Antibiotic-Induced Alterations of the Murine Gut Microbiota and Subsequent Effects on Colonization Resistance against Clostridium difficile

We selected a panel of seven antibiotics from six classes with the goal of differentially altering the microbiota and assessing their resistance to C. difficile colonization (Table 1). Following the cessation of antibiotics, each treatment group was given 1 day of recovery prior to challenge with C. difficile spores. One day after challenge, we determined the density of C. difficile in the animals’ feces. We observed reproducibly high levels of C. difficile colonization in mice treated with cefoperazone, metronidazole, and streptomycin (Fig. 1; see Fig. S1 in the supplemental material). We observed the most variation in the levels of C. difficile colonization in mice treated with ampicillin. None of the mice that received ciprofloxacin were colonized. In addition to administering ciprofloxacin by oral gavage, we provided ciprofloxacin by intraperitoneal injection (10 mg/ml). For both approaches, we provided 1 or 2 days of recovery from the antibiotic treatment. Regardless of the method, the resulting communities were resistant to C. difficile colonization. Only one of six mice receiving vancomycin was colonized with C. difficile. We suspected that this was due to residual vancomycin repressing C. difficile growth. In fact, 2 days after C. difficile challenge, C. difficile bloomed in this treatment group to a median of 9.1 × 10⁷ (interquartile range, 7.6 × 10⁷ to 1.1 × 10⁸) CFU/g feces. Furthermore, given 2 days of recovery after vancomycin treatment, there was no delay in C. difficile colonization to high levels, and on day 1 postchallenge, we observed a median of 3.0 × 10⁷ (interquartile range, 2.6 × 10⁷ to 3.6 × 10⁷ [n = 4]) CFU/g feces. These results suggest that although the gut tissue does not absorb vancomycin, the absence of C. difficile in the remaining five vancomycin-treated mice may have been due to residual antibiotics lingering in the environment. Overall, these antibiotic perturbations provided different levels of colonization by C. difficile, which led us to hypothesize that the resulting communities varied in their composition.

To test this hypothesis, we sequenced the 16S rRNA genes from the fecal communities of treated and untreated mice prior to C. difficile challenge to identify populations within the microbiota that conferred colonization resistance. All of the antibiotic treatments, except for the ciprofloxacin-treated mice (P = 0.09 by analysis of molecular variance [AMOVA]), resulted in distinct and reproducible changes to the structure of the microbiota relative to the untreated animals (P < 0.001 by AMOVA). The similarity in the structure of the microbiota in ciprofloxacin-treated and untreated mice suggests that a higher dose of ciprofloxacin may have been necessary to significantly perturb the microbiota to allow C. difficile to overcome colonization resistance. Comparisons of the microbiota between antibiotic classes indicated that their structures were significantly different from each other (P < 0.03 by AMOVA). The community structures of mice receiving β-lactams (i.e., cefoperazone and ampicillin) were not significantly different from each other (P = 0.37 by AMOVA). These results indicate that perturbing the gut microbiota with antibiotics resulted in nonoverlapping community structures that yielded significant variation in susceptibility to colonization when challenged with C. difficile.

On the basis of the C. difficile colonization levels in our seven antibiotic treatments, we hypothesized that titrating the dose of antibiotics that the mice received would result in smaller perturbations to the microbiota. Consequently, we expected a greater maintenance of resistance against C. difficile colonization in these titrated treatment groups. In addition to the previous treatments, we treated mice with lower concentrations of cefoperazone, streptomycin, and vancomycin (see Fig. S2 in the supplemental material). These antibiotics were selected because they are thought to target a broad spectrum of bacteria (i.e., cefoperazone), Gram-negative (i.e., streptomycin), and Gram-positive (i.e., vancomycin) bacteria. As expected in all mice receiving titrated doses of cefoperazone, C. difficile colonization levels decreased significantly (P < 0.02; Fig. 2). Titrating the dose of cefoperazone in the animals’ drinking water resulted in significant decreases in the relative abundance of an operational taxonomic unit (OTU) associated with the genus Escherichia (OTU 3) and a number of rare OTUs. We also observed increases in the relative abundances of OTUs associated with the family Porphyromonadaceae (OTU 5, 10, 11, 13, and 21; Fig. 2). Reducing the dose of streptomycin significantly reduced the colonization levels (P < 0.01; Fig. 2). Titrating the dose of streptomycin in the drinking water resulted in significant changes in the relative abundance of OTUs associated with the Porphyromonadaceae family (OTUs 2, 5, 6, 10, and 11), Alistipes genus (OTU 12), and Bacteroidales order (OTU 17). In addition to its anti-Gram-positive activity, vancomycin was also selected because although the community was quite different from untreated mice, we observed high levels of C. difficile colonization in only one mouse. We anticipated that lower doses might result in a community structure that would result in colonization. In fact, the 0.3- and 0.1-mg/ml doses of vancomycin resulted in similarly high levels of C. difficile colonization (P = 0.96). Seven OTUs were differentially represented across the three doses of vancomycin. Surprisingly, even though the colonization levels of C. difficile did not differ significantly between the mice receiving 0.1 and 0.3 mg/ml of vancomycin in their drinking water, four of the OTUs that had significantly different relative abundances were found only in mice receiving the lower dose. Three of these four OTUs were affiliated with members of the Porphyromonadaceae (OTUs 2, 5, and 6), and one was affiliated with a member of the genus Bacteroides (OTU 1). Two OTUs affiliated with the Akkermansia (OTU 6) and Lactobacillus (OTU 8) genera increased with increasing dose, and a third OTU affiliated with Escherichia (OTU 4) had a mixed response to the dose level. These results suggest that the context in which specific members of the microbiota are found is important in determining the overall resistance to C. difficile. For example, the relationship between the Bacteroides-affiliated OTU (OTU 1) and C. difficile colonization was positive in streptomycin-treated mice, and it was negative in cefoperazone-treated mice. In addition, cefoperazone- and streptomycin-treated mice had high levels of C. difficile, although the former had significantly higher levels of an Escherichia-affiliated OTU (OTU 3), which were absent in the streptomycin-treated mice. Together, these results suggest that individual populations were not sufficient to consistently predict colonization resistance. In light of such results, resistance is likely a product of the overall composition of the community.

In the experiments we have described thus far, we allowed the gut microbiota to recover for 24 h before challenging them with C. difficile. Several studies have demonstrated that perturbed communities can return to a “healthy” state in which resistance to C. difficile is restored (3, 17). To test the effect of recovery on colonization and gain greater insights into the populations that confer colonization resistance, we allowed the microbiota of the mice that received the full metronidazole and ampicillin treatment to recover for an additional 5 days (see Fig. S3 in the supplemental material). Among the metronidazole-treated mice, those with extended recovery had a 1.86 × 10⁶-fold reduction in colonization (P < 0.001; Fig. 3). In addition, 7 of the 14 mice given the longer recovery period had no detectable C. difficile 24 h after challenge. We detected six OTUs that were differentially represented in the two sets of metronidazole-treated mice (Fig. 3). Most notable among these six OTUs were two OTUs that affiliated with a member of the Barnesiella (OTU 2) and Escherichia (OTU 3). The relative abundance of this Barnesiella-affiliated OTU increased with extended recovery, and the relative abundance of this Escherichia-affiliated OTU decreased. Similar to the metronidazole-treated mice, the ampicillin-treated mice that were allowed to recover for an additional 5 days before challenge had a significant decrease in C. difficile colonization (P = 0.03). As before, we observed a similar increase and decrease in relative abundances for Barnesiella (OTU 2)- and Escherichia (OTU 3)-affiliated OTUs. However, untreated, fully resistant mice harbored significantly lower levels of the Barnesiella-affiliated OTU (OTU 2). Rather, untreated mice had high levels of various Porphyromonadaceae-affiliated OTUs (Fig. 1). These findings further confirm the context dependency of colonization resistance suggested by the results of our titration experiments.

To identify bacterial taxa that could be associated with resistance or susceptibility to C. difficile across the three sets of experiments, we measured the correlation between the relative abundance of each OTU on the day of inoculation with the level of C. difficile colonization 24 h later (Fig. 4). OTUs associated with providing resistance against C. difficile (n = 22) outnumbered those associated with susceptibility (n = 9). The Porphyromonadaceae-affiliated OTUs (ρ_average = −0.52; n = 11 OTUs) were consistently associated with low levels of C. difficile colonization. Among the three Proteobacteria-affiliated OTUs with a significant positive association with C. difficile colonization, the strongest was affiliated with the Escherichia (OTU 3; ρ = 0.54). By performing an OTU-based analysis, we were able to observe intrafamily and genus differences in association with C. difficile colonization. For example, the Lachnospiraceae have been associated with protection against C. difficile (18). Although within the Lachnospiraceae family we observed three OTUs that were associated with low levels of C. difficile colonization, one OTU was associated with high levels of C. difficile. In addition, we observed three significantly correlated Lactobacillus-affiliated OTUs (family Lactobacillaceae), two of which were associated with low levels of C. difficile and one was associated with high levels. The broad taxonomic representation of OTUs associated with low levels of C. difficile again suggests that a diverse community may be advantageous in preventing C. difficile colonization.

These three sets of experiments demonstrated that in certain contexts, individual OTUs could be associated with C. difficile colonization, but in other contexts, those OTUs had the opposite or no association. This suggests that colonization is a phenotype that is driven by multiple populations that act independently and possibly in concert to resist colonization. Correlation-based analyses cannot predict these types of context dependencies because they do not take into account the nonlinearity and statistical interactions between populations. Therefore, we used a regression-based random forest machine learning algorithm to predict the level of C. difficile colonization observed in the three sets of experiments using the composition of the microbiota at the time of challenge as predictor variables. The model explained 77.2% of the variation in the observed C. difficile colonization levels (Fig. 5). When we included only the top 12 OTUs based on the percent increase in the mean squared error when each OTU was removed, the resulting model explained 77.1% of the variation in the observed C. difficile colonization levels. The OTUs that were ranked as being the most important in defining the random forest model further validated the observations from the correlation-based analysis (see Fig. S4 in the supplemental material). According to the random forest model, colonization resistance was associated with OTUs that affiliated with the Porphyromonadaceae (OTU 15, 10, 6, 18, and 11), Lachnospiraceae (OTU 25), Lactobacillus (OTU 23), Alistipes (OTU 12), and Turicibacter (OTU 9) (Fig. 6). A loss in these populations, concurrently with a gain in OTUs affiliated with the Escherichia (OTU 3) or Streptococcus (OTU 90), was associated with an increased susceptibility to infection (Fig. 6). As we observed in the titration experiments, the relationship between an Akkermansia-affiliated OTU (OTU 4) and C. difficile indicated that wide variation in the relative abundance of Akkermansia was associated with different levels of C. difficile. There were different abundances of the Akkermansia-affiliated OTU in mice regardless of the level of C. difficile colonization. Finally, as indicated by the number of OTUs with relative abundances below the limit of detection, those mice could harbor different levels of C. difficile. These observations bolster the hypothesis that colonization resistance is context dependent.

We used 5- to 8-week-old C57BL/6 mice obtained from a single breeding colony maintained at the University of Michigan for all of our experiments. These mice were reared under specific-pathogen-free (SPF) conditions within the animal facility at the University of Michigan. All animal-related protocols and experiments were approved by the University Committee on Use and Care of Animals at the University of Michigan and carried out in accordance with the approved guidelines.

Mice were administered one of seven different antibiotics, including cefoperazone, vancomycin, metronidazole, streptomycin, ciprofloxacin, ampicillin, and clindamycin (Table 1). The route of administration depended on the specific antibiotic. Cefoperazone (0.5, 0.3, or 0.1 mg/ml), vancomycin (0.625, 0.3, or 0.1 mg/ml), streptomycin (5, 0.5, or 0.1 mg/ml), metronidazole (0.5 mg/ml), and ampicillin (0.5 mg/ml) were all administered in the mouse drinking water for 5 days. Ciprofloxacin (10 mg/kg of body weight) was administered via oral gavage, and clindamycin (10 mg/kg) was administered via intraperitoneal injection. Mice that had not received antibiotics were used as negative controls for these experiments, because C. difficile is unable to colonize mice that are not perturbed by antibiotics. The “No antibiotics” group in Fig. 1 and Fig. S1 in the supplemental material collectively refer to the microbiota of these untreated animals prior to their challenge with C. difficile, as well as the microbiota of animals prior to treatment with their respective antibiotic (i.e., the baseline for antibiotic-treated mice).

All antibiotic-treated mice were given 24 h to recover with untreated drinking water prior to C. difficile challenge. C. difficile strain 630Δerm spores were used in all experiments. Spores were prepared from a single large batch whose concentration was determined in the week prior to each C. difficile challenge (27). Spores were stored long term at 4°C. On the day of challenge, 10³ C. difficile spores were administered to mice via oral gavage. Immediately following this challenge, the remaining C. difficile inoculum was diluted in a series and plated to confirm the correct dosage.

Fecal samples were freshly collected for each mouse on the day of C. difficile challenge. On the day after challenge, another fecal sample was weighed and diluted under anaerobic conditions with anaerobic phosphate-buffered saline (PBS). The number of CFU was counted following 24-h growth on TCCFA plates at 37°C under anaerobic conditions (28).

Total bacterial DNA was extracted from each stool sample collected prior to challenge using the MOBIO PowerSoil-htp 96-well soil DNA isolation kit. We generated amplicons of the V4 region within the 16S rRNA gene and sequenced the fragments using an Illumina MiSeq as previously described (29).

These sequences were curated using mothur (v.1.35) as previously described (29, 30). Briefly, sequences were binned into operational taxonomic units (OTUs) using a 3% dissimilarity cutoff. Taxonomic assignments were determined by using a naive Bayesian classifier with the Ribosomal Database Project (RDP) training set (version 10) requiring an 80% bootstrap confidence score (31). In parallel to the fecal samples, we also sequenced a mock community where we knew the true sequence of the 16S rRNA gene sequences. Analysis of the mock community data indicated that the error rate following our curation procedure was 0.02%.

Complete scripts for regenerating our analysis and this paper are available at the online repository for this study (https://github.com/SchlossLab/Schubert_AbxD01_mBio_2015). Comparisons between the antibiotic-treated communities were made by calculating dissimilarity matrices based on the metric of Yue and Clayton (32). To avoid biases due to uneven sampling, the dissimilarity matrices were calculated by rarefying the samples to 1,625 sequences per sample. We then used analysis of molecular variance (AMOVA) to test for differences in community structure using 10,000 permutations (33). OTU-based analyses were performed using R (v.3.1.2). After subsampling the OTU frequency data to 1,625 sequences per sample, OTUs were considered for analysis if their average relative abundance within any treatment group was at least 1% (n = 38 OTUs). All OTU-by-OTU comparisons were performed using the Kruskal-Wallis rank sum test, followed by pairwise Wilcoxon rank sum tests. Comparison of log (base 10)-transformed C. difficile CFU/g feces between experimental groups was calculated using the Kruskal-Wallis rank sum test, followed by pairwise Wilcoxon rank sum tests. Spearman rank correlation analysis was performed between OTU counts and C. difficile CFU/g feces. All P values were corrected using a Benjamini and Hochberg adjustment with an experiment-wide type I error rate of 0.05 (34). Random forest regression models were constructed using the randomForest R package using 10,000 trees (35). To construct each tree, two-thirds of the samples were randomly selected to train the model, and one-third of the samples were selected to test the model. The regression was performed using the log (base 10) transformation of the number of CFU/g fecal material as the dependent variable and the 38 OTUs as predictor variables.

All 16S rRNA gene sequence data and metadata are available through the Sequence Read Archive under accession no. SRP057386.