally indicated with horizontal brackets in the figures). In some figures, a single horizontal line is used to summarize results of multiple pairwise comparisons with identical statistical results. A two-sample, unpaired Student's t test with equal variance was used for Fig. 5D, and a one-sample Student's t test was used for Fig. 3C and 5D. In all cases, the significance threshold was ␣ ϭ 0.05.

SUPPLEMENTAL MATERIAL
Supplemental material is available online only. FIG S1, PDF file, 0.03 MB. DATA SET S1, XLSX file, 0.02 MB.