Longitudinal clinical monitoring reveals widespread effects of medications on gut microbiome dynamics

Abstract Medications exert strong effects on the gut microbiome, impacting host physiology and affecting health outcomes. Studies have characterized associations between medications and the gut microbiome, primarily through in vitro analyses and, when conducted in humans, either using cross-sectional designs or focusing on specific medication classes. However, how medications directly reshape the microbiome within individual patients over time, especially in clinical settings, remains poorly understood. Here, we leverage longitudinal stool metagenomics, metabolomics, and electronic health records from 3,469 samples representing 1,122 patients at the University of Chicago Medical Center between 2020 and 2024. Using a longitudinal design where each patient serves as their own control, we compare patient microbiome composition before and after starting a medication while controlling for the effects of medical procedures, diagnoses, demographic variables, and other drugs taken simultaneously. In total, we identify 36,637 associations relating 138 of the most commonly given medications, such as oxycodone and pantoprazole, with 106 microbial genera, 204 microbial species, 627 microbial pathways, and 35 gut metabolites. We find, for example, that oral prednisone is associated with decreases in Parabacteroides and Enterobacter abundance and increases in microbial ergosterol and long-chain fatty acid biosynthesis pathways. Furthermore, the selective serotonin reuptake inhibitor sertraline is associated with increases in Alistipes abundance, increases in microbial pathways related to dopamine degradation, and increases in fecal concentration of the related metabolite kynurenic acid. Critically, we discover that medication effects are time-dependent, with microbiome recovery to baseline occurring over days to weeks. Our findings represent the first systematic characterization of medication-microbiome dynamics within individual patients in a clinical setting, establishing temporal patterns and mechanistic pathways that inform microbiome-guided therapeutic strategies. Description This repository contains contains data files and code to run the models and generate the figures in this paper, which should take about 5 minutes to run on a typical desktop computer. To generate the figures in this paper, extract the zip file and run:`python3 build_figures.py` Most data files are stored in the highly efficient Apache parquet format, which can be read in Python using the read_parquet function in pandas or in R using the read_parquet function in arrow. MRNs (medical records numbers) in data files have also been anonymized. Code Description build_figures.py Script to build figures used in this paper.paperfigures.py Contains code to build each of the figures in this paper and output them as PDFs.icd10.py Builds a table that maps ICD-10 codes to their text descriptions.main.py Runs the linear mixed models used in our analysis pipeline. pathways.py Contains functions to perform pathway enrichment analysisplots.py Helper functions for plotting.util.py Helper functions for use in the analysis pipeline. File Descriptions out/ output figures and supplementary tables produced by build_figures.py output/ Results from the linear mixed models, with columns indicating medication, feature, p-value, coefficient, significance (FDR < 0.1), estimated standard error, medication class, model duration, sample level type, model type, and whether the medication was a duration effect or not. output/results_2_to_10_days.csv Presence/absence results from our 2-10 day models. output/results_10_to_20_days.csv Presence/absence results from our 10-20 day models.output/results_20_to_30_days.csv Presence/absence results from our 20-30 day models.output/results_duration.csv Results from our duration models. data/ Intermediate data files generated from running our pipeline. data/table_demographics.parquet Anonymized demographics data with patient sex and birth year. data/table_microbes_clr.csv Data table containing CLR-transformed microbiome abundance data.data/table_pathways.parquet Data table containing microbial pathway abundances.data/table_pathways_general.parquet Data table containing microbial pathway abundances.data/metacyc_pathway_hierarchy.json Pathway hierarchy for humann3 metacyc pathways.data/hospital_visits.parquet List of all patient hospital visits, with accompanying clinical data.data/study_intervals.parquet List of all pairs of stool collections, with accompanying clinical data.data/table_microbes.csv Data table containing microbiome counts.data/table_microbes_relabund.parquet Data table containing microbiome relative abundances.data/medication_classes.parquet List of all medication classes.data/table_genus_metab_correl.parquet Correlations between microbial genera and metabolites.data/medications.parquet List of all medications included.data/samples.parquet De-identified table containing list of stool samples included in the study with collection dates.data/icd10cm-tabular-April-2024.xml ICD-10 codes reference table Dependencies Package Version python 3.10 numpy 1.23.5 pandas 1.5.3 matplotlib 3.6.3 scikit-learn 1.2.1 scikit-bio 0.5.8 seaborn 0.12.2 scipy 1.10.0 statsmodels 0.13.5 matplotlib 3.6.3 networkx 3.0