The emergence, maintenance, and demise of diversity in a spatially variable antibiotic regime

Abstract Antimicrobial resistance (AMR) is a growing global threat that, in the absence of new antibiotics, requires effective management of existing drugs. Here, we use experimental evolution of the opportunistic human pathogen Pseudomonas aeruginosa to explore how changing patterns of drug delivery modulates the spread of resistance in a population. Resistance evolves readily under both temporal and spatial variation in drug delivery and fixes rapidly under temporal, but not spatial, variation. Resistant and sensitive genotypes coexist in spatially varying conditions due to a resistance‐growth rate trade‐off which, when coupled to dispersal, generates negative frequency‐dependent selection and a quasi‐protected polymorphism. Coexistence is ultimately lost, however, because resistant types with improved growth rates in the absence of drug spread through the population. These results suggest that spatially variable drug prescriptions can delay but not prevent the spread of resistance and provide a striking example of how the emergence and eventual demise of biodiversity is underpinned by evolving fitness trade‐offs.

The effectiveness of antibiotic therapy to control infection is being steadily undermined by the combination of divestment in drug discovery, widespread occurrence of genetic resistance among microbes isolated from natural environments 1 , and continued evolution of resistant strains among all major pathogens 2 . Managing our existing arsenal of drugs to ensure they remain effective for as many people for as long as possible is therefore an urgent public health priority.
Beyond a blanket ban on prescriptions, there is little consensus on how to delay or prevent the emergence and spread of resistance. Using multiple drugs with distinct cellular and genetic targets is often suggested as the most effective treatment but sequential use can often select for multi-drug resistant strains 3 , while simultaneous delivery (as a drug cocktail) can be difficult for patients to tolerate 4 . An alternative approach is to exploit the use of drug-free sanctuaries or refuges in time or space to make it more difficult for selection to fix resistant strains. Sensitive strains that do not pay a fitness cost of resistance 5 are expected to outcompete resistant strains in drug sanctuaries, implying that resistance could be kept at manageably low levels if dispersal introduces sensitive strains at a sufficiently high rate [6][7][8] .
Sanctuaries have been used to good effect in managing resistance in agricultural systems 9,10 but their role in governing the spread of resistance in health settings remains unclear 11,12 . Theory suggests that the way sanctuaries are experienced in time and space can profoundly impact resistance evolution 7,11,13,14 . Periodic delivery of a drug in time generates regular cycles of strong antibiotic selection followed by periods of relaxed selection when the drug is either metabolized, excreted, or not in use 15,16 . Intermittent dosing or ward-wide use generates fluctuating selection that leads to the evolution of broadly adapted resistant types with high fitness in both the presence and absence of drug 17 . Drug delivery may also be spatially variable across wards in a hospital or among compartments (tissues or organs) within a host, generating divergent selection that can lead to the emergence of a single resistant generalist or, if selection is sufficiently strong relative to dispersal (for example, through the movement of patients and staff between hospital wards), the coexistence of more narrowly-adapted niche specialists that trade-off drug resistance with growth rate in the absence of drug 7,11,13 .
To evaluate the impact of drug sanctuaries on the emergence and spread of AMR, we tracked the evolution of resistance and fitness in twelve independently evolved, isogenic lines of P. aeruginosa strain PA14 propagated in environments that varied through time or space with sub-inhibitory concentrations of the commonly used fluoroquinolone antibiotic, ciprofloxacin. P. aeruginosa ranks among the top three drug resistant pathogens globally 18 and causes both acute infections of wounds and chronic infections of the respiratory tract, where it is a major source of morbidity and mortality in adults with cystic fibrosis (CF) 19,20 . We use drug concentrations that resemble that found in the sputum of CF patients undergoing fluoroquinolone treatment during exacerbation 21 . Controls were a permissive (PERM) environment consisting of drug free Luria-Bertrani (LB) medium and a constant selective environment (CONS) comprised of LB supplemented daily with sub-inhibitory concentrations (0.3 ug/ml) of ciprofloxacin sufficient to reduce population densities to 20% of that in the absence of drug. Temporally varying environments (TEMP) were constructed by transferring aliquots from each evolving population into either permissive (no drug) or selective (0.3ug/ml) conditions on alternating days. Spatially variable environments (SPAT) consisted of two subpopulations, or patches, one permissive and one selective, with equal volume aliquots from each mixed and redistributed into fresh drug-free and drug-supplemented media daily (see Methods and Supplementary Fig. S1). The SPAT treatment resembles a well-known model in population genetics in which genetic polymorphism can be maintained through negative frequency dependent selection, provided there is a strong fitness trade-off among genotypes across patches and population regulation occurs at the level of the local patch rather than the total population 11,[22][23][24] . In contrast to previous work 25-28 , we do not explicitly manipulate either fitness trade-offs or the manner of population regulation but instead allow these variables to evolve naturally over the course of the experiment.

Results and Discussion
After 20 days (~133 generations) of selection resistance failed to evolve in the absence of drug selection (  Table S2; mixed linear analysis of covariance between relative fitness (w) and day*resistance frequency, with random effects of individual isolate pair nested in population: slope = -0.41871, P < 0.0001). These results suggest that divergent selection leads to the stable coexistence of resistant and sensitive strains through negative frequency dependent selection, consistent with predictions from models for the maintenance of polymorphism in spatially variable environments [22][23][24] .
Despite the presence of strong negative frequency dependent selection, the polymorphism does not appear to be stable in our populations on evolutionary time scales.
Coexistence persists at day 40 in just three populations ( Fig. 2A, B, H) whereas four populations are fixed or nearly fixed for resistance ( Fig. 2C, D, I, J) and nearly lost in two (Fig. 2F, G). The fixation or loss of resistant strains could be due to stochastic variation generated by time-lagged frequency dependent selection 30 , perhaps associated with the daily serial transfer protocol we imposed in our experiment, or from high mutation supply rates generating complex dynamics among competing strains (clonal interference). To distinguish between these alternatives, we tracked the frequency of resistance between three independent pairs of resistant and sensitive isolates over time under spatially variable conditions (see Methods). In the absence of clonal interference we expect the frequency of resistance to tend towards an internal equilibrium without fluctuations, regardless of starting frequency. As expected, the frequency of resistant genotypes converges towards an equilibrium point between 0 and 1 within the first two days and remains relatively stable until day 8, after which frequencies diverge again, presumably as mutations are reintroduced into the population (Fig. 5). These results suggest that the fluctuations in the frequency of resistant and sensitive strains we observed in our original experiment are likely due to high mutation supply rates generating clonal interference (see also Maddemsetti et al. 2015 31 ), rather than stochastic effects associated with our transfer protocol.
High levels of clonal interference imply that there is a steady source of genetic variation on which selection can act, over and above the diversity supported by negative frequency dependent selection. If so, are these populations continuing to evolve or does negative frequency dependent selection act to prevent further evolution? To answer this question we focused attention on the how the trade-off between MIC and growth rate under drug-free conditions changed over the selection experiment. We find that the slope of this trade-off evolves to become shallower at day 40 than at day 20 (Supplementary Table S1; log(MIC)*day interaction: F = 0.1221 , P = 0.0382). Inspection of Figure 3A suggests that this effect is due to increases in the growth rate of resistant isolates in LB, rather than loss of resistance (Fig. 3A).  Table   S3; log(MIC)*day interaction: F = -0.0311, P = 0.6167). These results are consistent with the selection of second-site compensatory mutations that improve growth rate under drug-free conditions without compromising resistance 17,32-34 . As a consequence of this weakened tradeoff, the internal equilibrium frequency of resistant strains (i.e., resistance frequency when w = 1) increases on average from 59% at day 20 to 87% by day 40 (Fig. 4), implying that resistant genotypes are slowly spreading through most populations. A putatively stable polymorphism on ecological time scales can thus be readily undermined by selection which, in this case, leads to the evolution of broadly-adapted genotypes that are both resistant and have high fitness under drug-free conditions and the eventual loss of sensitive genotypes.

Conclusion
Taken together, our results provide a rather bleak prognosis for the use of drug sanctuaries to manage antimicrobial resistance. Perhaps not surprisingly, given the strong selection generated by antibiotics, drug sanctuaries by themselves cannot prevent the evolution of drug resistance. Nevertheless, the manner in which drug sanctuaries are experienced by a pathogen can impact the rate at which resistance spreads in a population: while temporal variation in drug sanctuaries does little to prevent the rise of resistance, spatially distributed sanctuaries can slow the rate at which resistance fixes in a population. In line with the predictions of evolutionary theory, the strong divergent selection imposed by spatial variation in drug delivery leads to the emergence of a genetic polymorphism between resistant and sensitive strains supported by negative frequency dependent selection. In contrast with classic models of selection in spatially variable environments, however, this polymorphism is quasistable in the long-term, being readily undermined by compensatory evolution that tends to generate generalist resistant strains with high fitness in the absence of drugs. In other words, even spatial variation in drug selection cannot prevent resistant strains from eventually replacing sensitive ones due to the evolution of drug resistant strains with reduced costs of resistance.
The implication here should be clear: in the absence of a continual pipeline of new drugs for treating infectious disease, the best we can do is slow the rate at which existing drugs lose efficacy. The use of drug sanctuaries in space may help, for a time, but natural selection will inevitably find a way to undermine our best strategies for preventing resistance.

Experimental evolution
A single colony of P. aeruginosa strain PA14 was grown overnight in Luria-Bertrani broth  Supplementary Fig. S1). Samples of the mixed population were frozen. Two populations from the SPAT treatment were excluded from final analyses due to contamination.

Minimum inhibitory concentration (MIC) and growth rate
We randomly selected eight colonies from each population at day 20, plus an additional eight colonies from the TEMP and SPAT treatments at day 40, by plating a sample of each population on agar and choosing those closest to an arbitrary point in the middle of the plate.
For each isolate, we assayed resistance as the MIC to ciprofloxacin by first reviving frozen cultures overnight in LB media and then inoculating 100ul of dense culture into 96-well plates containing LB supplemented with 0.0, 0.10, 0.20, 0.39, 0.78, 1.56, 3.12, 6.25, 12.5, 25, 50, or 100ug/mL ciprofloxacin, respectively, and incubated on an orbital shaker at 37 o C. Growth was scored by reading optical density (OD) at 600 nm after 48 hours. Log-transformed MICs were used for all analyses. Resistance is defined as an MIC exceeding 2ug/mL, equivalent to 4x the ancestral MIC (0.5ug/mL).
The growth of evolved SPAT isolates relative to that of PA14 was measured in LB with and without ciprofloxacin over 24 hours. In LB, 5uL of overnight culture was added to 195uL of LB and OD600 measured every 90 minutes. The same procedure was performed for ciprofloxacin growth assays, except we used 20uL of overnight culture with 180uL of LBciprofloxacin media (0.3ug/mL). Gen5 software (BioTek Instruments Inc., Winooski, VT) was used to estimate maximum growth rate during exponential phase for three replicates per isolate.
Growth rates are expressed relative to the mean growth rate of PA14 in LB when grown on the same 96-well assay plate.

Negative frequency dependent selection
Using reciprocal competitive invasion experiments, we estimated the strength of negative frequency dependent selection between resistant and sensitive isolates from all SPAT treatment populations that showed evidence of coexistence from the MIC assays. We first chose four random pairs of resistant and sensitive colonies from each population, then competed each pair by mixing pure culture samples from overnight cultures at ratios of 1:9, 1:1, and 9:1 by volume and allowing them to compete for two transfers (48 hours) under the same SPAT treatment transfer protocol as in the original selection experiment. We tracked the change in relative abundance of sensitive and resistant colonies by streaking 40 randomly selected colonies on 2ug/mL ciprofloxacin agar for both the initial (0 hrs) and final (48 hrs) populations.
Relative fitness, w, was estimated using the equation: where F initial and F final are the ratios of the frequency of resistant types to the frequency of sensitive types in the population, before and after competition, respectively. Doublings refers to the number of generations occurring between the initial and final measurements (~13 generations).

Dynamics of resistance
We assayed the frequency of resistance every four days for each replicate of the SPAT treatment by isolating 50 colonies at random on LB plates, streaking each colony on 2ug/mL ciprofloxacin agar, and visually inspecting each plate for growth after 24 hours at 37 o C.

Dynamics of invasion from rare
We