Dissecting the collateral damage of antibiotics on gut microbes

Antibiotics are used for fighting pathogens, but also target our commensal bacteria as a side effect, disturbing the gut microbiota composition and causing dysbiosis and disease1-3. Despite this well-known collateral damage, the activity spectrum of the different antibiotic classes on gut bacteria remains poorly characterized. Having monitored the activities of >1,000 marketed drugs on 38 representative species of the healthy human gut microbiome4, we here characterize further the 144 antibiotics therein, representing all major classes. We determined >800 Minimal Inhibitory Concentrations (MICs) and extended the antibiotic profiling to 10 additional species to validate these results and link to available data on antibiotic breakpoints for gut microbes. Antibiotic classes exhibited distinct inhibition spectra, including generation-dependent effects by quinolones and phylogeny-independence by β-lactams. Macrolides and tetracyclines, two prototypic classes of bacteriostatic protein synthesis inhibitors, inhibited almost all commensals tested. We established that both kill different subsets of prevalent commensal bacteria, and cause cell lysis in specific cases. This species-specific activity challenges the long-standing divide of antibiotics into bactericidal and bacteriostatic, and provides a possible explanation for the strong impact of macrolides on the gut microbiota composition in animals5-8 and humans9-11. To mitigate the collateral damage of macrolides and tetracyclines on gut commensals, we exploited the fact that drug combinations have species-specific outcomes in bacteria12 and sought marketed drugs, which could antagonize the activity of these antibiotics in abundant gut commensal species. By screening >1,000 drugs, we identified several such antidotes capable of protecting gut species from these antibiotics without compromising their activity against relevant pathogens. Altogether, this study broadens our understanding of antibiotic action on gut commensals, uncovers a previously unappreciated and broad bactericidal effect of prototypical bacteriostatic antibiotics on gut bacteria, and opens avenues for preventing the collateral damage caused by antibiotics on human gut commensals.

(Extended Data Fig. 3a). The newly established MICs also correlate well with available data 79 on antimicrobial susceptibility from databases such as EUCAST 24 or ChEMBL 27 (r s =0.69 and 80 r s =0.64, respectively), despite differences in strains and media used (Extended Data Fig. 3b).

81
Importantly, this new dataset considerably expands the available MICs, as much as by 80% 82 for non-pathogenic bacteria (Fig. 1c, Extended Data Fig. 3c). Altogether, the initial screen 83 and the new MIC dataset provide high-resolution information on the target spectrum of 84 antibiotics on commensal gut microbes, which we explored further.

99
Macrolides showed a strong impact on gut commensals and inhibited all tested microbes 100 ( Fig. 1d), except for the opportunistic pathogen C. difficile, which was resistant to all tested 101 macrolides and clindamycin (Extended data Fig. 2, red box). This is in line with the 102 associated risk of C. difficile infection after macrolide/clindamycin treatment 31 . Finally, 8 of 103 the 9 tested tetracyclines inhibited nearly all tested microbes, which is surprising in the light 104 of the gut microbiota being considered as reservoir for tetracycline resistance genes 32 . 105 the screen (Fig. 1d, f). In addition, MICs allow for comparisons with clinical breakpoints, i.e. more resistant to most antibiotic classes than previously reported for pathogens (aerobic 110 growth, Mueller-Hinton agar). Tetracyclines were the exception, inhibiting commensals at 111 significantly lower concentrations than what is reported for pathogens (Fig. 1f). Thus, 112 commensals might be considerably less resistant to tetracyclines than previously anticipated 113 and suggested by the detection of tetracycline resistance elements in fecal metagenomes.

114
Recent in-vivo studies have shown that β-lactams and macrolides have a strong and 115 long-lasting collateral impact on the gut microbiota composition and thereby on host health 5-8 .

116
As β-lactams exhibited strain-specific effects (Extended Data Fig. 1, 2, 4) and are known to 117 kill bacteria (bactericidal), they could irrevocably deplete specific members of the gut 118 microbiota, thereby explaining their differential and long-lasting effects on the community 119 composition. On the other hand, macrolides uniformly targeted all tested gut commensals 120 (Fig. 1d) and are textbook bacteriostatic antibiotics, i.e. inhibit bacterial growth, but do not kill 121 (at least at high numbers). In this case, the long-term community composition change is 122 more difficult to rationalize, as all community members are inhibited, but should be able to 123 regrow once drug is removed. Similarly, tetracyclines, another class of bacteriostatic 124 antibiotics that acted on nearly all gut microbes we tested, have known gastro-intestinal side-125 effects 18 , which are indicative of gut microbiome dysbiosis. We thus wondered at which level 126 macrolides and tetracyclines exert a differential effect on gut microbes. Although traditionally 127 both clinical use 34-37 and basic research 38,39 heavily rely on this distinction between 128 bactericidal and bacteriostatic antibiotics, there are reports of drugs changing their killing 129 capacity depending on the organism, drug concentration or medium tested 40,41 (and 130 increased evidence from meta-analyses that the distinction may have little relevance to 131 clinical practice 42,43 ). We therefore hypothesized that this bacteriostatic/bactericidal divide 132 may be less rigid for gut commensals, which are more phylogenetically diverse than the few 133 pathogens usually tested, and hence provide a level where the effect of these drug classes 134 on gut microbes becomes differential.

135
The standard way to determine whether an antibiotic has bactericidal or bacteriostatic 136 activity is to calculate time-kill curves, where the bacterial survivors are counted on agar at (ranging from 5 to 24 hours), the number of colony forming units (CFU)/ml of culture 139 decreases by more than 99.9%, the antibiotic is considered bactericidal 40 . We assessed the 140 survival of 12 abundant gut microbes over a 5-hour treatment of either a macrolide 141 (erythromycin or azithromycin) or a tetracycline (doxycycline) at 5 x MIC (Fig 2a +

152
In parallel, we excluded that the differences in killing capacity were confounded by growth 153 rate, growth phase or MIC of the bacterial species tested (Extended Data Fig. 8). We also 154 noticed that B. vulgatus and B. uniformis cultures decreased density in the presence of 155 erythromycin (Fig. 2d). We confirmed by time-lapse microscopy that this was due to lysis.

163
Knowing that drug combinations often have species-specific outcomes 12 , we 164 reasoned that we could identify drugs that selectively antagonize the effect of antibiotics on Prestwick library to identify antagonizing compounds to erythromycin or doxycycline on two 167 abundant and prevalent gut microbes, B. vulgatus and B. uniformis (Fig. 3a, Extended Data 168 Fig. 9). Of the 19 identified hits (Fig. 3b, Suppl. Table 4), we tested the 14 candidates with 169 the strongest activity in a concentration-dependent manner (Extended Data Fig. 10a). Nine 170 retained antagonistic activity over a broader concentration range, which we confirmed by 171 checkerboard assays (Fig. 3c). The antidotes that showed the strongest antagonisms were 172 the anticoagulant drug dicumarol, and two non-steroidal anti-inflammatory drugs, tolfenamic 173 acid and diflunisal. While dicumarol rescued B. vulgatus from erythromycin and diflunisal 174 from doxycycline, tolfenamic acid was able to protect B. vulgatus from both drugs. In 175 addition, these interactions were able to at least partially rescue the killing of B. vulgatus by 176 erythromycin and doxycycline (Extended Data Fig. 10b). We then probed two of these drugs 177 for their ability to protect other abundant gut commensals and confirmed that both dicumarol 178 and tolfenamic acid were able to counteract erythromycin on several species (

194
Antibiotics with preferential killing of some species may be the most detrimental to our gut

252
prevalence was defined as the percentage of samples with nonzero abundance; a 253 prevalence cut-off of 1% was chosen to classify species into "rare" and "common" species.

254
For all species in the MIC dataset, we manually assessed their status as pathogenic or non-255 pathogenic species using encyclopaedic and literature knowledge. Pathogenic species that 256 occur in more than 1% of healthy people (i.e. are designated as "common") were classified 257 as "potentially pathogenic species" that can, for example, cause diseases in 258 immunocompromised patients.

260
Killing curves and survival assay 261 Cells were precultured as described in the growth conditions section before being diluted to 262 an OD 578 =0.01 and grown for 2 h at 37°C (unless specified otherwise). Next, cells were 263 cells were detected using this method, a bigger volume of culture (up to 2 ml) was plated to 268 be able to detect CFUs. Agar plates were incubated overnight at 37°C and colonies were 269 counted the next day, either manually, for low CFU numbers, or using the Analyze Particles

333
Analysis pipeline and hit calling. All growth curves within a plate were truncated at the 334 transition time from exponential to stationary phase and converted to normalized AUCs using 335 in-run control wells (no drug) as described before 4 .We then calculated z-scores based on 336 these normalized AUCs, removed replicates with 8-fold differences in z-scores to eliminate 337 noise effects, computed mean z-scores across the two replicates and selected combinations 338 with mean z-scores > 3. This selection included 19 potential antibiotic antagonists and we 339 followed up on 14 of them (7 potential erythromycin and 7 potential doxycycline antagonists

619
Correspondence and requests for materials should be addressed to typas@embl.de or 620 l.maier@uni-tuebingen.de. a. Overview of antibiotics tested in initial screen at 20 µM concentration 4 and validated by MIC determination in this study. b. Principal component analysis based on AUCs from the initial screen on the effects of antibiotics on gut commensals. Antibiotic classes drive some separation at the phylum level, e.g. beta-lactams separate Bacteroidetes and macrolides/lincosamides/streptogramins separate Proteobacteria. c. Comparison of MICs from this study to MICs available from public databases. Species are classified as "common" or "rare" if they are present in the gut microbiome of more or less than 1% of 727 healthy individuals, respectively (see Methods). d. For the main antibiotic classes from the screen, the numbers of inhibited strains are shown (N as in a). 40 strains tested in total at a 20 µM antibiotic concentration. Boxes span the IQR and whiskers extend to the most extreme data points up to a max of 1.5 times the IQR. e. Number of inhibited strains per (fluoro-)quinolone drug generation. Number of tested drugs per generation is indicated in brackets on x-axis labeling. Boxplots as in panel d. f. MICs of drug-species pairs for the main antibiotic classes measured in this study are depicted next to EUCAST clinical (susceptibility) breakpoints for pathogens. Numbers of drug-species pairs (MICs; colored) and of antibiotic per class (EUCAST clinical breakpoints; grey) are shown in brackets. Boxplots as in panel, d, y-axis is log2 scale. The survival of 12 abundant gut microbe species was measured after a 5-hour treatment with a 5-fold MIC of erythromycin (ERY), azithromycin (AZI) or doxycycline (DOX). The survival was assessed by counting CFUs/ml before and after antibiotic treatment. The number of CFUs/ml before treatment was set as 100%. The detection limit for each experiment (gray bar) and the bactericidal threshold (shaded area) are indicated. Species are plotted according to phylogeny (IQTree, Methods) and in bold are noted the species that are used in later panels. The graph shows the mean+SD of 3 independent experiments. b. Time-kill curves of B. vulgatus, R. intestinalis and F. nucleatum after antibiotic treatments. Survival was assessed by CFU counting over a 5 hour-treatment of ERY, AZI or DOX. This graph shows the mean±SD of 3 independent experiments. Nd: non-detectable. Time-kill curves for the other tested gut microbes can be found in Extended Data  Schematic illustration of the screen concept: searching for antidote compounds that antagonize the antibacterial effect of erythromycin or doxycycline on commensal but not on pathogenic bacteria. b. Z-scores on bacterial growth (based on areas under the curve (AUCs)) for combinatorial drug exposure with antibiotic (ERY or DOX) and FDA-approved drug. Compounds that successfully rescued B. vulgatus and/or B. uniformis growth in the presence of the antibiotic (z-score > 3) are indicated in gray. The strongest hits (circles) were validated further in concentration-dependent assays (Extended Data Fig. 10a). For each antibiotic and each strain, ~1200 drugs were tested in two replicates. Boxplots are defined as in Figure 1d. c. For 9 of the validated antagonists, 8 x 8 checkerboard assays were performed to determine concentration ranges of the antagonistic interaction. Heat maps depict bacterial growth based on normalized median of AUCs of 4 replicates. All interactions were antagonistic, and pairs tested further in other commensal species are framed in bold. d. Checkerboard assays confirm the ability of tolfenamic acid to protect further gut commensals from growth inhibition by erythromycin. Heat map as in c, but for 2 replicates. Antagonistic interactions are framed in red. e. Checkerboard of tolfenamic acid with erythromycin reveal neutral interactions in S. aureus and E. faecium (aerobic conditions). Heat maps as in c, based on at least two independent experiments with two technical replicates each. f. Tolfenamic acid concentration-dependent rescue of commensal growth at clinical relevant erythromycin concentrations based on AUCs (anaerobic conditions). Erythromycin still retains its activity against pertinent pathogens such as S. aureus, E. faecium and S. pneumoniae (aerobic conditions). 0.625 µM correspond to ~0.5 µg/ml erythromycin, which is in the range of the MIC breakpoints for Staphylococcus (1 µg/ml), S. pneumoniae (0.25 µg/ml) and Streptococci groups A, B, C & G (0.25 µg/ml). Error bars depict standard deviation.

Extended Data Figure 2 -MICs for 17 species on 35 antimicrobials
Heat map depicts MICs for each drug-strain pair in µg/ml. Heat map color gradient is adjusted to the MICs concentration range tested on the respective MIC test strip. Black depicts sensitivity and light grey indicates resistance. Mean values across two biological replicates are shown (Suppl. Table 3). C. difficile is particularly resistant to all tested macrolides and clindamycin (red box). Figure 3 -MIC dataset validates antibiotic sensitivity profiles from the screen dataset and is consistent with publically available MICs. a. Receiver operating characteristic (ROC) curve analysis was performed to evaluate sensitivity and specificity of the screen 4 using the MIC dataset. Results from the screen were considered as validated if MICs were below/above the 20 µM antibiotic concentration that was tested in the screen (allowing a two-fold error margin). N is the number of antibiotics that we tested both in the screen and determined MICs for, AUROC is the area under the characteristic ROC. TN denotes true negatives, FP false positives, TP true positives, FN false negatives. b. Comparison including Spearman correlation coefficients of the MICs from this study to MICs from the ChEMBL 27 and EUCAST 24 databases. Panels in the upper row: comparison between all MICs that are shared between the two indicated datasets. Panels in the lower row: comparison of the 69 MICs that are shared across all three datasets. Despite experimental differences, our MICs correlate well with available EUCAST/ ChEMBL data. c. Number of the sum of new (this study) and already available MICs (EUCAST/ ChEMBL) per drug according to antibiotic class and prevalence/virulence of the bacterial species. The new dataset expands MICs across the board and specifically fills the knowledge gap on nonpathogenic species. The left panel shows an overlay of phase contrast and fluorescence microscopy images of propidium iodide (PI)-stained E. coli ED1a before and 5 hours after ERY, AZI or DOX treatments. The number of cells on each frame has no meaning, as cultures were concentrated before imaging; the scale bar is 10 µM. The right panel shows the corresponding quantification of live/dead-stained cells by flow cytometry with Syto9 on the xaxis (live cells) and PI on the y-axis (dead cells). Both the total number of measured events (n) and the percentage of cells found in each quadrant are indicated on the graphs. Figure 7 -Effect of oxygen and strain specificity on survival after doxycycline treatment a. The survival of E. coli ED1a was assessed after a 5-hour treatment with 5-fold MIC of DOX in the presence or absence of oxygen. Killing was similarly effective in both conditions. b. The survival of E. coli ED1a and E. coli BW25113 were assessed after a 5-hour treatment with 1, 2 and 5-fold MIC of DOX in MGAM medium in anaerobic conditions. The lab strain is more resistant to killing with doxycycline becoming boarder-line bactericidal at higher MICs. Figure 8 -Assessing potential confounding factors for the killing capacities of erythromycin, azithromycin and doxycycline a. Scatter plot of individual bacterial growth rates and percentage survival after a 5-hour treatment with 5-fold MIC of ERY, AZI or DOX treatments. r indicates the Spearman correlation coefficient. Tested species are color-coded here and in all panel thereafter as indicated in the bottom of this figure. Positive correlations for macrolides were tested further in b to check if changing growth rate in same species affects percentage killed. b. The survival of B. fragilis (blue) and F. nucleatum (beige) were assessed after a 5-hour macrolide treatment (5-fold MIC of ERY and AZI) at either 30°C (slow growth) or 37°C (fast growth) to test the effect of slowing down growth on survival. No significant change observed. This graph shows the mean±SD of three independent experiments. c. Scatter plot of MICs and percentage survival after a 5-hour treatment with 5-fold MIC of ERY, AZI or DOX treatments. r indicates the Spearman correlation coefficient. Doxycycline exhibited a strong and significant anti-correlation, that is that species which were more sensitive to doxycycline (lower MIC) were not killed when they were treated with 5-fold MIC concentrations. Thus, we tested further whether increasing the drug concentration in some of those sensitive strains decreased the % of survival (panel d). d. The survival of B. fragilis (blue) and F. nucleatum (beige) were assessed after a 5-hour treatment with increasing concentrations of DOX (5, 10 or 20-fold of MIC) to test whether higher concentrations of DOX induced more killing. This seemed not be the case. This graph shows the mean±SD of three independent experiments. e. To evaluate whether outgrowth of stationary phase and homogeneity of population affected our results, we selected two slow-growing strains, E. rectale and R. intestinalis and grew for 2 or 3 hours after being diluted from an overnight culture to an of OD578 0.01. Both strains were then treated for 5 hours with 5-fold MIC of ERY, AZI or DOX and their survival was assessed to test the impact of the growth phase on the percentage survival. Although slight differences were observed and 3h grown cultures were killed more effectively (presumably because more cells had exited stationary phase and were growing exponentially by then), the general trends remained the same. If anything, this means that we are underestimating the killing for slow-growers, since we performed all other experiments with 2 hours outgrowth. This graph shows the mean±SD of three independent experiments. f. The survival of 8 selected gut microbes was measured after treating cells in exponential phase (E -2 hours after dilution from an overnight culture) or in stationary phase (S -overnight growth) with 5-fold MIC of ERY for 5 hours to test the impact of the growth phase on the percentage survival. As expected, survival is higher in stationary phase for half of the strains, but in some cases stationary phase cells were as or more sensitive than exponentially growing cells -this is the case for B. caccae and F. nucleatum. This graph shows the mean±SD of three independent experiments. g. Same as in f but with DOX. Similar effects observed as in f, with more than half of strains becoming more resistant in stationary phase.