Non-responder phenotype reveals apparent microbiome-wide antibiotic tolerance in the murine gut

Antibiotic duration experiment

Twenty-eight-week-old female C57BL/6J mice from the same birth cohort were co-housed (5–6 mice per cage) prior to beginning the experiment, and then separated into individual cages 1 week prior to antibiotic treatment. Singly housed mice were exposed to 0.5 mg/mL³³ cefoperazone in their drinking water for 0, 2, 4, 8, or 16 days (Fig. 1A). Based on the literature, we calculated the minimum dose of cefoperazone based on the mean and standard deviation of water consumption by C57BL/6J mice ((7.7, mp 0.3,{mathrm{mL}}) per 30 g of body weight)⁴⁴. If the heaviest mouse in our study (~22 g) consistently consumed water at 2 SD below the mean (i.e. 5.5 of 0.5 mg/mL cefoperazone), they would still receive 125 mg/kg/day of cefoperazone, which is within the therapeutic dosing range for humans (100–150 mg/kg/day; although cefoperazone is administered to humans via intravenous injection)⁴⁵.

Power analyses were performed by sampling from beta-binomial distributions fit to an independent data set and processed with the same protocol as the data presented in this manuscript (see “Methods”). Beta-binomial distributions have recently been shown to recapitulate 16S amplicon data structure well and are highly flexible distributions that can account for the observed overdispersion in microbiome data without making assumptions about taxon–taxon correlations made by Dirichlet-multinomial distributions⁴⁶. The resulting abundance tables were used to evaluate the expected power of beta-diversity and differential taxon abundance tests as a function of sample number and effect size. From the power curves, we concluded that we can reliably detect differences in beta-diversity (PERMANOVA) with an R² as low as 0.05 and with as few as five samples per group (Fig. S1). Furthermore, we estimate that taxon abundance differences larger than twofold can be reliably detected with 5–10 samples per group using beta-binomial likelihood ratio test (abbreviated beta-binomial LRTs from here on, Fig. S1). 16S amplicon sequencing of the duration experiment demonstrated that most of the cefoperazone-treated mice showed altered gut microbiome communities during antibiotic treatment (Fig. 1). These mice showed pronounced turnover in community composition at the phylum level, with a near-complete loss of Bacteroidetes and Firmicutes (beta-binomial LRT false discovery rate FDR-corrected p < 1e−9) and a dramatic enrichment of Proteobacteria and Cyanobacteria (beta-binomial LRT FDR-corrected p < 1e−11; Fig. 1B). SILVA annotations showed that 98% of Proteobacteria and Cyanobacteria reads were identified as being derived from organelles (i.e. mitochondria and chloroplasts). Thus, these reads were likely derived from plant chloroplasts and mitochondria in the diet or host mitochondria. The observed shift towards mitochondrial and chloroplast reads in most of the antibiotic-treated samples is likely a consequence of antibiotic-induced collapse of bacterial biomass, which would increase the detection of the background contaminants, such as host- and diet-derived organelles. However, 6 of the 16 cefoperazone-treated mice in this duration experiment did not exhibit an enrichment in mitochondrial and chloroplast reads during antibiotic exposure (Fig. 1B, C). Thus, the microbiota in these mice appeared to be antibiotic tolerant. Consequently we designated individual mice showing [mitochondria+chloroplast] relative abundance (geqq 10%) (i.e. greater than the highest observed fraction observed in untreated controls) during antibiotic exposure as antibiotic “responders”, whereas mice showing [mitochondria+chloroplast] relative abundance <10% during antibiotic exposure were designated as antibiotic “non-responders”. The only duration where all mice responded to antibiotic treatment was the 4-day exposure (Fig. 1B). Overall, duration of exposure had no significant influence over the frequency of non-responder phenotypes (Fisher’s exact test p = 0.44). Prior to antibiotic treatment, there were no significant beta-diversity differences between responder and non-responder mice (PERMANOVA R² = 0.06, p = 0.37). Additionally, we found no differences in phylum-level abundances between controls, responders, and non-responders before antibiotic exposure (all beta-binomial LRT FDR-corrected p > 0.8). The microbiome composition of non-responder mice during exposure was similar to untreated control mice at the phylum level, with only the Tenericutes phylum showing a significantly lower abundance in non-responders (beta-binomial LRT FDR-corrected p = 0.007, Fig. 1B).

Initially, we had predicted that duration of exposure would be positively correlated with within-host amplicon sequence variant (ASV) extinction (i.e. ASVs present within a mouse initially, but not at the end of the experiment). To quantify the persistence, gain, or loss of ASVs, we tracked the presence or absence of sequence variants within the same mouse over time. ASVs present within a mouse at the first and final time points of the study were considered to be “persistent”, ASVs present in a mouse at the beginning of the experiment and missing at the last time point were considered “lost”, and ASVs absent at the beginning and present at the end were considered “gained”. Treatment duration had no significant effect on persistence, loss, or gain of ASVs (ANOVA p > 0.1). There was a significant increase in the proportion of lost ASVs and a significant decrease in persistent ASVs in responder mice, when compared to untreated control mice (Mann–Whitney U p < 0.02, Fig. 1E). Non-responder mice, however, showed no significant differences from controls in ASVs gained, lost, or persistent (all Mann–Whitney U p > 0.1, Fig. 1E). Thus, non-responder microbiota were apparently protected from phylum-level collapse of gut bacterial community structure and ASV loss following antibiotic treatment.

Seaweed diet experiment

Twenty-eight seven-week-old female C57BL/6J mice from the same birth cohort were co-housed prior to beginning the experiment (5–6 mice per cage), and then separated into individual cages 1 week prior to dietary treatments. Mice were assigned to four treatment groups: eight mice in group A (seaweed−/antibiotic+), six mice in group B (seaweed−/antibiotic−), eight mice in group C (seaweed+/antibiotic+), and eight mice in group D (seaweed+/antibiotic−). We included a larger number of replicate mice in the antibiotic treatment groups because we wanted to capture as many non-responders as possible. Half of the mice (treatment groups C and D) were given a 1% seaweed in normal chow diet and the other half (groups A and B) received a normal chow diet for 20 days (Fig. 2A). All mice were put on the same normal chow diet for six days prior to antibiotic treatment to rule out any potential direct effects of seaweed on antibiotic absorption during antibiotic treatment (e.g. potential binding and inactivation of the antibiotic by seaweed-derived compounds). On day 26, all mice continued on a normal chow diet and mice in treatment groups A and C were given 0.5 mg/mL cefoperazone in their drinking water for a period of 6 days, with mice in groups B and D acting as untreated controls (Fig. 2A). We hypothesized that prior exposure to antimicrobial secondary metabolites in red seaweed, such as polyphenols, bromophenols, and terpenes^37,38,39,47, may “harden” the gut microbiota against future damage from antimicrobials, and perhaps promote tolerance of cefoperazone. Stool pellets from the full set of replicate mice were sampled each day and frozen at −80 °C in glycerol with 0.1% l-cysteine (16 antibiotic-treated mice and 12 control mice) and a subset of these samples were processed for sequencing (Figs. 2A and S2).

Seaweed treatment had a very minor impact on the composition and diversity of mouse gut microbiomes (Fig. S2), similar to what we had observed previously³⁷. We identified the same non-responder and responder phenotypes as in the duration experiment, with 4 of the 16 mice exhibiting the non-responder phenotype (i.e. (geqq 10%) [mitochondria + chloroplast]; Fig. 2B). The seaweed diet had no effect on the frequency of the non-responder phenotype (Fisher’s exact test p = 1.0). To validate that the observed enrichment in organelle relative abundances indeed corresponded to a lower bacterial biomass, we measured total 16S gene copy numbers in each sample (i.e. a proxy for bacterial biomass; Fig. 2C). As expected, 16S copy number was inversely proportional to the fraction of mitochondria and chloroplasts (Spearman rho = −0.84, p = 2.7e−8). In particular, we found that responder mice showed lower fecal bacterial biomass following cefoperazone treatment than controls (Mann–Whitney U p = 4.4e−6) and non-responders (Mann–Whitney U p = 0.004), while non-responder microbiomes did not differ significantly from controls in biomass levels during antibiotic exposure (Mann–Whitney U p = 0.6, Fig. 2C). Thus, it indeed appears that the absence of appreciable bacterial biomass in a mouse gut results in an enrichment for host and dietary contaminants in stool 16S amplicon sequencing data.

We observed similar phylum-level compositions between controls and non-responders during antibiotic exposure, with a slight decrease in Bacteroidetes and a loss of Tenericutes in non-responders (beta-binomial LTR FDR-corrected p < 2e−3), and small differences in beta-diversity between controls and non-responder samples during treatment (R² = 0.07, PERMANOVA p = 0.03, all beta-binomial LRT p < 2e−3, Fig. 2D, E). As mentioned above, responder mice showed the same transition to Cyano- and Proteobacteria dominance and were well-separated from controls and non-responders in beta-diversity space (R² = 0.2, PERMANOVA p = 0.001, Fig. 2D, E). Seaweed treatment did not affect beta-diversity (R² = 0.01, PERMANOVA p = 0.3).

To exclude potential differences between responders and non-responders prior to antibiotic treatment, we sequenced an additional set of 44 archived fecal samples in a second batch, which included 16 non-responder, 16 responder, and 12 control samples, before and during antibiotic exposure (see “Methods”, Fig. S3). We observed no significant difference in beta-diversity between controls, responders, and non-responders prior to antibiotic treatment (PERMANOVA R² = 0.12, p = 0.14). Again all sample groups were similar at the phylum level prior to treatment (all beta-binomial LRT p > 0.36), with some slight variation in the abundance of the Erysipelatoclostridium genus between groups (beta-binomial LRT p = 0.02). Consistent with our prior data, we saw a large shift in beta-diversity during treatment between responders and non-responders (R² = 0.41, PERMANOVA p = 0.001, Fig. S3). Thus, the compositional shift in phylum abundances we observed was indeed triggered by antibiotic treatment and could not be explained with pre-existing differences in the microbiome composition (Fig. S3). Interestingly, one mouse initially classified as non-responder (A4) based on data from day 31 clustered with responders on day 29 (Fig. S3), which suggests that responders are able to transition to non-responder phenotypes during the course of antibiotic treatment.

We determined the abundance of cefoperazone in a subset of stool and blood plasma samples by Selected Reaction Monitoring (SRM), a targeted, quantitative mass spectrometry technique (Fig. S4). We observed that cefoperazone reached concentrations well above its observed MIC₅₀ for a wide range of 357 anaerobic bacterial strains⁴⁸ in responder mice during antibiotic treatment (Fig. S4A). Two days post antibiotic treatment, average stool antibiotic concentrations were lower in non-responders, although concentrations remained at similar levels as responders in at least one-third non-responder mice included in the sample subset for cefoperazone quantification (Fig. S4A). During treatment, cefoperazone concentrations in blood plasma were orders of magnitude lower than in stool samples (<10 ng/mL), and we observed a ~3-fold range in blood concentrations across antibiotic-treated responder mice (Fig. S4B). Due to time constraints and animal welfare concerns, blood collection was done by cheek bleed during antibiotic treatment on only the first three replicate mice from each treatment group (A1–3, B1–3, C1–3, and D1–3). Unfortunately, none of these replicate mice ended up being the non-responders. Furthermore, 16S and metatranscriptome data generation was prioritized, and all non-responder stool samples taken during antibiotic treatment were used up to generate DNA and RNA; hence, we lack cefoperazone data on those samples. However, as mentioned above, the concentration of cefoperazone was lower in two-thirds of the analyzed non-responder samples (compared to responder samples) taken 2 days following the end of antibiotic treatment, which suggests that most non-responders (but not all) may experience a reduced concentration of antibiotic in situ (Fig. S4). Whether these phenotypes are driven by host- or microbiome-associated heterogeneity, it is crucial that we understand the frequency and characteristics of this non-responder phenotype, especially in light of the fact that the cefoperazone treatment used in this study is an established experimental model of C. difficile infection in mice^33,34.

Temporal dynamics following antibiotic treatment

Shannon diversity declined for at least 2 days after antibiotic treatment in responder mice, and continued to decline for up to 4 days depending on the duration of treatment (Fig. 3A). Despite greater loss of species in responder mice (Fig. 1D), overall alpha diversity tended to recover over time in these mice after cessation of antibiotic treatment (2 days after treatment vs. 6–8 days after treatment Mann–Whitney p = 0.03, n = 18). Non-responder and control mice maintained relatively stable alpha-diversities over time (2 days after treatment vs. 6–8 days after treatment Mann–Whitney p = 0.93, n = 22), although antibiotic-treated, non-responder microbiota showed slightly lower alpha-diversities than controls (controls vs. non-responders Mann–Whitney p = 8e−8, n = 80; Fig. 3A). Despite the resilience of Shannon diversity in the responder mice over time, only a small number of these mice showed recovery of Bacteroidetes ASVs (Fig. 3B). In control and non-responder mice, Bacteroidetes was the dominant phylum over the entire time series. However, Firmicutes became the dominant phylum in responder mice following antibiotics, and in many mice there appeared to be a persistent loss of the Bacteroidetes phylum following recovery (Fig. 3B). Seaweed dietary treatment appeared to contribute somewhat to this loss in resilience, with none of the seaweed-fed responder mice showing recovery of the Bacteroides phylum (Fig. S2).

Mouse weights were measured daily in the seaweed diet experiment to ensure that there were no systematic differences in food and water consumption among mice. We saw no evidence for weight loss in antibiotic-treated mice relative to controls during antibiotic treatment, which would have suggested reduced water consumption (Wald test p = 0.17 and 0.4 for non-responders and responders, respectively). Furthermore, we saw no significant difference in weight change between responder and non-responder mice (Wald test p = 0.07), suggesting that patterns of food and water consumption were similar between these groups. While we saw no major differences in the gut microbiome structure between control diet (i.e. normal chow) mice and 1% seaweed-fed mice (Fig. 2C and Fig. S1), we did observe a difference in mouse weight loss several days after the end of antibiotic treatment (Fig. 3C). All responder mice showed a transient weight-drop between day 35 and 40 (Mann–Whitney U p = 0.03 for responders vs. controls; Fig. 3C). This weight-drop was more pronounced in mice-fed seaweed compared to mice on a normal chow diet (Mann–Whitney p = 0.02, Fig. 3C). However, none of these mice showed signs of illness or diarrhea. We do not have an explanation for this synergistic effect between seaweed diet and cefoperazone treatment on transient weight loss in mice several days following the cessation of antibiotic treatment, but we believe this to be an interesting research avenue to explore further.

Non-responder microbiomes exhibit an antimicrobial tolerance transcriptional program

To evaluate whether the occurrence of the non-responder phenotype might be associated with changes in gene transcription in the gut, we performed RNA sequencing on samples from 10 mice before antibiotic treatment (days 20 and 25) and 7 mice during antibiotic treatment (day 29). Because there was almost no bacterial biomass in responder fecal samples during antibiotic treatment (and, consequently, not enough RNA for sequencing), we compared non-responders to untreated controls on day 29. After RNA extraction, ribosomal depletion, and sequencing to a mean depth of 20 million reads per sample, we identified around 800,000 unique transcripts by de novo assembly (see “Methods”) ranging from 111 to >26,000 bp in length (longer contigs were polycistronic; see Fig. S5 for length and coverage distributions).

Power analyses were performed by sampling from negative binomial distributions fit to an independent metatranscriptomic sequencing data set and calculating power with DESeq2 (see “Methods” for more details). Due to the pooled variance inference in DESeq2, differential expression could be detected with as few as six samples (three per group) with an average log2-fold change of 2.5 (uniform between 0 and 5) as long as the mean log2 expression of the transcript was at least 5 (Fig. S1).

Transcripts were collapsed to orthologous protein clusters by alignment to the M5NR database⁴⁹. We were able to assign 61% of the original transcripts to functions in the SEED subsystem database⁵⁰. This allowed us to collapse transcript counts for each sample into SEED clusters, which yielded a total of 53,877 unique orthologous protein clusters. The majority of SEED clusters were detected in all 17 RNA-seq samples (Fig. S6). Control and non-responder communities could be easily distinguished by SEED cluster counts during antibiotic exposure (see Fig. 4A). In particular, about 21% of the variance in functional expression could be explained by non-responder vs. control status (Euclidean PERMANOVA p = 0.003) compared to 7% of explained variance in 16S beta-diversity (Bray-Curtis PERMANOVA p = 0.03). Thus, transcriptional differences appear to define the non-responder phenotype more than changes in community composition. Prior to antibiotic exposure, responder samples could not be distinguished from control samples (PERMANOVA p = 0.3).

After filtering out low abundance functions, differential expression testing between controls and non-responder communities was performed for each of the three time points sampled (see “Methods”). We observed that fewer than 0.5% of the SEED clusters were differentially expressed at an FDR-corrected p ≤ 0.05 2 days before antibiotic exposure (day 25), which fits our null-expectation that non-responder and responder mice are functionally similar before exposure (Fig. 4B). However, following antibiotic exposure (day 29), 27% of all SEED clusters were differentially expressed between untreated controls and non-responder communities (see Fig. 4C). This indicated a global transcriptional shift in non-responder microbiomes, mostly characterized by an upregulation of several functional groups in the non-responder mice (blue dots on left side of Fig. 4C).

The transcriptional program was most prominently characterized by an upregulation of efflux transporters and other antibiotic resistance defense mechanisms, and a down-regulation of motility and respiratory functions (Fig. 5). For instance, the SEED subpathway “Transporters in Models” was the most prominent subpathway in the differentially expressed functions, containing 82 significant hits (DESeq2 FDR-corrected p ≤ 0.05). Most of the significantly upregulated functions in the “Virulence, Disease and Defense” superpathway were also related to efflux pumps and their regulation (Fig. 5A). We also found large differences in respiratory pathways, albeit with a mixed pattern of up- and down-regulation (Fig. 5B). Some of these key respiratory pathways were downregulated by one to two orders of magnitude in the non-responder mice, which suggests an overall down-regulation of ATP synthesis. We found 146 differentially expressed protein clusters associated with carbohydrate metabolism, although these pathways did not show a clear pattern of regulation, with only 74/146 clusters having higher abundances in untreated controls. Additionally, we observed that 17 of the 19 flagellar motor and chemotaxis proteins were downregulated in the non-responder mice (Fig. 5C). All differentially expressed functions in the “Membrane Transport” superpathway were strongly upregulated in the non-responder mice, including components of TonB, which is known to be necessary for efflux transporter function⁵¹ (Fig. 5D). Together, these data are consistent with previous reports that upregulation of efflux transporters is accompanied by a concomitant reduction in growth rate^27,52. Indeed, the most striking respiratory difference we observed related to potential growth rate reduction was the down-regulation of three acetyl-CoA synthases, which were some of the most highly expressed functions in the untreated mice (Fig. 5E). Finally, we observed the upregulation of the entire vancomycin resistance locus, including the three efflux pumps Vex1–3 and the two-component system VncR and VncS (Fig. 5E). The induction of vancomycin cross-resistance by β-lactams has been described before^53,54 and might indicate that these loci confer general efflux-based tolerance to a range of antibiotics.

Conclusion

We found that nearly one-third of mice exposed to a high dose of the cephalosporin antibiotic cefoperazone, commonly used in mouse experiments to promote C. difficile colonization^33,34, did not exhibit the expected gut microbiome community turnover, biomass collapse, or species loss. The frequency of this non-responder phenotype did not depend on duration of antibiotic exposure or on prior exposure to antimicrobial phytochemicals present in seaweed, but did appear to increase as the concentration of cefoperazone in drinking water declined, as shown previously³³.

Despite very minor changes in community composition and ecological diversity between untreated and non-responder mice, we observe a striking difference in microbiome gene expression between these groups. Non-responder microbiomes show substantial down-regulation of central metabolism and motility, and upregulation of antimicrobial tolerance mechanisms. This combination of increased antibiotic tolerance and quiescence may protect gut communities from the extensive ecological damage observed in responder microbiomes. Furthermore, this may help to protect the gut from pathogen invasion by maintaining a more intact gut microbiota. However, further studies will be required to unravel the potential interplay between non-responder phenotypes and susceptibility to pathogen invasion.

We observed heterogeneity in stool and plasma cefoperazone concentrations, despite no significant differences in mouse weight gain between responders and non-responders (i.e. a proxy for changes in water or food consumption that might have been due to treatment). However, the average cefoperazone concentration in non-responder feces was lower than the average concentration in responder feces two days after antibiotic treatment ended, although one of the non-responder samples still showed cefoperazone levels equivalent to those measured in the responder samples. In one instance, we observed the same mouse switching between responder and non-responder status during the course of antibiotic treatment, which indicates that transitions between these states are possible.

While prior work has shown how isogenic sub-populations of cells and two-species communities can exhibit heterogeneous responses to antibiotics^26,27,28,29, the exact mechanisms underlying transitions to whole-community tolerance phenotypes in the mammalian gut are not yet clear and will require further study. Whether community-wide non-responder phenotypes are driven purely by phenotypic plasticity among isogenic populations of microbes, or whether they arise, in part, due to evolutionary adaptation, remains an open question. Prior work has suggested that antibiotic tolerance can facilitate the evolution of antimicrobial resistance⁵⁵, and exploring whether these non-responder phenotypes enable the fixation of resistance alleles is a topic for future population genomic work. Future work should also focus on identifying what factors tip microbiomes between non-responder and responder phenotypes, potential hysteresis of these phenotypes, and whether or not this transition point can be manipulated to protect commensal microbiota from antibiotic assault.

Source: Ecology - nature.com