Abstract

A few commonly used non-antibiotic drugs have recently been associated with changes in gut microbiome composition, but the extent of this phenomenon is unknown. Here, we screened more than 1,000 marketed drugs against 40 representative gut bacterial strains, and found that 24% of the drugs with human targets, including members of all therapeutic classes, inhibited the growth of at least one strain in vitro. Particular classes, such as the chemically diverse antipsychotics, were overrepresented in this group. The effects of human-targeted drugs on gut bacteria are reflected on their antibiotic-like side effects in humans and are concordant with existing human cohort studies. Susceptibility to antibiotics and human-targeted drugs correlates across bacterial species, suggesting common resistance mechanisms, which we verified for some drugs. The potential risk of non-antibiotics promoting antibiotic resistance warrants further exploration. Our results provide a resource for future research on drug–microbiome interactions, opening new paths for side effect control and drug repurposing, and broadening our view of antibiotic resistance.

Main

Pharmaceutical agents have both beneficial and undesirable effects. Studies on the mechanisms of action and off-target spectra of various drugs aim to improve their efficacy and reduce their side effects. Although many drugs have gastrointestinal side effects and the gut microbiome itself is pivotal for human health1, the role of the gut microbiota in these processes is rarely considered. Recently, consumption of drugs designed to target human cells and not microbes, such as antidiabetics (metformin2), proton pump inhibitors (PPIs)3,4, nonsteroidal anti-inflammatory drugs5 and atypical antipsychotics (AAPs)6, has been associated with changes in microbiome composition. A larger cohort study suggested that medication can alter gut microbiome composition more generally7. As it is unclear whether such effects are direct and go beyond the few drug classes studied, we systematically profiled interactions between drugs and individual gut bacteria. We aimed to generate a comprehensive resource of drug actions on the microbiome, which could facilitate more in-depth clinical and mechanistic studies, ultimately improving therapy and drug design.

A high-throughput drug screen on gut bacteria

To systematically map interactions between drugs and human gut bacteria, we monitored the growth of 40 representative isolates upon treatment with 1,197 compounds in modified Gifu anaerobic medium (mGAM) broth, which partially recapitulates the species relative abundances in human gut microbiomes8, under anaerobic conditions at 37 °C (Extended Data Fig. 1a). We used the Prestwick Chemical Library, which consists mostly of off-patent Federal Drug Administration (FDA)-approved compounds with high chemical and pharmacological diversity. Most compounds are administered to humans (1,079), and they cover all main therapeutic classes (Supplementary Table 1). Three quarters (835) of the compounds are human-targeted drugs (that is, have molecular targets in human cells), whereas the rest are anti-infectives: 156 with antibacterial activity (144 antibiotics, 12 antiseptics) and 88 effective against fungi, viruses or parasites (Fig. 1a). All compounds were screened at 20 μM, which is within the range of what is commonly used in high-throughput drug screens9.

Figure 1: Systematic profiling of marketed drugs on a representative panel of human gut microbial species.
Figure 1

a, Broad impact of pharmaceuticals on the human gut microbiota. Compounds from the Prestwick Chemical Library are divided into drugs used in humans, drugs used exclusively in animals (vet) and compounds without medical or veterinary use (non-drugs). Human-use drugs are further categorized according to targeted organism. Strain–drug pairs (that is, instances in which a drug significantly reduced the growth of a specific strain; see Methods) are highlighted with a vertical coloured bar in the matrix. Bacterial strains are sorted by drug sensitivity. The relative abundances of each strain in four cohort studies of healthy individuals are displayed on the right (boxes correspond to interquartile range (IQR) and central line to median relative abundance). b, Fractions of drugs with anticommensal activity by sub-category. Grey scale within bars denotes inhibition spectrum (the number of affected strains per drug). c, Correlation between species abundance in the human microbiome and drug sensitivity. For each strain (n = 40), the number of drugs that affect its growth is plotted against its median relative abundance in the human gut microbiome. Lines depict the best linear fit, rS the Spearman correlation and grey shading the 95% confidence interval of the linear fit. All drugs, and in particular human-targeted drugs, inhibit the growth of more abundant species more than that of less abundant species.

For our screen to be representative of the gut microbiome of healthy individuals, we selected a set of ubiquitous gut bacterial species (Supplementary Table 2). Prevalence and abundance in the human gut, and phylogenetic diversity, were our main selection criteria (Extended Data Fig. 1b), although we were occasionally constrained by strain unavailability or irreproducible growth in mGAM. In total, we included 40 human gut isolates from 38 bacterial species and 21 genera (Escherichia coli and Bacteroides fragilis were represented by two strains each), accounting together for 78% of the median assignable relative abundance of the human gut microbiome at genus level (60% at species level; Extended Data Fig. 1c). Most strains were commensals, covering 31 of 60 sequenced species detected at a relative abundance of 1% or more and prevalence of at least 50% in fecal samples from asymptomatic humans from three continents (Extended Data Fig. 1d). In addition, the set included four pathobionts (Clostridium difficile, Clostridium perfringens, Fusobacterium nucleatum and an enterotoxigenic strain of B. fragilis), a probiotic (Lactobacillus paracasei) and two commensal Clostridia species (C. ramosum and C. saccharolyticum). All 38 species are found in the gut of healthy individuals and are part of a larger strain resource panel for the healthy human gut microbiome8.

We screened all compounds in multiwell plates, measuring optical density over time to monitor growth, and quantifying the area under the growth curve (AUC) up to the time point at which controls with unperturbed growth transitioned to stationary phase (see Methods; Extended Data Fig. 2). We obtained at least three biological replicates per strain, and these replicates correlated highly (Extended Data Fig. 2c). We then tested for significant deviations from the normalized AUC distribution of samples with unperturbed growth, combining P values across replicates and correcting for multiple hypothesis testing on the complete matrix of compounds and strains (see Methods; Extended Data Fig. 2). Drugs that significantly reduced the growth of at least one strain (false discovery rate (FDR) < 0.01), were classified as hits with anticommensal activity (Supplementary Table 3a), reflecting their potential to modulate the human gut microbiota.

Of the 156 antibacterials tested, 78% were active against at least one species, typically with a broad activity spectrum (Fig. 1a, b). Inactive antibiotics belong mainly to the sulfonamides (which are inactive in our medium according to the manufacturer’s guidelines), aminoglycosides (which have compromised activity under anaerobic conditions10) and specific antimycobacterial drugs. Antibiotics are used to inhibit pathogens, but as expected, also target gut commensals. The medical importance of this collateral damage to the resident microbiome has recently been becoming clearer11. Nevertheless, to our knowledge, drug–microbiome species relationships have not previously been mapped at this scale.

Notably, 27% of the non-antibiotic drugs were also active in our screen. More than half of the anti-infectives against viruses or eukaryotes exhibited anticommensal activity (47 drugs; Fig. 1a, b). Antibacterial activity has been previously reported for many of these drugs, including the antifungal imidazoles12 (11 in our screen), but not for others (for example, the antivirals saquinavir and trifluridine). More noteworthy is the anticommensal activity of 203 (24%) of the human-targeted drugs. Most were effective against only a few strains, with the exception of 40 drugs that affected at least 10 strains. Fourteen of these had, to our knowledge, not been previously reported to have direct antibacterial activity (Supplementary Table 3b). Among known human-targeted drugs with anticommensal activity, auranofin has recently been reported to have broad-spectrum bactericidal activity13, and the ovulation stimulant clomiphene inhibits a conserved bacterial enzyme in the synthesis of an essential precursor for cell wall carbohydrate polymers14. Such drugs or their scaffolds can be used for repurposing towards broad-spectrum antibiotics, especially as many have minimal inhibitory concentrations (MICs) in the sub-microgram per millilitre range (Supplementary Table 4). By contrast, the microbial narrow-spectrum specificity of most human-targeted drugs could aid the development of microbiome modulators.

Bacterial species showed varied responses to drugs, with the abundant Roseburia intestinalis, Eubacterium rectale and Bacteroides vulgatus being the most sensitive, and γ-proteobacteria representatives being the most resistant (Fig. 1a). Overall, species with higher relative abundance across healthy individuals were significantly more susceptible to human-targeted drugs (P = 0.0012 based on Spearman correlation; Fig. 1c). This suggests that human-targeted drugs have an even larger impact on the gut microbiome, with key species related to healthy status15, such as major butyrate producers (E. rectale, R. intestinalis, Coprococcus comes) and propionate producers (B. vulgatus, Prevotella copri, Blautia obeum)16, and enterotype drivers (P. copri)17, being relatively more affected.

Dose relevance and validation of the drug screen

We sought to address how close the screening concentration (20 μM) was to drug concentrations in the terminal ileum and colon, where most gut microbes reside18. However, drug concentrations are systematically measured only in blood; there, human-targeted drugs have on average an order of magnitude lower concentrations than in our screen (Fig. 2a, Extended Data Fig. 3). We deduced colon concentrations on the basis of drug excretion patterns from published work, and small intestine concentrations on the basis of daily doses of individual drugs (Supplementary Table 1) and a measured example of duodenal concentrations for the well-absorbed drug posaconazole19 (see Methods). Based on these approximations, 20 μM was below the median small intestine and colon concentration of the human-targeted drugs tested here (Fig. 2a, Extended Data Fig. 3). Notably, human-targeted drugs that showed anticommensal activity had lower plasma and estimated small intestinal concentrations than ones with no such activity (Fig. 2a; P = 0.0061 and P = 0.0035, respectively, two-sided Wilcoxon rank sum test; we have fewer colon concentration estimates owing to data availability), suggesting that more human-targeted drugs would inhibit bacterial growth if probed at higher doses, closer to physiological concentrations. A case in point is metformin, which was recently identified as the key contributor to changes in the human gut microbiome composition of patients with type II diabetes2, but lacked anticommensal activity in our screen. Metformin reaches 10–40 μM in the plasma of treated patients with type II diabetes, but its small intestine concentration is 30–300-fold higher20, which matches our estimates of small intestine and colon concentrations (1.5 mM). When we probed for higher, more physiological intestinal metformin concentrations, 3 of 22 tested strains were inhibited at concentrations below 1.5 mM (Extended Data Fig. 4a).

Figure 2: Evaluating human-targeted drugs with anticommensal activity.
Figure 2

a, Estimated small intestine and colon concentrations and measured plasma concentrations of human-targeted drugs with (orange) and without (grey) anticommensal activity in our screen (see Methods; Extended Data Fig. 3). For both active and inactive compounds, the median estimated small intestine and colon concentrations are higher than the screened concentration (20 μM, black vertical lines), whereas plasma concentrations are lower. Non-hits in our screen generally reached higher plasma and small intestine concentrations (two-sided Wilcoxon rank sum test). Box plots: centre line, median; limits, upper and lower quartiles; whiskers, 1.5× IQR; points, outliers. b, Rarefaction analysis indicates that anticommensal activity would be discovered for more human-targeted drugs if we screened additional strains.

We also benchmarked our screen with an independent set of experiments, measuring IC25 (the drug concentration conferring 25% growth inhibition) for 25 selected drugs in a subset of up to 27 strains (see Methods). This analysis revealed excellent precision (94%), but slightly lower recall (85%) (Extended Data Fig. 5a, b). False negatives, that is, drugs with anticommensal activity missed in our screen, were due to specific chemicals that probably lost activity during screening (Extended Data Fig. 5d), and our stringent FDR cutoff for calling hits. Increasing this cutoff to 0.1 would almost double the fraction of drugs with anticommensal activity (Extended Data Fig. 5c). In addition, we found that more species were inhibited at higher concentrations (Extended Data Fig. 5d, Supplementary Table 4), and that IC25 values were mainly below the estimated gut concentrations and occasionally below plasma concentrations (Extended Data Fig. 6).

Furthermore, we screened only a representative subset of species, but the gut microbiome of an individual harbours hundreds of species and an even larger strain diversity21. Rarefaction analysis indicates that if more gut species were tested, the fraction of human-targeted drugs with anticommensal activity would increase (Fig. 2b).

In summary, we probed human-targeted drugs largely within physiologically relevant concentrations and our data are likely to under-report the impact of human-targeted drugs on gut bacteria.

Concordance with patient data

Having demonstrated that many human-targeted drugs inhibit gut bacteria in vitro at relevant doses, we searched for evidence that such effects manifest in vivo in the human gut. We reviewed all available clinical cohort data from metagenomics association studies and compared it to our screen if studies had enough statistical power and affected taxa that overlapped with those tested here. We found suitable studies for PPIs, AAPs, and seven further drugs, spanning altogether five different drug classes according to Anatomical Therapeutic Chemical (ATC) classification. All three PPI representatives in our screen exhibited broad anticommensal activity, similar to the microbiome changes that have been reported in patients taking PPIs3,4 (Fig. 3a): taxa with reduced abundance in patients exhibited reduced growth in our screen and taxa enriched in patients were rarely inhibited by PPIs in vitro (Extended Data Fig. 7a). This suggests that PPIs directly influence the gut microbiome composition, in addition to changing the stomach pH and thus affecting which bacteria reach the gut3,4. Concordance was similarly high for many microbe–drug associations identified in a large Flemish cohort7 for the immunosuppressive agent azathioprine, the antidepressant venlafaxine, the anti-inflammatory mesalazine, aminosalicylate, progesterone, oestrogens and amoxicillin; the only exception was another antibiotic, nitrofurantoin (Extended Data Fig. 7b, c). We also compared our data to a study that reported a reduction in Akkermansia levels in the gut of patients treated with AAPs6. Our screen included six of the ten AAPs investigated in that study. We found that Akkermansia muciniphila was more sensitive than other strains to these AAPs (P = 0.09; two-sided Wilcoxon rank sum test), while being more resistant to other human-targeted drugs (P = 0.0005, two-sided Wilcoxon rank sum test; Extended Data Fig. 7d). Finally, we found high concordance between a longitudinal microbiome study of patients taking metformin and our IC25 data for the same drug (Extended Data Fig. 4b).

Figure 3: Anticommensal activity of human-targeted drugs in vitro reflects patient data.
Figure 3

a, Changes in microbiome composition of patients taking PPIs are consistent with drug effects in our screen. Displayed are Spearman correlation coefficients between in vitro growth inhibition P values and changes in taxonomic relative abundance after PPI consumption for corresponding taxa from two studies (Twins UK4 and Dutch3 cohorts; 229 of 1,827 and 211 of 1,815 individuals had taken PPIs, respectively). The histogram represents the background distribution of correlations between the in vitro data for all human-targeted drugs and the in vivo response to PPIs; correlations with PPIs are highlighted by triangles. b, Human-targeted drugs with anticommensal activity in our screen had a significantly higher incidence of antibiotic-related side effects (orange trace shows cumulative distribution, n = 285 drug–side effect pairs) in clinical trials compared to drugs without activity (grey trace, n = 767; P = 0.002, two-sided Wilcoxon rank sum test). Dashed lines indicate the incidence of the same side effects upon placebo treatment, with no significant difference between active (n = 138) and inactive drugs (n = 474). c, Based on similarity to antibiotic-related side effects, we selected 26 candidate and 16 control drugs for testing for anticommensal activity. Although both candidate and control drugs inhibited bacterial growth at higher concentrations, candidate drugs had anticommensal activity at significantly lower doses than control drugs after normalizing for estimated intestine concentrations (P = 5.6 × 10−7, two-sided Wilcoxon rank sum test). Box plots as in Fig. 2a, n denotes number of drug-strain pairs.

Metagenomics association studies and our in vitro study have distinct limitations. We screened a subset of species, mostly one strain per species, out of the context of microbial communities and the host. Cohort studies can be underpowered or biased by methodological approaches and confounding factors, and may detect indirect effects. Nonetheless, we find high concordance between the effects of drugs in vitro and in humans, confirming clinical relevance and direct anticommensal activity for the aforementioned cases.

To assess the physiological relevance of our screen further, we investigated the registered side effects of these drugs in humans. We first identified side effects enriched in antibiotics for systemic use compared to those found in all other drugs in the SIDER database22. We identified 69 side effects that were enriched in antibiotics (see Methods; Supplementary Table 5). These antibiotic-related side effects occurred more often in clinical trials of human-targeted drugs with anticommensal activity than in trials of compounds that were inactive in our screen (P = 0.002, two-sided Wilcoxon rank sum test), whereas no significant difference was observed for placebo-treated patients (Fig. 3b). This suggests that the collateral damage of human-targeted drugs on gut bacteria can be detected by higher occurrences of antibiotic-like side effects in patients.

We then tested whether this side effect signature predicted anticommensal activity of human-targeted drugs, which we could have missed owing to the low drug concentration we used. We screened 26 candidate compounds that showed enrichment of antibiotic-related side effects and 16 that did not (control compounds) for effects on the growth of 18 bacterial strains (Extended Data Fig. 8), in concentrations up to 2.5 mM (Methods). Twenty-eight of these forty-two compounds inhibited the growth of at least one strain (Extended Data Fig. 8a–d), with both the fraction of active compounds and the number of affected strains being similar for both candidate and control compounds. However, when we normalized the measured IC25 by the estimated intestine concentration (based on the recommended single drug dose) to make amounts comparable between drugs, a significant difference was evident. Drugs that were predicted to be active had a median IC25 across all drug–strain pairs that corresponded to 4.3 drug doses, compared to 12 for control drugs (P = 5.6 × 10−7, two-sided Wilcoxon rank sum test; Fig. 3c). The IC25 corresponds to less than two drug doses in 34% of drugs with predicted activity, compared to just 8% for control drugs. Similarly, the IC25 is below the estimated colon concentration for 16/52 (31%) of candidate drug-strain pairs and for only 5/50 (10%) of control drug-strain pairs (Extended Data Fig. 8e).

In conclusion, human-targeted drugs with anticommensal activity have antibiotic-like side effects in humans, and for the few studies available, consumption of these drugs led to changes in taxa we also detected to be inhibited in vitro, implying that more drugs with anticommensal activity reported here will have an impact in vivo.

Features of drugs with anticommensal activity

Drugs from all major ATC indication areas exhibited anticommensal activity, with antineoplastics, hormones and compounds that target the nervous system inhibiting gut bacteria more than other medications (Extended Data Figs 9a, 10). Three ATC subclasses (antimetabolites, antipsychotics and calcium-channel blockers) were significantly enriched in hits (Extended Data Fig. 9a). Antimetabolites are used as chemotherapeutic and immunosuppressant agents, with their incorporation into RNA or DNA, or their interaction with synthesis enzymes being cytotoxic to human cells. Their molecular targets are often conserved in bacteria23, explaining the observed effects and raising the possibility that antibacterial effects may also be directly involved in the development of mucositis during chemotherapy24.

The enrichment in antipsychotics is intriguing, given that they target dopamine and serotonin receptors in the brain, which are absent in bacteria. Although phenothiazines are known to have antibacterial effects25, nearly all subclasses of the chemically diverse antipsychotics exhibited anticommensal activity (Extended Data Fig. 9b). These drugs targeted a significantly more similar pattern of species than expected from their chemical similarity (P = 2 × 10−19, permutation test; Extended Data Fig. 9c). This raises the possibility that direct bacterial inhibition may not only manifest as side effect of antipsychotics26, but also be part of their mechanism of action.

As different ATC indication areas contain chemically similar drugs, we investigated whether the chemical properties of drugs can influence their anticommensal activity (Extended Data Fig. 11a). To some degree, chemically similar human-targeted drugs had more similar effects in the screen than less similar drugs (Extended Data Fig. 11b). We tested several compound properties, including complexity, molecular weight, topological polar surface area (TPSA), volume and hydrophobicity (XLogP). Complex, heavier and larger compounds preferentially targeted Gram-positive bacteria, whereas Gram-negative bacteria were protected against such bulkier drugs by their selective outer membrane barrier (Extended Data Fig. 12). Owing to the vast number of chemical moieties present in drugs with anticommensal activity, we did not attempt an exhaustive enrichment analysis. Nevertheless, we did observe reactive nitro-groups being enriched in drugs with anticommensal activity (P = 6.4 × 10−6, Fisher’s exact test), indicating that local chemical properties may confer antibacterial activity.

Human-targeted drugs may boost antibiotic resistance

There is a strong correlation between resistance to antibacterials and resistance to human-targeted drugs in our data that cannot be explained simply by general cell envelope composition, as there is no clear division between Gram-positive and Gram-negative bacteria (Fig. 4a). We reasoned that more specific but common mechanisms could confer resistance to both drug groups. To test this hypothesis, we selected TolC, known to efflux several antibiotics in E. coli and other bacteria27, as a prominent representative of a general resistance mechanism against antibiotics. We profiled an E. coli ΔtolC mutant and its parental wild type (BW25113) against the Prestwick Chemical Library. E. coli lacking TolC not only became more sensitive to antibacterials (22 hits more than wild type), but also became equally more sensitive to human-targeted drugs (19 additional hits; Fig. 4a, Supplementary Table 6). This effect is not specific to E. coli or TolC, as a more antibiotic-resistant B. uniformis strain (HM-715) was also equally more resistant to human-targeted drugs (Fig. 4a).

Figure 4: Antibiotic resistance mechanisms protect against human-targeted drugs.
Figure 4

a, Susceptibility to antibacterial agents and human-targeted drugs correlates across the 40 tested strains (Spearman correlation, rS = 0.6 and a line depicting the nonlinear least-squares estimate of the odds ratio, OR = 0.06), suggesting common resistance mechanisms against both drug types. Knockout of a major antibiotic efflux pump (tolC) in the laboratory E. coli strain BW25113 (which behaves like the other two commensal E. coli strains in the screen) makes E. coli equally more sensitive to both antibacterials and human-targeted drugs. Two antibiotic-resistant isolates of B. fragilis (black square, HM-20) and B. uniformis (black diamond, HM-715) were screened in addition to the main screen, with only the latter showing a similar increase in resistance towards human-targeted drugs. b, Chemical genetic screen of an E. coli genome-wide overexpression library in seven non-antibiotics; all screens except for metformin were performed in ΔtolC background to sensitize E. coli to these drugs. Genes that when overexpressed significantly improved the growth of E. coli in the presence of at least one of the drugs are shown here; genes in bold have been previously associated with antibiotic resistance. Among them are genes encoding for transporters from different families: DMT (drug metabolite transporter), MFS (major facilitator superfamily), MATE (multidrug and toxin extrusion), SMR (small multidrug resistance) and ABC (ATP-binding cassette). Growth is measured by colony size (median n = 4)40, colour depicts the normalized size difference from the median growth of all strains in the drug (more than sixfold difference), and dot size the significance (FDR-corrected P < 0.1). Control denotes the growth of the library without drug.

While our data support a strong role for common general resistance mechanisms, there are also outliers to this trend, the most prominent being C. difficile and P. distansonis (Fig. 4a). For both, strong antibiotic resistance28 contrasted with relatively weaker resistance to human-targeted drugs. Similarly, an antibiotic-resistant B. fragilis strain, HM-20, was not equally resistant against human targeted drugs (Fig. 4a). These examples make the important distinction between specific antibiotic resistance mechanisms, which are irrelevant for resistance to human-targeted drugs, and more predominant, general mechanisms, which confer resistance to both drug groups.

To elucidate mechanisms conferring resistance against human-targeted drugs more systematically, we used a chemical genetic approach29 and screened a genome-wide overexpression library in E. coli against seven non-antibiotics (six human-targeted drugs and niclosamide, an antiparasitic) that showed broad anticommensal activity in our screen. As wild-type E. coli was one of the most resistant gut species (Fig. 4a), we used the ΔtolC mutant, which is sensitive to most of these drugs, allowing us to probe further resistance mechanisms. For all tested drugs except metformin, overexpression of tolC rescued E. coli growth, as expected. Furthermore, we identified a number of diverse transporter families that contributed to resistance against these drugs (Fig. 4b). Many of them have previously been linked to antibiotic resistance30,31,32,33. Resistance was also acquired by overexpression of transcription factors (for example, rob, which controls efflux pump expression34), the ribosome maturation factor rrmA, which plays a role in resistance to the antibiotic viomycin35, and detoxification mechanisms (nitroreductases modify nitro-containing antibiotics36). For methotrexate, we validated the known primary target in bacteria (E. coli dihydrofolate reductase)37, illustrating the potential of this approach to identify bacterial mechanism of action of human-targeted drugs29.

All of these results point to an overlap between resistance mechanisms against antibiotics and against human-targeted drugs, implying a hitherto unnoticed risk of acquiring antibiotic resistance by consuming non-antibiotic drugs.

Discussion

We report a systematic drug screen against a reference panel of human gut bacteria. Twenty-seven per cent of non-antibiotics (24% of human-targeted drugs) inhibited the growth of at least one species. As we demonstrated, this is likely to be an underestimate owing to stringent thresholds for calling hits and the limited selection of bacterial strains screened. Many of the direct in vitro effects described here may translate into microbiome shifts in vivo, because (i) we used concentrations within the range of what is estimated to be found in the human gut for many drugs; (ii) our observations agree with the few clinical microbiome studies for which medication has been recorded; and (iii) the side effects of anticommensal drugs in humans resemble those of antibiotics. Thus, our results underscore the necessity of accounting for potential medication-related confounding effects in future microbiome disease association studies. Moreover, one could speculate that pharmaceuticals, used regularly in our times, may be contributing to a decrease in microbiome diversity in modern Western societies38.

Although the antibacterial potential of human-targeted drugs has been profiled repeatedly in the quest for new antimicrobials, previous efforts have focused on pathogenic and often multi-drug-resistant (MDR) bacterial species9,13,14. We demonstrate that some of these species or their commensal relatives are the most drug-resistant in our screen (for example, γ-proteobacteria: Bilophila wadsworthia and E. coli were affected by 2 and 4–7 human targeted drugs, respectively), that many human-targeted drugs have species-specific effects, and that resistance mechanisms to antibiotics and human-targeted drugs partially overlap (thus, MDR species may be more resistant to human drugs too). Together, these findings explain why previous efforts have failed to register how many human-targeted drugs can inhibit bacteria.

Many pharmaceuticals influence the human gut microbiota. As gut bacteria, in turn, can also modulate drug efficacy and toxicity39, the emerging drug–microbe network could guide therapy and drug development. The resource described here opens up new avenues for translational applications in mitigating drug side effects, improving drug efficacy, repurposing of human-targeted drugs as antibacterials or microbiome modulators, and controlling antibiotic resistance (see Supplementary Discussion). However, before any translational application can be pursued, our in vitro findings need to be tested rigorously in vivo (in animal models, pharmacokinetic studies and clinical trials) and understood better mechanistically.

Methods

Bacterial strains and growth conditions

Bacterial isolates used in this study were purchased from DSMZ, BEI Resources, ATCC and Dupont Health & Nutrition, or were gifts from the Denamur Laboratory (INSERM). All strains were recovered in their recommended rich medium (resource and literature). The screen and validation experiments were performed in mGAM (HyServe GmbH & Co.KG, Germany, produced by Nissui Pharmaceuticals)41, as almost all species could grow robustly in this medium in a manner that is reflective of their gut abundance8. Because we selected for robust growth, potential positive effects of drugs on growth could not be detected. Only one strain was grown in Todd-Hewitt Broth (Sigma-Aldrich), one in a 1:1 mixture of mGAM and gut microbiota medium42 and, for one strain, mGAM was supplemented with 60 mM sodium formate and 10 mM taurine (see also Supplementary Table 2). All media were pre-reduced at least 1 day before use under anoxic conditions in an anaerobic chamber (Coy Laboratory Products Inc.) (2% H2, 12% CO2, rest N2) and all experiments were performed under anaerobic conditions at 37 °C unless specified otherwise. No statistical methods were used to predetermine sample size.

Species selection

To select a representative core of species in the human gut microbiome, we analysed 364 fecal metagenomes from asymptomatic individuals from three continents43,44,45,46. Species were defined and their abundance quantified as previously described47,48. A core set of 60 microbiome species was defined (Extended Data Fig. 1b–d), and from this core, 31 species were selected for this screen. Seven additional species were selected for reasons explained in the main text.

Screen of the Prestwick Chemical Library

Preparation of screening plates. The Prestwick Chemical Library was purchased from Prestwick Chemical Inc. with compounds coming dissolved in dimethyl sulfoxide (DMSO) at a concentration of 10 mM. Compounds were re-arrayed to redistribute the DMSO control wells in each plate and to minimize the total number of 96- and 384-well plates (4 × 384-well plates or 14 × 96-well plates). At the same time, drugs were diluted to a concentration of 2 mM to facilitate further aliquoting, and these plates were stored at −30 °C. For each experimental batch (10 replicates in 96-well plates; 20 replicates in 384-well plates), we prepared drug plates in the respective growth medium (2× for 96-well plates, 1× for 384-well plates), and stored them at −30 °C until use (maximum 2 months). Before inoculation, plates were thawed and pre-reduced in the anaerobic chamber overnight. The Biomek FXP (Beckman Coulter) liquid handling system was used for all rearranging and aliquoting of the library compounds.

Inoculation. Strains were grown twice overnight to make sure we had a robustly and uniformly growing culture before inoculating the screening plates. For 96-well plates, the second overnight culture was diluted to fresh medium in order to reach 2× the desired starting optical density (OD) at 578 nm. Next, 50 μl of this diluted inoculum was added to wells containing 50 μl of 2× concentrated drug in the respective culture medium using a multichannel pipetter. The final drug concentration was 20 μM and each well contained 1% DMSO. We inoculated 384-well plates with a 384 floating pin replicator VP384FP6S (V&P Scientific, Inc.), transferring 1 μl of appropriately diluted overnight culture to wells containing 50 μl of growth medium, 1% DMSO and 20 μM drug. For bacterial species that reached lower OD in overnight cultures we transferred twice 1 μl of appropriately adjusted OD culture. For both 96- and 384-well plates, the starting OD was 0.01 or 0.05, depending on the growth preference of the species (Supplementary Table 2).

Screening conditions. After inoculation, plates were sealed with breathable membranes (Breathe-Easy) to prevent evaporation and cross-contamination between wells, and incubated at 37 °C without shaking. Growth curves were acquired by tracking OD at 578 nm with a microplate spectrophotometer (EON, Biotek). Measurements were taken every 1–3 h after 30–60 s of linear shaking, initially manually but later automatically using a microplate stacker (Biostack 4, Biotek), fitted inside a custom-made incubator (EMBL Mechanical Workshop). We collected measurements for 16–24 h. Each strain was screened in at least three biological replicates.

Normalization of growth curves and quantification of growth. Growth curves were analysed by plate. All growth curves within a plate were truncated at the time of transition from exponential to stationary phase. The end of the exponential phase was determined automatically by finding the peak OD (using the median across all compounds and control wells, and accounting for a small increase during the stationary phase) and verified by inspection. Using this time point allowed us to capture the effects of drugs on lag phase, growth rate and stationary phase plateau (Extended Data Fig. 2a). Time points with sudden spikes in OD (for example, caused by condensation) were removed, and growth curves were discarded completely if they had too many missing time points (Extended Data Fig. 2a). Similarly, growth curves were discarded if the OD fell far outside the normal range (for example, caused by coloured compounds). Three compounds had to be completely excluded from the analysis, as they caused aberrant growth curves: Chicago sky blue 6B, mitoxantrone and verteporfin.

Growth curves were processed by plate to set the median OD at the start and end time points to 0 and 1, respectively. Then we defined reference compounds across all replicates as those that did not reduce growth significantly for most drugs, had measurements for >95% of all replicates, and for which the final OD was >0.5 for more than 142 out of 152 replicates. We used these reference compounds as representatives of uninhibited growth. As wells containing reference compounds outnumbered control wells within a plate, we used control wells only later to verify the P value calculation (Extended Data Fig. 2d). After identifying reference compounds, we rescaled growth curves such that the median growth of reference compounds at the end point was 1.

While growth curves in control wells and most wells with reference compounds followed the expected logistic growth pattern, a variety of deviations were observed for drugs that influenced growth. To quantify growth without relying on assumptions about the shape of the growth curve, we calculated the area under the curve (AUC) using the trapezoidal rule. Although we set the median starting OD to 0, the ODs of individual wells deviated from this. We used two methods to correct for this and determine the baseline for each growth curve (Extended Data Fig. 2a). First, a constant shift was assumed, subtracting the same shift from all time points of the growth curve such that the minimum is zero. Second, an initial perturbation was assumed that affects initial time points more than later time points (for example, condensation). To correct this, we first subtracted a constant shift as above, and then rescaled the curve such that a time point with an uncorrected OD of 1 also had an OD of 1 after correction. AUCs were calculated for both scenarios, rescaled such that the AUC of reference compounds was 1, and then for each compound the baseline correction that yielded an AUC closest to 1 (that is, normal growth) was selected.

AUCs are highly correlated to final ODs, with a Pearson correlation of 0.95 across all compounds and replicates. Nonetheless, we preferred to use AUCs to decrease the influence of the final time point, which will contain more noise than a metric based on all time points.

Identification of drugs with anticommensal activity. We detected hits from normalized AUC measurements using a statistical method that controls for multiple hypothesis testing and varying data quality. We fitted heavy-tailed distributions (scaled Student’s t-distribution49) to the wells containing reference compounds for each replicate and, separately, to each individual plate. These distributions captured the range of AUCs expected for compounds that did not reduce growth, and represented the null hypothesis that a given drug did not cause a growth defect in the given replicate or plate. We calculated one-sided P values from the cumulative distribution function of the fitted distribution. Within a replicate, each compound was associated with two P values: one from the plate on which it was measured, and one for the whole replicate. Of those two, the highest P value was chosen (conservative estimate) to control for plates with little or high noise, and varying levels of noise within the same replicate.

The resulting P values were well-calibrated (that is, the distribution of P values was close to uniform with the exception of a peak at low P values, Extended Data Fig. 2d) and captured the distribution of controls, which were not used for fitting the distribution and kept for validation. We then combined P values for a given drug and strain across replicates using Fisher’s method. Lastly, we calculated the FDR using the Benjamini–Hochberg method50 over the complete matrix of P values (1,197 compounds by 40 strains). After inspecting representative AUCs for compound–strain pairs at different FDR levels, we chose a conservative FDR cut-off of 0.01.

Drug indications, dose, and administration. We annotated drugs by their primary target organism on the basis of their WHO ATC classification, or, if there were uncertainties, based on manual annotation. Compounds were classified as: antibacterial drugs (antibiotics, antiseptics), anti-infective drugs (acting against protozoa, fungi, parasites or viruses), human-targeted drugs (that is, drugs whose mechanism of action affects human cells), veterinary drugs (used exclusively in animals), and finally non-drugs (which can be drug metabolites, drugs used only in research, or endogenous substances). If a human-use drug belonged to several classes, the drug class was picked according to this order of priority (from high to low): antibacterial, anti-infective, and human-targeted drug. This ensured that drugs used also as antibacterials were not classified in the other two categories.

Drugs from the Prestwick Chemical Library were matched against STITCH 4 identifiers51 using CART52. Identifiers that could not be mapped were annotated manually. Information about drug indications, dose and administration was extracted from the ATC classification system and Defined Daily Dose (DDD) database. Dose and administration data were also extracted from the Drugs@FDA resource. Doses that were given in grams were converted to mol using the molecular weight stated in the Prestwick library information files. When the dose guidelines mentioned salt forms, we manually substituted the molecular weight. Dose data from Drugs@FDA stated the amount of drug for a single dose (for example, a single tablet). Analysing the intersection between Drugs@FDA and DDD, we found that the median ratio between the single and daily doses was two. To combine the two data sets we therefore estimated the single dose as half of the daily dose (Supplementary Table 1).

In general, it is difficult to estimate effective drug concentrations in the intestine, as those depend on the dose, the speed of dissolution, uptake and metabolization by human cells and by bacteria, binding to proteins, and excretion mechanisms into the gut. To estimate gut concentrations of drugs based on their dose with a simple model, we relied on an in situ study for posaconazole19. When 40 mg (57 μmol) of the drug is delivered to the stomach in either an acidic or a neutral solution, the maximum concentration in the duodenum reaches 26.3 ± 10.3 or 13.6 ± 5.8 μM, respectively. This is equivalent to dissolving the drug in 300 ml (240 ml of water to swallow the pill as recommended for bioavailability/bioequivalence studies plus ~43 ml resting water in the small intestine53) and an absorption rate of 90%. We collected doses for as many human-targeted drugs as we could find and used the above assumption to estimate small intestine concentrations. To estimate colon concentrations, we relied on reported fecal excretion data (Supplementary Table 1, gathered from DrugBank 5.054 and across the literature) assuming a single daily dose, 24 h transit time55 and a volume of distribution in the colon of 0.6 l56 (Extended Data Fig. 3).

IC25 determination and screen validation

To validate our screen, we selected 25 drugs including human-targeted drugs (19), antiprotozoals (3), one antiparasitic, one antiviral and one ‘non-drug’ compound. The human-targeted drugs spanned five therapeutic classes (ATC codes A, G, L, M and N). Our selection comprised mostly drugs with extended antibacterial activity in our screen (19 drugs hit >10 strains). This bias ensured that we could also evaluate false positives. We chose 15 strains to test IC25s (that is the minimal concentration of drug that causes 25% growth inhibition), spanning different phyla (5) and including both sensitive (E. rectale, R. intestinalis) and resistant species (E. coli ED1a).

Compounds for validation were purchased from independent sources (Supplementary Table 1) and dissolved at 100× starting concentration in DMSO. Twofold serial dilutions were prepared in 96-well U-bottom plates (as for the screen). Each row contained a different drug at eleven twofold dilutions and a control DMSO well in the middle of the row (in total eight drugs per plate). These master plates were diluted to 2× assay concentration and 2% DMSO in mGAM (50 μl) and stored at −30 °C (<1 month). For the assay, plates were pre-reduced overnight in the anaerobic chamber, and mixed with an equal volume (50 μl) of appropriately diluted overnight culture (prepared as described for screening section) to reach a starting OD578 of 0.01 and a DMSO concentration of 1% across all wells. OD578 was measured hourly for 24 h after 1 min of shaking. Experiments were performed in two biological replicates.

Growth curves were converted to AUCs as described above, using in-plate control wells (no drug) to define normal growth. For each concentration, we calculated the mean across the two replicates. We further enforced monotonicity to conservatively remove noise effects: if the AUC decreased for lower concentrations, it was set to the highest AUC measured at higher concentrations. The IC25 was defined as the lowest concentration for which a mean AUC of below 0.75 was measured. In 68% of cases, IC25s were equal between replicates and in a further 22%, there was a twofold change between replicates, which is within the twofold error margin reported for inhibitory concentrations57. Additionally, MIC as listed in Supplementary Table 4 was defined as the lowest concentration for which the AUC dropped below 0.1. In the large-scale screen, we detected significant growth reductions, which do not necessarily correspond to complete growth inhibition (Extended Data Fig. 2b). To ensure comparability between the results of the validation procedure and the screen, we used the IC25 metric for benchmarking. As inhibitory concentration calculations are known to have a twofold error margin57, we considered an IC25 of 10–40 μM as being in agreement with the screening result (Extended Data Fig. 5a, b). A higher number of false negatives implies that more human-targeted drugs are likely to have anticommensal activity.

Analysis of side effects

Side effects of drugs were extracted from the SIDER 4.1 database22 using the mapping between Prestwick compounds and STITCH 4 identifiers described above. In SIDER, side effects are encoded using the MedDRA terminology, which contains lower-level terms and preferred terms. Of these, we used the preferred terms, which are more general. We excluded rare side effects that occurred for fewer than five drugs from the analysis. Drugs with fewer than seven associated side effects were discarded58. In a first pass, we identified side effects associated with antibiotics in SIDER, by calculating for each side effect its enrichment for systemic antibiotics (ATC code J01) versus all other drugs using Fisher’s exact test (P value cut-off: 0.05, correcting for multiple hypothesis testing using the Benjamini–Hochberg method). Antibiotics are typically administered in relatively high doses, and some of the enriched side effects might therefore be caused by a dose-dependent effect (for example, kidney toxicity). We therefore used an ANOVA (type II) to test whether the presence of side effects for a drug was more strongly associated with it being an antibiotic or with its (log-transformed) dose. Side effects that were more strongly associated with the dose were excluded from the list of antibiotic-related side effects.

Data on the incidence rates of side effects in patients was also extracted from SIDER 4.1. As different clinical trials can report different incidence rates, we computed the median incidence rate per drug–side effect pair. As SIDER also contains data on the incidence of side effects upon placebo treatment, we were able to ensure the absence of systematic biases.

Experimental validation of side effect-based predictions. Selected candidate and control compounds belonged to multiple therapeutic classes (ATC codes A, B, C, G, H, L, M, N, S for candidate compounds and A, C, D, G, H, M N, R, S, V for control compounds). Compounds of interest were purchased from independent sources (Supplementary Table 1) and if possible, dissolved at 5 mM concentration in mGAM. Lower concentrations were used when the solubility limit was reached. Solutions were sterile filtered, and three fourfold serial dilutions were arranged in 96-well plates, aiming at covering a broad range of drug concentrations. Inoculation and growth curve acquisition was performed as described for the IC25 determination experiments.

Chemical genetics in E. coli

Conjugation of the TransBac overexpression plasmid library into E. coli ΔtolC. The TransBac library, a new E. coli overexpression library based on a single-copy vector59 (H.D. and H.M., unpublished resource) was conjugated in the BW25113 ΔtolC::Kan strain. The receiver strain (BW25113 ΔtolC::kan) was grown to stationary phase in LB medium, diluted to an OD578 of 1, and 200 μl was spread on an LB plate supplemented with 0.3 mM diaminopimelic acid (DAP). Plates were dried for 1 h at 37 °C and then a 1536 colony array of the library carried within a donor strain (BW38029 Hfr (CIP8 oriT::cat) dap- 60) was pinned on top of the lawn. Conjugation was carried out at 37 °C for ~6 h, and the first selection was done by pinning on LB plates supplemented with tetracycline only (10 μg/ml) and growing overnight. Two more rounds of selection followed on LB plates containing both tetracycline (10 μg/ml) and kanamycin (30 μg/ml) to ensure killing of parental strains and select only for tolC mutants carrying the different plasmids.

Chemical genetic screen. The screen was carried out under aerobic conditions on solid LB Lennox medium (Difco), supplemented with 30 μg/ml kanamycin, 10 μg/ml tetracycline, the appropriate drug, and 0 or 100 μM IPTG. Drugs were used at the following sub-inhibitory concentrations for the tolC mutant: diacerein 20 μM, ethopropazine hydrochloride 160 μM, tamoxifen citrate 20 μM, niclosamide 1.25 μM, thioridazine hydrochloride 40 μM, methotrexate 320 μM, or for the wild type: metformin 100 mM. The 1536 colony array of BW25113 ΔtolC::kan mutant carrying the TransBac collection was pinned on the drug-containing plates, and plates were incubated for 16–38 h at 37 °C. In the case of metformin we used the version of the TransBac library in which each plasmid complements its corresponding barcoded single-gene deletion mutant59, since we did not need to use the ΔtolC background to sensitize the cell. Growth of this library was determined at 0 and 100 mM metformin (both in the presence of 0, 50 and 100 μM IPTG). All plates were imaged using an 18-megapixel Canon Rebel T3i and images were processed using the Iris software40.

Data analysis. We used colony size to measure the fitness of the mutants on the plate. For standardization of colony sizes, we subtracted the median colony size and then divided by a robust estimate of the s.d. (removing outliers below the 1st and above the 99th percentile). We found edge effects affecting up to five rows and columns around the perimeter of the plate. We therefore first standardized colony sizes across the whole plate using only colony sizes from the inner part of the plate as reference. To remove the edge effects, we subtracted from each column its median colony size, and then from each row its median colony size. Finally, we standardized the adjusted colony sizes using the whole plate as reference. The distribution of adjusted colony sizes was right-skewed (that is, more outlier colonies with larger sizes), suggesting a log-normal distribution. At the same time, the presence of outliers suggested that a logarithmic equivalent of the Student’s t-distribution with variable degree of freedom49 would be more suitable. We fitted such a distribution for each plate and calculated P values for both tails of the distribution. This approach assumes that the overexpression of most genes does not affect growth in response to drug treatment. P values were combined using Fisher’s method across replicates and IPTG concentrations (since we noticed that different IPTG concentrations resulted in largely the same results—that is, plasmids are leaky). We corrected for multiple hypothesis testing for each drug individually using the Benjamini–Hochberg method50.

Analysis of common resistance mechanisms.To determine a relationship between the number of human-targeted drugs (h) and the number of antibacterial drugs (a) that affect each strain, we determined the odds ratio (OR):where H = 203 and A = 122 are the numbers of human-targeted and antibacterial drugs that show activity against any strain, respectively. We computed the nonlinear least-squares estimate for OR using the following equation:

Data availability

Data are available from FigShare: http://dx.doi.org/10.6084/m9.figshare.4813882. All data generated during this study are included in this published article and its Supplementary Information files.

Code availability

Scripts for analysing data and generating figures are available at https://git.embl.de/mkuhn/drug_impact_gut_bacteria. A snapshot of the repository has been deposited together with the data.

References

  1. 1.

    , & Intestinal microbiota in health and disease. Nature 535, 47 (2016)

  2. 2.

    et al. Disentangling type 2 diabetes and metformin treatment signatures in the human gut microbiota. Nature 528, 262–266 (2015)

  3. 3.

    et al. Proton pump inhibitors affect the gut microbiome. Gut 65, 740–748 (2016)

  4. 4.

    et al. Proton pump inhibitors alter the composition of the gut microbiota. Gut 65, 749–756 (2016)

  5. 5.

    & The influence of non-steroidal anti-inflammatory drugs on the gut microbiome. Clin. Microbiol. Infect. 22, 171–179 (2016)

  6. 6.

    , , , & Interaction between atypical antipsychotics and the gut microbiome in a bipolar disease cohort. Pharmacotherapy 37, 261–267 (2017)

  7. 7.

    et al. Population-level analysis of gut microbiome variation. Science 352, 560–564 (2016)

  8. 8.

    et al. Nutritional preferences of the human gut bacteria reveal their metabolic idiosyncasies. Nat. Microbiol. (2018)

  9. 9.

    et al. Combinations of antibiotics and nonantibiotic drugs enhance antimicrobial efficacy. Nat. Chem. Biol. 7, 348–350 (2011)

  10. 10.

    , , & Bacterial uptake of aminoglycoside antibiotics. Microbiol. Rev. 51, 439–457 (1987)

  11. 11.

    Antibiotic use and its consequences for the normal microbiome. Science 352, 544–545 (2016)

  12. 12.

    , & Imidazoles as promising scaffolds for antibacterial activity: a review. Mini Rev. Med. Chem. 13, 1812–1835 (2013)

  13. 13.

    et al. Auranofin exerts broad-spectrum bactericidal activities by targeting thiol-redox homeostasis. Proc. Natl Acad. Sci. USA 112, 4453–4458 (2015)

  14. 14.

    et al. Antagonism screen for inhibitors of bacterial cell wall biogenesis uncovers an inhibitor of undecaprenyl diphosphate synthase. Proc. Natl Acad. Sci. USA 112, 11048–11053 (2015)

  15. 15.

    , , , & Machine learning meta-analysis of large metagenomic datasets: tools and biological insights. PLOS Comput. Biol. 12, e1004977 (2016)

  16. 16.

    , , & From dietary fiber to host physiology: short-chain fatty acids as key bacterial metabolites. Cell 165, 1332–1345 (2016)

  17. 17.

    et al. Enterotypes of the human gut microbiome. Nature 473, 174–180 (2011)

  18. 18.

    , & Gut biogeography of the bacterial microbiota. Nat. Rev. Microbiol. 14, 20–32 (2016)

  19. 19.

    , , & Supersaturation and precipitation of posaconazole upon entry in the upper small intestine in humans. J. Pharm. Sci. 105, 2677–2684 (2016)

  20. 20.

    , & Metformin and the intestine. Diabetologia 51, 1552–1553 (2008)

  21. 21.

    et al. Genomic variation landscape of the human gut microbiome. Nature 493, 45–50 (2013)

  22. 22.

    , , & The SIDER database of drugs and side effects. Nucleic Acids Res. 44, D1075–D1079 (2016)

  23. 23.

    , & Antibacterial activities of antineoplastic agents. Antimicrob. Agents Chemother. 28, 437–439 (1985)

  24. 24.

    , , & Chemotherapy-induced modifications to gastrointestinal microflora: evidence and implications of change. Curr. Drug Metab. 10, 79–83 (2009)

  25. 25.

    & Phenothiazines as anti-tubercular agents: mechanistic insights and clinical implications. Expert Opin. Investig. Drugs 20, 1665–1676 (2011)

  26. 26.

    et al. The antipsychotic olanzapine interacts with the gut microbiome to cause weight gain in mouse. PLoS One 9, e115225 (2014)

  27. 27.

    , & The challenge of efflux-mediated antibiotic resistance in Gram-negative bacteria. Clin. Microbiol. Rev. 28, 337–418 (2015)

  28. 28.

    , & ESCMID Study Group on Antimicrobial Resistance in Anaerobic Bacteria Antimicrobial susceptibility of Bacteroides fragilis group isolates in Europe: 20 years of experience. Clin. Microbiol. Infect. 17, 371–379 (2011)

  29. 29.

    , & Chemical genetics in drug discovery. Curr. Op. Syst. Biol. 4, 35–42 (2017)

  30. 30.

    et al. NorM, a putative multidrug efflux protein, of Vibrio parahaemolyticus and its homolog in Escherichia coli. Antimicrob. Agents Chemother. 42, 1778–1782 (1998)

  31. 31.

    et al. Antibiotic susceptibility profiles of Escherichia coli strains lacking multidrug efflux pump genes. Antimicrob. Agents Chemother. 45, 1126–1136 (2001)

  32. 32.

    et al. Phenotypic landscape of a bacterial cell. Cell 144, 143–156 (2011)

  33. 33.

    , & New substrates on the block: clinically relevant resistances for EmrE and homologues. J. Bacteriol. 194, 6766–6770 (2012)

  34. 34.

    , , & Activation of multiple antibiotic resistance and binding of stress-inducible promoters by Escherichia coli Rob protein. J. Bacteriol. 177, 1655–1661 (1995)

  35. 35.

    & Identification of the rrmA gene encoding the 23S rRNA m1G745 methyltransferase in Escherichia coli and characterization of an m1G745-deficient mutant. J. Bacteriol. 180, 359–365 (1998)

  36. 36.

    , , & Reduction of polynitroaromatic compounds: the bacterial nitroreductases. FEMS Microbiol. Rev. 32, 474–500 (2008)

  37. 37.

    et al. Dihydrofolate reductase: x-ray structure of the binary complex with methotrexate. Science 197, 452–455 (1977)

  38. 38.

    et al. The microbiome of uncontacted Amerindians. Sci. Adv. 1, e1500183 (2015)

  39. 39.

    , , & The microbial pharmacists within us: a metagenomic view of xenobiotic metabolism. Nat. Rev. Microbiol. 14, 273–287 (2016)

  40. 40.

    et al. A tool named Iris for versatile high-throughput phenotyping in microorganisms. Nat. Microbiol. 2, 17014 (2017)

  41. 41.

    , & Cultivation-based multiplex phenotyping of human gut microbiota allows targeted recovery of previously uncultured bacteria. Nat. Commun. 5, 4714 (2014)

  42. 42.

    et al. Extensive personal human gut microbiota culture collections characterized and manipulated in gnotobiotic mice. Proc. Natl Acad. Sci. USA 108, 6252–6257 (2011)

  43. 43.

    et al. A human gut microbial gene catalogue established by metagenomic sequencing. Nature 464, 59–65 (2010)

  44. 44.

    et al. A metagenome-wide association study of gut microbiota in type 2 diabetes. Nature 490, 55–60 (2012)

  45. 45.

    Human Microbiome Project Consortium. Structure, function and diversity of the healthy human microbiome. Nature 486, 207–214 (2012)

  46. 46.

    et al. Identification and assembly of genomes and genetic elements in complex metagenomic samples without using reference genomes. Nat. Biotechnol. 32, 822–828 (2014)

  47. 47.

    , , & Accurate and universal delineation of prokaryotic species. Nat. Methods 10, 881–884 (2013)

  48. 48.

    et al. MOCAT2: a metagenomic assembly, annotation and profiling framework. Bioinformatics 32, 2520–2523 (2016)

  49. 49.

    Bayesian estimation supersedes the t test. J. Exp. Psychol. Gen. 142, 573–603 (2013)

  50. 50.

    & Controlling the false discovery rate: a practical and powerful approach to multiple testing. J. R. Stat. Soc. B 57, 289–300 (1995)

  51. 51.

    et al. STITCH 4: integration of protein-chemical interactions with user data. Nucleic Acids Res. 42, D401–D407 (2014)

  52. 52.

    et al. CART-a chemical annotation retrieval toolkit. Bioinformatics 32, 2869–2871 (2016)

  53. 53.

    et al. Quantification of gastrointestinal liquid volumes and distribution following a 240 mL dose of water in the fasted state. Mol. Pharm. 11, 3039–3047 (2014)

  54. 54.

    et al. DrugBank 4.0: shedding new light on drug metabolism. Nucleic Acids Res. 42, D1091–D1097 (2014)

  55. 55.

    & How to interpret a functional or motility test—colon transit study. J. Neurogastroenterol. Motil. 18, 94–99 (2012)

  56. 56.

    et al. Fasting and postprandial volumes of the undisturbed colon: normal values and changes in diarrhea-predominant irritable bowel syndrome measured using serial MRI. Neurogastroenterol. Motil. 26, 124–130 (2014)

  57. 57.

    & Setting and revising antibacterial susceptibility breakpoints. Clin. Microbiol. Rev. 20, 391–408 (2007)

  58. 58.

    , , , & Drug target identification using side-effect similarity. Science 321, 263–266 (2008)

  59. 59.

    et al. GenoBase: comprehensive resource database of Escherichia coli K-12. Nucleic Acids Res. 43, D606–D617 (2015)

  60. 60.

    et al. Identification of essential genes and synthetic lethal gene combinations in Escherichia coli K-12. Methods Mol. Biol. 1279, 45–65 (2015)

  61. 61.

    et al. Metformin alters the gut microbiome of individuals with treatment-naive type 2 diabetes, contributing to the therapeutic effects of the drug. Nat. Med. 23, 850–858 (2017)

  62. 62.

    et al. Recent developments of the chemistry development kit (CDK) - an open-source java library for chemo- and bioinformatics. Curr. Pharm. Des. 12, 2111–2120 (2006)

  63. 63.

    et al. PubChem substance and compound databases. Nucleic Acids Res. 44, D1202–D1213 (2016)

Download references

Acknowledgements

We thank P. Beltrao (EBI), K. C. Huang (Stanford) and F. Cabreiro (UCL) for feedback on the manuscript; F. Rippmann (Merck KGaA) for pointing to the delayed onset of antipsychotics; S. Wicha (University of Hamburg) for discussions on drug concentrations; J. Overington (Medicines Discovery Catapult) for help with drug plasma concentrations, and members of all four laboratories for fruitful discussions (in particular T. Hodges for suggestions on the manuscript and M. Driessen for experimental support). We thank the EMBL mechanical workshop for the custom-made incubator. We acknowledge funding from EMBL and the Microbios grant (ERC-AdG-669830). L.M. and M.P. were supported by the EMBL Interdisciplinary Postdoc (EIPOD) programme under Marie Sklodowska Curie Actions COFUND (grant 291772). A.Te. and A.R.B. were supported by a Sofja Kovaleskaja Award of the Alexander von Humboldt Foundation to A.Ty.

Author information

Author notes

    • Mihaela Pruteanu

    Present address: Institute for Biology, Humboldt University Berlin, 10115 Berlin, Germany.

    • Lisa Maier
    • , Mihaela Pruteanu
    •  & Michael Kuhn

    These authors contributed equally to this work.

Affiliations

  1. European Molecular Biology Laboratory, Genome Biology Unit, 69117 Heidelberg, Germany

    • Lisa Maier
    • , Mihaela Pruteanu
    • , Anja Telzerow
    • , Exene Erin Anderson
    • , Ana Rita Brochado
    • , Keith Conrad Fernandez
    •  & Athanasios Typas
  2. European Molecular Biology Laboratory, Structural and Computational Biology Unit, 69117 Heidelberg, Germany

    • Michael Kuhn
    • , Georg Zeller
    • , Kiran Raosaheb Patil
    • , Peer Bork
    •  & Athanasios Typas
  3. Graduate School of Biological Sciences, Nara Institute of Science and Technology, 630-0101 Ikoma, Japan

    • Hitomi Dose
    •  & Hirotada Mori
  4. Max-Delbrück-Centre for Molecular Medicine, 13125 Berlin, Germany

    • Peer Bork
  5. Molecular Medicine Partnership Unit, 69120 Heidelberg, Germany

    • Peer Bork
  6. Department of Bioinformatics, Biocenter, University of Würzburg, 97024 Würzburg, Germany.

    • Peer Bork

Authors

  1. Search for Lisa Maier in:

  2. Search for Mihaela Pruteanu in:

  3. Search for Michael Kuhn in:

  4. Search for Georg Zeller in:

  5. Search for Anja Telzerow in:

  6. Search for Exene Erin Anderson in:

  7. Search for Ana Rita Brochado in:

  8. Search for Keith Conrad Fernandez in:

  9. Search for Hitomi Dose in:

  10. Search for Hirotada Mori in:

  11. Search for Kiran Raosaheb Patil in:

  12. Search for Peer Bork in:

  13. Search for Athanasios Typas in:

Contributions

The study was conceived by K.R.P., P.B. and A.Ty., designed by L.M., M.P., G.Z., A.R.B. and A.Ty., and supervised by K.R.P., P.B. and A.Ty. In vitro screening was established by M.P. and performed by L.M., M.P., A.Te. and K.C.F. Follow-up and validation experiments were conducted by L.M., M.P. and E.E.A. H.D. and H.M. constructed and provided the Transbac library. Data preprocessing was performed by M.K. and G.Z.; statistical analyses by M.K.; data curation by L.M., M.K. and E.E.A.; data interpretation by L.M., M.P., M.K., G.Z., K.R.P., P.B. and A.Ty. L.M., M.K., G.Z., K.R.P., P.B. and A.Ty. wrote the manuscript with input from M.P. and A.R.B.; L.M., M.K. and G.Z. designed figures with input from K.R.P., P.B. and A.Ty. All authors approved the final version for publication.

Competing interests

EMBL has filed two patent applications on repurposing compounds identified in this study for the treatment of infections and for modulating the composition of the gut microbiome, and on the use of the in vitro model of the human gut microbiome to study the impact of xenobiotics (Tentative European patent application numbers EP 18156520 and EP 18155278, respectively). Authors L.M., M.P., M.K., G.Z., K.R.P., P.B. and A.T. are listed as inventors.

Corresponding authors

Correspondence to Georg Zeller or Kiran Raosaheb Patil or Peer Bork or Athanasios Typas.

Reviewer Information Nature thanks K. Lewis, H. B. Nielsen, G. Wright and R. Xavier for their contribution to the peer review of this work.

Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Supplementary information

PDF files

  1. 1.

    Life Sciences Reporting Summary

  2. 2.

    Supplementary Information

    This file contains a Supplementary Discussion.

Excel files

  1. 1.

    Supplementary Table 1

    This file shows the compounds used in this study. Sheet 1 shows the compounds of the Prestwick Chemical Library, their assignment to therapeutic classes according to the ATC classification system, chemical and physical properties, single doses (which are half of the daily doses), plasma concentrations, estimated small intestine concentrations, route of excretion, estimated colon concentrations and conversion of the molar concentration of the screen (20 μM) into μg/ml. Sheet 2 shows the chemicals used in this study.

  2. 2.

    Supplementary Table 2

    This file contains the strains used in this study, their medium requirements and starting ODs for drug screening in 96- or 384-format.

  3. 3.

    Supplementary Table 3

    This file contains drugs with anticommensal activity in our screen. Sheet 1 shows the adjusted p-values for the impact of 1197 drugs on anaerobic growth of 40 human gut bacteria (see Methods). Sheet 2 shows the literature evidence and lowest measured MICs for antibacterial activity of the top 40 human-targeted drugs that affect more than 10 strains in our screen.

  4. 4.

    Supplementary Table 4

    This file contains IC25 determination which shows 25 selected drugs in a subset of up to 27 strains.

  5. 5.

    Supplementary Table 5

    This file contains the antibiotic-related side effects in human-targeted drugs (see Methods).

  6. 6.

    Supplementary Table 6

    This file contains the screen results of additionally screened strains (Figure 4), which shows the E. coli ΔtolC, its parental background (BW25113) and of B. fragilis HM-20 (NT5057) and B. uniformis HM-715 (NT5065).

About this article

Publication history

Received

Accepted

Published

DOI

https://doi.org/10.1038/nature25979

Further reading

Comments

By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.