## Introduction

Bacteria employ many appendages for movement and dispersal in their ecological niches1. Of these, flagella-driven motility is the fastest, promoting the colonization of favorable niches in response to environmental signals, and contributing significantly to the pathogenic ability of some species2,3,4,5. Flagella are used both for swimming individually through liquid, and swarming collectively over surfaces3,6,7. An unexpected and clinically relevant property of swarms is their nongenetic resistance to antibiotics at levels lethal to free-swimming planktonic cells of the same species8,9,10,11,12. The resistance depends on high cell densities associated with swarms12, and is phenomenologically similar to adaptive resistance of bacterial biofilms13. While several terminologies such as resistance, tolerance, and persistence have been used to describe bacterial resistance phenotypes observed under various settings14,15, a shared attribute of a majority these phenotypes is their association with reduced metabolism and slow bacterial growth16. In stark contrast, bacterial swarms are metabolically active and grow robustly, suggesting that different resistance mechanism/s may be operative here. To distinguish the adaptive resistance seen in swarms from that in other reported cases of this phenomenon, we are calling swarming-specific resistance SR.

In a study of several bacterial species, swarms that successfully colonized media containing antibiotics nonetheless experienced substantial cell death12. Programmed bacterial cell death, where death of a subpopulation benefits the community by providing nutrients, has been reported by several studies17,18,19,. Cell death incurred during swarming, however, appeared to provide a protective benefit, leading to speculation that a dead-cell layer at the bottom of the swarm might shield live cells swarming on top by acting as a physical barrier against antibiotic penetration12. In a study examining the 3D architecture of a swarm in the presence of different antibiotics, dead cells were not always found at the bottom, ruling out a barrier role20. Instead, live cells were seen to migrate away from the antibiotic source, leading to a suggestion that cell death emanating around this source releases an avoidance “signal” that contributes to the resistance20.

The impetus for the present study came from our observation of heterogeneity in antibiotic-induced cell killing in a swarm, where a distinct subpopulation was more susceptible to killing. To test if death of this subpopulation contributes to SR, pre-killed cells were introduced into the swarm, where they indeed enhanced SR. In a study focused on Escherichia coli (E. coli), we have used this observation to purify the SR-enhancing factor. We show that the dead cells release a “necrosignal” in the form of AcrA, which is a periplasmic protein component of a tripartite RND efflux pump. AcrA binds to the TolC component of this pump on the outer membrane (OM) of live cells to activate immediate efflux, as well as stimulates expression of multiple pathways that overcome antibiotic stress, pathways that are already enhanced in the swarm. Our finding of heterogeneity in cell killing within the swarm is suggestive of altruistic features that ultimately increase survival of the swarm. Finally, we demonstrate that necrosignaling operates in multiple bacterial species and exhibits species specificity.

## Results

### Death of a subpopulation may be linked to SR

Initially, we monitored the antibiotic susceptibility of planktonic vs. swarm cells of E. coli. Exposure to a broad concentration range of the antibiotic kanamycin (Kan) produced distinct survival patterns (Supplementary Fig. 1). At low concentrations (e.g, Kan2.5; 2.5 μg/ml), and during the early phase of growth (0.5 h), swarm cells were surprisingly more susceptible to killing than planktonic cells, a trend maintained for up to Kan10 (Supplementary Fig. 1a, b). This trend reversed at both higher Kan concentrations as well as later times, where planktonic cells became more susceptible. We hypothesized that the higher initial susceptibility of swarm cells could be indicative of heterogeneity in the population, with a subpopulation that was more sensitive. This rationale was tested in silico by simulating the survival patterns of each subpopulation as a first-order decay where the survival “rate” is analogous to that population’s susceptibility to the antibiotic (Supplementary Fig. 1a and “Methods”). A homogeneous cell population with a singular response rate reasonably modeled the planktonic data (MSE range: 10−6–10−3), but failed to model the swarm data (MSE range: 10−3–10−2) (Fig. 1a). Conversely, when a heterogeneous population composed of two subpopulations differing in initial rates of death were introduced into the simulations, the resultant curves successfully fit the experimental data as determined by MSE scores 10–1000 times lower (MSE range: 10−7–10−5) compared to the homogeneous population. The fits between the two population models for both swarming and planktonic cells were determined to be statistically significant by the Wilcoxon rank sum test (Fig. 1a). Taken together, the data indicate the existence of two distinct subpopulations within the swarm, one more susceptible to the antibiotic. Given that the experimental MIC for Kan in swarm cells was almost double that of planktonic cells (Supplementary Fig. 1c), we hypothesized that the more susceptible population might serve as an early distress signal to elicit resistance in the swarm community.

### Dead cells release AcrA as a necrosignal

To directly test whether cell death acts as a signal, the border-crossing assay12 was primarily employed as diagrammed in Fig. 2a, where in a divided petri plate containing swarm media, only the right chamber has the antibiotic. To test if cell death enhances SR, pre-killed cells were applied to the right chamber, now containing antibiotic at a concentration higher than the swarm’s normal tolerance (Fig. 2b). WT E. coli inoculated on the left, swarm over the right chamber with Kan25 (not shown) but not with Kan50 (Fig. 1 b1). When E. coli cells killed by Kan250 were applied to the right chamber, the WT population could colonize Kan50 (Fig. 1 b2). Although cells killed with Kan promoted migration over Kan50, the enhanced resistance was independent of the killing method (Supplementary Fig. 2a), with the exception of heat (Supplementary Fig. 2b). The response to killed cells was sustained, in that the swarm retained its capacity for resistance even after exiting a zone of dead cells (Supplementary Fig. 3).

The heat-sensitive nature of the SR-factor (Supplementary Fig. 2b) suggested that it might be isolatable. To this end, cell extracts prepared from Kan250-treated cells were assayed, and showed activity in the supernatant fraction (Supplementary Fig. 2b). The activity was resistant to DNaseI and RNaseI (Supplementary Fig. 2c), but sensitive to protease (Fig. 1 b3). A 30% ammonium sulfate precipitate, when resuspended in buffer and applied as lines, promoted the swarm to track along these lines (Fig. 1 b4). We will henceforth refer to this active factor as the necrosignal, and its ability to promote SR as “necrosignaling”.

We found necrosignaling to be operative in other bacterial species as well (Fig. 1e, Supplementary Fig. 2d). However, except for E. coli and Salmonella, where killed cells of one species promoted a reciprocal response in the other (Fig. 1e, maroon areas), species-specificity of the response was evident in Bacillus subtilis, Pseudomonas aeruginosa, and Serratia marcescens (Fig. 1e, blue areas). Given that E. coli and Salmonella have an interchangeable response, we used both bacteria to purify and determine the common identity of the necrosignal (Supplementary Fig. 4). MS/MS analysis of the active fractions obtained after the final purification step yielded five common proteins (Supplementary Fig. 5; AcrA, UspE, BaeR, YhdC, and Crp). All subsequent experiments were performed with E. coli.

Deletion and overexpression analysis helped narrow down the necrosignal candidate (Fig. 1c, d). In both analyses, candidate strains were pre-killed and applied to the right chamber. Only ΔacrA abolished the enhanced resistance response to Kan50 (Fig. 1c; Δcrp is not included because this deletion severely represses swarming21). Conversely, when overexpressed, AcrA (pacrA) promoted swarming over Kan70, but so did BaeR (pbaeR) (Fig. 1d). BaeR is a positive regulator of acrAD operon22. Given that ΔbaeR did not abolish the response, AcrA is most likely the necrosignal. Purified AcrA (flanked by His- and FLAG-epitope tags; see “Methods” and Supplementary Fig. 4d), showed a concentration-dependent SR response, plateauing at Kan70 (Fig. 1f). The estimated number of AcrA molecules per cell required to elicit a response increased exponentially with increasing antibiotic (Fig. 1f, inset) reaching a plateau, suggesting specificity (nonspecific binding is generally linear23).

When purified AcrA was added to a planktonic culture treated with Kan near its MIC99 (see Supplementary Fig. 1c for MIC analysis), only a modest increase in survival was observed, unless cell density was increased (Fig. 3). Even at high planktonic cell density, swarm cells had a better survival response, indicating that high cell density is important but not sufficient to account for swarm resistance, and that the physiological state of the swarm might be an important contributor as well.

### Mechanism of AcrA necrosignaling

AcrA is the periplasmic component of the RND efflux pump AcrAB-TolC (and AcrAD-TolC)24. To elucidate whether the activity of this surprising candidate for the necrosignal was independent of the OM component TolC, we carried out both deletion and complementation analysis with tolC. Absence of TolC in the active swarm annulled the resistance response to AcrA applied on the right (Fig. 4 a1), while complementation with TolC restored it (Fig. 4 a2). However, absence of TolC in killed cells applied on the right did not affect the response outcome (Fig. 4 a3). Taken together, these results indicate that the necrosignaling activity of AcrA is dependent on the presence of TolC in the live cells, and suggest that AcrA released from dead cells may bind to TolC from the outside.

We next sought to identify residues critical for the response in both AcrA and TolC. For TolC, we took advantage of reports that TolC serves as a receptor for certain colicins25 and bacteriophages26,27, using residues important for phage binding26 as a guide for our mutagenesis (Fig. 4b). As shown in Fig. 4c, a subset of the residues tested were also essential for the SR response. For AcrA (Fig. 4b), deletion of 72 and 75 residues from its C- and N- terminus respectively, did not affect its activity, but perturbation of its helix-turn-helix (HTH) motif or nearby residues eliminated it (Fig. 4c), suggesting that the HTH region of AcrA is critical. (Primary data for results summarized in Fig. 4c are provided in Supplementary Fig. 6).

To test binding of AcrA to TolC exposed on the OM, we marked AcrA with Qdot705-labeled anti-FLAG antibody, and labeled the membrane with FM 1-43. Microscopy images show localization of AcrA to the membrane (Fig. 4d and Supplementary Fig. 7a). Deletion and complementation of tolC led to loss and restoration of AcrA binding, respectively (Fig. 4e and Supplementary Fig. 7b). Critical TolC residues as deduced from plate assays (Fig. 4c), were verified in the binding assay by the observation that tolCS257A but not tolCR55A supported AcrA binding (Fig. 4e and Supplementary Fig. 7b). The extracellular localization of AcrA was confirmed by trypsin digestion (Fig. 4e and Supplementary Fig. 7c; see Supplementary Fig. 7d for Qdot705 control). Taken together, both genetics and microscopy corroborate that AcrA binds to TolC externally to elicit necrosignaling.

### Mechanism of SR

To determine if necrosignaling impacts gene expression, RNA-seq analysis was performed. Expression of 566 genes, many of which ontologically classified to motility, membrane function, and energy metabolism, showed specific changes in swarm cells (Fig. 5a and Supplementary Fig. 8). Efflux pumps and transport functions were upregulated, with further increases in expression upon addition of antibiotics and of AcrA. Expression of OM porin genes were down in the swarm (Fig. 5a) and those of ROS catabolism genes were up (Fig. 5a). These results were verified in several ways.

The Nile Red assay for efflux28,29 confirmed a higher rate of efflux, stimulated by AcrA (Fig. 5b). The Alamar Blue assay for membrane permeability29,30 showed lower permeability, unaffected by AcrA (Fig. 5c). Next, we directly measured the intracellular concentration of antibiotics using two methods—spectrometry employing the fluorescent antibiotic Fleroxacin31, and disc-diffusion assay for the thermostable Kanamycin32. The concentration of both antibiotics was significantly lower in swarm cells compared to planktonic cells (Fig. 5d, e), consistent with the data in Fig. 5b, c.

The RNA-seq data were further validated by genetics. Overexpression of selected genes involved in ROS catabolism, iron transport, and efflux pumps increased SR, whereas overexpression of porins decreased it (Supplementary Fig. 9a, c). Deletions of many efflux pump components and their regulators also decreased SR (Supplementary Fig. 9b, d).

Taken together, the data show that E. coli swarms are preprogrammed to be less permeable, more prepared to efflux, and to tolerate antibiotic stress by upregulating ROS catabolism pathways. Antibiotic exposure and the resultant cell death and release of AcrA enhances these pathways further. AcrA also stimulates immediate efflux from TolC pumps. Figure 6 summarizes these results.

## Discussion

Bacterial swarms alter the expression of genes controlling both morphology and physiology10,33,34,35,36,37,38. We show here that E. coli swarms are intrinsically programmed to survive encounters with antibiotics by modulating at least three different pathways as proactive shields: efflux, ROS catabolism, and membrane permeability. The first two pathways are upregulated, while the third is downregulated. The upregulation of multiple efflux pump components has also been reported in Pseudomonas10,39 and Salmonella35. ROS stress is known to be generated by antibiotics40,41,42, high cell densities43,44, and increased energy utilization45, the latter two being hallmarks of swarms12,21,46. Densely packed and highly motile swarm cells utilize more energy46 via extensive utilization of TCA cycle21,46, which leads to intracellular ROS stress via Fenton reactions through agents such as FADH2 and menaquinone47. Superoxide dismutase (SodB) and catalases (KatE, KatG) (Fig. 5a) protect the cells from damage to the Fe–S clusters, which are co-factors/prosthetic groups of various proteins48. The importance of iron homeostasis during swarming10,21,35,36,49,50, essential for the regeneration Fe–S clusters of ROS catabolizing enzymes51, was corroborated in our study as seen by the upregulation of multiple genes involved in uptake (feoAB, fecABCDE, fhuABCD; see Supplementary Fig. 8c) and metabolism of iron (ahpC, rsxABCDE, sdhA; see Fig. 5a). A decrease in membrane permeability has also been reported for Salmonella and Proteus swarmers36,46.

Our study shows that despite encoding an intrinsic program for surviving antimicrobials, swarms deploy an added security measure by exploiting the death of a subpopulation (induced by antibiotic or other insults; Fig. 1b and Supplementary Fig. 2a) to communicate a state of emergency to the group to further activate efflux and ROS pathways. The novel mechanism by which cell death enhances tolerance to antibiotics is by the release of AcrA, a periplasmic constituent of TolC efflux pumps, which binds to TolC on live cells from the outside (Fig. 4d). AcrA interaction with TolC immediately stimulates efflux in the population (Fig. 5b) and upregulates the expression of a multitude of genes for a sustained SR response (Fig. 5a). How AcrA triggers these events, and whether AcrA is also the signal that induces an avoidance response to antibiotics in the swarm20, will be avenues of future research.

A comparison of AcrA from the bacterial species used in this study shows a high degree of sequence conservation (Supplementary Fig. 10a, b), the E. coli and Salmonella proteins being most closely related (Supplementary Fig. 10c), explaining their interchangeable necrosignaling response (Fig. 1e). Interestingly, TolC is absent in many bacteria including Bacillus. In contrast, AcrA is highly conserved across the Bacterial domain (Supplementary Fig. 11) suggesting it might have other binding partners, i.e., more necrosignaling modules may be found in nature. The existence of a specific signal-receptor survival module (AcrA:TolC) is expected to benefit its own species in a multispecies ecological community (Fig. 1e).

The surprising aspect of our findings is that a normal component of an RND drug efflux pump, AcrA, moonlights as a necrosignal directed at similar pumps. Another curious finding is the structural homology of AcrA to colicin E3 (Fig. 7a–c, RMSD = 0.85 Å), which interacts with BtuB (vitamin B12 receptor), a β-barrel OM protein similar to TolC52,53. We have used the colicin homology to suggest possible modes of AcrA binding to TolC (Fig. 7d, e). The best fit has striking similarities with colicin E3–BtuB interaction54, suggesting perhaps a case of molecular mimicry. While we do not know whether colicins and AcrA share a common evolutionary history, we speculate that such mimicry may have evolved in environmental niches where release of AcrA via colicin-induced cell death resulted in a competition for extracellular binding to TolC on live cells.

Death of a bacterial subpopulation for the overall survival of community and its resemblance to altruism, has been the subject of earlier studies19,55,56. While a variety of functions for cell lysis have been reported—recycling nutrients during bacterial starvation57, exchange of genetic material58, and activation of systems that deliver toxins to confer fitness in inter-species co-cultures59—the cell death response we report here has several features not shown in earlier studies. For example, the existence of a heterogeneous population with respect to antibiotic susceptibility (Fig. 1a and Supplementary Fig. 1a), and the advantage it confers, is akin to a bet-hedging survival strategy19. High cell densities provide another selective advantage given that a significant fraction of the swarm must be killed for SR to be observed (~1/2 at Kan25; Supplementary Fig. 1a), an observation that should be considered for clinical antibiotic regimen guidelines60 and in ecological studies involving microbial interactions61,62. A large motile population, that can continuously acquire new territory, and that has a strategy to overcome stressful environments and survive mass extinction events promoted by the event itself, will also increase its chances of propagating beneficial mutations such as antibiotic resistance. We note that a response to cell death that benefits the living is not confined to bacteria, but is widespread in nature, as seen in insects (necromones)63,64, fishes65, birds66, and mammals67, possibly even in eukaryotic tumors68.

## Methods

### Strains, plasmids, and growth conditions

Bacterial strains, plasmids, and primers used in this study are listed in Supplementary Table 1. The strains were propagated in LB (10 g/L tryptone, 5 g/L yeast extract, and 5 g/L NaCl) broth or on 1.5% Bacto agar plates for single colony isolation. Antibiotics for marker selection were added as follows: Kan (Kanamycin) 25 μg/ml, Amp (Ampicillin) 100 μg/ml, and Cam (Chloramphenicol) 30 μg/ml. Antibiotic concentrations are indicated throughout with superscripts.

Gene deletions were achieved by transferring the Kan deletion marker from donor strains in the Keio collection69 to E. coli MG1655 using P1 transduction70 (P1vir). For overexpression analysis, genes were expressed from ASKA library71 pCA24N plasmid, induced with 0.1 mM IPTG. For AcrA purification, the periplasmic portion of AcrA without its lipoprotein signal peptide72 was cloned in pTrc99a plasmid with an N- and C-terminal His and FLAG-epitope tags, respectively. The tolC gene was cloned in pTrc99a, and site-directed mutagenesis was performed as described73.

### Motility assays

All assays used LB as the nutrient medium. For swarm assays, plates were prepared with Bacto agar for all bacteria except for E. coli (which requires Eiken agar; Eiken Chem. Co. Japan) at the following agar concentrations: E. coli (0.5%), S. enterica and P. aeruginosa (0.6%), S. marcescens (0.8%), and B. subtilis (0.7%). For E. coli and S. enterica, 0.5% glucose was included. Poured swarm plates were dried overnight at room temperature prior to use. Plates were incubated at 30 °C for E. coli and S. marcescens, and at 37 °C for all others.

Border-crossing swarm assay12 is illustrated in Fig. 2. Four microliters of mid-log phase bacterial culture were inoculated in the left no-antibiotic chamber and allowed to dry by leaving the lid of the petri dish open for 15 min. The plates were then incubated at 30 °C or 37 °C for 24 and 16 h, respectively, which is the time it took for bacteria in control plates to colonize the entire no-antibiotic right chamber. All plates were photographed with a Canon Rebel XSI digital camera using the “bucket of light” as a light source74.

### Preparation of dead cells and extracts

A 10 ml culture of E. coli cells (0.6 OD600) was treated for 30 min with any one of the following antibiotics: Kan250, Gen50, Amp500, and Cip25. Cells were harvested by centrifugation (12000 × g, 3 min, 4 °C). The pellet was once washed with sterile HEPES buffer (1.2 mM CaCl2, 1.2 mM MgCl2, 2.4 mM K2HPO4, 20 mM HEPES, 115 mM NaCl, pH of 7.4) and resuspended in the same buffer. Efficiency of killing was monitored by CFU counts on LB agar. Fifty microliters of the dead (i.e., no detectable CFUs) cell suspension was used for the border-crossing assay. To prepare extracts from the killed cells, the protocol above was scaled up to a 50 ml culture, the final cell pellet resuspended in 1 ml of HEPES buffer, and the cell suspension lysed by passing 3 times through a SLM AMINCO French press (1200 psig units). The lysed extract was precipitated by centrifugation (12000 × g, 5 min, 4 °C), and the supernatant used for further purification.

Enzyme treatments of cell extracts: 100 µL were treated with 10 µl of one of the following for 30 min at 37 °C: DNaseI (Thermo, 10 µg/ml), RNase A (Thermo, 10 µg/ml), or Proteinase K (Sigma, 10 µg/ml).

### Identification of the necrosignaling factor

The border-crossing assay was used to detect activity at all steps.

1. Cell extract preparation: Four litre cultures (0.6 OD600) of E. coli/S. enterica cells were treated with Kan250 for 30 min. The cells were harvested by centrifugation (5000 × g, 10 min, 4 °C). The pellet was washed once with HEPES buffer (1.2 mM CaCl2, 1.2 mM MgCl2, 2.4 mM K2HPO4, 20 mM HEPES, 115 mM NaCl, one SIGMAFAST™ protease inhibitor tablet, pH of 7.4). The cell pellet (3.8 g) was resuspended in 15 ml of HEPES buffer and then lysed by six passages through a French Press (1200 psig units). The resultant extract was centrifuged (10000 × g for 10 min at 4 °C) to collect the supernatant (50–70 mg/ml protein). 2. Ammonium sulfate precipitation: the supernatant was treated with ammonium sulfate (AS) (30% for E. coli and 35% for S. enterica) overnight (O/N) in a cold room (4 °C), while keeping it in a tumbling mode using a Barnstread thermolyne LabquakeTM rotisserie shaker. The precipitate was collected by centrifugation (12000 × g for 30 min at 4 °C), the pellet was resuspended in 5 ml of HEPES buffer and dialysed (O/N) using HEPES as the exchange buffer (~5 mg/ml protein). Fifty microliters each of initial cell extract and the AS pellet were used to test for activity. 3. Q sepharose anion exchange chromatography: 3.5 ml of the AS fraction was loaded on a Q sepharose anion exchange chromatography column (10 cm × 0.8 cm2 = 8 ml), and after the flow-through was collected, the column was washed with HEPES buffer (1 column volume). The proteins were manually eluted (0.5 ml/min) using a 50–500 mM NaCl gradient in a total of fifteen-1 ml fractions (~2 column volumes). Fractions 5 and 6 showed activity for both strains. Their protein concentrations were ~1.5 mg/ml. 4. CM cation chromatography: the active fractions from step 3 were pooled and loaded onto a CM cation chromatographic column (7.5 cm × 0.8 cm2 = 6 ml). The flow-through was collected, followed up a 1-column volume wash with Phosphate-buffered saline (NaCl 137 mM, KCl 2.7 mM, Na2HPO4 10 mM, KH2PO4 1.8 mM, pH 7.2). The proteins were manually eluted (0.5 ml/min) with a NaCl gradient of 50–250 mM using ~2 column volumes (9 fractions × 1.5 ml each = 13.5 ml total). The first four fractions (50–125 mM NaCl) showed activity, and had protein concentrations in the range of 0.39–0.1 mg/ml.

### Mass spectrometric analysis

Active fractions from the CM cation step were used to perform ESI-MS in a Thermo Orbitrap Fusion hybrid mass spectrometer. The proteins were Trypsin digested on the column before Tandem mass spectra were extracted. All MS/MS samples were analyzed using Sequest (Thermo Fisher Scientific; version IseNode in Proteome Discoverer 1.4.1.14) and X! Tandem (CYCLONE 2010.12.01.1). Sequest and X! Tandem were searched with a fragment ion mass tolerance of 0.80 Da and a parent ion tolerance of 10.0 PPM. Carbamidomethyl of cysteine was specified in Sequest and X! Tandem as a fixed modification. Glu->pyro-Glu of the N-terminus, ammonia-loss of the N-terminus, Gln->pyro-Glu of the N-terminus and oxidation of methionine were specified in X! Tandem as variable modifications. Oxidation of methionine was specified in Sequest as a variable modification.

Scaffold (version Scaffold_4.8.4, Proteome Software Inc., Portland, OR) was used to validate MS/MS based peptide and protein identifications. Peptide identifications were accepted if they could be established at greater than 16.0% probability to achieve an FDR less than 1.0% by the Peptide Prophet algorithm75 with Scaffold delta-mass correction. Protein identifications were accepted if they could be established at greater than 95.0% probability to achieve an FDR less than 5.0% and contained at least four identified peptides. Protein probabilities were assigned by the Protein Prophet algorithm76. Proteins that contained similar peptides and could not be differentiated based on MS/MS analysis alone, were grouped for parsimony.

### Purification of AcrA

BL21(DE3) harboring the His-AcrA-FLAG construct described above (referred to as simply AcrA henceforth) was cultured overnight in LB containing ampicillin100. An overnight culture was diluted 1:100 into 1 l of fresh LB, and incubated at 37 °C till 0.6 OD600 was reached. Protein Expression was induced with the addition of 1 mM IPTG for 4 h. The cells were harvested by centrifugation, washed, resuspended in buffer [10 mM Tris-HCl (pH 8.0), 100 mM NaCl, one SIGMAFAST™ protease inhibitor tablet, 1 mg/ml lysozyme, 1 mM β-ME, 50 mM imidazole], and sonicated on ice. The supernatant was collected after centrifugation (12000 × g, 30 min, 4 °C). Protein was purified using HisTrap HP (1 ml) column (GE) according to the manufacturer’s protocol with slight modifications. A 5-step gradient of imidazole (100, 200, 300, 400, and 500 mM) was used to elute the bound protein manually (2 column volumes for each step) and 1 ml fractions were collected. The fractions were dialyzed against two changes of buffer A (10 mM Tris-HCl (pH 8.0), 100 mM NaCl, 1 mM EDTA) overnight. All fractions were active.

### RNA-seq analysis

E. coli cells were grown in 10 ml broth (planktonic) culture (0.6 OD600) in absence of antibiotic and harvested with or without the treatment of antibiotic (see Figure legends for specific experiments). E. coli swarm cells from the right chamber in the border-crossing assay (with or without antibiotics) were resuspended in 3 ml LB medium and harvested. The harvested cells were resuspended in 1 ml ice-chilled LB medium, kept on ice for 2 min, and pelleted by centrifugation (4 °C, 10000 × g, 2 min). Every experimental condition tested had two biological replicates. Total RNA was isolated from these samples using Qiagen RNeasy® Protect Bacteria Mini Kit and following the enzymatic lysis method. The RNA was then used for library preparation using NEBNext Ultra II Library Prep Kit and sequenced on an Illumina NextSeq 500 platform (SR 75) yielding a total of 267.1 million reads for 10 samples. Sequence quality was determined by FastQC v0.13 and MultiQC77. The raw files were processed (Cutadapt78), aligned (Bowtie279), mapped (Samtools80, Bedtools81) to reference genome (GenBank ID U00096.3), normalized and analyzed (DESeq282), and visualized (R studio, Microsoft Excel, and Graphpad). The gene enrichment analysis was performed by using DAVID83,84.

### Microscopy

The Kan marker from ΔtolC and ΔacrA strains were removed using a pCP20 based method85 to generate ΔtolCΔK and ΔacrAΔK respectively. To prepare the AcrA probe, DYKDDDDK Tag mouse monoclonal antibody (FG4R) (Thermo Fisher) was labeled with a quantum dot (Qdot 705) using a click chemistry kit (SiteClick™ Qdot™ 705 Antibody Labeling Kit, Thermo Fisher)86. 100 μl of either E. coli planktonic culture in LB medium (0.6 OD600, treated with Kan25 for 30 min) or swarm cells collected from border-crossing assay plates (Kan25, as described in RNA-seq section) were incubated with membrane dye FM1-43 (0.1 mg/ml, FilmTracer™ FM™ 1–43, Thermo Fisher) for 15 min at 37 °C in the dark. 100 ng of purified AcrA was added to these cell suspensions and incubated for 15 more min. The cells were then pelleted, washed, resuspended in 100 μl of 1× PBS (pH 7.4), and Qdot 705-labeled antibody was added (1:1000) for 15 min incubation at room temperature. The excess antibody was removed by washing the cells by PBS buffer. Cells were then visualized using a light microscope (BX53F; Olympus, Tokyo, Japan) and cellSens software (v1.6) with minimal adjustments and alignments in Adobe Photoshop. FM1–43 was visualized using a GFP filter (excitation wavelength, 460–480 nm; emission 495–540 nm). To reduce quenching and avoid bleed-through because of overlapping excitation or emission wavelengths, the Qdot705 was visualized using a bandpass filter AT-Qdot 705 filter (Chroma, 39018) that allows excitation wavelength of 400–450 nm (425 CWL) and emission wavelength of 685–725 nm (705 CWL). After incubation with AcrA for the microscopy experiments, a subset of the swarm cell sample (200 μl) was used to determine the localization of extracellularly added AcrA87 with modifications. The cells were pelleted with gentle centrifugation (1500 × g, 10 min, 4 °C) and resuspended in protease buffer (50 mM Tris, 7.5 mM CaCl2, pH 8.8). The suspension was divided into two equal halves. One half was lysed with sonication (three cycles of 30 s on and 30 s off) and supernatant was collected. The un-lysed cells and the supernatant of the lysed sample were incubated with trypsin (50 μg/ml, Thermo Fisher) for 30 min at RT. The protease was then inactivated with 0.1 M PMSF. These samples were used for western blot analysis with DnaK as cytoplasmic control using mouse monoclonal anti-DnaK antibody (abcam).

### Efflux pump assays

Planktonic and swarm cells were harvested as for RNA-seq analysis. Nile Red and Alamar Blue assays were performed as described29 with modifications. Each sample (1 ml) was centrifuged for 10 min at 5000 × g; the pellet was suspended in potassium phosphate buffer (PPB, 20 mM potassium phosphate, 1 mM MgCl2, pH 7.0) to obtain OD600 of 1.0.

Nile Red assay. Samples were incubated with 10 mM CCCP (prepared in 50% DMSO) for 30 min at RT followed by addition of 10 mM Nile Red (VWR) at 37 °C and shaken at 200 rpm for 30 min. These cells were then kept in room temperature for 15 min without shaking, harvested via centrifugation, and supernatant was carefully and completely discarded. The resultant pellet was resuspended in PPB to obtain OD600 of ~1.0; a 10× dilution of that suspension was transferred to a quartz cuvette to measure fluorescence (excitation at 552 and emission at 636 nm) using a QuantaMaster spectrofluorometer every 10 s for 100 s with intermittent manual stirring. The efflux of Nile Red was triggered by addition of 100 μl of 1 M glucose followed by measuring fluorescence every 10 s for 200 s. In a subset of samples, 100 ng of purified AcrA was added along with the glucose. The obtained absorbance (A636) was normalized with OD600.

Alamar Blue assay. 20 μl of Alamar Blue liquid reagent (Invitrogen) was added to 180 μl of cell suspension. The fluorescence of the samples was measured (excitation at 565 nm, emission at 590 nm) every 5 min for 30 min followed by every 10 min for 30 more min. In a subset of samples, 20 ng of purified AcrA was added before the addition of Alamar Blue.

Fleroxacin assay. Planktonic and swarm cells were harvested as before. A cell suspension of ~0.6 OD600 was prepared in 2 ml LB and each sample was divided into two equal halves. Appropriate amounts of Fleroxacin (VWR) were added to one half, and the cells were incubated for 30 min while shaking at 37 °C31. The other half was used as the untreated control (990 μl) and to calculate CFU (10 μl). The cells were pelleted, washed once, resuspended in 1 ml of 50 mM sodium phosphate buffer (pH 7.2), and sonicated to lyse (3 rounds of 30 s on and 30 s off). The supernatant was used to determine fluorescence at 442 nm with excitation at 282 nm. The actual Fleroxacin fluorescence was calculated by subtracting the fluorescence of the untreated sample from fluorescence of treated ones. The concentration of Fleroxacin was determined by comparing with a standard curve of known Fleroxacin concentrations.

Disc-diffusion assay. Swarm cells were collected from the right chamber of a border-crossing plate containing the indicated Kan concentrations in Fig. 5e, 2 h after the cells crossed the border, and planktonic cultures were incubated for 2 h in the same Kan concentrations. 100 μl of collected cells were kept at 95 °C for 15 min, chilled in 4 °C for 5 min, and centrifuged at 10,000 × g for 5 min. The supernatants were spotted on filter discs, dried for 2 h, and placed on LB plates. The supernatant was used for deposit on discs. Kan is a thermostable antibiotic32. The CFUs were estimated for the cultures used for cell extract preparation. The following formula was applied to calculate intracellular [Kan]:$$([ {\rm{Kan}} ]{\rm{estimated}}\,{\rm{from}}\,{\rm{diameter}}\,{\rm{of}}\,{\rm{disc}})/ ({{\rm{CFU}}\,{\rm{of}}\,{\rm{culture}}})$$.

### Killing curves

An O/N culture of E. coli was sub-cultured (1 ml) to obtain 0.6 OD600. The swarm cells were collected from a regular swarm plate using 1 ml of LB and finally diluted to 0.6 OD600. These cell suspensions were incubated with various concentrations of Kan for 2 h at 37 °C with shaking at 200 rpm. A fraction of cells (100 μl) were collected every 30 min and dilution plating was used for measuring CFU counts. To confirm CFU counts, another fraction of cells (100 μl) from some samples were stained with LIVE/DEAD Bac Light Bacterial Viability Kit (Thermo Fisher) for 15 min by adding 5 μl of SYTO9 and PI dye each. The cells were visualized using a fluorescent microscope (BX53F; Olympus, Tokyo, Japan) with GFP and RFP filters for SYTO9 and PI dyes respectively.

### External AcrA docking on TolC

AcrA and TolC structures were obtained from PDB88 (ID: 5NG5)(rcsb.org). One AcrA molecule was docked on to TolC tripartite complex using the HADDOCK2.2 web server89 (https://haddock.science.uu.nl/services/HADDOCK2.2/) with default parameters. The docking was guided by excluding the region of TolC that normally interacts with AcrA in the periplasm (Gln142-Thr152 and Ala360-Val372), but the whole AcrA structure was used. The resulting top four models had HADDOCK scores ranging from −172 to −139, z-scores from −2.5 to −0.4, and RMSD of 0.7 to 2.3 Å.

### AcrA sequence comparison and distribution

AcrA sequences from five organisms studied here were used for multiple sequence alignment using Clustal Omega90 algorithm and visualized by Jalview91. The sequences were obtained from NCBI protein database (https://www.ncbi.nlm.nih.gov/protein). A cladogram was generated by neighbor-joining method without distance corrections using Clustal Omega showing relatedness of these AcrA sequences.

E. coli AcrA and TolC sequences were used separately as inputs in the homology-based protein sequence alignment algorithm phmmer in HMMER92 web server (https://www.ebi.ac.uk/Tools/hmmer/) with a significant e value and hit cutoffs of 10−20 and 0.03, respectively. The hits were then visualized in HMMER taxonomy, and counts for each taxon in the Bacterial domain were used for generating a heat map (Supplementary Fig. 11).

### Modeling bacterial response to antibiotics

The percent survival of the bacterial population for a given antibiotic concentration as a function of time was modeled as a set of ordinary differential equations. Initially, rate of change in a specific bacterial population was modeled as simple first-order reaction given by Eqs. 13.

$$A\mathop{\longrightarrow}\limits^{{k_a}}D,$$
(1)
$$\frac{{{\mathrm{d}}A_n}}{{\mathrm{d}}t} \,=\, - k_a\left[ A \right]_n,$$
(2)
$$[A]_{n + 1} \,=\, [A]_n \,-\, k_a\left[ A \right]_n.$$
(3)

The population live cells, A, will convert into dead cells, D, with some rate of change, $$k_A$$. Here we have simplified all the processes that lead to changes in the bacterial population to a single term $$k_a$$ that ideally represents a convolution of both bacterial growth and death. Time-step updates to the population of A cells is given by Eq. 3. The percent survival for a single population of cells is given by Eq. 4.

$$S_n \,=\, [A]_n/\left[ A \right]_0.$$
(4)

The percent survival of cells, Sn, is given by dividing the remaining cells at any time-point n, [A]n, by the initial starting population, [A]0. A Gibbs sampler was built to search for suitable parameters for the rate of change for the population based on reducing the mean-square error (MSE) between the simulation and experiment over 100,000 iterative steps. Simulations for a homogenous population exhibited deviation between experiment and simulation results as measured by MSE for swarming cells (10−3–10−2) and significantly better results for planktonic cells (10−6–10−3) (Fig. 1a).

The failure of the simulation to reproduce the experiment was hypothesized to be the result of the simplicity of the starting assumption that there is one population of cells with a singular response rate to the antibiotics. A more complex approach was developed using a heterogeneous starting population composed of two subpopulations (A and B) that follow the same first-order kinetics shown in Eqs. 13 but are allowed to differ in rates of change and relative starting population to one another. The survival percentages were calculated using Eq. 5.

$$S_n \,=\, \frac{{[A]_n \,+\, [B]_n}}{{[A]_0 \,+\, [B]_0}}.$$
(5)

Using the Gibbs sampler to explore the independent rates for both subpopulations, the simulation was much better able to replicate the experimental results for swarming cells with an MSE range of 10−7–10−5 typically conferring a hundred to ten-thousand fold better fit over a homogenous simulation (Fig. 1a). The differences between the populations usually converged into a solution where cells of one population were 10–100 times faster at dying, and are referred to as the fast-dying (FD) cells. Typically, the slow-dying cells (SD) where twice as numerous in starting populations than the FD cells (50–70%). In addition, magnitude of the difference of rates of death, $$\left| {k_A \,-\, k_b} \right|$$, between the FD and SD cells decreased as a function of antibiotic concentration. In contrast, simulations on heterogeneous swimming cells showed a drastically lower rate of improvement compared to homogenous cells with MSE ranges ranging 10−7–10−3, typically indicating a 2–100-fold improvement in fit. The differences in MSE scores between the homogeneous and heterogeneous models are statistically significant for both swarming (p value: 6.5 × 10−4) and planktonic cells (p value: 4.5 × 10−3). Taken as aggregate, there is sufficient evidence that two populations of bacterial cells exist within the swarming populations that differ in initial concentrations and susceptibility to kanamycin.

### Gibbs sampler

Bacterial models where built using Eqs. (3) and (5) while determining the percent survival at discrete time steps for comparison to experimental data and MSE calculation. Each model had 2 × N variables that could be altered to determine a better fit, where N is the number of subpopulations. The variables for each subpopulation are the initial size of the population (A0, B0, C0, etc.) referred to as starting_size and the rate of change (ka, kb, kc, etc.) referred to as death_rate. The Gibbs sampler was initiated with equal starting_size and random death_rates chosen from a uniform distribution between 0.1 and 0.001 for each subpopulation. A resulting initial MSE for the model was calculated and referred to as old_MSE. To converge on a reasonable value for each variable and reduce the MSE, the following Gibbs-sampling algorithm was employed as follows for 100,000 iterations:

1. (1)

Choose a variable randomly. This variable is referred to as old_variable.

2. (2)

The chosen variable selects a proposed value for the variable from a normal distribution centered on the current value of the variable with a standard deviation one-tenth the size of the current value. This value will be referred to as new_variable.

3. (3)

Estimate the new_MSE using the new_variable instead of the old_variable. Unselected variables will remain the same.

4. (4)

If the new MSE is lower than the old MSE:

set old_variable = new_variable.

Set old_MSE = new_MSE.

Variables were chosen at random for each iteration, so as not to introduce an order selection bias to determining variable selections.

### Statistics and reproducibility

Unless mentioned otherwise, all the experiments were performed with three biological replicates.

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.