We begin by recapitulating the known effects of antibiotic stress on bacterial survival and biofilm formation. Fig 1 (panels A, B) shows the effects of antibiotics on Pseudomonas aeruginosa where we use optical density as a qualitative indicator of cell number. As expected, increasing the concentration of three widely used clinical antibiotics—ciprofloxacin, rifampicin and tetracycline—monotonically reduces the optical density of shaking cultures. However, as previously shown, this monotonic decrease is not seen when the same concentrations are applied to standing cultures. In standing cultures, the liquid is not agitated and cells can readily attach to the edge of the well and establish biofilms. Under this condition, biofilm formation actually increases for many concentrations of the antibiotics, until the concentrations become so high that toxicity dominates [13].

Applying sub-lethal concentrations of antibiotics then leads to an increase in biofilm formation. Can we understand these widely reported biofilm responses to antibiotics in terms of ecological competition? We tested this by mixing natural isolates of P. aeruginosa in standing cultures and studying the impact on biofilm formation. Consistent with the hypothesis that biofilm formation is a response to ecological competition, we find that biofilm formation is increased in mixed cultures compared to pure cultures, both in a broad scale survey that mixed combinations of 22 natural isolates (Fig 1C) and in defined pairwise comparisons in which we show that the strength of the biofilm response depends on the relative proportion of each strain at the start of the co-cultures (Fig 1D).

The observed biofilm response is consistent with a competitive response. However, an alternative hypothesis is that the strains cooperate with one another to produce more biofilm (see Introduction). In order to better assess whether the strains were cooperating to make more biofilm, we labeled five of our P. aeruginosa natural isolates with constitutive fluorescent proteins (cyan fluorescent protein, CFP, and yellow fluorescent protein, YFP) by putting a single copy of the desired fluorescent protein gene at a specific site in the genome [50]. We then set up a round-robin tournament, in which the five strains were mixed in all combinations, and measured biofilm formation, planktonic growth, and the relative frequency of the two strains (Fig 2). This revealed a strong competitive effect with one strain typically dominating after 24 h (median proportion of dominant strain in biofilm = 0.98, Wilcoxon-signed rank test versus 0.5, n = 10, p < 0.01). Thus, the biofilm response is a manifestation of competition between strains rather than a cooperative synergy in which both strains benefit from the presence of one another [34].

There are some cases in which biofilm formation in the mixture is less than that expected from the average of the two starting strains (Fig 2). However, accounting for which strain actually makes the final biofilm reveals that mixing nearly always causes an increase in biofilm formation by the dominant strain. For example, mixtures of 7 and 19 result in less biofilm than monocultures of strain 19. However, strain 7 eliminates strain 19 in co-cultures, producing more biofilm than it does when alone (Fig 2). More generally, we can calculate the expected biofilm formation from the pure culture data using a weighted average, which weights by the frequencies of the two strains in the biofilm after competition. This reveals that the mixed culture biofilm normally increases relative to the expectation based upon pure cultures (9/10 mixes increase, Wilcoxon signed rank test of expected/observed biofilm against a null of 1, n = 10, p < 0.01), while the amount of cells in the planktonic phase decreases (n = 10, p < 0.01). Moreover, the increases in biofilm and decreases in plankton that we observe upon mixing strains are significantly negatively correlated across the different strain mixes (Spearman rank correlation, n = 10, r = -0.782, p < 0.01). This correlation agrees with recent observations on multispecies interactions during biofilm formation [31].

The biofilm response is not an idiosyncrasy of our particular growth medium. S1 Fig shows the effect of varying levels of nutrients, buffering, and soluble iron for two genotype mixtures. Despite the success of strains in pairwise competition being strongly dependent on the environment, the biofilm response is detected across all conditions in at least one of the two strain combinations, but typically in both (S1 Fig).

We next focus on the cue that drives the increase in biofilm formation in co-cultures. We first examined whether the biofilm response requires cell–cell contact by growing pairs of strains on either side of a 0.4 μm permeable membrane (Fig 3A). Increased biofilm is observed across the membrane in most cases. In one case, there is a strong inhibitory effect on both biofilm and planktonic growth, which is consistent with the responder strain losing in co-culture (specifically, strain 4 is inhibited and outcompeted by strain 1, Fig 2). We also found that the biofilm response is recapitulated by the addition of cell-free culture supernatant of different strains. We focus on strain 4, which responds robustly to the addition of the supernatant from other strains (Fig 3B). The nature of the response is dose dependent: lower concentrations tend to induce biofilm formation while high concentrations inhibit both planktonic growth and biofilm formation.

In order to identify the active component within the supernatant, we first performed anion exchange chromatography on the supernatant of strain 69 as it generates a robust response in strain 4 (Fig 3B). The active fraction that induced biofilm formation in strain 4 was further separated based on size by gel filtration. A single fraction (C9) appeared to contain the bulk of the active component as it strongly inhibited growth in strain 4, while the two flanking fractions C8 and C10 (i.e., lower concentrations of the effector) induced the biofilm response (Fig 4A). We subjected fraction C9 to protein mass spectrometry (see Methods). Of the proteins identified, three are notable for their known role in microbial interactions: homologs (PA14_8070, PA14_8090, PA14_8210) of the R and F pyocins in P. aeruginosa strain PA14 (S1 Table) [51]. Given that PA14 is a well-characterized strain, we next sought to explore how the pyocins of PA14 induce the biofilm response in our natural isolates.

We first evaluated the effects of cell-free culture supernatant from PA14 on biofilm formation in the five strains from the round robin tournament. Strain 19, which was isolated from a cystic fibrosis infection, is a very poor competitor (Fig 2) and showed no detectable growth in our supernatant assays. It was therefore excluded from further analysis. Among the remaining strains, increased biofilm was observed in only strain 4 (Fig 4B), and we confirmed that direct mixtures of PA14 and strain 4 also lead to increased biofilm formation (S2 Fig). Why is the biofilm response absent in the other three strains? To address this question, we grew the four strains in shaking culture in the presence or absence of the supernatant from PA14 and followed their growth over time. The results obtained show that the strains that do not show the biofilm response to the supernatant of PA14 also are not harmed by it (Fig 4C). This correlation between toxicity and increased biofilm helps explain the lack of response in the three strains. Cell damage appears to be required for the biofilm response.

We next studied the effect of various mutants of pyocin production in PA14 on their ability to damage and induce the biofilm response in strain 4. We focused particularly on a strain of PA14 that has the repressor of pyocins, prtR, mutated to a version that is constitutively active (S3 Table) [52]. We added cell-free supernatant of the PA14 prtR mutant to strain 4 and compared the effect to the addition of WT PA14 supernatant by propidium iodide staining. Propidium iodide preferentially enters cells with envelope damage (see Methods), and this assay confirmed that the pyocins produced by PA14 cause cell damage to strain 4 (S3 Fig), which agrees with published literature on pyocin-driven lethality [51]. Moreover, additional experiments showed that the pyocins of PA14 are the major effectors of both population growth inhibition (Fig 4C) and the biofilm response of strain 4 (Fig 4D).

Focusing on the biofilm response, the use of mutants that lack only one class of pyocin suggested that F pyocins have some effect on biofilm formation at high concentrations, but it is particularly the R pyocins that strongly stimulate biofilm formation. This is, furthermore, consistent with the importance of cell damage because R pyocins are known to be efficient killers that can contribute significantly to competition among P. aeruginosa strains [53,54]. To further explore the potential of cell damage being required for the biofilm response, we screened 12 natural isolates for inhibition by the supernatant of PA14 in shaking culture and found that three strains were inhibited like strain 4. We randomly selected two of the inhibited strains for further study in the biofilm assay. This revealed that both exhibit a biofilm response comparable to that of strain 4 and, again, the response occurs in a pyocin-dependent manner (S3 Fig). We also tested the response of the strain PAK under the same conditions. This serves as a good control case since PAK is known to be inhibited by the pyocins produced by PA14 [53]. Again, we find evidence of pyocin-driven biofilm formation (S4 Fig).

These data further support the idea that pyocin-driven damage under these conditions is sufficient to drive the biofilm response. We note, however, that this does not exclude the possibility that there are additional mechanisms by which P. aeruginosa strains damage and induce biofilm in other susceptible strains. Indeed, the link between pyocin-based cell damage and the biofilm response is consistent with experiments using clinical antibiotics, in which antibiotic resistance mutants require higher concentrations of antibiotics to elicit the biofilm response [10]. To confirm this effect in our assays, we tested a rifampicin resistant mutant of PAO1, with defined mutations in rpoB [55] and found that it does not respond to the same concentrations of rifampicin that trigger the biofilm responses in the wild type (S5 Fig). In sum, we see comparable effects on biofilm formation from diverse forms of cell damage. This suggests that cell damage from a diverse range of bacteria interactions, both between strains and between species, has the potential to promote biofilm formation.

Our experiments so far have utilized the crystal violet microtiter dish assay of standing cultures [48]. Although this experimental approach is widely used to quantify biofilm responses to antibiotics [13], it comes with two notable limitations. Firstly, indirect proxies are used to estimate biofilm formation (crystal violet staining) and planktonic cell numbers (optical density). Secondly, the presence of a large planktonic population in the standing culture means that the planktonic cells can affect the biofilm, and vice versa, in an undefined manner throughout the experiment. We therefore sought to recapitulate our results in a more defined assay using flow cells and direct imaging of the biofilms (Fig 5A).

We begin by studying one of our focal combinations, strain 1 (PAO1) and strain 4. As we find in the standing culture assay (Figs 1D and 2B), PAO1 dominates strain 4 and largely excludes it from the biofilm after 24 h (Fig 5B). Moreover, we observe an increase in the volume of biofilm formed by PAO1. Our observations in the less-defined crystal violet assays, therefore, are recapitulated in flow cell biofilms: strain-mixing results in a single strain dominating and an increase in biofilm formation. Do we also observe this biofilm response in strain 4, which was the focus of detailed analyses in the previous section? In order to study this, we needed to create a condition where strain 4 is not eliminated from the biofilm. Preliminary tests with our two focal mixtures (1 + 4, 4 + 69, Fig 1D) revealed that strain 4 will persist with strain 69, if it is over-represented at inoculation. We, therefore, tested strain 4 mixed with strain 69 in the flow cells at a 3:1 initial ratio. Under these conditions, strain 4 dominates the biofilms and largely excludes strain 69 and, most importantly, we see a clear increase in biofilm formation in strain 4 here when compared to its single-strain counterpart (Fig 5C).

Experiments using flow cells have a separate attachment phase at the beginning of the experiment to seed the surface with bacteria (see Methods). Counting attached cells after the attachment phase revealed that strain 4 attaches more in the presence of strain 69 but strain 1 does not attach more in the presence of strain 4 (Fig 5). Both initial attachment and subsequent biofilm accumulation, then, contribute to the response to strain mixing but while we always see increased biofilm formation, we do not always see an initial difference in attachment (also see next section). In sum, the flow cell experiments provide validation from a second assay that P. aeruginosa strains increase biofilm formation in response to the presence of competing strains.

Our data suggest that bacteriocins produced by natural isolates of P. aeruginosa act as a cue that drives biofilm formation. This raises the possibility that the widely reported biofilm responses to clinical antibiotics [10–17] also have their origins in responses to ecological competition from strains and species. But does the addition of antibiotics, or cell-free culture supernatant from competitors, also result in increased biofilm formation in a microfluidic device?

Experiments with a single concentration of antibiotic revealed that the addition of cell-free culture supernatant or clinical antibiotics commonly results in cell detachment at the initial stage of the experiment. This is consistent with the effects of supernatants and antibiotics in the crystal violet assay being strongly concentration dependent, whereby lower concentrations cause the biofilm response and higher concentrations cause growth inhibition (Fig 3). We, therefore, developed a more robust method to study the effects of cell-free supernatant and antibiotics on biofilms in microfluidic devices. We moved to use microfluidic platforms with two inlet channels in which only one of the two inlets contains the antibiotics (Fig 6A).

The goal of this experimental approach was to create a gradient inside the main chamber with a wide range of antibiotic concentrations, from sub-lethal to lethal (see Methods). Our protocol always allows some parts of the chamber to contain negligible amounts of the antibiotic, which ensures that not all cells detach when antibiotics are flowed through. Importantly, the cells will experience a wide range of concentrations of antibiotic in a single experiment, and we can directly observe how they respond. Under these conditions, we observe a large increase in biofilm formation upon the addition of multiple classes of antibiotics in multiple strain backgrounds (Figs 6, S6 and S7). Our data show that antibiotics do indeed promote biofilm formation in flow cells. In addition, they support the value of studying gradients for understanding cellular responses to stress. Such gradients are common in biofilms [56,57], but their effects can be missed by static culture assays.

We next asked whether pyocins are also capable of promoting biofilm formation in our dual inlet experiment. In support of this, we find that flowing the cell-free culture supernatant of PA14 over strain 4 results in increased biofilm formation, but not when we use the supernatant of the pyocin null-mutant (Fig 6). What drives the increase in biofilm? Increased biofilm accumulation could occur through stronger attachment and better retention of cells throughout the experiment, or through the acceleration of cell division. In order to test the impact of pyocins on attachment, we allowed strain 4 cells to attach in the presence of cell free supernatant of PA14 and compared this to the effect of cell free supernatant of the PA14 pyocin null-mutant (S8 Fig). This shows that pyocins do indeed lead to increased attachment. This phenotype is not simply sedimentation of damaged cells. We flush unattached cells out before imaging and, importantly, we see increased attachment in the presence of pyocins both on the bottom and the top of the channel (S9 Fig). Moreover, propidium iodide staining shows that, concurrently, there is increased cell damage from the supernatant of PA14 than with the supernatant of the pyocin null mutant (S8 Fig), consistent with what we found for shaking culture (S3 Fig). We see, therefore, that pyocins from competitors both cause cell damage (S3 and S8 Figs) and an increase in biofilm formation (Figs 6, S4, S8 and S9).

The attachment experiments (S8 and S9 Figs) suggest that the biofilm response may occur via an increased propensity for cells to attach, and subsequently persist, in the biofilm. An increased rate of cell division could also contribute to the biofilm response in some or all strains. Direct assessments of cell division in three-dimensional biofilms is challenging and was not possible with our current methods. However, there is evidence from Escherichia coli that sub-lethal levels of antibiotics can increase growth rate [58]. More generally, there is evidence for widespread transcriptional and metabolic changes in response to antibiotics [16,59–66], which have recently been the focus of intense debate [62,67–70] but still lack evolutionary explanation. If these changes are also involved in the responses to other bacterial strains and species, it would suggest that many of the physiological and metabolic changes seen upon antibiotic use have their natural origins in competition sensing [41] and the ecology of the species concerned.

A total of 22 strains of Pseudomonas aeruginosa were used, including both clinical and environmental isolates from the Kolter strain collection (S2 Table). A subset of these strains was subject to detailed analyses, including the well-studied clinical strain PAO1 (here referred to as strain 1), and strains 4, 7, 19, and 69. These five strains were chromosomally tagged with constitutively-expressed fluorescent tags (cfp or yfp) using the mini-Tn7 system [50]. The pyocin-null prtR^S153A (constructed as in [52]), producing constitutively active repressor, and its parental strain PA14 were kindly provided by Jon Penterman. The R and F pyocin mutants are from the PA14 Non-Redundant Transposon Insertion Mutant Set [79]. We found no differences in our assays between the two sources of PA14 strains and only show the data from the Penterman source in the figures. Other strains are described in S3 Table.

These assays were based on standard protocols [48]. Briefly, cells from overnight cultures of P. aeruginosa were sub-cultured and from these, exponential phase cells were inoculated in peg-lidded 96-well plates (Nunc, conical bottom) in 100 μl tryptone broth (10 g Bacto triptone per 1 l water) at starting optical density of 0.0025 (A₆₀₀), and allowed to grow at 22°C. After 24 h, 50 μl of medium was removed and the optical density at A₆₀₀ was measured to estimate the density of planktonic cells. Pegs were then washed, stained with 0.1% crystal violet, washed again, and then the remaining stain was dissolved with 200 μl of 33% acetic acid. Finally, the optical density at A₅₉₅ was measured and used as a proxy for biofilm formation. For direct strain-mixing experiments (Figs 1C, 1D and 2), we used plates without pegs because this allows cells to be more easily removed and sampled from the biofilms when required (see below). The only substantive difference is that the biofilm on the edge of the well is assessed using crystal violet rather than the biofilm on a peg. We confirmed that the two assays give very similar results (S10 Fig). While the experiments presented in the main text were done in tryptone broth, the biofilm response was also investigated in other standard bacterial growth media, including Luria broth (LB), Pseudomonas minimal medium (PMM) and casamino acid (CAA) medium, along with different concentrations of iron or buffer, which did not have qualitative effects in the cellular responses to mixing (S1 Fig).

Three antibiotics belonging to different structural families, namely ciprofloxacin, rifampicin, and tetracycline, were chosen to recapitulate the effect of sub-inhibitory concentrations of antibiotics on biofilm formation of P. aeruginosa [13]. Stock solutions of each antibiotic (obtained from Sigma Chemical Company) were stored at -20°C and diluted to the desired concentration on the day of each experiment. All antibiotics experiments used tryptone broth as the growth medium. Biofilm formation and growth in shaking culture were assayed as detailed in the sections “Static Biofilm Assays” (above) and “Growth Curves” (below), respectively. For the dual-flow microfluidic experiments (Fig 6) we used 10x the concentration that inhibited growth in shaking culture, as shown in Fig 1. These are 1 μg/mL for ciprofloxacin, 50 μg/mL for rifampicin, and 1,000 μg/mL for tetracycline.

The initial survey of how strain mixing affects biofilm formation was based on 22 strains of P. aeruginosa. Biofilm formation after 24 h was evaluated in 22 single strains, 22 pair-wise combinations, eight combinations of five strains, and two of ten strains. Strains were evenly mixed based upon the optical density of stationary phase cultures. The strain combinations in the mixtures were chosen to ensure a near-even representation of the strains across the mixtures. Each strain occurred twice in the pairs, twice in the mixtures of fives (with one strain appearing only once), and once in a mixture of ten (with one strain excluded). Sixteen replicates were performed in each experiment for each strain or mixture, to give a total of 432 biofilm assays in each experiment. All treatments were performed on two different days.

The second set of experiments mixed five fluorescently-labeled strains in all combinations. We assessed biofilm formation and planktonic cell density in 96 well plates after 24 h. Samples of planktonic cells and biofilm cells were taken from each treatment to record the frequency of the two strains using fluorescence microscopy. The biofilm sample was taken by vigorously shaking the washed plates in phosphate buffer for one hour (experiments revealed that this treatment gives identical frequency estimates as shaking with 500 μm glass beads and removing the whole biofilm). In addition, direct images of the biofilms were taken by growing strains in 2 ml of media containing 25 x 25 mm² glass coverslips in 6-well plates under conditions otherwise identical to the 96-well plate assay. The bacteria form a biofilm on the coverslip slightly below the air-water interface, which is then visualized by fluorescent microscopy (Fig 2).

All experiments were repeated twice on different days. Comparison of these data with the pair-wise mixtures in the first experiments (above) confirmed that the mixing responses were not affected by fluorescent labeling. All methods were the same as before with the addition that we diluted overnight cultures of each strain and grew them for 2 h to obtain exponential phase cells to inoculate the biofilm assays (Fig 2). This did not affect biofilm formation, but it helped reduce variability in the strain frequency in the mixtures. For the two large-scale strain-mixing experiments (Figs 1C and 2), we present data that is the combination of the two repeats from different days as variability between days was relatively high.

Biofilm formation was also monitored under laminar flow conditions using the BioFlux 200 system (Fluxion Biosciences, South San Francisco, CA, United States). Experiments were performed essentially as described elsewhere [80]. First, microfluidic channels (350 μm x 75 μm x 4.89 mm for width, height, and length, respectively) were primed with growth media, here tryptone broth, to prevent air bubbles from entering the channels on subsequent loading steps, and just after, inoculated with cells at the exponential phase. The concentration of these cultures was adjusted so that the optical density at 600 nm in tryptone broth was 0.25. Then, after the attachment period without flow (20 min at 22°C), fresh medium was pumped from inlet wells through the channels to outlet, firstly with a flow rate of 40 μl/h for 20 min to flush out planktonic cells, and then with a constant flow rate of 4 μl/h until the end of the experiments (24 or 48 h). To address the effect of pyocins on cellular attachment we allowed cells to attach in the channels for 20 min with tryptone broth made with 10% cell-free supernatant, either from strain PA14 WT or a mutant that has the repressor of pyocins, prtR, mutated to a version that is constitutively active (S3 Table). We then flushed out planktonic cells, as explained above, and imaged to assess attachment.

For consistency, we used only one plate type for all microfluidic experiments, the “Invasion” plate of Fluxion Biosciences that has two parallel inlets and one outlet (Fig 6). When studying co-cultures, the two inlets contained the same medium, and when studying the effect of antibiotic gradients we added the required concentration of the antibiotic (or cell-free culture supernatant) to one of the inlets only, thus creating a gradient of the compound being tested in the main chamber. Each experiment was repeated twice, on different days, and the images of biofilms taken come from the middle of the channels, where the channels are the flattest, when accessing the effects of strain mixing. In all treatments, images were taken at the same distance from the respective inlets. When accessing the effect of antibiotics or cell-free supernatant we imaged directly downstream of the inlets, as shown in Fig 6, where the gradient is the steepest.

To access the effects of diffusible substances in the biofilm response, two sets of experiments were performed. Firstly, we used 96-well Transwell plates (Corning) with two different inlays separated by a 0.4 μm pore size membrane. In these experiments, the strain in the bottom of the well was inoculated at an optical density of 0.0025 (A₆₀₀) in 75 μl of media. This bottom strain was the tester strain in which biofilm formation was analyzed. The top strain was the inducer and was inoculated at a higher density of 0.1 (A₆₀₀) with 50 μl media. These revealed cell-to-cell contact is not strictly necessary for the biofilm response (Fig 3A). To investigate this further, we studied responses of strains to cell-free supernatants from other strains. The supernatant of the strains were obtained as follows. Exponential phase cells of all strains were inoculated in 6-well plates in 4 ml of tryptone broth at starting optical density of 0.025 (A₆₀₀), and allowed to grow at 22°C for 21 h. Cultures were then spun down (3,000 g during 10 min) and filtrated using 0.22 μm filters (Millipore). The spent media assay followed the same methods for the biofilm assay, except that exponential phase cells of the responder strain were inoculated in media containing (2.5/25%) cell-free culture supernatant from itself or others.

Fractionation of the cell free supernatant was performed by GE 1 ml resource Q anion exchange chromatography (buffer A: 50 mM Tris-HCl [pH 7.5], 2% glycerol, 2 mM β-mercaptoethanol; buffer B: buffer A + 1 M NaCl) followed by GE Superdex 200 10/300 GL gel filtration in SEC buffer. The active fraction, as determined by the crystal violet assay (Fig 4), was subject to mass spectrometry by the FAS Center for Systems Biology Core Facilities, Harvard University, US. Briefly, the sample was reduced, carboxyamidomethylated and digested with trypsin. Peptide sequence analysis of each digestion mixture was then performed by microcapillary reversed phase high-performance liquid chromatography coupled with nanoelectrospray tandem mass spectrometry (μLC-MS/MS) on an LTQ-Orbitrap mass spectrometer (ThermoFisher Scientific, San Jose, CA, US).

Exponential phase cells from overnight cultures were inoculated in flat-bottomed 96-well plates (Nunc) in 200 μl of tryptone broth at the starting optical density (A₆₀₀) of 0.0025, as for the biofilm assays, and allowed to grow at 22°C, shaking, inside a plate reader (Tecan, Infinite M200 Pro). Test treatments were supplemented with 25% of cell-free supernatant from PA14, the respective pyocin-null, or distilled water (control). The optical density at A₆₀₀ was measured automatically every 500 s. When required, cell damage was assessed by staining shaking cultures or attached cells in a microfluidic device with propidium iodide (at the final concentration of 3 μl/ml)[81]. Propidium iodide-stained cells were visualized using epifluorescence microscopy with DsRed-based filters.

Images of BioFlux biofilms were obtained using a Zeiss confocal laser scanning microscope (LSM 700), a digital camera (AxioCam MRm), and the associated Zen 2011 software (blue or black edition, for fluorescence or confocal microscopy, respectively). Objectives used were 10x, 20x, or 40x, as specified in the figure legends.

A MATLAB (MathWorks) script was used to determine the number of attached cells on the surface of microfluidics, along with the volume of biofilms after 24 h and 48 h of incubation. The volume calculation uses the area of fluorescence signal of biofilms in each confocal slice, which is then summed across the slices in z-plane.