Drug-mediated metabolic tipping between antibiotic resistant states in a mixed-species community

Abstract

Microbes rarely exist in isolation, rather, they form intricate multi-species communities that colonize our bodies and inserted medical devices. However, the efficacy of antimicrobials is measured in clinical laboratories exclusively using microbial monocultures. Here, to determine how multi-species interactions mediate selection for resistance during antibiotic treatment, particularly following drug withdrawal, we study a laboratory community consisting of two microbial pathogens. Single-species dose responses are a poor predictor of community dynamics during treatment so, to better understand those dynamics, we introduce the concept of a dose-response mosaic, a multi-dimensional map that indicates how species’ abundance is affected by changes in abiotic conditions. We study the dose-response mosaic of a two-species community with a ‘Gene × Gene × Environment × Environment’ ecological interaction whereby Candida glabrata, which is resistant to the antifungal drug fluconazole, competes for survival with Candida albicans, which is susceptible to fluconazole. The mosaic comprises several zones that delineate abiotic conditions where each species dominates. Zones are separated by loci of bifurcations and tipping points that identify what environmental changes can trigger the loss of either species. Observations of the laboratory communities corroborated theory, showing that changes in both antibiotic concentration and nutrient availability can push populations beyond tipping points, thus creating irreversible shifts in community composition from drug-sensitive to drug-resistant species. This has an important consequence: resistant species can increase in frequency even if an antibiotic is withdrawn because, unwittingly, a tipping point was passed during treatment.

Antimicrobial resistance poses a formidable challenge for medicine with resistance to all but the most recently discovered antibiotics encountered in clinical and agricultural practice¹. Seeking behavioural changes in antibiotic prescription to control resistance is a field of active theoretical, laboratory and clinical research. Importantly, it has been mooted that resistance could be eliminated using evolution-aware strategies that reverse the arrow of time². But is drug resistance reversible? And if not, why not?

Antibiotic cycling, whereby different antibiotics are prioritized and restricted through time, can lead to the reversal of resistance if resistant microbes pay the price for their abilities to resist by having reduced fitness when drugs are not around^3,4. This idea has been tested clinically, with mixed outcomes. Restricting use can reduce resistance^5,6 though not always⁷ and, perversely, increases in resistance have been observed following drug restrictions^6,8. Thus clinical strategies that cycle antibiotics have unpredictable effects: they can work^9–13 but sometimes they fail^14,15. It is unclear why a self-evidently worthwhile strategy of antibiotic withdrawal would not reduce resistance. An absence of fitness costs of resistance¹⁶ is one potential explanation but, in microbial communities, as we now explain, there is another.

Our explanation is this. For simplicity, imagine a microbial community dominated by two species, S and R. Assume the former is sensitive to an antimicrobial and the latter is resistant. Suppose S can invade, and displace, R in the absence of drug and R can invade and displace S in the presence of drug; in this case the drug-resistance phenotype of the community is reversible. However, if, now in a different community, R invades and displaces S in the presence of drug but, in the drug’s absence, the community exhibits a frequency-dependent bistability¹⁷ whereby either R or S can dominate, then this community need not have reversible resistance. Here, application of drug forces R to become dominant and the inability of S to always re-invade following withdrawal could cause resistance not to reverse. Now a tipping point is said to occur when S can no longer invade and we provide theoretical mechanisms and microbial data demonstrating how the irreversibility of resistance can arise through tipping.

Metagenomic analyses are rapidly improving our understanding of microbial communities. We now know that antibiotics affect communities in load¹⁸ and in diversity¹⁹ and, intriguingly, the removal of antibiotic sometimes^20,21 but not always^22,23 restores the community to its original, pre-treated composition. However, selection for resistance within communities is poorly understood because key pharmacological indicators, such as the minimal inhibitory concentration (MIC), dose responses, between- antibiotic drug interactions and costs of resistance are measured in single-species assays. These assays ignore the antibiotic’s true context: while microbes can exist as single-species populations, in bloodstream infections, say, most real-world microbes thrive in communities. Why, therefore, should a single-species understanding of microbial responses to antibiotics completely explain resistance progression on the skin, in the gut or a hospital ward?

To support this view, here we show that single-species resistance measures can be poor predictors of resistance in a synthetic microbial community both during treatment, and after antibiotic withdrawal, because of a hitherto unobserved phenomenon: communities can have tipping points when abiotic parameters such as treatment duration, antibiotic dose and nutrient availability vary. For example, clinicians vary dosing regimens^24–26 and treatment duration²⁷ of critically ill patients while nutrient availability, in the form of glucose concentration, can vary from 0.01 to 0.28% in urine²⁸ to 0.1 to 2.7% in blood with substantial daily variation^29,30. The impact of this on drug resistance is unknown.

We explore resistance and abiotic variation in the simplest possible community of two species, Candida albicans and Candida glabrata. Both are commensal microbes found together in the microbiota of healthy individuals but they are also opportunistic pathogens causing mucosal infections³¹ and life-threatening disseminated infections among immunocompromised patients³². Difficult to diagnose, Candida infections are associated with high mortality rates, ranging from 46 to 75% for candidiasis in the bloodstream^33–35. Strikingly, as many people die each year from the top ten invasive fungal diseases, including candidiasis, as do from tuberculosis or malaria³³. Apart from its substantial impact on human health, our community treatment model is suited to studying drug withdrawal dynamics because C. glabrata infections are relatively unresponsive to the most frequently used antifungal drug fluconazole, while C. albicans is sensitive to fluconazole³⁶. Monoculture dose-response assays demonstrate this for our study strains at clinical doses (Supplementary Fig. 2).

We determined temporal dynamics of the Candida species empirically by monitoring their relative frequencies. For this we co-cultured both in shaken, 96-well plates containing liquid growth media supplemented with glucose and fluconazole. The plates were inoculated with a mixture of fluorescently labelled C. albicans, at proportion f, and C. glabrata, at proportion 1 – f. After 24 h of growth (also known as one season) densities and frequencies of each species were determined using flow cytometry and a fixed volume sample (3.3%) of the community was transferred to a new 96-well plate containing fresh media, marking the beginning of a new season (Methods).

Applying single-species logic to this community, the resistant species, C. glabrata, should dominate in the presence of enough drug. Indeed, there is evidence this is predictive of clinical outcomes: the use of fluconazole prophylaxis was found to influence the proportion of C. albicans and C. glabrata isolated from the blood of patients with candidemia³⁷, leading to an increase in C. glabrata frequency. Fluconazole withdrawal should then shift the community towards the sensitive species, C. albicans. Importantly, we can replicate both these observations using our community: under fluconazole treatment, the drug resistant C. glabrata dominates and when the drug is removed, the sensitive C. albicans subsequently recovers. This creates a repeatable, cyclical dynamic as the drug is repeatedly applied and withdrawn (Fig.1a).

Fig. 1. C. albicans and C. glabrata dynamics.

a, Lab data: C. albicans and C. glabrata are inoculated at 50–50 proportions into growth in SC media supplemented with 0.1% glucose and propagated in the presence of fluconazole (Flu) for two seasons and the C. albicans frequency subsequently decreases. After two seasons, fluconazole is withdrawn and C. albicans recovers. When fluconazole is later re-applied for three seasons, C. albicans again decreases in frequency, and so the cycle repeats. Grey boxes mark seasons undergoing fluconazole treatment. Data are mean ± 95% confidence interval, n = 3, raw data shown. b, Lab data: the initial C. albicans frequency (f) on the x axis versus the final frequency each season obtained using the laboratory microcosm on the y axis, also known as Φ(f). Grey line marks the diagonal where f = Φ(f) and the solid red circle denotes the intersection of the diagonal and the data interpolant. Four exemplars are shown. Data are mean ± 95% confidence interval, n = 3; glucose (Glu) and fluconazole (Flu) given in the legend. c, Theoretical example: how to read Φ(f) to understand dynamics. Left: starting at timepoint 0, the dynamics follow Φ_on while treatment proceeds, it then follows Φ_off when treatment stops. The sequence of treatments here is: on, on, on, on, off, on, off, on, on, off. Numbers represent seasons and grey line marks the diagonal where f = Φ(f). Solid dots represent instances where drug is not used while open circles represent instances when drug is used. Right: how C. albicans reversibly increases and decreases in frequency according to whether drug is used (open circles) or not (solid dots). Open circles and solid dots are colour coded according to C. albicans frequency for ease of linking with Φ(f) dynamics (left).

But is this the only dynamic possible following fluzonazole withdrawal? To answer this, we now systematically explore community dynamics under different abiotic conditions by applying ideas from microbial population biology³⁸. Recalling C. albicans are inoculated into the microcosm at proportion f, suppose F denotes the frequency of C. albicans after one season, so R(f) = F/f denotes the change in C. albicans relative frequency. Now, F = f×R(f) is a ‘single-season frequency change map’ that gives the frequency of C. albicans after one season and we will write F as a mathematical function, calling it Φ, thus Φ (f) = f × R (f). Repeated applications of Φ to frequency values can therefore be used to determine the C. albicans frequency after any number, n ≥ 1, of seasons, provided a given initial (inoculum) frequency f₀ is known. So, f₁ = Φ (f₀) is the frequency of C. albicans after one season, f₂ = Φ(Φ(f₀)) = Φ(f₁) is the frequency of C. albicans after two seasons, f₃ = Φ(Φ(Φ(f₀))) = Φ(f₂), and so on. In population dynamics theory, it is common to write season number as a subscript n, so f_n+1 = Φ (f_n) is a shorthand representation of the season-by-season dynamics at season n. It follows by definition that Φ(0) = 0 because with no C. albicans in the inoculum, it cannot appear subsequently. Similarly, Φ(1) = 1 must hold as a closed community containing only C. glabrata initially must always do so (see Supplementary Section 2 for details).

What should a biologically reasonable Φ look like? Figure 1b shows four lab-derived exemplars and we also use a bottom-up mathematical model that builds theoretical Φ functions (Fig. 1c and Supplementary Section 3). The latter can incorporate many microbial life history traits and environmental variables but here we focus on antimicrobials (at concentration a) and extracellular nutrient concentrations, say g, denoting the carbon source glucose. We restrict attention for now to just a, ignoring g dependence, and write Φ(f, a) to emphasize this.

We now ask how the community responds to an antimicrobial by introducing a sequence of antimicrobial dosages, a_n, so that f_n+1 = Φ (f_n, a_n) where the treatment can change with each season. In the clinic, a_n might be one of two extremes, either a high dosage above the drug’s MIC or zero when treatment stops. Figure 1c shows two theoretically constructed Φ functions, Φ_on and Φ_off, motivated by this clinical context:

$$f_{n+1}=[\begin{array}{c}{\Phi }_{\text{on}}\left(f_n\right):\text{if}\text{antimicrobial}\text{is}\text{applied}\text{(so}{a}_n\ge \text{MIC)}\\ {\Phi }_{\text{off}}\left(f_n\right):\text{if}\text{antimicrobial}\text{is}\text{not}\text{applied}\text{(so}{a}_n=0)\end{array}$$

Empirical Φ functions (Fig. 1b) strongly resemble their theoretical counterparts (Fig. 1c) and, in these figures, both theory and data exhibit reversible resistance.

However, theory-derived Φ_on and Φ_off provide information about when not to expect reversible resistance (Fig. 2a–d) and the shape of these two functions is all-important. If Φ_on satisfies Φ_on(f) – f for all f between 0 and 1, then f_n+1 = Φ_on(f_n) < f_n follows, meaning the frequency of C. albicans decreases each season when drug is applied (Fig. 2a). Conversely, if Φ_off(f) > f for all f then C. albicans increases each season after treatment withdrawal, whence resistance is reversible (Fig. 2b). Points of separation, or separatrixes, between these two cases arise when there are frequencies, f, for which Φ_off(f) = f (Fig. 2c,d). As a consequence, for our purposes, a tipping point, f_u, satisfies Φ_off(f_u) = f_u and other technicalities (Supplementary Sections 2 and 4). This condition allows either C. albicans or C. glabrata to dominate in the absence of drug and, as a result, drug treatment can coerce the microcosm towards either possibility when treatment stops (Supplementary Section 4).

Fig. 2. Population dynamics theory states that one can deduce multi-season frequency dynamics from the ‘cobweb diagram’ determined from the initial C. albicans frequency plotted versus the final frequency each season.

The changes in these can be used to define a two-dimensional dose response mosaic. a, Φ_on lies below the diagonal so C. glabrata outcompetes C. albicans. b, Φ_off lies above the diagonal line of equal frequencies, so C. albicans outcompetes C. glabrata. c, Φ_off is such that there exists a special frequency, f_c, which lies on the diagonal line, and Φ_off(f)>f for 0 < f < f_c while Φ_off(f) < f for f_c < f < 1–this is a stable coexistence state. d, Φ_off is such that there exists a special frequency, f_u, which lies on the diagonal line, and Φoff(f) < f for 0 < f < f_u while Φ_off(f) > f for f_u < f < 1. In this case, either species can dominate depending on their initial frequencies: if the initial frequency of C. albicans is smaller than f_u then C. albicans loses out in competition to C. glabrata, otherwise C. glabrata loses out–this is ‘bistable exclusion’. e, A theoretical example of a three-season treatment, which stops short of the tipping point (marked ‘tip’) with seasons 4–9 continuing without the drug being applied. f,g, A theoretical example of a four-season treatment that goes beyond the tipping point, causing a divergence in the trajectory following drug withdrawal (f); the dynamics of tipping is illustrated in g. In e,f, solid dots represent instances where drug is not used while open circles represent instances when drug is used. In g, solid dots represent instances where drug is used while open circles represent instances when drug is not used. h, Theoretical two-dimensional dose-response mosaic. The mosaic describes the equilibrium outcome of competition in the Candida community as glucose and fluconazole are varied. C. albicans wins the competition inside orange squares, C. glabrata wins inside the blue squares and bistable exclusion occurs in the grey squares. Drug on–off treatments that encounter the grey squares may exhibit tipping (for example, ABAB, FEFE and CDCD treatment sequences); treatments that stay inside the orange and blue squares will exhibit reversible resistance (for example, a FBFB sequence). i, Theoretical Φ functions at points A–F in the dose-response mosaic, with dots highlighting the location of the special frequency f_u for each Φ that crosses the diagonal. Grey line represents the diagonal, throughout.

Using a fixed-glucose stochastic model f_{n + 1} = Φ(f_n, a_n) + σ_n, where σ_n is small-variance noise and the form of Φ is defined in Supplementary Section 3, we show that resistance need not be reversible because a tipping point is encountered in a theoretical four-day treatment (Fig. 2f,g) that is not encountered if treatment terminates at three days (Fig. 2e). Figure 2h,i then incorporates glucose dependence, g, and explores a model Φ(f, a, g) in different abiotic conditions by systematically varying g and antibiotic dose, a (Supplementary Sections 3.2 and 4). This is impossible to do in empirical microbial ecologies but computational simulations (Fig. 2h) show the dominant Candida species–C. albicans, C. glabrata or neither–in the (a, g)-plane (this is the ‘dose-response mosaic’). This computational analysis shows the dosage at which tipping occurs depends on glucose availability and we will therefore now also manipulate glucose availability in our empirical microcosm.

So does the laboratory treatment community also have tipping points when antibiotic dose or else glucose availability vary? This is difficult to assess directly for several reasons. First, our modelling framework is general but simple and so is not able to accurately pinpoint tipping points in an empirical context. Second, theoretical tipping mechanisms require an unstable fixed point (f_u) under drug-free conditions. These are hard to identify empirically because observations move away from unstable fixed points and so these points, if present, cannot be detected directly in longitudinal data, we can only infer their presence. Other warning signals of tipping exist^39,40, for example, so-called critical slowing down^41,42, slow recovery from perturbations^43,44, an increase in autocorrelation⁴⁵, an increase in the variation of fluctuations^46,47 or time series skewness⁴⁸. We therefore chose variance increases because modelling indicates between-replicate variance (BRV) should increase sharply at a tipping point (Fig. 3).

Fig. 3. The dose-response mosaic shows tipping points are encountered in many ways.

a, A tipping point can be encountered, as fluconazole concentration is varied (or, for example, by varying glucose concentration, see Supplementary Fig. 10 for details). Four, fixed-dose treatments start on season 5 and end on season 12 (grey box) at a 0.95% (by volume) glucose dose. First, the community converges towards C. albicans domination in the absence of drug (orange dots). C. glabrata starts to dominate as drug is applied, but it rescinds when treatment ends (brown dots) and the community returns to its pre-treatment composition and then continues towards C. albicans dominance with more seasons. However, a tipping point appears at just high-enough fluconazole dose (dark blue dots) whereby the post-treatment trajectory diverges from the previous outcome (at a slightly lower drug dose) and C. albicans is lost as the seasons pass. Royal blue dots show trajectories at dosages well above the tipping point. b, Introducing additive stochastic noise to simulations from a shows that replicate trajectories diverge at the tipping point, creating large variations between frequency trajectories that had identical drug dosage regimes and initial Candida frequencies. c,d, A signature we can seek in empirical data: BRV spikes at the tipping point (c), causing a large season-by-season change in BRV (Δ BRV here taken to be the mean change in standard deviation) that is significantly positive at the tipping point (d).

We sought antibiotic tipping experimentally for three treatments, α, β and γ, that are designed to explore the dose-response mosaic as fully as possible. In the α treatment, glucose is held constant but the drug steadily withdrawn; in the β treatment, glucose and drug are held constant; and in the γ treatment, glucose is reduced while the drug is withdrawn (Fig. 4a). For these treatments, glucose varied between 0.1 and 4.0% mirroring previous in vitro Candida experiments^49,50 and fluconazole varied between 0 and 3 μgml^–1, mirroring previous in vitro drug-adaptation studies⁵¹.

Fig. 4. Exploring the dose-response mosaic.

a, Three laboratory treatments with different dynamics: treatments α (red) and γ (black) withdraw fluconazole (Flu) but β (blue) keeps it at constant levels (dosages represented as circle sizes). b,c, Laboratory experimental trajectories of treatments (b) and the corresponding BRV (c) as indicated by the kernel density estimate of the distribution of final-season C. albicans frequency differences (48 replicates for α treatment, 96 for β and 55 for γ). The trajectories of treatments α and β have low BRV in species frequencies at all times whereas treatment γ has high BRV at the end of treatment. The trajectories show why: community dynamics for treatment β maintain a steady state and C. albicans sweeps through the community during treatment α following drug withdrawal. However, in treatment γ, trajectories of different replicates vary markedly beyond season 6 whereby either species can dominate by season 14, despite all replicates having close to 50–50 initial composition (Supplementary Fig. 11 has additional data). d, Data from the Candida community shows mean BRV increases significantly, approximately eightfold for treatment γ on season 6. Treatments α and β also have significant increases on occasion, but by no more than threefold. Data are mean ± 95% confidence interval. e, Taking a conservative Bonferroni-corrected significance at the level P < 0.001 in an F-test using linear regression (see Methods), significant changes in mean BRV are shown as circled dots. The largest increase (approximately eightfold) is significant and occurs in treatment γ on season 7. Data are mean ± 99.9% confidence interval.

Treatments α and β lead to reversible resistance (Fig. 4b) whereby the C. albicans frequencies on the last season have almost unimodally distributed BRV (Fig. 4c). However, treatment γ exhibits characteristics of tipping: BRV statistics of C. albicans frequencies on the last observed season approximate a uniform distribution (Fig. 4c) and mean BRV spikes on season 6 (approximately eightfold increase, Fig. 4d,e). The rapid divergence of replicate trajectories (Fig. 4b and Supplementary Fig. 12) that forms a uniform distribution of treatment outcomes for γ (Fig. 4c) in a manner consistent with theory (Supplementary Fig. 9) means that many community trajectories have not returned to their inoculum positions, in contrast to reversible resistance (Fig. 1a) where they have.

Discussion

The reversibility of resistance is often conceptualized through resistance costs¹⁶, a property that ensures resistance genes are lost following drug withdrawal due to a fitness reduction of the mutants that carry them. However the analogy of resistance costs between species is difficult to define. For example, without the drug, in our community the Candida species have different metabolism⁵² from which complex, density- and frequency-dependent ecological interactions such as cheating and cooperation can result⁵³. Indeed, the myriad ecological interactions present in natural communities are necessarily perturbed by an antibiotic drug, so a model of resistance progression in which resistant and susceptible microbes differ by a single allele will have limited explanatory power here. Thus we invoke tipping as a new mechanism for understanding the dynamics of drug resistance following exposure to antibiotics.

The explanation behind the tipping-induced irreversibility of resistance is this: if a community could persist in multiple configurations in the absence of drug¹⁷, antibiotics, indeed, any abiotic perturbation, might push the community into the ‘basin of attraction’ of the most resistant configuration from all those available. So even if treatment stops, resistance species’ frequencies could increase. Our mathematical models illustrate just two basins of attraction, one above the tipping point and one below (Fig. 3a), but real-world communities may well have more.

Unfortunately, the key ingredient for tipping, multistability, is known to be difficult to demonstrate in real communities¹⁷, but if present, we then know the removal of drug can create an uncertain future for that community. Figures 3b and 4b show in theory and in data how some of those divergent futures pan out; some return whence they came, others move towards a new configuration of the community. This is the defining property of multistability¹⁷ and if this new configuration comprises more species that are less susceptible to the drug than were there before treatment, resistance in the community will increase even though treatment has stopped.

Our microcosms highlight just one treatment consistent with this theory (Fig. 4), but what other treatments might do this? Indeed, data show that not all drug treatments induce tipping (Fig. 4). However, mathematics answers the question: any co-variation of abiotic environment and drug, whether stochastic, cyclical, gradual or abrupt, that guides the community into a region of the dose-response mosaic that exhibits multistability (Fig. 2h, grey zone) creates the right conditions for tipping. Our empirical data provides one example of this (treatment γ) from the infinitely many treatments we could have tested, and the mathematical model we present undergoes tipping with this type of treatment (Supplementary Fig. 9) and for many more besides. The theoretical treatment examples we provide (Fig.3 and Supplementary Fig. 10) illustrate, perhaps, the simplest possible abiotic variation that can exhibit tipping, namely the abrupt cessation of a constant-dose drug treatment (Fig. 3) of the kind given to patients in the community.

To conclude, we argue that single-species logic is insufficient to understand resistance in microbial communities. Particularly lacking is a theory of how abiotic variation promotes resistance and yet this is relevant to patients. For example, infections involving C. glabrata are more frequently found in diabetic patients with high blood glucose levels than in patients with lower glucose levels^54,55, indicating that nutrient availability may play a role in clinical resistance, just as it does in our community. Our observations may also indicate potential for alternative therapeutic rationales for polymicrobial infections. Diet is known to alter the host microbiota^56–58 and so fashioning specific environments by manipulating nutrients might tip the balance of competition in favour of drug-susceptible species and render an infection more amenable to treatment. There is a precedence for this idea^59–61.

Methods

Strains and assay medium

The strains Candida albicans ACT1–GFP and Candida glabrata ATCC2001 were used throughout this paper. The C. albicans ACT1–GFP strain is SBC153⁶² has green flourescent protein (GFP) integrated into the genome under the control of the native actin promoter (ACT1p), with transformants initially selected through nourseothricin resistance (NAT1). The C. glabrata ATCC2001 strain is the wild-type reference strain obtained from the American Type Culture Collection. The assay medium throughout was synthetic complete (SC; 0.67% w/v yeast nitrogen base without amino acids, 0.079% w/v synthetic complete supplement mixture (Formedium)).

Measuring growth of C. albicans and C. glabrata in isolation in the absence/presence of drugs

Overnight cultures in yeast extract peptone dextrose (YPD) media of C. albicans and C. glabrata were diluted, counted on a hemocytometer and adjusted to 2 × 10⁷ cells ml^–1 in SC medium containing 2% (weight/volume) glucose. Sterile plastic microdilution plates containing 96 flat-bottomed wells were utilized. Stock solution of fluconazole was diluted in SC medium and dispensed in 75 μl volumes into six replicate wells to yield nine twofold serial dilutions of fluconazole with final concentrations ranging between 0 and 64 μg ml^–1. For each drug concentration, three of the six replicate wells were filled with an additional 75 μl from the C. albicans strain suspension while the remaining three wells were filled with 75 μl from the C. glabrata strain suspension. The plate was sealed with a transparent adhesive seal and two holes were punctured over each well by means of a sterile needle. The plate was incubated at 30 °C with shaking over 24 h and growth monitored by measuring the absorbance of the cell suspensions at 650 nm (A₆₅₀). Absorbance units were converted into number of cells per ml by means of calibration curves prepared for each Candida species.

Drug susceptibility dose-response for C. albicans and C. glabrata

Standard microdilution susceptibility testing was performed in sterile 96-well flat-bottom microtitre plates with the following modifications. Fluconazole was diluted from a 2 mg ml^–1 stock solution to a dilution series ranging from 512 to 0 μg ml^–1 in SC media containing 1% or 4% glucose (weight/volume), at a volume of 75 μl per well. Overnight cultures in YPD of C. glabrata and C. albicans were counted by haemocytometer and diluted to 10⁵ cells ml^–1 in SC broth containing either 1% or 4% glucose. For each species, 75 μl of cell suspension was added per well, resulting in a final volume of 150 μl, which contained 5 × 10⁴ cells, a final drug concentration of 256, 192, 128, 96, 64, 48, 32, 16, 8, 4, 2 or 0 μg mL^–1 of fluconazole, and either 1% or 4% glucose. The plate was sealed with a transparent adhesive seal and two holes were punctured over each well, by means of a sterile needle, for aeration. Plates were incubated for 48 h at 30 °C after which final absorbance was measured at A₅₉₅. All samples were assayed in technical triplicate.

Competition of C. albicans and C. glabrata in the absence/presence of drugs

Overnight cultures in YPD of C. albicans and C. glabrata were diluted, counted on hemocytometer, and adjusted to 2 × 10⁷ cells ml^–1 in SC medium containing either 0.1, 2 or 4% (weight/volume) glucose. The two Candida species were then mixed to achieve a range of starting ratios (0, 10, 30, 50, 70, 90 and 100% C. albicans) The cell suspension was then diluted 1:1 in SC media containing two times the desired fluconazole concentration (0, 0.5 or 2 μg ml^–1) to a final volume of 150 μl in a 96-well flat-bottom microtitre plate. Plates were sealed using sterile adhesive films, two holes were punched in the seal above each well using a sterile needle and incubated overnight in an orbital shaker set at 30 °C, 180 r.p.m. Each condition was repeated in triplicate. After 24 h incubation period (one season) plates were opened, the contents of the well were vigorously pipetted to achieve a homogenous cell suspension, and relative frequency of C. albicans ACT1–GFP was determined by flow cytometry in the following way (Supplementary Fig. 1). Cellular fluorescence from GFP was determined quantitatively with a FACSAria flow cytometer (Becton Dickinson) equipped with a 20 mW, 488 nm argon ion laser. All samples were suspended in PBS and briefly sonicated to disperse potential cell clumps before analysis. Typically, 10,000 cells were analysed per competition sample with the following settings: forward scatter, 150 V, log mode; side scatter, 200 V, log mode. GFP was detected on a 530/30 filter (600 V, log mode) and sample acquisition was performed using BD FACSDiva software. Initially, a sample consisting of C. albicans ACT1–GFP cells only was detected and gated to contain 99–100% of all measured events as positive for GFP fluorescence. All events occurring within the gate during subsequent analysis of competition samples were considered to be C. albicans ACT1–GFP cells. For any given competition sample, the frequency of gated events was calculated by means of FlowJo software and was taken to be the population percentage of C. albicans ACT1–GFP within the sample.

Long-term competition of C. albicans and C. glabrata for different drug regimes

Overnight cultures in YPD of C. albicans and C. glabrata were diluted, counted by hemocytometer and adjusted to a 1:1 ratio of each species, at a final concentrations of 2 × 10⁷ cells ml^–1 in SC medium containing either 0.1% or 4% (weight/volume) glucose. The cell suspension was then diluted 1:1 in SC media containing two times the desired fluconazole concentration (3 μgml^–1) and the matching glucose concentration, to a final volume of 150 μl in a 96-well flat-bottom microtitre plate. Plates were sealed using sterile adhesive films, two holes were punched in the seal above each well using a sterile needle and incubated overnight in an orbital shaker set at 30 °C, 180 r.p.m. Each treatment was repeated in triplicate. After 24 h incubation period (one season), plates were opened, the contents of the well were vigorously pipetted to achieve a homogenous cell suspension, and 5 μl of cell suspension was added to a new well containing 145 μl of SC media, which contained either the same, or a reduced concentration of glucose and fluconazole. Specifically, for treatment α, the cultures were maintained in 0.1% glucose throughout, while the fluconazole concentration was adjusted each season (3, 2, 1, 0.5, 0 μgml^–1). For treatment β, the glucose concentration remained at 0.1% and fluconazole concentration at 0.5 μgml^–1 throughout the duration of the experiment. Lastly, for treatment γ, both glucose and fluconazole concentrations were reduced each season, glucose decreasing from 4 to 0.1% (4, 2, 1, 0.5, 0.1) and fluconazole decreasing from 3 to 0 μg ml^–1 (3, 2, 1, 0.5, 0). Note that 5 μl represents the smallest volume transfer that ensures the accuracy of pipetting is maintained. The relative frequency of C. albicans was monitored either by flow cytometry or colony forming units (CFUs), as described above. For daily monitoring, 9–12 replicate biological samples were measured, while at the endpoint all replicates were analysed (48 replicates for treatment α, 96 replicates for treatment β and 55 replicates for treatment γ).

Oscillatory (a repeated on–off) drug treatment

Overnight cultures in YPD of C. albicans and C. glabrata were diluted, counted by hemocytometer and adjusted to a 1:1 ratio of each species, at a final concentrations of 2 × 10⁷ cells ml^–1 in SC medium containing 0.1% (weight/volume) glucose. The cell suspension was then diluted 1:1 in SC media containing two times the desired fluconazole concentration (0, 2 or 4 μg ml^–1) to a final volume of 150 μl in a 96-well flat-bottom microtitre plate. Plates were sealed using sterile adhesive films, two holes were punched in the seal above each well using a sterile needle and incubated overnight in an orbital shaker set at 30 °C, 180 r.p.m. Each condition was repeated in triplicate. After 24 h incubation period (one season), plates were opened, the contents of the well were vigorously pipetted to achieve a homogenous cell suspension, and 5 μl of cell suspension was added to a new well containing 145 μl of SC media that contained the same drug concentration as the day before, for three days. From day 3 to day 14, cells were cultured without drug. On day 14, all conditions were treated with 2 μgml^–1 fluconazole for an additional three days (to day 17) at which point drug was again omitted from culturing to the end of the experiment. All wells were passaged daily; on days with data points shown in Fig. 1a, a volume of suspension was removed for sampling. To monitor relative frequency of each Candida species, CFUs were enumerated by plating on YPD agar either with or without the C. albicans strain selection nourseothricin (NAT) at 200 μgml^–1. Briefly, cell suspension from overnight culture was diluted to roughly 200 CFU per 100 μl and plated on both YPD and YPD + NAT plates; each well was plated in duplicate. Plates were incubated at 30 °C for 48 h and colonies counted, with the percent C. albicans being determined by the ratio of NAT-resistant cells to total cells on untreated YPD plates.

Intracellular fluconazole accumulation

The accumulation of fluconazole for both C. albicans and C. glabrata was analysed in energized cells in the presence of glucose using the protocol described in ref. ⁶³. Cells were incubated with [³H]-FLC (specific activity 740 GBa mmol^–1, 20 Ci mmol^–1, 2 × 10⁴ CPM pmol^–1, 1 μCi μl^–1, 50 μM fluconazole (FLC); custom synthesis by Amersham Biosciences). Cells were grown overnight in CSM complete medium at 30 °C to an optical density typically between OD₆₀₀ 6.0 and 8.0, unless otherwise noted. Cells were subsequently collected by centrifugation (3,000 × g, 5 m) and washed three times with yeast nitrogen base (YNB) complete (1.7 g yeast nitrogen base without amino acids or ammonium sulfate, 5 g ammonium sulfate per litre, pH 5.0) without glucose (for starvation) and without supplementation, unless otherwise noted. Cells were resuspended at an OD₆₀₀ of 75 in YNB for 2–3 h for glucose starvation. Reaction mixes consisted of 250 μl of YNB, 200 μl of cells (75 OD) and 50 μl of [³H]-FLC (1/100 dilution of stock). The resulting [3H]-FLC concentration is 50 nM (0.015 μg ml^–1), which is significantly below the MIC for all strains. Samples (100 μl) were removed at various time points and placed into 5 ml stop solution (YNB + 20 mM (6 μg ml^–1) FLC), filtered on glass fibre filters (24 mm GF/C, Whatman) pre-wetted with stop solution and washed with 5 ml of stop solution. Filters were transferred to 20 ml scintillation vials. Scintillation cocktail (Ecoscint XR, National Diagnostics) was added (15 ml) and the radioactivity associated with the filter was measured with a liquid scintillation analyser (Tri-Carb 2800 TR, Perkin Elmer) and normalized to CPM/1 × 10⁸ cells. Rate of [³H]-FLC uptake was determined by incubating samples in the presence of increasing concentrations of unlabelled FLC (unless otherwise noted) and samples were analysed for [³H]-FLC accumulation at designated time points.

Computational methods

Numerical simulations of a theoretical community model were obtained using Matlab’s differential equation solvers to generate the season-by-season dynamical map Φ(f). Differential equations were parameterized using data from C. albicans and C. glabrata, as detailed in the Supplementary Information.

To determine tipping points, we sought significant increases in BRV defined as follows. If $\mathcal{F}={\left\{f_j\right\}}_{j=1}^n$ is the set of observed C. albicans frequencies, expressed as values between zero and one (although some figures express this as a percentage), the set of between-replicate differences is then ${\left\{\left|f_j-f_k\right|\right\}}_{j,k=1,j>k}^n$ and mean BRV is the mean of this set; note, this is one form of set radius. As the frequency of C. glabrata is G_j = 1 – f_j and G_j – G_k = 1 – f_j – (1 – f_k) = f_k – f_j, C. glabrata frequency data have the same between-replicate differences as C. albicans.

Kernel density estimates were obtained for distributions of BRV values (Fig. 4a) using a kernel estimation algorithm implemented in Matlab⁶⁴. To test for significant season-by-season differences in BRV, BRV_n and BRV_n+1 observed at seasons n and n + 1, we applied linear regression to test (with P < 0.001) against the null hypothesis of a constant mean BRV between those seasons, testing for a non-zero slope parameter from the regression. This is written Δ BRV and this change was found to be largest for season 6 of treatment γ (Fig. 4b).

Reporting Summary

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

Data Availability

All experimental data generated during this study can be found at https://doi.org/10.24378/exe.345.