The ultimate guide to bacterial swarming: An experimental model to study the evolution of cooperative behavior

SUMMARY

Cooperation fascinates biologists since Darwin. How did cooperative behaviors evolve when they carry a fitness cost to the cooperator? Bacteria have cooperative behaviors that make phenomenal models to take on this age-old problem from both proximate (molecular) and ultimate (evolutionary) angles. We delve into Pseudomonas aeruginosa swarming, where billions of bacteria move cooperatively across distances of centimeters in a matter of a few hours. Experiments with swarming unveiled a strategy called metabolic prudence that stabilizes cooperation, showed the importance of spatial structure and revealed a regulatory network that integrates environmental stimuli and direct cooperative behavior similar to a machine learning algorithm. The study of swarming elucidates more than proximate mechanisms: it exposes ultimate mechanisms valid to all scales, from cells in cancerous tumors to animals in large societies.

INTRODUCTION

Bacteria are not solitary organisms. Each bacterium lives surrounded by thousands, millions, billions of others in crowded communities. The gut of an adult person contains more than 100 trillion bacteria of as many as 50 phyla (Ericsson and Franklin 2015). Bacteria packed in these communities compete and cooperate in webs of ecological dealings (Coyte et al. 2015) that resemble larger ecosystems such as jungles (Hibbing et al. 2010). Bacteria in less diverse communities, such as the symbiotic luminescent bacteria that colonize the light organ of bobtail squids (Lupp and Ruby 2005), depend on social interactions between individuals of the same species for their survival. They communicate with each other using molecular signals (Ng and Bassler 2009), stick to each other to form biofilms (Parsek and Greenberg 2005) and move in packs called swarms (Kearns 2010).

The study of bacterial social behaviors, Sociomicrobiology (Greenberg 2010), was mostly at the fringe until late 1990’s when breakthroughs like the role of quorum sensing in biofilm formation (Davies et al. 1998) brought it into mainstream microbiology. Its initial focus was on function: what molecular mechanisms—genes, signaling molecules or regulatory pathways—underlie these remarkable behaviors? The focus changed to evolutionary causation in the late 2000’s: how did natural selection favor bacterial social behaviors, especially those costly to an individual bacterium that benefit a whole bacterial community? Understanding cooperative behavior is one of science’s major questions (Pennisi 2005). Work with eukaryotic microbes such as the amoeba Dictyostelium discoideum and the yeast Saccharomyces cerevisiae was already showing results (Strassmann et al. 2000; Greig and Travisano 2004) and bacteriologists joined social evolutionary theorists to search for new answers (Jiricny et al. 2014; Nadell et al. 2013; Diggle et al. 2007; Özkaya et al. 2018; Velicer et al. 2000; Jessup et al. 2004; Kümmerli et al. 2009; Rainey and Rainey 2003).

Sociomicrobiology thrives today, with its own conferences including a ASM specialized meeting called “mechanisms of interbacterial cooperation and competition” (Lories et al. 2017). Its success is special because molecular microbiology addresses functional mechanism and social evolution addresses evolutionary mechanism, and the two can differ profoundly even when they address the same behavior. Mayr explained the difference in his essay Cause and Effect in Biology (Mayr 1961): functional studies seek physiological, molecular or environmental factors underlying a behavior, i.e. proximate mechanisms; evolutionary studies ask how the behavior evolved, i.e. ultimate mechanism. Sociomicrobiology unites the two camps to seek both proximate and ultimate mechanisms underlying bacterial social behavior.

Swarming motility is a cooperative bacterial behavior. Other reviews have already discussed proximate mechanisms (Kearns 2010); here we use social evolutionary theory to discuss ultimate mechanisms (Box 1).

BOX 1: SEMANTICS OF SOCIAL EVOLUTION AND THE PROBLEM WITH COOPERATION.

Here we define the social evolution terminology used in this review to avoid misinterpretation. Semantics can be contentious; terms like altruism and selfishness have strong connotations for human behavior, and each of us may have a subjective definition. We start with a formal definition of evolution: Evolution is the change of gene frequency within a gene pool. In other words, anytime genes change in frequency within a population there is evolution. Many assays that microbiologists use in the lab—antibiotic selection used as an intermediate steps to engineer bacteria for example—are experiments with evolution, even if we don’t realize it (van Ditmarsch and Xavier 2014). Evolution acts mainly through four processes: Natural selection, where genotypes that encode heritable traits which improve survival and reproduction become more frequent within a population; Mutation, a process that introduces genetic variation often through DNA copy errors during replication; Migration, which happens because populations are not hermetically closed, and migrating individuals carry genetic material in and out of the population; Genetic drift, a random process by which some individuals die or reproduce regardless of their genotype, is especially important in small populations or during bottlenecks (Chen et al. 2017).

Fitness is the change in gene (or allele) frequency within the population that is due to natural selection. This is an important: many microbiologists use the term fitness when they mean physiological growth rate. Growth rate is important for fitness, but is not its only component.

A social behavior is any adaptive behavior of an organism that affects the fitness of one or more other organisms of the same species (Robinson et al. 2008). This definition is strict, and not everyone uses it; some researchers include interspecies behaviors within the definition, which we won’t do here. There are four types of social behaviors: mutualism, selfishness, altruism and spite (Foster 2011). Each of these is carried out by one individual, the actor, and affects the fitness of at least one other individual, the recipient. Behavior types differ in how they impact the fitness of the actor (positively or negatively) and how they impact the fitness of the recipient (positively or negatively) producing a 2X2 matrix (Fig. 2).

The term cooperative behavior is often used as a synonym for altruism, the social behavior that costs the actor and benefits the recipient. Explaining this type of behavior is a challenge. If evolution is survival of the fittest this selfish dog-eat-dog world should ruthlessly expunge altruists and erase their genes from the genetic pool. Eusocial species of insects, such as bees and ants, prove the contrary: cooperation can indeed evolve and remain stable in nature. These organisms live in cooperative colonies with division of labor where most individuals are altruistically infertile, and only a few, for example the queen bee, reproduce and propagate their genes to the next generations. Hamilton (Hamilton 1964a; Hamilton 1964b) proposed a famous rule to explain when altruistic behavior should evolve: r>C/B. The C in Hamilton’s rule represents the cost to the actor, the B represents the benefit to the recipient and the r, called relatedness, is the probability that the actor and the recipient share a gene for cooperative behavior by common descent. Hamilton’s rule conveys the powerful intuition that an altruistic behavior with a high cost-to-benefit ratio can still be favored by natural selection if it happens mostly between relatives (when r larger than C/B).

Relatedness is often low in bacterial populations where microbes mix with each other constantly (Andersen et al. 2015). If bacteria show behaviors that seem cooperative, those behaviors should have mechanisms to stabilize cooperation even in conditions where relatedness is low. For example, some species of bacteria secrete iron-scavenging molecules called siderophores costly to produce. Siderophore producers increase the chances of benefiting a relative by using specific siderophores that only bacteria with a compatible importer take up, which increases the chances that they are related (Kümmerli et al. 2014). The main message is that bacteria provide models to study the molecular mechanisms that allow cooperation to evolve and remain stable. We learn more about the evolution of cooperation, but what we learn can be directly applied in microbiology. For example, pathogenic bacteria rely on cooperative secretions for their virulence; understanding how cooperative secretions evolved and remain stable can lead to new strategies to fight pathogenic bacteria (e.g. Griffin et al. 2004; Boyle et al. 2013; Xavier 2016; Allen et al. 2014).

Here we focus on swarming in Pseudomonas aeruginosa, a spectacular collective behavior where colonies spread on agar surfaces in beautiful branching patterns (Fig. 1A, Supplementary video 1). P. aeruginosa is a species well studied because it causes opportunistic hospital infections (Klevens et al. 2007) and chronic infections of cystic fibrosis patients (Smith et al. 2006). Social evolution studies in P. aeruginosa swarming benefit from plenty of molecular knowledge including tools for gene engineering, databases of sequenced genomes, transposon libraries and mathematical models (Du et al. 2011). Those tools gave us proximate mechanisms such as the importance of flagella assembly and rotation, of the rhamnolipid surfactant that lubricates the surface and of its quorum sensing regulation. Experiments that explore ultimate mechanism aim to understand more completely this remarkable bacterial behavior, and hopefully we can learn more about how cooperative behavior evolved.

Figure 1. P. aeruginosa swarms on soft agar using its flagella and rhamnolipid secretion; this cooperative behavior benefits the colony which can harvest more nutrients and grow to larger numbers.

(A) The swarm started to expand around 5 h and reached the edge of a petri dish before 24 h (Supplementary video 1). (B) Swarming on soft agar enabled P. aeruginosa to use available nutrients and increase to a population size of more than 15 billion cells. The strain inoculated on the same nutrient medium but in hard agar cannot swarm, and yields less than 3 billion cells. Deletion either flgK or rhlA gene prevented swarming even on soft agar, resulting in final population sizes of less than 5 billion cells. (C) Two P. aeruginosa swarms expel each other through rhamnolipid secretion (Supplementary video 2).

SWARMING IN P. AERUGINOSA: COOPERATIVE BACTERIAL BEHAVIOR

Swarming is a bacterial social behavior where billions of bacteria migrate together over a surface. Many species of bacteria swarm, although the definition and the mechanics of the swarming behavior can differ (Kearns 2010). In P. aeruginosa, swarming can be studied in the lab using Petri dishes with semi-solid agar (Box 2). The percentage of agar needs to be low (0.5 % w/v, the so-called soft-agar) compared to the usual 1.0% or 1.5% used in most microbiology assays (the so-called hard-agar). The nutrient composition is also important: swarming is sensitive to nutrients such as phosphate (Rashid and Kornberg 2000), the source of carbon used (Shrout et al. 2006), iron concentration (Boyle et al. 2015) and the ratio of nutrients such as carbon-to-nitrogen (Xavier et al. 2011). The assay starts by spotting bacteria at the center of a soft-agar Petri dish, which is then followed by incubation at 37°C. After ~5 h the colony grows and spreads over the agar towards the edge of the dish. During this time the spreading can be photographed repeatedly to produce time-lapse video (Supplementary video 1). The colony stops spreading after ~20 h; its end shape branches out in ~0.5 cm wide tendrils, each made of billions of bacterial cells (Fig. 1A).

BOX 2: EXPERIMENTAL SOCIAL EVOLUTION WITH P. AERUGINOSA SWARMING.

a). A simple recipe for swarming

Prepare stocks of 250 mL bottles with 100 mL of agar at 1.25%. Sterilize in autoclave, let cool down and solidify and keep for stock. Every time you need to make swarming plates, melt the agar in a microwave and add the following ingredients: 93.75 mL of sterilized Millipore water, 50 mL of the 5x stock buffer solution, 250 μL of the 1M magnesium sulfate solution, 25 μL of the 1M CaCl₂ solution, 6.25 mL of a casamino acid stock solution at 200 g/L. Make sure agar is cool enough before adding the casamino acids by feeling temperature with your hand. The 5x stock solution is made by mixing the following into 1L Millipore water and then sterilizing: 64 g Na₂HPO₄•7H₂O (or 33.9 g if anhydrous), 15 g KH₂PO₄ and 2.5 g NaCl. Pour exactly 20 mL of molten agar medium onto each 9 cm wide Petri dish. It should be enough for 12 plates. Let cool down for 30 min and then flip upside down, then let cool down for ~15 h (overnight). To inoculate the swarming plate take an overnight culture of P. aeruginosa grown in LB liquid medium, wash twice by spinning down the cells and suspending in 1X PBS and then use a P10 pipette to spot 2 μL of the culture at the center of the swarming plate. You should let the 2 μL drop hang from the pipette tip and gently touch the agar surface. After inoculation let the droplet dry by leaving the plate open on your bench until the water is absorbed by the agar. Close the plate and incubate at 37°C for 24 h. The plate should be upside down in the incubator. This protocol is robust, but consider other protocols and recommendations for swarming reproducibility (Morales-Soto et al. 2015; Tremblay and Déziel 2008). Take specially care with these points:

• Set the microwave power to 20% when melting the stock bottles of agar, and then heat until completely melted, shaking occasionally to release the heat. Make sure the agar has melted well before adding the other reagents.

• Let the plates cool down for 30 min and then flip the plate over. The plates should not be left upright overnight.

• Do not use plates more than 15 h old. If the plates dry too much that means the agar concentration is higher and the colony cannot swarm.

• The colony cannot swarm if you add too much water either. The final agar concentration should be 0.5%. At 0.3% agar concentration, it is already too wet for swarming and the bacteria will swim through the agar. This is called a swimming assay (Deforet et al. 2014).

• Make sure you use Bacto casamino acids and the Bacto agar both from BD. Other brands tested in our hands have not worked well.

• Use a stir bar when mixing agar solution

b). Using the Petri dish as a fighting arena

There are many great tools to study molecular (proximate) mechanisms of P. aeruginosa swarm, such as targeted genetic engineering (Hmelo et al. 2015), transposon insertion library and many others. Here we provide a few ideas to study ultimate mechanisms. Those ideas still borough tools from classic molecular microbiology which very convenient to monitor temporal and spatial dynamic of a population.

1. Fluorescent labeled cells: Label bacteria with constitutively expressed fluorescent proteins inserted into a phage attachment site without disrupting bacterial gene function. Using two fluorescent proteins whose spectra can be well separated by a detector is key. Microscopy or colony forming unit (CFU) assays could later on be used to quantify the types mixed. Don’t forget to swap the colors for control experiments: fluorescent proteins may have different maturation time and even different impacts on your cells.

2. Inducible promoters: we can modulate the timing and investment in a social behavior by tuning the expression level of genes or operons encoding the behavior.

3. Whole genome sequencing: The outcome of an evolutionary experiment which produced mutations can be unpredictable. Sequencing is now affordable enough to be used routinely. Other omics platforms such as transcriptomics and metabolomics help reveal other proximate mechanisms behind the mutation that enabled its evolution.

Happy Swarming!

How can use the simple swarming assay for experiments on the evolution of cooperative behavior? Swarming is a cooperative behavior: it benefits the population as a whole, which by spreading out can harvest more nutrients from the agar than if it was prevented from swarming. This is easily shown by spotting side-by-side the same amount of bacteria in soft agar (0.5% agar) and hard agar (1.5% agar), both containing exactly the same nutrients. When the two plates are incubated at 37°C for the same exact time (24 h), the colony spotted in hard agar does not swarm and stays restricted to a circle with ~0.5 cm diameter at the center of the Petri dish. This colony, restricted from moving, depletes all the nutrients available locally, and reaches a final population of only ~2 billion cells. The colony spotted in soft agar starts swarming at ~5 h and spreads to a final radius of ~8 cm; this colony reaches a final population of ~17 billion cells, more than 8 times the number of the non-swarming colony (Fig. 1B) (Xavier et al. 2011).

Molecular studies such as genetic screens and transcriptomic studies pinpointed functions required for swarming, including flagellar motility, rhamnolipid synthesis, cell-cell signaling and metabolic regulation (Overhage et al. 2007; Yeung et al. 2009; Tremblay et al. 2007; Heurlier et al. 2004; Caiazza et al. 2005). But since those studies where concerned with proximate mechanisms they only compared the swarming phenotypes of isogenic mutants in monoclonal populations. Most of those studies did not investigate whether those functions could be exploited by cheaters, which could be addressed for example by mixing wild type and mutant in the same colony and using neutral markers such as fluorescent proteins to distinguish the two (Box 2). Mixing experiments, although simple, go beyond the proximal mechanisms and address the ultimate mechanisms of cooperative behavior.

THE IMPORTANCE OF BEING FLAGELLATED

P. aeruginosa has a single flagellum located at one of its poles. The collective flagella forces of billions of individual bacteria power swarming. Mutants without a flagellum cannot swim nor swarm. These loss-of-function mutants occur in any of several genes essential for flagellum assembly and regulatory pathways: genes for the flagellum machinery components such as the filament protein fliC (Köhler et al. 2000) and the hook protein flgK (Overhage et al. 2007), genes for directing flagellum assembly to the pole such as fhlF (Murray and Kazmierczak 2006), and genes that mediate the level of the intracellular signalling molecule cyclic diguanylate (c-di-GMP) such as flgZ and dipA (Mattingly et al. 2018; Baker et al. 2016).

Since swarming cooperation results form the cooperative flagella forces of individual cells, which is a process that spends energy, can it be exploited by flagella-less mutants who avoid the cost of synthesizing a flagellum and save energy by not having to turn its motor? We can address this question by checking whether swarming is prone to free-riding by loss-of-function mutants in flagella genes. The mutant ΔflgK grows faster than its isogenic wild type in well-mixed liquid culture, because swimming is unnecessary there and the mutant saves valuable energy and metabolic resources. However, the ΔflgK loses the swarming competition because it cannot free-ride on the wild type (Fig. 3A). The cooperative forces generated by flagellated cells are not a public good, but rather a private good that only benefits flagellated bacteria that do their share of the work.

Figure 3. Swarming was robust against flagella-less cheaters, but flagella-less mutants could block swarming when they started at large frequencies.

(A) The wild type (in red) swarmed well when mixed with a flagella-less ΔflgK (in green) when the starting ΔflgK frequency did not exceed 1:1; the ΔflgK could not free-ride and was left behind at the nutrient-exhausted colony center. (B) The ΔflgK could compromise swarming when its initial frequency was 5:1 or higher; at 50:1 it blocked the wild type from swarming entirely.

Interestingly, in competitions mixing ΔflgK and wild type at a high initial frequency of 50:1 the ΔflgK created a wall that blocked the wild type from swarming (Boyle et al. 2017). The flagella motility in swarming cooperation is robust to cheating, but those mutants can spitefully prevent the colony from swarming when they outnumber the wild type (Fig. 3B).

Experimental evolution provided more evidence that swarming resists cheating from flagella-less mutants: if flagella-less mutants could cheat they should emerge spontaneously from wild type populations if they were passaged long enough in swarming plates. But instead these evolutionary experiments produced mutants with more flagella, not less; these multi-flagellated mutants called hyperswarmers swarm even better than the wild type (van Ditmarsch et al. 2013).

The evolution of hyperswarmers is reproducible. When P. aeruginosa is passaged repeatedly to a new swarming plate every day for a period between 5 and 10 days (Fig. 4A) the single flagellated wild type (Fig. 4B) evolves into hyperswarmers each time we do the experiments (Fig. 4C). The exact number of daily passages cannot be predicted because mutation happens randomly. But once hyperswarmers do evolve they change the shape of the swarming colony. Wild type swarms have branches and cover only a fraction of the Petri dish; hyperswarmers spread over the entire Petri dish (Supplementary video 3). All hyperswarmers evolved in the lab so far have point mutations in the same gene, fleN, which regulates the number of flagella per cell (Dasgupta et al. 2000). They don unnatural numbers of flagella: between 2 and 5(Fig. 4D). The fleN point mutations they show seem attuned for optimal swarming, unlike a loss-of-function mutation such as a transposon insertion in fleN (Dasgupta et al. 2000) or a ΔfleN deletion (van Ditmarsch et al. 2013) which wipes out swimming and swarming because of too many flagella.

Figure 4. Repeated rounds of growth in soft agar evolved multi-flagellated hyperswarmers.

A) Scheme of experimental evolution which started from a single flagellated ancestor strain (B) which was repeatedly passaged to fresh soft agar plate everyday for ~9 days (C) and evolved into a robust hyperswarming phenotype with multi-flagellated bacterium (D).

Hyperswarmers spread faster on soft agar (>4 mm/h compared to the wild type’s ~3 mm/h) but this speed comes with a trade-off. Through liquid they swim slower (~35 mm/h compared to the wild type’s ~55 mm/h) and clumsier (making wider turns each time flagella switch their rotation) (Deforet et al. 2014). They pay a small but measurable cost for their many flagella, growing ~10% slower than the wild type in well mixed liquid culture. Most importantly, they are worse at making biofilms, a key P. aeruginosa trait in natural environments and clinical settings. The inability to make good biofilms explains why we have failed to find fleN-mutated hyperswarmers outside the laboratory (van Ditmarsch et al. 2013).

Hyperswarmers seem to have their advantage only on swarming plates. In this spatially structured environment, hyperswarmers move rapidly to the edge of the expanding colony where they get prime access to fresh nutrients (Supplementary video 4). Hyperswarmers out-swarm wild type bacteria and get ahead, and this compensates for the cost of many flagella. When the colony was mixed deliberately hyperswarmers no longer won the competition (Deforet et al. 2017).

This advantage granted by spatial structure also happens in other systems with very different scale. In ecology, the speed at which a species invades a new territory tends to become faster even when faster speed carries a physiological cost. This happened with species as different as the cane toads invading Northeastern Australia (Phillips et al. 2006) and the pines invading South African fynbos (Reich et al. 1994; Richardson et al. 1990). But it also happens when cancer cells spew from invasive tumors (Biddle et al. 2011). Faster speeds come with costs in all these cases. These systems are hard to study because field data are difficult to obtain and high-quality in vivo data of study cancer invasion are rare. Experiments with hyperswarmers bacteria provide a tractable model to investigate invasion (Deforet et al. 2017).

In conclusion, the hyperswarmer experiment exemplifies well the benefit of combining proximate and ultimate tools. It show that P. aeruginosa can have many flagella and swarm better, which is a proximate insight; it showed that multi-flagellated cells overcome their isogenic ancestor despite a cost and explained that we don’t find multi-flagellated P. aeruginosa in environmental and clinical settings because they can’t make good biofilms, which are ultimate insights. This experimental model may be used to test new theories and develop general principles beyond swarming bacteria (Deforet et al. 2017).

METABOLIC PRUDENCE STABILIZES COOPERATION

Another essential feature of P. aeruginosa swarming is the production and secretion of biosurfactants that lubricate the agar surface and allow the colony to slide over the top (Déziel et al. 2003; Caiazza et al. 2005). The surfactants are generally called rhamnolipids but really consist of three substances: 3-(3-hydroxyalkanoyloxy) alkanoic acids (HAAs), mono-rhamnolipids and di-rhamnolipids. HAAs are the lipid moieties of rhamnolipids only, so they lack a rhamnose unit; their synthesis is catalyzed by the rate-limiting enzyme RhlA (Déziel et al. 2003) which takes intermediates away from fatty acid synthesis and into rhamnolipid synthesis (Zhu and Rock 2008). Mono-rhamnolipids are HAAs with a single unit of rhamnose, which is added by enzyme RhlB. Di-rhamnolipids have a second unit of rhamnose, which is added by enzyme RhlC (Chong and Li 2017). The genes encoding RhlA and RhlB locate in the same operon called rhlAB, and RhlC locates in another operon with a transporter of unknown function (Rahim et al. 2001).

These three secreted substances affect swarming in different ways. A mutant lacking rhlA cannot secrete any of the three surfactants and therefore cannot swarm. Mutants lacking rhlB or rhlC secrete some of the surfactants and can swarm, but their colony shapes look different from the wild type (Caiazza et al. 2005). An interesting study separated the three components by chromatography and tested each one individually (Tremblay et al. 2007). HAAs acted as colony repellents, the mono-rhamnolipids did not attract nor repel and the di-rhamnolipids acted as colony attractants. The fraction of the three components varied with the distance to the colony edge. This should affect how bacteria respond to another colony, and may explains the fascinating branching pattern and the ability of two colonies to repel each other (Deng et al. 2014) (Fig. 1C, Supplementary video 2). It remains unclear how the ratio of these molecules influences the evolutionary stability of the swarming phenotype.

Rhamnolipid synthesis only starts when the rate-limiting enzyme RhlA is expressed. The transcriptional regulation of rhlAB implements the decision to draw resources away from metabolism and into rhamnolipid production. This decision is important; once bacteria secrete the surfactants they become a public good and could be exploited by cheaters. P. aeruginosa invests 20% or more of its own dry weight in rhamnolipids (Guerra-Santos et al. 1984; Xavier et al. 2011). The rhlAB operaon operon is tightly regulated (Reis et al. 2011); untimely expression (de Vargas Roditi et al. 2013) or overexpression (Boyle et al. 2017) could be costly.

A deletion mutant ΔrhlA lacks rhamnolipid production and cannot swarm in monoculture: incapable of lubricating the soft agar surface, ΔrhlA cells stay restricted at the inoculation site where they exhaust the nutrients and the colonies stop at ~2 billion cells (Fig. 1C). But this mutant can swarm when mixed with the wild type. Experiments with fluorescently labeled bacteria showed that a ΔrhlA colony inoculated on a swarming plate next to a wild type colony was initially incapable of swarming, but when the wild type swarm came into contact with the non-motile ΔrhlA colony, these mutant cells began to swarm using the public surfactants provided by the wild type cells (Fig. 5A, Supplementary video 5) and grew to >5 billion. These surfactants are a public good, benefitting mutant cells that do not contribute to the rhamnolipid pool (Xavier et al. 2011) (Box 3).

Figure 5. P. aeruginosa uses a strategy called metabolic prudence to avoid cheating by free-riders that do not produce rhamnolipids.

(A) ΔrhlA could free-ride on a wild type swarm. (B) Experimental evolution of ΔrhlA:wild type mixture (initially 1:1) showed that the wild type is robust to cheating because its numbers did not decrease over time. (C) This robustness was absent in a strain that could be induced to express rhlAB constitutively (also initially at 1:1 ratio with ΔrhlA) because this strain lacked metabolic prudence. (D) Metabolic prudence ensures that cells use their carbon source for growth when they have all reagents needed to make new biomass (e.g. nitrogen source or iron) and start rhamnolipids synthesis to consume excess carbon when growth is limited by another nutrient (e.g. once nitrogen or iron are depleted).

BOX 3: A METAPHOR FOR ROBUST SWARMING COOPERATION.

Swarming cooperation requires both rhamnolipid secretion and flagella motility, but these two traits are very different. Both are robust to cheaters but for distinct reasons. Rhamnolipid-deficient mutants can free-rider but the wild type avoids exploitation because metabolic prudence nullifies rhamnolipid costs (Xavier et al. 2011; de Vargas Roditi et al. 2013). Flagellum-deficient bacteria cannot even join the swarm and the wild type leaves them behind at the colony center without nutrients (Boyle et al. 2017). The cooperative forces created by the flagella is not a public good. It is more like a private insurance, where the insurance company creates a financial pool from the premium charged to each policyholder; when disaster strikes the company pays the insured but not the general public. Rhamnolipids resemble more a tax system, where everyone is supposed to contribute but even those who evade taxes still benefit from public infrastructure created with tax-payers’ money. To stretch the metaphor further, the tax-payers (metabolically-prudent rhamnolipid producers) only pay their taxes when they can afford to do so (have excess carbon), and in this way sustain themselves in the population.

However, free-riding does not mean winning, and the ΔrhlA did not outcompete wild type bacteria. When the initial mixing of wild type and ΔrhlA started at a 1:1 ratio, the ratio was stably maintained even after several days, and swarming remained robust (Fig. 5B). Swarming resisted invasion by free-riders, but how did the wild type manage to sustain its numbers if it had to produce and secrete massive amounts of rhamnolipids?

The answer turned out to be in the regulation of rhlAB. Experiments conducted with a fluorescent reporter P_rhlAB:GFP monitored rhlAB expression in media where the ratio of carbon and nitrogen ratio was carefully controlled. P. aeruginosa only expressed rhlAB in media prepared with a high cabon-to-nitrogen ratio, where nitrogen ran out first and the excessed carbon remained. Bacteria only start producing rhamnolipids when they have more carbon than they can use for growth and cell division. At that time they can afford to express rhlAB, and direct carbon-rich metabolic intermediates toward rhamnolipid production (Fig. 5D) (Xavier et al. 2011).

This strategy, called metabolic prudence, delays the investment into a public good to times when it becomes metabolically affordable. Metabolic prudence should work with other growth-limiting nutrients besides nitrogen, such as phosphate or iron because exhausting these nutrients alsos triggers rhlAB expression (Mellbye and Schuster 2014). The ultimate insight is that living organisms can reduce the cost of cooperation if they cooperate only when they have excess resources. The excess resource in the case of rhamnolipids was the carbon source, and when cells stop rhlAB expression immediately once they ran out of carbon (Boyle et al. 2015).

The expression of rhlAB requires not only excess carbon but also quorum sensing. rhlAB is under the direct regulation of the RhlI/RhlR quorum sensing system, which is itself under the regulation of the LasI-LasR forming a hierarchical quorum sensing network (Latifi et al. 1996). The ΔrhlIΔlasI double-mutant produces none of the two auto-inducer molecules yet is able to respond when they are provided exogenously. Experiments with this mutant showed that the amplitude of rhlAB expression when all nutrients area available depends only on the concentration of quorum signals in the medium (Boyle et al. 2015). In fact rhlAB expression integrates quorum sensing and nutrient information (Boyle et al. 2015; Xavier et al. 2011; Mellbye and Schuster 2014) and this way combines information on the growth rate of the individual cell and the state of the population as a whole.

There is clearly a benefit to minimizing the cost of rhamnolipid production by delaying its synthesis to times of excess carbon, when bacteria could use rhamnolipids to search for the nutrient that is lacking, whether this is nitrogen or another one. The role of quorum signaling in the evolutionary stability of this phenotype is less clear. There is little doubt that quorum signaling informs cells on more than population numbers, especially for a phenotype that depends on the spatial environment such as swarming. Quorum sensing signals could report on the diffusivity of the environment (West et al. 2012) for example. The two signals in the hierarchical structure governing rhlAB expression could help cells better resolve their social and physical environments due to the differing decay rates and diffusivities of the auto inducer molecules, 3-oxo-C12-HSL and C4-HSL for the Las and Rhl systems respectively (Cornforth et al. 2014). Aggregates of P. aeruginosa in a population density as low as 5,000 cells may communicate across more than 100 μm via 3-oxo-C12-HSL (Darch et al. 2018), the less stable of the two molecules. Spatial and temporal differences between diffusion of these two molecules may be important in strategies that prevent cheater invasion. Moreover, the quorum sensing network is more complex than previously thought: recent studies showed that RhlR regulates its downstream genes independently of its canonical quorum sensing signal using an auto-inducer produced by PqsE (Mukherjee et al. 2018; Mukherjee et al. 2017).

The regulation of rhlAB endows P. aeruginosa with a capacity to determine whether it has enough resources to invest and whether it is part of a population that is large enough. This hypothesis could be tested with experiments that ask an ultimate question: Would bacteria that produce rhamnolipids in a constant rather than regulated way, succumb to cheater mutants? Cosntitutive rhamnolipid producers were engineered by starting with the rhamnolipid-less ΔrhlA mutants and inserting a P_BAD:rhlAB construct inducible by L-arabinose. These rhamnolipid producers were severely outcompeted when mixed with ΔrhlA non-producers; their frequencies decreased sharply each day from their initial 1:1 ratio and the frequency of ΔrhlA non-producers increased (Fig. 5C). After 4 days the mixed colony was overrun by ΔrhlA non-producers and stopped swarming. As expected, when rhamnolipid production was constitutive and not prudent swarming was costly and therefore prone to invasion by cheaters (Xavier et al. 2011). The only way to avoid the extinction of constitutive rhamnolipid cooperators was through high-relatedness. Experiments showed that a high enough r could overcome the high C/B of constitutive cooperation, which is consistent with Hamilton’s rule. But bacteria live in environments where mutation and mixing constantly reduce relatedness. Metabolic prudence minimizes C/B and reduces the relatedness required for stable swarming cooperation (de Vargas Roditi et al. 2013).

METABOLISM AND SOCIAL BEHAVIOR

The sophisticated regulation of swarming cooperation reminds us that bacteria don’t need a brain to implement social strategies, a fact noted by Hamilton and Axelrod, two pioneers of social evolution theory (Axelrod and Hamilton 1981). Bacteria use networks of molecular processes to implement those strategies and metabolism—the currency of all cellular processes—plays a key role.

In bacteria, a process called catabolite repression ensures that genes for utilizing specific nutrients are expressed only when needed (Brückner and Titgemeyer 2002). The major catabolism repressor Crc of P. aeruginosa regulates the expression of metabolic enzymes that optimize metabolism for growth in specific conditions (Sonnleitner et al. 2012). Crc is itself repressed by a two component system, CbrA/CbrB, which responds to the nutrients available at a given time.

The ΔcbrA mutant overproduces rhamnolipids, but pays a fitness cost so high that it cannot swarm. Experimental evolution starting with ΔcbrA quickly evolved compensatory mutants that recovered swarming. In >80% of the cases the recovery occurred through spontaneous loss-of-function mutations in crc. This agreed with the known mechanism where the cbrA deletion should have de-repressed crc to deregulate catabolite repression and imbalance the entire metabolism. The compensatory mutations in crc restored the balance, suppressing the negative fitness impact of the cbrA deletion and lowering rhamnolipid production to wild type levels and restoring some level of swarming—although not completely. Targeted metabolomics confirmed that the crc compensation re-balanced the perturbed metabolism of ΔcbrA: the ΔcbrA mutation had caused profound changes in dozens of metabolites, but the crc mutation restored those metabolites back to wild type levels (Boyle et al. 2017).

For the <20% of the mutants found without mutations in crc but with restored swarming, we found they all had mutations in hfq, a gene coding for a RNA chaperone that modulates small RNAs which regulate metabolism of alternative nutrient sources (Sánchez-Hevia et al. 2018). hfq mutants recovered their swarming even more than crc mutants. A social interaction assay developed to quantify the costs and benefits of swarming cooperation revealed more important differences in the ultimate consequences of the crc and hfq mutations. First, the assay showed that ΔcbrA behaved extremely altruistically, with a high fitness cost a giving a large benefit to a ΔrhlA non-swarming recipient. A mutant with a compensatory mutation in crc still behaved altruistically, but less so: both the fitness cost to the actor and the benefit to the recipient lowered compared to ΔcbrA. Finally, the mutant with a compensatory mutation in hfq behaved mostly similar to the wild type: no fitness cost to the actor while still benefiting the recipient. The hfq mutation had recovered robust swarming cooperation (Fig. 2).

Figure 2. The four types of social behavior, showing P. aeruginosa mutants according to the type of social behavior they display: mutualism, altruism, selfishness or spite.

Wild type bacteria benefit themselves by swarming but also others, such as the rhamnolipid deficient mutants, without that impacting their fitness; this is an example of mutualistic cooperation. The hyperswarmer (faced against the wild-type) moves to the front of swarm and consumes fresh nutrients, selfishly increasing its own frequency and harming the wild-type. The ΔflgK (faced against the wild-type) is spiteful: at low frequencies it cannot free-ride and is left behind, but drags down the wild type or even blocks it if its initial frequency exceeds 1:1. The constitutive rhamnolipid producer P_BAD:rhlAB, the deletion strain ΔcbrA and the compensatory mutant ΔcbrA crc* overproduce rhamnolipid at different levels, but for each one the secretion is highly costly to their own fitness and those behaviors altruistically help ΔrhlA swarm to increase its frequency. The ΔcbrA hfq* helps ΔrhlA without a cost to itself, which is a mutualistic behavior similar to the wild type.

What could explain the different behaviors of ΔcbrA crc and ΔcbrA hfq? The metabolomic analysis of the hfq mutant revealed that its metabolism had failed to recover the wild type state. Instead, the hfq mutation showed an alternative metabolome composition which could have compensated the lack of proper catabolite repression caused by the cbrA deletion (Boyle et al. 2017). Interesting questions remain as to why hfq mutants were less frequent among evolved clones, and whether they would still have evolved if nutrient conditions had been different.

Bacteria have many other examples where the metabolic state of an organism conditions its social behavior. For example, E. coli cells communicate with others through quorum sensing more often when they have more nutrients, and suppress cell-to-cell communication when starved by sequestering external signals (Xavier and Bassler 2005). Metabolism powers all cellular processes, and is key to organismal behavior, including social behavior. Natural selection, when all else is equal, should favor organisms that execute a behavior at the lowest C/B ratio, and the conservation of metabolic resources should be a major factor in reducing cost. Therefore, metabolism plays a central role in the feedback from genes to social behavior—a proximate link—and back through natural selection—an ultimate link (Boyle et al. 2017).

Examples beyond bacteria suggest that the impact of metabolism on social behaviors is conserved through the tree of life (Rittschof et al. 2015). In humans, low glucose levels in the brain conditions aggression (Siever 2008) and possibly other social behaviors. A study of parole decisions among judges in Israel (Danziger et al. 2011) found that the rate of favorable decisions dropped sharply during a session. Surprisingly, favorable decisions rose again after each meal, then started dropping again as the judges endured each session without eating. This curious observation raises important questions on the role of physiological factors in decisions taken by one individual (the judge) that influence another (the felon candidate to parole).

MACHINE LEARNING AND EVOLUTION OF THE C-DI-GMP NETWORK

P. aeruginosa makes biofilms by producing of a dense matrix of extracellular polymers and sticking to surfaces. Biofilms are important for virulence and increase the resilience of P. aeruginosa to antimicrobials (Høiby et al. 2010). Biofilms and swarming are two cases of surface-associated social behavior where dynamics of cooperation and conflict play major roles (Nadell et al. 2009). But biofilms are static and swarms are motile, so they represent alterative ways to live on a surface. When bacteria reach a surface, how do they decide between the two behaviors?

P. aeruginosa and many species of bacteria rely on the intracellular messenger c-di-GMP (Bis-(3’-5’)-cyclic dimeric guanosine monophosphate) to make the vital decision between staying in one place to form a biofilm, or to leave that place in search of better conditions. c-di-GMP is at the center of a signaling network shaped like a bow-tie (Fig. 6A). The network has many inputs—the stimuli—and many outputs—the downstream phenotypes that it regulates (Hengge 2009). Several protein machineries trigger the synthesis and the destruction of intracellular c-di-GMP, and many of these machineries respond to extracellular stimuli, such as the presence of a solid surface or a specific chemical compound. When c-di-GMP is high, bacteria enter ‘biofilm mode’ by upregulating biofilm genes and repressing motility genes and they; when c-di-GMP is low, bacteria enter a ‘motile mode’ by repressing biofilm genes and expressing flagella instead (Hickman and Harwood 2008).

Figure 6. The c-di-GMP network functions as a classifier that determines whether bacteria should stay in that environment to form a biofilm or move away.

(A) The c-di-GMP signaling network is shaped like a bow-tie. (B, E) Wild type cells could classify the environment properly and form biofilm when they are supposed to. (C) The hyperswarmer with mutation in fleN fails to respond to high c-di-GMP level and therefore is locked in swarming phenotype. (D) Mutation in the wspF gene resulted in high c-di-GMP levels even in swarming conditions and locked the mutant in a biofilm phenotype. Additional mutation to fleN* could change its phenotype to either mimic wild type (fleN*dipA*, B, E), or mimic wspF* (fleN*wspF*, D)

Bow-tie networks exist in many biological systems including the toll-like signal receptor network in immune systems (Oda and Kitano 2006). Natural selection favors bow-tie networks when the information flowing from many inputs to many outputs can be compressed to a few internal network nodes (Friedlander et al. 2015). Bacteria utilize the c-di-GMP network to compress the stimuli sensed by several protein machineries, and implement the inverse regulation of biofilm and swarming (Caiazza et al. 2007).

Isolates of P. aeruginosa differe in their intracellular c-di-GMP levels. One could expect based on the inverse-regulation model (Caiazza et al. 2007) that the isolates with lowest c-di-GMP should form less biofilm and swarm better than those with high c-di-GMP. However this is not the case: c-di-GMP levels measured across a panel of 28 P. aeruginosa clinical isolates showed no apparent correlation between the c-di-GMP, biofilm and swarming (Yan et al. 2017). What could explain such discrepancy between clinical and laboratory observations?

Mathematical modeling provided an explanation: the bow-tie network architecture was resembles an algorithm used in machine learning called ‘logistic regression’. The many protein complexes that make and break c-di-GMP could serve as inputs for the classification task that determines whether an environment is suitable to form a biofilm, or if bacteria should move away (Fig. 6B, E). A network produced by natural selection across many changing environments would be well adapted to the stimuli encountered in the past, and this is similar to the way that a machine learning algorithm learns a classification task from a large training dataset. The network is tuned through advantageous mutations in the input and output proteins, and incrementally improves the bacterium’s ability to make a biofilm or swarm depending on the environment sensed (Yan et al. 2017).

The analogy between natural selection and machine learning helped clarify the discrepancy between clinical and laboratory data. Clinical strains had diverged evolutionarily and their networks adapted to the distinct histories of environments experienced since a common ancestor. Laboratory mutants, on the other hand, diverged recently through engineered mutations or through experimental evolution. Mutants produced by strong selection in the lab have their networks locked in a single mode: biofilm or swarming. The coupling between c-di-GMP and the two social behaviors is strong in those mutants.

There are many examples of mutants in c-di-GMP network genes which are locked in either biofilm or swarming. For example, some of those mutants occur in the Wsp protein complex. This transmembrane machinery senses the presence of a surface and activates a c-di-GMP synthase, WspR, which then raises the c-di-GMP levels (Huangyutitham et al. 2013). This complex is thought to provide bacteria with a ‘sense of touch’. Loss-of-function in the de-methylator WspF over-activates the complex, and the mutant cells behave as if they were constantly touching a surface: c-di-GMP levels are high, they form robust biofilms and cannot swarm. Other mutations—such as loss-of-function in the transmembrane complex WspA—disable the complex and cells loose their sense of touch: c-di-GMP levels stay low, and cells cannot form good biofilms but swam robustly (Yan et al. 2017).

The bow-tie model also explains why fleN-mutated hyperswarmers are poor biofilm formers (Fig. 6C). The FleN protein forms a complex with FleQ, a transcriptional regulator of flagella that responds to high-levels of intracellular c-di-GMP by repressing flagella genes and expressing extracellular polymers to make a biofilm (Hickman and Harwood 2008). The point-mutations in fleN make the FleN-FleQ complex less sensitive to high levels of c-di-GMP, and hyperswarmers stay locked in motile mode even in conditions where they should form a biofilm (van Ditmarsch et al. 2013). The effect of the fleN mutation could be overcome by acquiring additional mutations such as in dipA or wspF, both of which function in sensing or transducing environmental signals to c-di-GMP synthesis, and restored biofilm formation (Fig. 6D, E). The effect of wspF mutation is so strong that the fleN*wspF* locked cells in biofilm mode (Yan et al. 2017).

Interestingly, nutritional factors can modulate the behavior of c-di-GMP network components, suggesting a wider influence of environmental factors on the decision between biofilm and swarming (Mattingly et al. 2018).

Besides providing an explanation for the central role of c-di-GMP in bacterial social behaviors, the analogy between evolutionary tuning of a molecular network and the training of a machine learning classifier has general implications (Yan et al. 2017). Understanding how simple biochemical processes can evolve to become complex decision-making machines is a key problem in biology (Watson and Szathmáry 2016). The theoretical concepts of machine learning can shed new light on the evolution of biological complexity.

CONCLUSION

Full understanding of a behavior requires proximate and ultimate mechanisms (Mayr 1961). Proximate studies with bacteria can reveal the fine-scale molecular, cellular and population-level function of a cooperative bacterial behavior. Ultimate studies can explain how natural selection favors and stabilizes that cooperative behavior even it comes with a fitness cost to the actor. Other reviews focused on the proximate mechanisms of swarming in the past (Kearns 2010). Swarming in P. aeruginosa requires rhamnolipid secretion, flagella motility and even other sub-cellular functions such as the pili, and regulatory pathways with complex architecture such as hierarchical quorum sensing and the c-di-GMP network that regulates transitions between motile and biofilm modes across many species of bacteria. Proximate mechanisms provide the basis for ultimate studies that address questions which proximate studies cannot: How did swarming evolve and remain stable if cheaters could theoretically emerge through spontaneous loss-of-function mutations? How does evolution produce complex networks that learn through incremental mutations? Many intriguing questions remain. We hope this review stimulates new swarming experiments or even experiments with other common microbiology assays that seek ultimate causality.

It is fair to question whether common microbiology assays reveal anything about evolution. Those who focus on molecular mechanism may deem this approach unless it proposes a new molecular mechanism or improves an existing one. Evolutionary experiments with microbes occasionally do produce new proximate insight, as when P. aeruginosa spontaneously evolved fleN-mutated hyperswarmers (van Ditmarsch et al. 2013). Hyperswarmers have many flagella, which agrees with an insertion mutant in fleN constructed earlier (Dasgupta et al. 2000). But the insertion mutant could not move, and the hyperswarmers swarmed better than the wild type. The discovered fleN-mutations will no doubt be important (Kearns 2013) even though the regulation of flagella is an old problem, studied exhaustively by molecular bacteriologists.

A possible caveat is that laboratory selections are strong. Are experimentally-evolved phenotypes trivial consequences, expected given the selection applied? Repeated culture of P. aeruginosa in swarming plates always leads to the evolution of multi-flagellated hyperwarmers; we know this because we repeated the experiment many times (van Ditmarsch et al. 2013. But this was unpredictable before we conducted the first experiment because P. aeruginosa swarming relies on many cellular traits. Any of those traits could have been improved by spontaneous mutations and laboratory selection. Nothing guarantees that selection in the laboratory should always converge, and there are counter examples (Kawecki et al. 2012). The reproducible evolution of hyperswarmers was unexpected and, therefore, noteworthy.

Experiments with swarming cooperation provide ultimate insights even without new revelations new molecular insights. At the proximate level metabolic prudence level is simply regulation of gene expression. At the ultimate level it is a decision-making strategy that stabilizes cooperation. Looking at microbial phenomena beyond their proximate mechanisms can help solve on of biology’s great puzzles: the evolution of cooperative behavior.

Supplementary Material

supporting video 1

Supporting video SV1: Time lapse of P. aeruginosa PA14 wild type swarming on 0.5% agar.

supporting video 2

Supporting video SV2: Two wild type colonies repelled each other during swarming.

supporting video 3

Supporting video SV3: Hyperswarming mutant evolved from P. aeruginosa wild type lost branching pattern and spread over the entire petri dish

supporting video 4

Supporting video SV4: The hyperswarmer could fast move to the edge of a swarm during competition with wild type (cells fluorescently labeled: hyperswarmers: green; wild type: red)

supporting video 5

Supporting video SV5: rhlA^- mutant which cannot synthesize rhamnolipid could free ride on the swarm of the wild type.