Complex and unexpected outcomes of antibiotic therapy against a polymicrobial infection

Model overview and parameters

Our mathematical model of the CF lung microbiome dynamics, originally developed in [20], is based on knowledge of the physiology and interactions among community members from experimental data and evidence in the literature. The model setting is a mucus-plugged tube, open to the air at the top and sealed at the bottom, mimicking a lung bronchiole. This setting is meant to pair with a previously established experimental microcosm called the WinCF system [21], which we use below for experiments. There is an important spatial component to the model, as oxygen penetration from the open top of the tube is constant and shapes the community structure. The consequences of these chemical gradients were first modelled in our initial study [20]. The community members are classified as either “pathogens”, representing classic CF pathogens, or “fermenters”, representing other anaerobic organisms commonly encountered in CF airways. These classifications are a significant simplification, but they can be considered as guilds, in that their individual members have similar inherent properties defined by their core metabolism, antibiotic resistance, and niche occupancy [20]. The definition of classic pathogens and anaerobic fermenters is also clinically relevant, as the former are those assayed in clinical labs for antibiotic resistance to inform treatment decisions, whereas anaerobic fermenters are not cultured or tested for susceptibility in most clinical labs. Classifications of each microbiome member into these guilds are available in Tables S2–S4. Fermenters reside in low oxygen areas and utilize sugars to produce acids [20] (Fig. 1). Pathogens, principally, but not exclusively, Pseudomonas aeruginosa, occupy high oxygen regions where they aerobically respire and utilize amino acids as a carbon source producing ammonium, which increases the surrounding pH [20] (Fig. 1). Pathogens can also respire anaerobically, with nitrate as an electron acceptor (Fig. 1). In addition to increasing the surrounding pH, they produce inhibitor molecules (such as phenazines and quinolones) that inhibit the growth of fermenters [20] (Fig. 1). This model is hereon referred to as the “mathematical model”.

Predicting and modelling outcomes of antibiotic therapy

To better conceputalize and compare our modeling and experimental results, we first theoretically predicted the outcomes of antimicrobial therapy against the two guilds using three theoretical drugs: one with fermenter coverage (denoted T_f), one with pathogen coverage (denoted T_p), and one with broad spectrum coverage (denoted T_w). This approach is hereon referred to as the “theoretical prediction”. To further enable comparison to experimental data we outline characteristics of the two guilds we expect to observe in the experiments. Firstly, the growth of anaerobic fermenters is positively correlated with an increase in gas production (bubble formation in the WinCF system) [21]. Second, an increase in P. aeruginosa positively correlates with an increase in its inhibitor molecule (e.g., Quinolone HHQ) and P. aeruginosa does not produce gas in the WinCF system [21]. Thirdly, based on Tables S1–S4 and the CF microbiome literature, fermenters are more diverse than pathogens [2, 43, 44]. These characteristics of our theoretical prediction enable direct comparison to microbiome measures of experimental results, such as alpha diversity, beta diversity, pathogen relative abundance, fermenter relative abundance and total bacterial load (TBL).

With our theoretical prediction we expect the following outcomes when communities are exposed to antibiotics: (1) community resistance, (2) community death, (3) pathogen death, and (4) fermenter death (Fig. 2A–E). In both the complete absence of an antibiotic and community resistance, we expect TBL, pathogens, fermenters, HHQ, and gas production measures to increase until reaching carrying capacity (Fig. 2B). The opposite, community death (treatment with T_w) results in both microbial entities failing to grow (Fig. 2C). T_w treatment would not change alpha or beta diversity, as we would simply measure the initial inoculum due to total community death. Outcomes 1 and 2 have a degree of uncertainty due to the fact that it is difficult to assume the community would not change from the inoculum without an antibiotic present, but it is expected that T_w would have less impact on microbiome diversity than T_p or T_f (Fig. 1C). Treatment with T_p results in an anaerobic fermenter bloom, increasing alpha and beta diversity along with gas production and a decrease in HHQ production (Fig. 2D). Finally, in the case of T_f treatment, fermenter abundance and gas production would decrease while HHQ abundance would increase (Fig. 1E). Treatment with T_f will also result in a decrease in alpha diversity and an increase in beta diversity because of changes in community structure when the diverse anaerobic fermenters are killed (Fig. 2E).

The theoretical prediction was then tested with the mathematical model hereon referred to as “model iteration 1”. Importantly, our model parameters can use relative abundance data of the two guilds as input. Therefore, we used the sputum microbiome data of all 24 subjects as inputs for model interation 1 (Fig. 2F–H). The outputs were in line with our theoretical prediction and showed that the fermenter drug would reduce the fermenter load, with little effect on the pathogens, the pathogen drug vice versa, and the broad-spectrum antibiotic would kill both (Fig. 2F–H). However, model iteration 1 did produce some unexpected results. The TBL of the T_w decreased to similar levels as T_f and T_p, indicating similar levels of killing whether there was selection against a single guild or the whole community (Fig. 2H). In addition, the TBL and Pathogen/Fermenter log-ratio were variable, indicating the carrying capacity and community dynamics were predicated upon characteristics of this initial sputum inoculum (Fig. 2F–H). Our theoretical prediction (Fig. 2A–E), in tandem with model iteration 1 (Fig. 2F–H), provided a platform for comparison to the in vitro antibiotic experiments with the WinCF system described below.

Experimental results of antibiotic therapy against the lung microbiome

We examined the effects of antibiotics (n = 11) on the CF sputum microbiome cultured in a lung bronchiole microcosm (WinCF system, n = 24) using a combination of 16S rRNA gene amplicon sequencing, metabolomics, and qPCR analysis and compared to our theoretical prediction and model iteration 1. This is hereon referred to as the “antibiotic experiment”. The antibiotics were chosen to represent the main chemical classes commonly used in CF clinics and included: amoxicillin, azithromycin, aztreonam, ciprofloxacin, colistin, doxycycline, levofloxacin, meropenem, metronidazole, bactrim (a combination of sulfamethoxazole/trimethoprim), and tobramycin. Each of the 24 sputum samples were used as an incoculum in ASM treated with one of 11 different antibiotics cultured at 37 °C for 48 h (Table S1) and compared to a no-treatment control. WinCF tubes were also inoculated with this media/sputum/antibiotic mixture to quantify gas bubble production from fermentation (as described in [21]). The antibiotic concentration for each drug was variable and chosen to match the measured concentrations in the blood or sputum of pwCF in pharmacokinetic studies (Table S1). The most prominent genera across all samples after growth were Pseudomonas, Streptococcus, Veillonella, Haemophilus, Fusobacterium, Prevotella, Staphylococcus, Achromobacter, and Neisseria (Fig. S2). A principal component analysis (PCA) biplot, examining the top five factors by percent contribution, showed the primary genera driving community differentiation were Pseudomonas, Streptococcus, and Staphylococcus (Fig. S3). The effects of antibiotics and individual patients on the composition of the communities were compared via PERMANOVA (Table S7). Tested separately, both antibiotic and subject source had a highly significant effect on the community structure (p < 0.001). However, the nested effect and interactions between antibiotic and the patient did not (p = 1). Thus, the changes in the ASV composition were the result of both the antibiotic and the subject’s initial community separately, but there were not universal responses across subjects for each drug.

We visualized the changes in our microbiome and physiology measures compared to the no antibiotic control in the context of the theoretical prediction (colored areas in Fig. 3) to aid the identification of outcomes that did or did not match the predictions. All measures had significantly different changes across antibiotics according to a Kruskal-Wallis test, except for HHQ abundance (Tables S8, S9). Alpha diversity (Shannon index) showed a general decrease compared to an untreated control when the antibiotic was applied (Fig. 3A), but this depended on the antibiotic. amoxicillin and meropenem resulted in the strongest decreases in alpha diversity, being significantly lower than the other treatments (Table S10), which changed little and had instances of increases in diversity (Fig. 3A and Table S10). Beta diversity (weighted UniFrac distance) comparisons of treatment samples to the no antibiotic control enabled quantification of the degree of microbiome change due to treatment. amoxicillin and meropenem had the highest beta diversity increases, with the latter being significantly higher than 8 others (Table S11) and azithromycin the lowest (Tables S9, S11), though there was significant variability within each drug limiting the statistical significance across the different treatments. The variability in the antibiotic experiment showed that although some drugs had smaller impacts than others all antibiotics impacted the microbiome composition with some unique responses for particular patients (Figs. 3B and S4). Plotting Pathogen/Fermenter log-ratio changes compared to the control enabled the quantification of dynamics between the two guilds and direct comparisons to the theoretical prediction and model iteration 1. Again, amoxicillin (significantly higher than 7 of 10) and meropenem (higher than 8 of 10) increased the relative abundance of pathogens compared to fermenters. Significant decreases in this ratio were observed with aztreonam, tobramycin, and ciprofloxacin (Fig. 3C and Tables S8 and S12). An unexpected result not identified by theoretical predictions or model iteration 1 was observed when comparing TBL changes between treatment samples and controls. Overall, the rRNA gene copy number (a measure of total bacterial abundance using qPCR) did not change significantly across the different antibiotics, except for meropenem, which significantly reduced this ratio compared to 8 of 10 treatments (Fig. 3D, Tables S8 and S13). Interestingly, despite the decrease in alpha diversity and increase in beta diversity compared to the control, amoxicillin did not have a significant decrease in TBL. Furthermore, all drugs had samples that increased in total bacterial abundance (i.e. values above 1 in Fig. 3D). Specifically, 17.8% of all samples showed a 20% increase in rRNA gene copies and 6.8% increased by 40% (Fig. 3D and Tables S17, S18). Therefore, despite the presence of an antibiotic meant to inhibit bacterial growth, the total carrying capacity increased in many samples of the antibiotic experiment, but this phenomenon was not driven by a specific drug. HHQ abundance changed dynamically with antibiotic treatment (greater than 2-logs) and these changes were mostly driven by the individual subject source not a specific antibiotic (Tables S14, S15 and Figs. S4a, S4e), meaning that there was a more personalized response to the production of this P. aeruginosa associated-metabolite. Finally, gas production, our measure of microbial fermentation in the WinCF system, showed an overall decreasing trend compared to the control, most pronounced from meropenem, doxycycline and amoxicillin, but few comparisons were significant due to extensive variation within each treatment (Fig. S3a and Table S16). Similarly to the increases in TBL, but this time predicted by the model, increases in the gas production were seen in the experiment and all antibiotics had at least one instance of an increase compared to the no-treatment control (Fig. S4a and Table S16).

Characterizing outcomes of antibiotic therapy against the CF lung microbiome

To better quantify and characterize outcomes from the antibiotic experiment, microbiome measures of interest were plotted against the UniFrac distance from the control sample (Fig. 4). Four outcomes observed from this experiment matched the theoretical predictions and model iteration 1 including: 1) community resistance, 2) community death, 3) pathogen death, and 4) anaerobe death (outcome definitions quantified in Table S17). Outcomes five and six were not predicted and were defined as 5) niche replacement events and the 6) release of community level inhibition. The most common outcome was 1) community resistance, which encompassed 44.6% of all samples tested (Fig. 4, quantified outcome definitions available in Tables S17, S18). This may indicate that the CF lung microbiome has an inherent antibiotic resistance due to decades of exposure and the propensity of its constituents to grow as biofilms [45, 46]. Community death (outcome 2), occurred 17.8% of the time. Cases of community death with little change in beta diversity were rare, indicating that comprehensive antibiotic killing most often results in a community structure change compared to a no antibiotic control. Both pathogen death (8%) and fermenter death (17%) outcomes were observed in our experiments (Fig. 4 and Tables S17, S18). Anaerobe death outcome was driven by meropenem and amoxicillin as shown in Fig. 3C, whereas, pathogen death was not driven by any particular drug. Niche replacement (outcome 5) occurred when the TBL of the sample did not change but the UniFrac distance was above 0.4, which encompassed 6.4% of samples (Fig. 4b, d). This outcome may reflect the diverse nature of the fermenter guild; when a certain species is killed, another can take its place, maintaining the fermentative nature of the community but resulting in a community structural change. The release of community level inhibition (outcome 6) was defined as an increase in TBL (>40%), which occurred in 6.8% of samples. The microbiomes of outcome 6 were predominantly dominated by pathogens compared to the control samples (Fig. S7). We found this outcome to be especially interesting, with potential clinical relevance; we therefore performed follow up experiments to understand it further.

Other interesting data relationships were found in these experiments (Fig. S8) though they were not defined as outcomes. For example, the changing UniFrac distance and change in alpha diversity were negatively correlated (Fig. S8a). A large increase in UniFrac distance (over 40% increase), was generally associated with takeover by a particular ASV, driving this phenomenon (Figs. S7 and S9). According to prevalence measures of theses samples the prominent genera in these instances were Pseudomonas and Streptococcus (Fig. S9). In the cases of meropenem and amoxicillin, UniFrac distances were increased while the Shannon indices were decreased, due to the killing of diverse anaerobic community, but there were fewer cases of an increase in alpha diversity and a significant microbiome change (observed in 3 samples only) indicating a kind of buffering of the microbiome by the diverse anaerobic community (Fig. S7a). The increase in TBL characterizing outcome 6 was rarely associated with an increase in alpha diversity (Table S17). Finally, similar to a phenomenon described in CF sputum [31], when the microbiome alpha diversity increases the metabolome diversity decreases, likely reflecting consumption of different metabolites by a more diverse microbiome (Fig. S7c).

Model iteration 2 and experimental validation to explain increase in TBL

Because model iteration 1 did not predict the interesting outcome 6, we altered its parameters to determine if we could observe an increase in TBL in the presence of an antibiotic, hereon referred to as “model iteration 2”. In model iteration 1, parameter λ in the function g2(Z) was set to 0.1, which represents pH driven inhibition of fermenters on pathogen growth. Due to the inverse relationship of this parameter, reducing it to 0.05 increased the strength of inhibition, resulting in an increase in TBL for some subjects, akin to that observed in our experimental outcome 6 (Fig. 5A). This only occurred in T_f treatments in model iteration 2, corresponding to a bloom in pathogens after killing of anaerobes. Furthermore, this phenomenon was only present in modelled samples that initially contained much lower populations of the fermenter guild compared to pathogens and is dependent on the spatial structure driven by oxygen gradients that is an inherent property the modeled system (Figs. 1 and 5A). This finding suggests that outcome 6 in the antibiotic experiment may be driven by an antibiotic mediated release of community level inhibition driven by the effect of low pH from fermenters on pathogens and the inhibition of anaerobes by oxygen [20]. Thus, we set out to explore this phenomenon in more detail experimentally.

A simple in vitro experiment was performed where three antibiotics, meropenem (T_w), tobramycin (T_p), and metronidazole (T_f), were added at 2.048 mg/L in ASM media inoculated with two representative communities obtained from pwCF: P1 and P2 (n = 10 replicates) (Fig. 5B–F). The three drugs were selected based on their common uses against CF infections based on pathogen and/or anaerobic coverage, but we acknowledge that their effects are not exclusive to these organisms. Community P1 did not contain P. aeruginosa via culturing on cetrimide agar, whereas the bacterium was isolated from the sputum of P2. This provided a unique opportunity to test the predictions from model iteration 2 on the outcomes of a community with or without P. aeruginosa. A lower concentration of antibiotics was chosen to avoid widespread killing of the communities. We examined the following: rRNA gene copies, approximate pH (based on RGB color values inferred from phenol red buffered media standards) and 16S rRNA gene amplicon sequencing (Fig. 5). This is hereon referred to as the validation experiment. The validation experiment reproduced outcome 6, where both the number of rRNA gene copies were higher when the antibiotic was present than in the no treatment control for both P1 and P2 (Fig. 5C). In contrast to model iteration 2, this only occurred in treatment T_w (paired t-test, p = 0.000831) (Fig. 5). Accordingly, this increase in TBL corresponded to an increase in pH of the cultures, validating the association of the anaerobe induced fermentation with an inhibition of the communities’ total carrying capacity (p = 1.69 × 10⁻⁹, Fig. 5B–E). In fact, there was a strong positive correlation between the TBL and media pH overall (Fig. 5B). Furthermore, P2 reached a higher bacterial load overall than P1 in the validation experiment, indicating that the pathogen’s presence drove the community to a higher carrying capacity (Fig. 5E). The lower growth in community P1 shows that a community of primarily anaerobic fermenters struggles without the aerobic pathogen present. Microbiome profiles of these follow up experiments validated the predictions of model iteration 2 and initial findings of outcome 6 (Fig. 5F, G). Meropenem killed the anaerobic community (primarily Streptococci) and the increase in TBL was driven by a bloom of Pseudomonas (P2 community) and Staphylococcus (P1 community) to a higher level than the communities’ inherent carrying capacity (Fig. 5F, G). This experiment was subsequently repeated (n = 5), with the same results observed (Fig. S10). It was interesting that a similar increase in TBL occurred from a community without a dominant pathogen (P1, Fig. 5G). We hypothesize that this result is due to the importance of both oxygen and pH in the governing dynamics. With very low levels of the pathogen guild, the community struggles to grow due to high oxygen penetration. When the anaerobes are inhibited by antibiotics, even low levels of an initial pathogen can begin to bloom, as they are not inhibited by oxygen or the antibiotic, and this leads to an increase in total carrying capacity.

Antibiotic effects at the strain level in pwCF

To explore similar phenomena in outcomes 5 and 6 from pwCF treated with antibiotics we sequenced the metagenomes of sputum samples collected from subjects immediately prior to and during antibiotic treatment (n = 6) (Table S19). To minimize the effects of multiple therapies at once, a common occurrence in CF therapeutics, these samples were selected based on the treatment provided being the only known antibiotic prescribed to the subject at the time. Metagenomes were analyzed at the strain level and TBL was examined using qPCR. Overall, there was no significant decrease in TBL (Fig. 6A, Wilcoxon rank-sum test, p = 0.095), but alpha diversity significantly decreased (Fig. 6B, Wilcoxon rank-sum test, p = 0.045). Analysis of the rank abundance changes of the microbiome at the strain level showed that all six subjects had dynamic changes in their sputum microbiomes associated with antibiotic treatment despite little decrease in TBL (Fig. 6C). Thus, like outcome 5, and indicative of outcome 6, dynamic community changes occur in pwCF with minor changes in TBL.

Source: Ecology - nature.com