Antibiotic resistance evolves more quickly in monoculture than in obligate co-culture

First, we tested the rate at which antibiotic resistance evolved in monocultures of E. coli and S. enterica as well as obligate co-cultures of the two species. We established six replicate cultures of each monoculture, and the co-culture. Each culture was distributed along an antibiotic gradient of either ampicillin or rifampicin, with cells initially inoculated into each antibiotic concentration along the gradient. After 48 hours of growth, we transferred cells from all wells in the gradient in a 1/200 dilution to fresh medium in the same antibiotic concentration, as well as double that concentration (Fig 1A). The same antibiotic gradient and transfer regime was used for both monocultures and co-cultures. Replicates may ultimately end up experiencing different maximum antibiotic concentrations, but this is because of differences in evolved resistance allowing them to move faster up the gradient, rather than externally imposed differences in gradients. We transferred populations for 20 transfers (approximately 150 generations). At each transfer, we measured total population density (by OD600) and calculated MIC₉₀ based on the density in wells along each gradient. After the final transfer, three colonies of each species were isolated from each replicate population and their MIC₉₀ values measured to assess the correlation between specific mutations and growth rate, yield, and MIC. Due to contamination, we removed one replicate of rifampicin-evolved S. enterica in monoculture and conservatively only used data out to transfer 10 (75 generations) for the ampicillin-evolved lines. There was no significant difference in E. coli population size in monoculture and co-culture (p = 0.43), while in S. enterica the co-culture was 42% smaller than the monoculture (p < 0.001, S1 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic of experimental setup and expectations.

A. Monocultures of E. coli (blue) and S. enterica (yellow), as well as cross-feeding co-cultures (green), were distributed in six replicate populations along an antibiotic gradient in 96-well plates. The antibiotics tested included rifampicin and ampicillin, and the concentration of antibiotic increased twofold at each well. 96-well plates were incubated with shaking at 30°C for 48 hours, then cells were transferred to fresh medium and antibiotic in a new plate. The transfer regime was to transfer 1μL of culture to a fresh well containing the same concentration of antibiotic at which it had previously grown, and 1μL to a fresh well containing one concentration step higher antibiotic. At each passage, the OD600 of the plate was measured, as well as species-specific fluorescence (CFP for E. coli, YFP for S. enterica). B-C. Hypothesis on why time under selection may be sufficient to explain MIC differences between monocultures and co-cultures. B. In monoculture, each species is under selection at every time step, thus selecting for increasing resistance with each passage. C. In obligate co-culture, only the more antibiotic-sensitive species is under selection at a given time, and effective co-culture resistance requires an increase in MIC in both species. This leads to the slower rise in resistance in the co-culture.

https://doi.org/10.1371/journal.ppat.1008700.g001

We found that resistance evolved more quickly in monoculture populations than in co-cultures (Fig 2A and 2B). We used a mixed effects model to analyze the rate at which MIC doubled in each replicate (i.e. the response variable was log₂(MIC)). The model included a fixed effect of treatment, a fixed effect of time, and a fixed effect of the interaction between treatment and time (it also included a random effect of replicate to control for structure resulting from repeated measures, see Methods). If monocultures and co-cultures evolved resistance along the same trajectory, but only began at different points along that trajectory, then we would expect to observe a significant main effect of treatment on log₂(MIC), but not a significant interaction between treatment and time. A significant increase in the doubling rate unaffected by treatment would appear as a significant main effect of time but no interaction. Finally, a difference in the doubling rate of MIC by treatment would be apparent as a significant treatment:time interaction, regardless of either significant main effect. In rifampicin, growth in each monoculture was associated with a significantly higher increase in MIC per transfer than growth in co-culture (treatment:time interaction term p = 0.0328 for E. coli, treatment:time interaction term p = 0.0100 for S. enterica). In ampicillin, resistance in both species also rose more quickly in monoculture than co-culture, even over just ten transfers. The per-transfer increase in MIC of both monocultures was higher than that of co-cultures (treatment:time interaction term p = 0.007 for E. coli, treatment:time interaction term p = 0.0272 for S. enterica) (Fig 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

Resistance evolves more slowly in co-culture-evolved populations vs. monoculture-evolved populations. Six replicate populations each of monocultures and co-cultures were evolved along a rifampicin gradient (A-B) or an ampicillin gradient (C-D); gradients were identical for monocultures and co-cultures. Population MICs for each species (E. coli A, C; S. enterica B, D) were measured each transfer and the resulting MICs plotted. Statistical analysis was performed using a mixed-effects model with a randomized slope for each replicate within a culture type. The fitted slopes for each treatment are indicated by the dashed lines. P-values are for the interaction term between passage and culture type. Error bars represent the standard deviation of MIC among the six replicates.

https://doi.org/10.1371/journal.ppat.1008700.g002

Similar mechanisms of rifampicin resistance evolve in monoculture and obligate co-culture

To determine whether cross-feeding selected for different resistance mechanisms we sequenced resistant populations at the end of the experiments. We extracted genomic DNA from the well that grew at the highest concentration of antibiotic for each replicate population. For each resistant population we created a list of mutations and their frequencies, excluding any mutations also observed in antibiotic free controls (S1 Table and S2 Table). For further analysis, we focused on genes that acquired mutations in more than one replicate within a treatment, as parallel evolution is a signature of adaptation [29].

Under rifampicin pressure, the genes that acquired mutations in co-culture were a subset of those that acquired mutations in monoculture for both species (Fig 3A and 3B). The most clearly identifiable resistance-associated mutation was rpoB, a component of RNA polymerase and the most common mutational target for rifampicin resistance [30]. Mutations in this gene arose in both species, in four out of the six replicates each in monoculture and co-culture (Fig 3A) and were strongly tied to higher levels of resistance (Fig 3C and 3D). All rpoB mutations were nonsynonymous single base pair substitutions. Interestingly, E. coli monoculture lines all evolved different rpoB mutations, but all coculture E. coli evolved the same S574Y mutation (S1 Table). A mutation in a prophage tail-specific protein, prc, of E. coli was also repeatedly observed in both monoculture and co-culture (S2 Fig); all mutations were short (2-11bp) deletions (S1 Table). The overlap in mutations suggests that co-cultures and monocultures evolved rifampicin resistance along similar adaptive trajectories, albeit at different rates.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Resistance-associated mutations in rifampicin-resistant evolved populations A-B. Lists of mutations which arose in E. coli (A) or S. enterica (B) monocultures and co-cultures. The number in each box represents the number of independent replicates in which the putative resistance mutation was observed. Additional mutations that occurred in only one population, and details on the nature of these mutations, may be found in S1 Table. C-D. Rifampicin MICs for isolates with wild-type vs. mutant rpoB genes in E. coli (C) and S. enterica (D). Isolates were obtained from passage 20 populations by streaking onto selective medium and picking isolated colonies. Each data point represents the average MIC for three isolates obtained from a single population. For each species- culture type combination, there are six populations total, and the statistical comparisons represent MIC comparisons between populations with wild type vs. mutant alleles. Note that there are only five S. enterica monoculture populations due to possible contamination in one population. MIC₉₀ values for isolates were defined as the lowest concentration of antibiotic which decreased growth by greater than 90% by 48 hours at 30°C. Additionally, monoculture lines evolved more mutations than co-culture lines. Mutations in mdoG and mdoH were observed in both monocultures of E. coli and S. enterica under rifampicin selection, but at much lower frequencies in co-culture (S1 Table, S2 Fig). All but one of these mutations were insertions or deletions of 1-6bp causing frameshift mutations (S1 Table). Both mdoG and mdoH likely influence cell membrane permeability [31,32]. The mutations were not associated with any changes in monoculture or co-culture growth rates (S3 Fig). Taken together, the pattern of rifampicin resistance mutations suggests that populations were moving along the same evolutionary trajectory in monoculture and co-culture, but progressed further along that trajectory in monocultures.

https://doi.org/10.1371/journal.ppat.1008700.g003

Different mechanisms of ampicillin resistance arise in monocultures vs. obligate co-cultures

In contrast to our results from rifampicin, the mutational spectra that we obtained from sequencing ampicillin-resistant populations suggests different resistance mechanisms arose in monoculture vs. co-culture. We saw a greater variety of genes acquiring adaptive mutations across replicates in ampicillin relative to rifampicin treatments. Nevertheless, we observed distinct mutational signatures in monocultures and co-cultures.

Very few genes acquired mutations more than once in the E. coli replicates. Two co-culture replicates evolved unique mutations in proQ, a regulator of efflux pumps [33] (Fig 4A, S1 Table). In monoculture, a gene associated with stress response (rne) contained an identical mutation in two populations (S1 Table). Given the relative paucity of mutations there is at best only a weak signature of differential resistance mechanisms evolving in E. coli in monoculture and co-culture.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4.

Resistance- associated mutations in ampicillin-resistant evolved populations A-B. Lists of mutations which arose in E. coli (A) or S. enterica (B) monocultures and co-cultures. The number in each box represents the number of independent replicates in which the putative resistance mutation was observed. Additional mutations that occurred in only one population, and details on the nature of these mutations, may be found in S1 Table. C. Image of co-culture populations containing ftsI mutations on Petri plates with LB pH = 7 (left) and pH = 4.7 (right). Blue colonies are E. coli which metabolize X-gal to a blue color; white colonies are S. enterica. No white colonies were observed on LB at pH = 7. D. MICs of isolates from libraries containing ftsI mutations in pH = 4.7 medium. Isolates were obtained from passage 10 populations by streaking onto selective medium and picking isolated colonies. Each data point represents the average MIC of three isolates from a single population. For each species- culture type combination, there are six populations total, and the statistical comparisons represent MIC comparisons between populations with wild type vs. mutant alleles MIC₉₀ values for isolates were defined as the lowest concentration of antibiotic which decreased growth by greater than 90% by 48 hours at 30°C. Each point represents the average MIC of three isolates from a single population. P = 0.0603, Mann-Whitney U.

https://doi.org/10.1371/journal.ppat.1008700.g004

Compared to E. coli, there was a larger number of S. enterica genes that acquired mutations only in monoculture or in coculture, but not both. (Fig 4B, S1 Table). Mutations in efflux pump genes (acrB, ramR) and cell permeability (ompR) arose in S. enterica monoculture but not in co-culture. The ompF mutations in monoculture were caused by an IS10 insertion and the yoaE mutations by an intergenic inversion (S2 Table). None of these mutations were associated with a significant increase in resistance except for ramR, wherein mutants had a higher median MIC than wild-type isolates (p = 0.0491, S4 Fig). Conversely, mutations in ftsI, a penicillin-binding protein (PBP3), were only observed in co-culture. ftsI is thought to be an essential gene and is known to be a target site for β-lactam antibiotics [34]. Interestingly, the mutations we observed in ftsI in our whole-genome sequencing were a combination of point mutations (D534Y at 75% frequency in rMM158) and mutations which should have ablated gene function (+A at 100% frequency in rMM127 and Q142* in rMM158) (S1 Table). We found that we could not isolate S. enterica clones with these mutations from evolved populations spread on typical LB agar, except for rare isolates that also had a suppressor mutation in the same codon as the Q142* mutation site which eliminated the stop codon (S1 Table).

As ftsI is an essential gene [35], it was perhaps not surprising that isolating loss-of-function mutations was difficult. However, this then raised the question of how ftsI loss-of-function mutants could rise to high frequency in the co-cultures. Castanheira et al. recently demonstrated that the lethality associated with ftsI loss in S. enterica could be mitigated by growing S. enterica under acidic conditions (pH<5.8); under these conditions, a second PBP3, which they called PBP3_SAL, is expressed [35]. In light of this, we plated co-culture populations containing Q142* or +A mutations in ftsI on LB of pH 4.7. On acidic plates we were able to obtain S. enterica isolates at roughly equal frequencies to what we would expect in the population, and Sanger sequencing demonstrated that these isolates did contain the loss-of-function Q142* or +A mutations and no secondary repressor mutations. These isolates did not show detectable growth in isolation unless the growth medium was acidified. They were associated with increased MICs versus isolates with wild-type ftsI, though this difference was not statistically significant due to low power (Fig 4D, p = 0.0603). Perplexingly, the co-culture media in which the ftsI mutants evolved did not show detectible pH changes, suggesting some novel induction mechanism for PBP3_SAL in our co-cultures. Interestingly, ftsI mutant isolates had co-culture growth rates comparable to wild-type when paired with an E. coli ancestor (S5 Fig). This suggests that ftsI knockout mutations were non-viable in monoculture but conferred little cost in co-culture.

A model suggests that interdependency is sufficient to generate differences in the rate of evolution of resistance

To test the effect of interdependence on evolution of antibiotic resistance in the absence of species-specific biological details, we developed a simple model. The model is based upon the haploid Wright-Fisher model with selection. Each species was kept at a density of 1000 individuals per well (regardless of how many species were present in a well) and transferred every ten generations. Genotypes within a species had two alleles: one conferring a growth rate, which affected their reproductive success between generations, and one conferring antibiotic resistance, which determined whether they could grow at all in the well. Each generation mutations could occur which could alter either growth rate or MIC. The transfer regime mimicked our experiments and allowed populations to progress out along an antibiotic gradient as they evolved resistance. At the beginning of growth after a transfer, at least one genotype from each inter-dependent species must be resistant to the present antibiotic concentration for any growth to occur. This interdependency allowed us to model cross-feeding, wherein one species cannot grow without metabolites supplied by the other, but is also representative of other types of mutualism. Importantly, this model allowed us to control for population size and physiology between monocultures and co-cultures, which we were unable to do in our experimental setup. This model therefore allowed us to test whether interspecies interdependence was sufficient to drive differences in rate of resistance evolution.

We first examined the relationship between the number of interdependent species and the rate of evolution of antibiotic resistance. We found that antibiotic tolerance (the highest antibiotic concentration at which a species grew) evolved more slowly as more interdependent species were simulated (Fig 5A, linear mixed-effects model, t = -9.4, p < 1e-7). When the mutation rate was reduced in all populations, cross-feeding more strongly suppressed the rate at which resistance evolved (Fig 5B, one-way ANOVA, F(2,297) = 6.9, p = 0.0012). Cross-feeding also reduced the number of mutations that accumulated within a population (S6 Fig, F(2,102) = 20.98, p < 1e-7).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Simulation model of the evolution of antibiotic resistance in single-species vs. multi-species obligately dependent communities.

A. A simulated evolution experiment with 35 replicate populations shows that increasing the number of interdependent species in a consortium results in a slower increase in the average MIC over time. Error bars represent the standard deviation among replicates. B. Effect of lower mutation rate on resistance evolution in model. Lower mutation rates result in bigger differences in the relative tolerance of one-species versus two species systems. C. Distributions from which MIC-altering mutations were randomly pulled. The x-axis is relative to the maximum MIC change possible in one mutation. The “null” model used a uniform distribution. The “neg exp” used a truncated negative exponential distribution with rate parameter 2.3, which kept 90% of random numbers less than the max MIC change of 1.0. D. Mean MIC increase conferred per mutation as a function of the mutation distribution and the number of species.

https://doi.org/10.1371/journal.ppat.1008700.g005

We next tested how the distribution of mutations influenced the mechanisms of resistance. We simulated evolution when all mutations were equally likely versus when mutations which conferred a greater resistance were less common, using a negative exponential distribution (Fig 5C). Under a uniform distribution there was no difference in the average size of resistance mutations accumulated in monocultures and co-cultures (Fig 5D). However, when big-effect mutations were rare, populations in co-cultures acquired significantly smaller effect mutations than populations in monoculture (Fig 5D, two-way ANOVA, interaction between # of species and mutant distribution: F(1,111) = 5.087, p = 0.026; Tukey's HSD finds only neg. exp with 2 species is different from other treatments). Our results demonstrate that when the frequency of mutations is not uniform, populations in monoculture and co-culture will sample different sets of mutations.

Bacterial strains and media

The Escherichia coli and Salmonella enterica strains used in this study have been described previously [57]. Briefly, the E. coli str. K12 was the ΔmetB strain from the Keio collection [58] that was mated with an hfr line to reinsert the lac operon (29). The S. enterica LT2 mutant was selected and engineered to excrete methionine [59]. Each strain is fluorescently labelled with a genomic integration of a fluorophore; E. coli is labelled with cyan fluorescent protein (CFP) and S. enterica is labelled with yellow fluorescent protein (YFP). Bacteria were grown in minimal Hypho medium containing phosphate and varying amounts and types of carbon and nitrogen as previously described [57]. Co-culture media contain 2.78mM of lactose, E. coli monoculture medium 2.78mM of lactose and 0.020mM of methionine, and S. enterica medium contained 5.55mM of glucose.

Experimental evolution

Three different culture conditions (E. coli monoculture, S. enterica monoculture, and obligate co-culture) were each evolved across two antibiotic gradients (rifampicin, Chem-Impex International Inc. 00260, and ampicillin, Fisher BP1760). Six replicate populations were evolved for each antibiotic- culture condition combination. Each antibiotic gradient began with an antibiotic-free control well and increased twofold with each subsequent well, starting at 0.25μg/mL and ending at 64μg/mL. Stocks of antibiotics were prepared at the beginning of the experiment such that 2μL could be added to 198μL of fresh growth medium and cells to produce the desired antibiotic gradient. Monocultures and co-cultures were exposed to the same antibiotic gradients throughout the experiment. When populations evolved to grow at 64μg/mL, the upper end of the gradient was increased to 1024μg/mL and the lowest concentrations were removed. Fresh antibiotic stocks were prepared and filter-sterilized immediately before the experiment began.

Initial bacterial cultures were inoculated from DMSO freezer stocks in 10mL of monoculture minimal Hypho medium for approximately 48 hours at 30°C, to stationary phase (OD~0.4). Cells were then distributed into 96 well cell culture plates. For monocultures, 2μl of bacterial cells (~2x10⁵ cells) were inoculated into 196μl of the appropriate monoculture medium with 2μL of antibiotic stock that would confer the correct gradient concentration. In co-cultures plates, 1μl of E. coli and 1μl of S. enterica were inoculated into 196μL medium. By inoculating 2μL cells total into the growth medium for both monocultures and co-cultures, we were able to maintain starting cell densities in both cases and avoid inoculation effects. The plates were then incubated for 48 hours at 30°C with shaking at 450 rpm. After each growth phase, cells were then transferred to a new 96-well plate with fresh media. 1μL culture from wells of each antibiotic concentration were transferred to a new well with the same antibiotic concentration, and 1μL cells were transferred to a new well that was one step higher on the antibiotic gradient (see Fig 1). This transfer occurred whether or not there was visible growth. This regime challenged bacterial populations with double the antibiotic concentration that they were exposed to at the previous transfer, but also allowed for populations to be maintained throughout the experiment if resistance was not acquired during a given transfer. Lack of adaptation within a transfer did occur over the course of the experiment, as evidenced by any point where MIC did not increase between transfers. Additionally, by consistently inoculating 2μl total of cells into fresh media, we were able to avoid any inoculum effects [44,60] that might have altered antibiotic tolerance in a given passage. The 96-well plate from each growth phase was frozen down in 10% DMSO for future analysis. After each 48-hour growth period, each plate was also placed onto a Tecan InfinitePro 200 plate reader where OD600 and species-specific fluorescence measurements were obtained. These readings were used to calculate the 90% minimum inhibitory concentration (MIC₉₀) for each replicate; growth at a given antibiotic concentration was confirmed if the well OD600 was above 10% of the OD600 of the antibiotic-free control well for a given replicate. Statistical analysis of rate of resistance evolution between monocultures and obligate co-cultures for each antibiotic was performed in R. A linear mixed- effects model with a random slope for each replicate within a treatment group was used to test the relationship between MIC and transfer, culture type (monoculture or co-culture), and a transfer-culture type interaction.

Sequencing

To identify mutations which conferred antibiotic resistance in each evolved population, the most resistant population of each replicate in each antibiotic-growth condition combination was whole-genome sequenced. Two antibiotic-free populations per antibiotic-growth condition combination were also sequenced to identify any mutations which may have arisen during passaging but are not related to antibiotic resistance. Each population to be sequenced was scraped from 10% DMSO freezer stock and grown up in 10mL Hypho at the appropriate antibiotic concentration for 48 hours at 30°C. gDNA was then extracted from each population using Zymo Quick-gDNA Miniprep Kit (11-317C). The gDNA was then used to prepare Illumina sequencing libraries according to the Nextera XT DNA Library Prep Kit protocol. Libraries were submitted to the University of Minnesota Genomics Center for QC analysis and sequenced on an Illumina Hi-Seq with 125bp paired-end reads.

Sequence analysis was performed using BreSeq [61] to align Illumina reads to reference E. coli and S. enterica genomes as previously described [62]. Briefly, mutation lists for resistant populations were filtered such that variation between our ancestral strains and the reference genome were removed, as well as any mutations which also arose in the antibiotic-free populations. Mutation lists were then assembled for each population (S1 Table) and any known functions described (S2 Table). Mutations which rose to above 50% in frequency were examined further for their possible role in conferring antibiotic resistance (S1 Table).

Isolation and phenotyping of isolates

At the final transfer of each evolution experiment, the well in each replicate displaying growth at the highest antibiotic concentration, according the MIC₉₀, was identified as the ‘resistant population’. These resistant populations, as well as the corresponding antibiotic-free control populations from each replicate, were plated onto minimal Hypho medium containing X-Gal (5-Bromo-4-chloro-3-indolyl β-D-galactopyranoside, Teknova X1220), which allowed us to differentiate between β-galactosidase-positive E. coli and β-galactosidase-negative S. enterica through a colorimetric change. Three isolates of each species each from the resistant and antibiotic-free wells were selected from these plates and frozen down in 10% DMSO for further analysis. We did not genotype individual isolates, as the frequencies of most mutations of interest were sufficiently high that the isolate genotypes should reflect that of the population (S1 Table). The same three isolates from each population were then used to assess growth rate, yield, and MIC differences associated with specific population-level mutations.

For ampicillin-evolved populations in which we were unable to obtain S. enterica isolates on pH neutral medium, we prepared acidic Hypho plates by adding 6M HCl to the growth medium until the pH reached ~4.7. Growth medium was then autoclaved and X-gal was added as described above. For growth of these isolates in liquid medium, 6M HCl was added to autoclaved Hypho medium to a pH of ~4.7.

Evolved isolates were scraped from DMSO freezer stocks onto solid agar plates and grown up overnight, then inoculated into 96- well plates containing the appropriate species-specific Hypho and grown at 30°C for 48 hours at 450rpm to acclimatize them to minimal media growth and standardize cell density. 2μL of stationary-phase cells was then used to inoculate a new 96-well plate containing 198μL monoculture growth medium. These plates were then placed in a Tecan InfinitePro 200 for 48 hours at 30°C with shaking at 300rpm; OD600, CFP, and YFP were measured every 20 minutes. After 48 hours, 1μL from the wells of monoculture isolate plates were transferred to new 96-well plates. These plates contained 198μL of fresh co-culture Hypho medium and 1μL of either ancestor strain of E. coli or S. enterica. If the monoculture isolate was E. coli, the ancestor strain of S. enterica was added, and vice-versa. The plates were then placed into a Tecan InfinitePro 200 for 48 hours at 30°C, and OD600 and florescent data was obtained every 20 minutes. Growth rates and yields were then calculated using Baranyi curves in R software using an in-house script. Statistical analyses and graphs were prepared in Stata version 14.

Evolved isolates were prepared for inoculation as described above for beginning monoculture growth rate experiments. Based on the MIC₉₀ of the population from which they originated, 2μL of each isolate culture was then inoculated across an antibiotic gradient such that the antibiotic concentration equaling the population MIC₉₀ was in the middle of the gradient. An antibiotic-free control well was also included for each gradient. If the gradient used was insufficient to calculate MIC (e.g. cells grew at all concentrations, or at none of them), the experiment was repeated with the antibiotic gradient shifted up or down as necessary. MIC of isolates was determined using MIC₉₀ as described above. The average MIC of isolates from each population can be found in S7A Fig (rifampicin-evolved populations) and S7B Fig (ampicillin-evolved populations). Statistical analyses and graphs were prepared in Stata version 14.

Evolution model

We used an evolution simulation to examine the rate at which antibiotic resistance evolved as a function of the number of interdependent species and other variables. There were two time scales in this model. The “within-transfer” time scale, and the “between-transfer” time scale. Within a transfer, evolution was simulated in any given “well” using a modified haploid Wright-Fisher simulation with selection. A well was initiated with N individuals (default N = 1000) of each species. If this was the first transfer, these individuals were clones of a genotype with growth rate (s) = 1 and MIC = 0. The well had a pre-determined antibiotic concentration. Evolution occurred over a predetermined number of generations. Each generation, the population was fully replaced. The new population was picked from the previous generation depending on the frequency of each genotype and its growth rate. The unscaled fitness of individual i was determined by . This was scaled to get expected frequency . In other words, . The next generation was then created by generating N random numbers from a uniform distribution on [0,1] and choosing genotypes from the previous generation based on their W_j. Prior to calculating w, each genotype’s MIC was compared to the antibiotic concentration in the well. If the genotype’s MIC was less than the antibiotic concentration, its s = 0 for the fitness calculation.

Once the new population was generated, some of the new individuals may be mutated. Each simulation used a pre-determined mutation rate u (default = 0.001 mutations per individual per generation). We generated N random numbers from a uniform distribution on [0,1]. Any random numbers less than 0.5 meant those individuals gained a mutation. In all cases, mutations could be either a growth rate mutation or an MIC mutation with a 50:50 chance. Growth rate mutations were simulated by adding a random number from a normal distribution with mean 0 and standard deviation 0.1 to the current growth rate (i.e. deleterious growth rate mutations were equally as likely as advantageous). MIC mutations altered the MIC of the individual, as described in the results. We tracked the genealogy of all individuals.

To simulate interdependence, we simulated >1 species per well. Each species had N individuals, and the dynamics of each species were independent of the other species, except that there must be individuals of all species with an MIC greater than the antibiotic concentration for any species to grow.

The second time scale—the transfer scale—was simulated to approximate the wet-lab experiments. Individuals were transferred from one well to two wells: the same antibiotic concentration, and the next higher antibiotic concentration. Therefore, wells received individuals from two wells. If there were surviving species in both source wells, N/2 individuals were randomly chosen from each well to populate the new well. If there were surviving species in only one source well, all N individuals were transferred from that well to the new well. If no source wells had surviving species, the destination well was sterile. Simulations were conducted in R.

Statistical analyses of the simulation data: We analyzed the rate of evolution of antibiotic tolerance as was done in the wet-lab experiments. A linear mixed- effects model with a random slope for each replicate within a treatment group was used to test the relationship between the response variable MIC and explanatory variables transfer, culture type (monoculture or co-culture), and a transfer-culture type interaction. A one-way analysis of variance tested the amount of mutations observed in the most-tolerant population, with an independent variable of the # of interdependent species. To calculate the relative tolerance of two-species simulations versus one-species simulations (Fig 5C), we first found the mean and standard error in the tolerance in each set of simulations. Then, the mean of the two-species community was divided by the mean of the one-species community. The error was propagated using (mean_relative_tolerance)*sqrt((SEM[two species]/mean_tolerance[two_species])^2 + (SEM[one species]/mean_tolerance[one_species])^2). The average MIC mutation size (Fig 5D) was the average MIC in a surviving population divided by the average number of mutations in that population. In all cases, since species were functionally equivalent, we arbitrarily chose one species to assess. Statistics were performed in R. Example code to run simulations is available as supplementary files (S1 Code).