Impact of commonly used drugs on the composition and metabolic function of the gut microbiota

The human gut microbiota has now been associated with drug responses and efficacy, while chemical compounds present in these drugs can also impact the gut bacteria. However, drug–microbe interactions are still understudied in the clinical context, where polypharmacy and comorbidities co-occur. Here, we report relations between commonly used drugs and the gut microbiome. We performed metagenomics sequencing of faecal samples from a population cohort and two gastrointestinal disease cohorts. Differences between users and non-users were analysed per cohort, followed by a meta-analysis. While 19 of 41 drugs are found to be associated with microbial features, when controlling for the use of multiple medications, proton-pump inhibitors, metformin, antibiotics and laxatives show the strongest associations with the microbiome. We here provide evidence for extensive changes in taxonomy, metabolic potential and resistome in relation to commonly used drugs. This paves the way for future studies and has implications for current microbiome studies by demonstrating the need to correct for multiple drug use.

4. The authors perform systematic analysis of drug co-administration and find that there is a strong correlation between steroid and beta sympathomimetic inhalers (first paragraph of results). In general, it does not become clear, how this data on drug co-administration is used in the subsequent association analysis. It seems that only the number of administered drugs has an impact on microbiome composition. 5. Given the resource-character of this work, raw sequencing data, including metadata, should be made freely accessible (without required permission) from one of the common databases. We thank the reviewer for bringing out this topic, which indeed was lacking in our discussion.
As the reviewer has pointed out in this comment, the use of medication can be indicative of health conditions, and therefore, it becomes challenging to study the relation between changes in the gut microbiota composition and the medication usage. In addition, usage of medication is commonly complemented with changes in the lifestyle (for example, diet) that can also have an impact on the microbial composition in the gut. This complex relation is also relevant in the study of host-disorders. While most of the studies typically consider the use of antibiotics as an excluding or correcting factor in their analyses, the effect of other commonly used medications is still underestimated.
In the current version of the manuscript we have summarised additional potential confounding effects in the discussion. In addition, and following the suggestions of Reviewer 2, we have expanded the discussion on the identified associations and their relation with previous published findings (see answers 2.2 and 2.3) .
Lines 343 -358 "The complex interaction between the use of medication, the gut microbiota and confounding factor, poses several limitations in our study. Firstly, the cross-sectional nature of this study cannot identify causality in the observed associations. Second, the use of medication by itself is indicative of changes in the health condition of the host, that may also be accompanied by changes in lifestyle, which are both known to influence the microbiome composition in the gut. Although the effect of diet and lifestyle are relevant contributors of the microbial composition, they explain a relatively small proportion of the interindividual variation which we and others have shown previously1-3. In our current analysis, we controlled for the effect of age, sex, sequencing depth and body mass index (BMI). The latest is known to be related with diet and lifestyle, and therefore, we expect to capture part of this effect when correcting for BMI.
In this study we identified 6 drugs to be associated with pathways or taxonomy when taking the use of other drugs into account. An important consideration is that we can divide these identified drugs into two groups, namely first the drugs which are only prescribed in one disease, for example metformin in type 2 diabetes (T2D). Therefore, to assess the real effect of the drug we would need to compare patients with T2D not using metformin with those using it. Unfortunately, the number of T2D patients not using metformin is very limited in our cohort, since metformin is the first choice of drug in this patient group in the Netherlands.
The second group of drugs are those who are prescribed for numerous indications, this could either be for numerous diseases or for multiple symptoms, for example the proton-pump inhibitors (PPIs). These show one of the strongest associations with the microbiota in our study, are used for various indications: gastroesophageal reflux (GERD), but also in the context of bloating or the prevention of ulcers in therapies involving other drugs like NSAIDs.
Another fact to consider is that these drugs can be sold over the counter in the Netherlands.
In this case, and besides this heterogeneity and multiple confounding factors, the earlier findings using sequencing data were later replicated in longitudinal and in-vivo studies4 -8, showing that although the known limitation of cross-sectional sequencing studies, this kind of research can be useful in the discovery of drug-microbiota associations. We agree with the reviewer and now we have added more details on the specificities of the medication categories and dosages, however, data regarding the duration was not available in our cohorts. Due to the size of the table, this information can now be found in the updated Supplementary table 1 for convenience.
To compile this information, we had to go back to the original data source, which included questionnaires in the case of the population cohort and medical records in the case of the patients cohorts (both IBD and IBS). For that, and due to privacy restrictions, we had to request consent to the participants to access the data and ask medical doctors to retrieve that information from the electronic patient files. Unfortunately, this process has delayed our reply to the reviewers.

Medication classification
Following the comment of the reviewer we revised the classification of medication subtypes and added this information in the Supplementary tables 1. In brief, our categorization follows the ATC database classification, classifying each drug based on their indications. Moreover, we reviewed these groups creating sub-categories based on the chemical structure or working mechanisms. For example, antidepressant drugs were divided into 3 groups: SSRIantidepressants, tricyclic antidepressants and a general category that represented the remaining antidepressant drugs. Regarding aspirin use, we did not categorize it in the NSAID group. In the Netherlands, this drug is prescribed as a platelet aggregation inhibitor which leads to lower dosages of the drug used in the Netherlands (80 milligrams/day) (higher doses of aspirin are needed to function as a painkiller and those are not prescribed in the Netherlands). Regarding the question on antibiotic usage, in our cohort there were 30 antibiotic users, for which the most prevalent ones were tetracyclines (n=9), penicillines (n=7) and fluorquinolones (n=6) (Supplementary table 1).
Moreover, and as the reviewer suggested we investigated if there was a differential effect depending on specific medication type, e.g. comparing tetracycline users with non-antibiotic users and penicillin users with non-antibiotic users. Strikingly, we found that different PPIs showed a similar effect, while in the category of antibiotics the associations were mostly derived from tetracyclines users. The relatively low numbers of antibiotic users prevented us however to identify major differences between the different types of antibiotics. All results are now summarized in the Supplementary table 12.
To clarify how each medication was classified we have now added the following text in the method section in lines 381 -385, as well as the detailed information on Supplementary table 1: Table 1)."

Medication dosage
Regarding medication dosage, we retrieved the dosages from participants for those microbiome-associated drugs of the multivariate analyses as suggested by Reviewer 1. This was the case for the drugs SSRI-antidepressants, alpha-blockers, antibiotics, laxatives, proton pump inhibitors and metformin. However, differences in doses could not be tested on the antibiotics, alpha-blockers, SSRI-antidepressant or laxatives due to the standardized prescription dosages (almost all participants using the same dosages). In the case of PPIs and metformin users, since most of the participants were using comparable doses, users were separated into two categories "high dosages users" and "low dosages users". For PPI users, dosages less than or equal to 20 mg/day were considered as a low dosage and higher than 20 mg/day was considered as a high dosage. For metformin this cutoff was set at less than 1000 mg/day for the low dosage users.
In total, 46 pathways associated with PPI use showed a dosage effect, however, no significant associations were observed in metformin users (FDR<0.05, Supplementary table 12).
The following text has been added in the main manuscrip in lines 459 -472t: "Methods:

Individual medication and dosage-dependent effects
Statistically significant medication-microbiome associations were further assessed for the differential influence of drug types within the same category and the prescription dosages.
Medication subtypes were analysed if they were present in at least 5 participants. To evaluate the effect of each medication subtype, the abundance of the associated microbial features was compared between users of a drug subtype and participants not using drugs belonging to the same category. An example of that is the comparison between tetracycline users to participants not using antibiotics. Due to the distribution of the data referring to medication doses (Supplementary table 12), samples were grouped into two categories: "high dose" and "low dose" of each particular drug. For PPI's this threshold was set to a minimum of 40 mg/day for the high dosage group and for metformin this minimum was set at 1000mg/day. Users of laxatives, alpha-blockers, SSRI-antidepressants or antibiotics of our cohort, reported similar prescription patterns or the subtypes within this medication categories showed major differences in dosages. Therefore, we were unable to analyse dosages in these medication categories. Differences between groups were tested using non-parametric t-test (wilcoxontest).  We thank the reviewer for the suggestion. In the revised version of the manuscript, we have added a deeper characterization of all drug-pathways associations by: a) For each pathway associated to a medication category we explored which bacteria were contributing to the pathways abundance. Next, we compared these values between medication users and non-users. To do so, we retrieved the bacterial contribution of each pathway from HUMAnN2 default output. We filtered those pathways missing in more than 90%

Results
in each cohort and normalized as described in the method section of our manuscript. A nonparametric t-test (Wilcoxon-test) was performed to evaluate if the bacterial contribution of each pathway differed between users and non-users. Resulting p-values were adjusted for multiple testing using Benjamini-Hochberg calculation. b) Investigating which gene families are implicated in each associated pathway: Gene families involved in specific pathways were retrieved using the humann2_unpack_pathways script which is provided with the software. Filtering and normalization were performed as described above and differential abundance between users and non-users were tested using Wilcoxontest. c) Updated figure 2 to make easier the interpretation of the data.
Regarding the associations with PPI use, the 125 microbial associated pathways were predicted from 201 known bacterial genomes. After filtering pathways which were at least present in 10% of the samples of each individual cohort, 3174 were considered for analysis.
Consistently with the observations in the taxonomic analysis, Streptococcus species were the top contributors in the differential abundance of those pathways in all three cohorts. For example, 29 organisms were found to contribute to the pathway involved in the L-arginine biosynthesis via the acetyl cycle (MetaCyc ID ARGYSYNBSUB). Nonetheless, in PPI users the increased abundance of this pathway was mainly linked to Streptococcus mutans(FDR < 0.05 in the three cohorts). At the gene-family level, more than 30.000 Uniref90 gene families were identified to be involved in the 125 PPI-associated pathways. Our analysis at this level revealed a similar pattern as previously described: being Streptococcus genes enriched in the gut microbiota of PPI users.
We repeated this type of analyses for each associated pathways and provided the data in the "Taxonomic contribution to metabolic pathways Pathways that were shown to be associated with medication use in the multivariate metaanalysis were further investigated. To estimate the bacterial contribution to each pathway we calculated the species-level stratified abundances using the HUmann2 pipeline. Gene families were also extracted using the humann2_unpack_pathways script. Values were transformed to relative abundance and log-transformed as described above. For each medication category associated with changes in the metabolic potential of the gut microbiota, the differential abundances in the stratified pathways and gene families were tested using the Wilcoxon signed-rank test. Significant levels were adjusted for multiple testing applying the Benjamini-Hochberg correction."

15)
We thank the reviewer for noticing the mistake. The references have now been adjusted.

Thyrax and ferrum should be listed by their generic names
We have now listed them by their generic names: 'thyrax' has been replaced by 'levothyroxine' and 'ferrum' has been replaced by 'iron preparations'.

Opiates and melatonin are misspelled
We have corrected figure 1 and revised the spelling in the manuscript.

It would be nice to discuss the null acetaminophen results in the context of the PNAS acetaminophen study from 2009
We thank the reviewer for pointing us to this interesting study. We have added the reference in the discussion section at lines 338 -342: "Although an interaction between acetaminophen and the gut microbiota has been described We agree with the reviewer that the definition of bacterial activity is not clearly stated. Indeed, bacterial activity was not directly measured and the results previously linked to "bacterial activity" referred to the metabolic potential, calculated as the pathways inferred from metagenomic sequencing alignments. We have adjusted the sentence to make it more accurate. It now reads at line 129 -131 : We agree with the reviewer's correction. We have removed the expression "essential pathways" from the main text. Regarding the changes associated with PPI use we saw that the enrichment of Streptococcus and Veillonella sp. is also reflected in the gene abundance and pathway contribution, suggesting the functional changes are mainly consequence of this enrichment. For example, the purine deoxyribonucleosides degradation pathway, which represents a mechanism described in E.coli in which purines are utilized as a source of carbon and energy, was detected in more than 20 different bacterial species in each cohort. However, when comparing PPI users vs non-users, the significant enrichment was only observed in those pathways Consistently with the changes observed in the gut microbiota of oral steroids users, the pathways significantly enriched were identified to belong to Methanobrevibacter smithii.
In the current version of the manuscript, we expanded the discussion on pathways finding, for PPI users see question 1.4, in the case of metformin we discuss the findings on question 2.3. We agree with the reviewer that the statement was not accurate enough. In our meta-analysis, the abundance of E.coli was not significantly associated with the use of metformin. Although an increased abundance is observed in the population cohort, this effect was not replicated in the other two cohorts. Changes in the microbiome composition associated with gastrointestinal disorders (such as IBD or IBS) may explain the heterogeneity of this effect. Although the association between the increased abundance of E.coli and the use of metformin has been shown previously, in-vitro and in-silico experiments could not demonstrate the direct association.

PPI, laxatives
In contrast, significant associations were observed between functional changes (predicted from metagenomic data) and metformin use.
We agree with the reviewer that, although we see an enrichment on certain genes and pathways predicted from E.coli strains pan-genomes, we cannot determine if the change in the metabolic potential in the gut microbiota of metformin users is solely due to E.coli.
Therefore, metatranscriptomics and metabolomic approaches combined with culturomics and detailed analyses for specific species is the preferred approach to disentangle this association.
We have now edited the manuscript accordingly to the previous explanation. We changed the results section title and added the following text at lines 197 -206: "Metformin use is associated to changes in enterobacteria metabolic potential  Table 19)." In addition, in the discussion section we have now added the following text at lines 324 -342: We want to thank the reviewer for pointing out that the methodology was not completely clear.
Indeed, we report that certain medication groups, including the case of steroids and beta sympathomimetic inhalers, are frequently prescribed together in our cohort. We therefore considered investigating the potential interacting effect and/or stratifying participants based on combination of medications In the first part of the result section we wanted to highlight the co-administration patterns of certain drugs in order to better characterize the drug usage in our cohorts and therefore, provide support for the interpretation of the results of the univariate drug-microbiota associations. In addition, we also focus in two broad microbiota metrics: richness (shannon index) and overall microbial composition (Bray Curtis dissimilarities matrix). We then showed that, in our cohort, the use of individual drugs do not alter the richness or the overall composition of the gut microbiota, with the exception of proton-pump inhibitors, which are significantly associated with changes in the Bray-Curtis dissimilarities. Due to the fact that the use of multiple drugs could mean the exposure of the intestinal microbiota to several external compounds but also a combination of effects in the host, we hypothesize that the number of administered drugs (number of drugs that a participant was taking at the time of fecal sampling collection) could have an impact on the richness and the overall composition. Interestingly, we show that the number of medications used by the host is associated with changes in the microbial composition which are probably also indicative of its health status.
In the second part of the study, we focus on the specific drug-microbiota association using two different models. Due to the multiple medication combination (more than 500), it was not possible to estimate co-administration effects. To correct for this (possible) effect, first, we considered the association between bacteria and each drug individually and then, we added all other medication categories in the same model to account for the confounding effect of coadministered drugs.
In the current version of the manuscript we have clarified the methods section at lines 421 - 5. Given the resource-character of this work, raw sequencing data, including metadata, should be made freely accessible (without required permission) from one of the common databases.
We fully agree with the reviewer and we are committed with the goal of making all research FAIR. All data used to carry out this research is available in the European Genome-phenome Archive. Due to the current privacy and IRB regulations at the time of sample collection, data can be freely shared with academic institution but with an obligated control step on the data access, consisting on reviewing applications before allowing access. Notice that the three datasets used in this study depend of three different institutions: LifeLines Deep population cohort (LifeLines), case control IBS cohort (Maastricht University Medical Centre) and the IBD cohort (University Medical Center of Groningen), therefore the access policy on the data depends on each of the owners.