Disentangling microbiota-dependent and -independent impacts of antibiotics on host physiology

Abstract

Antibiotics are believed to affect host physiology by altering the gut microbiota. However, it remains unclear which effects of antibiotics are microbiota dependent and which are microbiota independent. Here we present a resource encompassing host and microbiota responses to antibiotic treatment and a computational tool with which to analyse them. Metagenomic and transcriptomic analyses of conventional and germ-free mice revealed that antibiotic administration routes and their combinations affect host and microbiota responses. We developed CompoRes—a computational tool that distinguishes the microbiota-independent impact of antibiotic-induced perturbations, revealing that antibiotics directly impact host physiology. We found that diurnal expression of major histocompatibility complex class II in the mouse colon is controlled by glucose via feeding. This pattern was lost during antibiotic treatment as vancomycin altered feeding behaviour in a microbiota-independent and evolutionarily conserved manner. Together, our study establishes a direct, non-microbial pathway by which antibiotics reshape host immunity and circadian physiology.

Using a computational tool called CompoRes, the authors show that antibiotics can alter feeding behaviour, immunity and circadian regulation, revealing a non-microbial pathway by which antibiotic treatment reshapes host physiology.

Main

The environment, including medications, affects and shapes host physiology^1,2. It is becoming increasingly apparent that many of the effects of the environment on host physiology are mediated via the gut microbiota³. This involvement of the gut microbiota in host response makes it difficult to discern whether the environment is acting on the host directly or through the gut microbiota. Germ-free animals can be used to test whether a certain perturbation acts directly on the host, and gnotobiotic animals can be used to test whether the effect is mediated by a defined group of microorganisms^4–6. Although current tools partially address whether a host phenotype is correlated with their microbiota composition^7–9, a tool that rigorously predicts whether an environmental factor acts directly on the host or via the microbiota from observational data is currently not available.

Antibiotics are an example of such environmental factors. Discovered by Alexander Fleming nearly 100 years ago¹⁰, antibiotics were initially used to treat infectious diseases in humans. Currently, antibiotic exposure is nearly unavoidable, due to their extensive use in clinical practice and agriculture^11,12. Antibiotic treatment is also a common tool in biological studies and is mostly used to deplete or alter the gut microbiota of laboratory animals^11,12. For example, recent works have shown that gut microbiota composition oscillates in a circadian pattern, which directly affects the circadian oscillation of host immune effectors and metabolism. Depletion of the gut microbiota via treatment with a cocktail of antibiotics disrupts this diurnal oscillation of transcripts in the gut epithelium^13,14. In this example, antibiotics disrupt normal physiological function by altering the gut microbiota.

Although antibiotics were developed to target microorganisms, it is becoming apparent that they can affect host cells regardless of their effect on the microbiota. Antibiotics can directly affect cultured cells^15–17, induce antiviral immune responses in germ-free mice¹⁸, affect host metabolism¹⁹, inhibit mitochondrial protein synthesis²⁰ and induce endoplasmic reticulum stress in intestinal cells²¹. Endoplasmic reticulum stress goes on to inhibit mucus secretion by intestinal goblet cells, predisposing the host to the development of intestinal inflammation²¹. Thus, antibiotics are an example of an environmental factor that can affect host physiology, acting both via the microbiota and directly on host cells.

As antibiotics are widely used in the clinic and research, it is essential to understand exactly how they affect host physiology. For this Resource, we leveraged the far‑reaching effects of antibiotics on host and microbiota to develop computational tools that predict whether a given perturbation impacts host physiology directly or via the microbiota. Applying these tools suggested that systemic administration of vancomycin disrupts diurnal gene expression patterns in the gut in a microbiota‑independent manner. To test this prediction, we treated germ-free mice with vancomycin and discovered that vancomycin affected mouse feeding behaviour. These behavioural changes shaped the oscillation of the antigen presentation machinery in the colonic epithelium in a microbiota-independent manner.

Results

The route of antibiotic administration impacts its effect on the gut microbiota

Antibiotics can be administered via multiple routes in the clinic and in experiments. Yet, whether the administration route affects treatment outcomes is unclear. To test this, we treated mice reared in a specific pathogen-free (SPF) barrier facility with four commonly used antibiotics (that is, vancomycin, ampicillin, metronidazole and neomycin), either separately or combined (a four-antibiotic mix). Antibiotics were administered orally via gavage or systemically via intraperitoneal (i.p.) or intravenous (i.v.) injection. Mimicking real-world conditions, mice were treated twice a day for 3.5 days (to reflect real-world observations and the shift towards shorter regimen periods^22–26), after which stool and colonic tissue samples were collected (Fig. 1a). Faecal samples were also collected before treatment, to establish a baseline.

Fig. 1. Analysis of faecal microbiota composition and host response to antibiotic treatment in the gut.

a, Schematic depicting the experimental procedure (n = 6). b, Analysis of effective numbers of genera using the Shannon diversity index²⁷. c, Pairwise community-level comparison of the effect of antibiotic treatment versus the administration of PBS as a control, using Adonis2. Each symbol represents a group. The dashed line indicates an adjusted P value of 0.05. Comparisons were made between treatment and PBS groups with matching administration routes. For PBS groups, oral administration was compared with i.p. administration, i.p. administration with i.v. administration and i.v. administration with oral administration. Statistical significance was determined by pairwise permutational multivariate analysis of variance (Adonis2; vegan R package; 9,999 permutations). P values from pairwise comparisons were corrected for multiple testing using the Benjamini–Hochberg false discovery rate method. d–f, Principal coordinates analysis (PCoA) of microbial composition, calculated with Aitchison distance, in antibiotic-treated mice based on administration route (oral (d), i.p. (e) or i.v. (f)). g, PCoA of microbial compositional data as in d–f in vancomycin-treated mice. The shaded ellipses represent 2-s.d. confidence regions based on the data’s covariance. Each dot represents a faecal sample from a single mouse. h, Total number and fold change of significant (P < 0.05) DEGs in antibiotic-treated mice. The dot colour indicates the number of DEGs (two-sided independent-samples t-test versus PBS control; uncorrected P < 0.05). The dot size indicates the mean absolute log₂[fold change] (LFC) of those genes. i, Enhanced and suppressed Gene Ontology (GO) clusters in antibiotic-treated mice using GSEA. Differentially expressed GO clusters were grouped and aggregated by top-level annotation; enrichment indicates the fraction of GO terms up- or downregulated in each category (more details provided in the ‘GSEA’ section in Methods). Amp, ampicillin; d₀, before treatment; d₄, after treatment; Met, metronidazole; Mix, a combination of the four antibiotics; Neo, neomycin; p.o., oral administration; Van, vancomycin. Schematic in a created in BioRender; Bel, S. https://biorender.com/cmusc4c (2026).

Source data

We first analysed the bacterial composition of faecal samples using 16S ribosomal RNA (rRNA) amplicon sequencing. We calculated the effective number of genera (using the Shannon diversity index²⁷) and performed a pairwise community-level comparison of the effect of the antibiotic treatment using Adonis2 (ref. ²⁸) (Fig. 1b,c). We detail our observations on the effects of ampicillin, metronidazole, neomycin and a mix of these three antibiotics on microbial composition in Extended Data Fig. 1. Herein, we focus on the results relating to vancomycin. Contrary to our expectations, we found that vancomycin treatment had a significant and unique effect on the effective number of genera and bacterial composition, regardless of the administration route (Fig. 1b–g and Supplementary Table 1). This was surprising, as vancomycin is a large hydrophilic molecule that should cross the gut mucosa poorly²⁹. However, mass spectrometry analysis confirmed that systemic i.p. administration of vancomycin resulted in noticeable amounts of vancomycin in the small intestinal and colonic lumen, whereas oral administration showed low systemic absorption, as expected²⁹ (Extended Data Fig. 1a–d). Furthermore, the effects of vancomycin on bacterial composition were distinct for each administration route. Whereas oral administration of vancomycin led to an increase in Enterobacteriaceae (P value = 0.0002; q value = 0.0007; log₂[fold change] = 5.3), systemic treatment increased Tannerellaceae (P value = 0.002; q value = 0.057; log₂[fold change] = 5.2); both are Gram negative (Extended Data Fig. 1i,j). Thus, contrary to common notion^29,30, systemic vancomycin treatment leads to detectable amounts of the antibiotic in the gut lumen, explaining the changes to bacterial composition.

Extended Data Fig. 1. Compositional analysis of antibiotic-treated mice and mass spectrometry.

a-d, Mass spectrometry detection of vancomycin in the kidney (a), liver (b), colon (c) and small intestine (d) of mice treated as indicated. Each dot represents a mouse. e-h, Principal coordinates analysis of microbial compositional data, calculated with Aitchison distance, in mice treated with ampicillin (e), metronidazole (f), neomycin (g) or a mix of antibiotics (h). Shaded ellipses represent 2-s.d. confidence regions based on the data’s covariance. Each dot represents a faecal sample from a single mouse. d₀, before treatment; d₄, after treatment. i,j, Taxonomic analysis of faecal microbiota at the family (i) and genus levels (j) based on 16S rRNA amplicon sequencing. Each bar represents the microbiota composition from a single mouse.

Source data

Changes to gut microbiota composition do not predict alterations in gut gene expression

Next, we assessed the impact of antibiotics on host gene expression. We performed RNA sequencing (RNA-seq) of colons from antibiotic-treated mice (as above) and employed Gene Ontology³¹ enrichment analysis using gene set enrichment analysis (GSEA³²). We detail our observations on the effects of ampicillin, metronidazole, neomycin and a mix of these antibiotics on host gene expression in Supplementary Table 4. Of all of the antibiotics studied here, i.p. vancomycin treatment affected the largest number of transcripts (Fig. 1h). GSEA analysis showed that i.p. vancomycin treatment affected a large number of Gene Ontology pathways (Fig. 1i)—more than were affected by i.v. injection (Fig. 1i). We found this interesting as i.p. and i.v. administration of vancomycin had similar effects on the gut microbiota (Fig. 1g and Extended Data Fig. 1i,j). This suggests that the effect of systemic vancomycin treatment is somewhat decoupled from microbiota compositional changes, and that changes to microbiota composition may not fully predict the effects of perturbations on host response.

Different antibiotics induce unique changes to gut gene expression

The antibiotics used here have overlapping microbial targets³⁰. Thus, after exploring the impact of individual antibiotics on gut gene expression, we aimed to determine which effects were unique to specific antibiotics. We trained a random forest classifier on transcriptional data from the colons of mice treated with the different antibiotics (including the four-antibiotic mix) via each of the three administration routes separately (Fig. 2a–c). Of all of the conditions and treatments, the random forest algorithm most correctly classified the i.p. treatment with vancomycin, followed by oral administration of the four-antibiotic mix, and i.p. treatment with neomycin (Extended Data Fig. 2a–c). This represents the unique effects that these antibiotics had on colonic gene expression.

Fig. 2. Random forest analysis reveals the distinct effects of different antibiotics and administration routes on host transcription.

a–c, Heatmaps showing the transcriptional changes of the top 400 genes that distinguish between the treatment groups for oral (a), i.p. (b) and i.v. (c) treatment with antibiotics. Each column represents a mouse. To avoid clutter, only selected gene names are shown as row labels; the complete gene list is provided in the Source Data for Fig. 2. Gene importance rankings are provided in Supplementary Table 6. d–f, GO pathway analysis, using Enrichr³⁴, for genes enriched following i.p. vancomycin (d), i.p. neomycin (e) and i.v. neomycin administration (f), clustered by the random forest (RF) classifiers. These three clusters are also highlighted in b and c. Each dot represents an enriched pathway. g, Heatmaps depicting the transcriptional changes of the top 400 genes that distinguish between the administration routes on the basis of the classifier. Each column represents a mouse. Gene importance rankings are provided in Supplementary Table 7. h, Expression levels of the indicated circadian-related genes in the colons of mice treated as indicated (n ≥ 4). Statistical significance was determined by one-way analysis of variance. The numbers on the charts indicate P values for comparisons with the relevant control. Counts per million (CPM) were determined using Voom/Limma. Each dot represents a mouse. NFAT, nuclear factor of activated T cells.

Source data

Extended Data Fig. 2. Random forest classification of host response to antibiotics in the colon.

a-c, Confusion matrices using random forest classifiers for colonic transcriptional data from mice treated orally (a), intraperitoneally (b) and intravenously (c) with the indicated antibiotics. Numbers in the squares of the matrices indicate the percentage of predictions for each category against the true classifications. Diagonal entries represent the accuracy of predictions for each category (true positives), while off-diagonal entries indicate the model’s misclassifications.

Source data

Pathway analysis of the genes selected by random forest to be the most discriminatory for i.p. vancomycin treatment revealed that these transcripts belong to a signalling pathway that is activated when antigens are presented via major histocompatibility complex (MHC) class II molecules to T cells³³ (Fig. 2d), indicating that i.p. vancomycin treatment alters colonic immune responses. Furthermore, i.p. and i.v. treatment with neomycin induced the expression of a unique cluster of genes involved in antiviral immune responses (Fig. 2e,f). Neomycin-induced antiviral responses in the host were previously described¹⁸. This illustrates that different antibiotics affect host transcription in distinct manners. Interestingly, many clusters of genes that were unique to certain antibiotics when administered individually were lost when the four antibiotics were combined (Fig. 2a–c).

Antibiotic administration routes affect host gene expression in the gut

Next, we wanted to determine whether different administration routes result in distinct effects on gut gene expression. We trained a random forest classifier on the colonic transcriptome of each antibiotic separately, based on administration route (Fig. 2g). We found that despite the distinct transcriptional profiles associated with different routes of antibiotic administration, all three administration routes of ampicillin and vancomycin had a strong and similar effect on the core clock genes Ciart, Nfil3, Per3 and Dbp, which control diurnal gene expression (Fig. 2h). Indeed, the effects of oral antibiotics on diurnal gene expression in the gut have been shown^13,14, but not those of systemic delivery. This was unexpected as the effects of vancomycin on the gut microbiota were distinct for oral versus systemic administration routes (Fig. 1g). Surprisingly, administering the four-antibiotic mix suppressed the effects of ampicillin and vancomycin on these core clock genes for all administration routes (Fig. 2h). Thus, different antibiotic administration routes produce distinct effects on the host transcriptome in some respects but have similar effects on core clock genes (for ampicillin and vancomycin), which are lost when antibiotics are combined.

Treatment with antibiotic pairs does not result in an additive effect

The results above suggest that combinations of antibiotics can mask the effects of these antibiotics when administered individually. Although the effects of antibiotic drug–drug interactions on bacterial killing have been studied^34,35, their effects on the host are unclear. As the four-antibiotic mix we used above is uncommon in the clinic, we explored these effects by pairwise testing drug–drug interactions in a set of clinically relevant pairs. We treated mice orally with vancomycin + metronidazole (used for Clostridioides difficile infections³⁶) or neomycin + metronidazole (used in surgery preparation³⁷). As for the four-antibiotic mix, combining metronidazole with vancomycin led to attenuation of the gene induction seen for individual administration (Extended Data Fig. 3a). Furthermore, combining neomycin and metronidazole suppressed ribosomal genes, even compared with the control group (Extended Data Fig. 3b). Thus, combining different antibiotics, even in pairs, can result in non-additive and counterintuitive effects on the host (Extended Data Fig. 3c–f).

Extended Data Fig. 3. Combination treatment effect on host transcription.

a-b, Heatmaps showing the genes that are suppressed in combined oral treatment with vancomycin and metronidazole (a) and neomycin and metronidazole (b) versus each antibiotic separately. c-d, Deviation from additivity computed using pydeseq2. Statistical testing was performed using DESeq2 (pydeseq2) with a factorial model including three binary design factors: Has-A, Has-B and Is-Combo. The Is-Combo coefficient tests whether the observed transcriptional effect of the antibiotic pair (metronidazole and vancomycin (c) or metronidazole and neomycin (d)) deviates from the additive expectation of the two single treatments. Wald tests were used for inference. P values were adjusted for multiple comparisons using the Benjamini–Hochberg (BH) method as implemented in pydeseq2. Tests are two-sided. e-f, Inter-replicate heterogeneity for metronidazole and vancomycin (e) or metronidazole and neomycin (f) (n ≥ 7). Centre line: median. Box bounds: 25th percentile (Q1) to 75th percentile (Q3). Whiskers extend to the most extreme data point within 1.5× the IQR from the box.

Source data

Use of CompoRes as a computational tool to associate microbiota compositional changes with gut gene expression

Our results indicate that microbial composition changes do not necessarily correlate with host gene expression changes. This prompted us to speculate that some antibiotics affect the host directly, regardless of their effects on the microbiota. Indeed, we and others have recently shown this to be the case for specific antibiotics^18–21. This line of investigation requires a tool to integrate microbial compositional data with host parameters and predict whether a change to microbial composition can explain host parameters. To understand whether antibiotics affected gut gene expression independent of the antibiotic-driven changes in the microbiota, we developed a composition–response (CompoRes) tool to account for the taxon–gene co-variation.

Existing methods^8,38 address this problem partially, but are limited by binary classification, a reliance on phylogenetic clustering regardless of actual taxon co-variation, or the absence of rigorous statistical significance estimation. Briefly, CompoRes preprocesses compositional data (handling zero-read taxa) and infers taxon correlations using SPARCC³⁹, hierarchically clustering them into operational compositional units (OCUs). Log-ratios of all OCU pairs are then correlated with gene expression profiles, and data shuffling yields a CompoRes P value quantifying the significance of these correlations (Fig. 3a and Methods). A low P value (high −log₁₀[P value]) indicates a significant association between gene expression and microbial composition.

Fig. 3. Vancomycin disrupts gut circadian gene expression in a microbiota-independent manner.

a, Schematic of the CompoRes algorithm. After preprocessing microbial compositional data, taxa are hierarchically clustered to form OCUs, regression is computed for each gene and the highest correlation coefficient is recorded. The data are then shuffled multiple times to generate a null hypothesis, estimating the correlation P value (Methods). b, Schematic of the experiments comparing DEGs from SPF versus germ-free (GF) mice treated systemically (by i.p. injection) with PBS or vancomycin (left), and DEGs from mice treated orally (p.o.) with PBS or vancomycin versus GF mice, which received an FMT from the orally treated mice (right)²¹. c,d, Results of CompoRes analysis, showing the relative frequency of P values (c) and a violin plot of the data (d) for correlations between microbiota composition and DEGs, as portrayed in b. Low P values indicate high correlation between differences in gene expression and changes in microbial composition. The curves show kernel density estimates of CompoRes −log₁₀[P values] for i.p. and p.o. routes. Statistical significance was determined by two-sample, one-sided Welch’s t-test. e, CompoRes analysis of all DEGs from neomycin-treated mice labelled in GO as having antiviral response functions. Each dot represents a DEG. The x axis presents CompoRes P values, with a small −log₁₀[P value] equivalent to a large P value, meaning that gene expression was not correlated with microbial compositional changes. The y axis presents DEG P values. f, CompoRes values (the same as in e) for autophagy-related DEGs from SPF mice treated with i.p. vancomycin. g, CompoRes values (the same as in e and f) of circadian-related DEGs from SPF mice treated with vancomycin as indicated. In c–g, the dotted lines indicate P = 0.05. DEG P values were calculated using DESeq2. Schematics in b created in BioRender; Bel, S. https://biorender.com/p5tyubd (2026).

Source data

To validate CompoRes, we used synthetic data (Extended Data Fig. 4a,b) and experimental data for which the ground truth (that is, a perturbation exclusively affecting the host directly or exclusively acting via the microbiota) was known. We used transcriptional data from germ-free mice treated with i.p. vancomycin (antibiotic perturbation that can affect the host only directly) and data from germ-free mice that received a faecal microbiota transplantation (FMT) from SPF mice that were treated orally with vancomycin (antibiotic perturbation that can affect the host only via changes to the microbiota)²¹ (Fig. 3b). CompoRes was used to analyse differentially expressed genes (DEGs) in both SPF and germ-free mice treated with i.p. vancomycin. This determined high CompoRes P values, indicating that their differential expression could not be explained by changes to microbiota composition (Fig. 3c). Conversely, DEGs in both SPF mice treated orally with vancomycin and germ-free mice that received an FMT from those orally treated mice showed low CompoRes P values, indicating that their differential expression can be explained by changes to the gut microbiota (Fig. 3c). Comparing the P values of the two groups of genes (i.p. versus oral administration) directly validated the ability of CompoRes to distinguish between perturbations that affect the host in a microbiota-independent versus microbiota-dependent manner (Fig. 3d). Moreover, neomycin-treated SPF mice did not exhibit correlations between microbiota composition and antiviral gene expression (Fig. 3e), as was previously shown¹⁸, supporting CompoRes’s ability to identify microbiota-independent perturbations. Finally, CompoRes analysis of autophagy-related genes revealed that their differential expression in response to vancomycin treatment did not correlate with microbial changes (Fig. 3f), consistent with previous findings²¹. Thus, CompoRes predicted germ-free outcomes solely from SPF data, testing hypotheses in silico before germ-free experiments were performed. Moreover, CompoRes works equally well when the microbiota composition is derived from metagenomic or 16S sequencing (Extended Data Fig. 4).

Extended Data Fig. 4. CompoRes validation and use.

a, Analysis of the CompoRes P values (as detailed in the Methods) calculated on synthetic data generated based on the CompoRes analysis of Ciart expression correlation in the orally administered vancomycin experiment. 13 synthetic datasets were generated with the observed correlation and 13 without, then CompoRes was run on each of the 30 datasets and a P value was calculated from shuffled data acting as the null hypothesis. Centre line (red): median. Red circle marker: mean. Box bounds: 25th–75th percentile (IQR). Whiskers extend to the most extreme data point within 1.5× the IQR from the box. Points beyond the whiskers are individual outlier markers. b, By considering sliding P value cutoff values for the data shown in a, one may generate an ROC curve and calculate the area under the ROC curve (AUROC) for the classification between correlated and uncorrelated data. c-e, Calculation of AUROC values for three clock genes (Nfil3 (c), Ciart (d) and Dbp (e)) in the orally administered vancomycin condition at various levels of gene expression noise. Box plot of RMSE values used as synthetic noise levels. Centre line: median. Box bounds: 25th–75th percentile (IQR). Whiskers extend to the most extreme data point within 1.5× the IQR from the box. Points beyond the whiskers are shown as individual outlier markers. f-g, Demonstration of CompoRes results for Ciart for the orally administered (f) and i.p. (g) vancomycin conditions. The yellow curves are the best correlation coefficients between these genes and compositional balances. The purple shaded area represents the distribution of correlation coefficients based on shuffled data (1,000 times shuffled), acting as the null distribution. Shaded dark: 25th and 75th percentiles, Shaded light: 5th and 95th percentiles. The centre represents the median. In f, the correlation for oral administration is separated from the 5th percentile and therefore has a P value of <5%. In g, the correlation for i.p. injection is within the 5th percentile and therefore has a P value of >5%. h, Comparison, for each host gene, of the CompoRes P value for association with microbiome composition when the microbial community is quantified by 16S rRNA amplicon profiling (y axis) or by whole-metagenome sequencing followed by MetaPhlAn 4 profiling (x axis). Error bars, shaded areas and centre are as in f and g. The two modalities show substantial agreement both at the level of binary significance calls (red dashed lines; P = 0.05) and across the full continuous range of inferred P values. Genes identified as significantly associated with microbiome composition in one modality also tend to be identified in the other, and the overall P values are strongly correlated (Pearson’s r = 0.579; P < 1 × 10⁻¹⁰⁰). Point colour indicates local point density. i, CompoRes analysis for the circadian genes Per2 and Clock following oral vancomycin treatment, using microbial community composition inferred either from 16S rRNA amplicon data (top) or shotgun metagenomic sequencing (bottom). In both data modalities, expression of Per2 and Clock is significantly associated with microbiota composition (P < 0.05), consistent with Fig. 3g. These genes are highlighted because the sampling time point was compatible with their circadian regulation. Thus, both 16S- and metagenomics-based community profiles support the same biological conclusion: expression of these clock genes tracks microbiome variation after orally administered vancomycin treatment (n ≥ 6).

Source data

After validating, we applied CompoRes to identify perturbations that affect the host response independent of microbiota composition. We chose to focus on the effects of antibiotics on the expression of core clock genes, as our observations above indicated that vancomycin administration via different routes has distinct effects on the gut microbiota (Fig. 1g) but similar effects on the expression of core clock genes (Fig. 2h). CompoRes revealed that alteration in the expression of core clock genes was correlated with compositional changes to the microbiota (low CompoRes P value) in mice treated orally with vancomycin (Fig. 3g), as previously shown^13,14,40,41. In contrast, these same genes showed low correlation with microbiota composition (high CompoRes P values) in mice receiving vancomycin systemically via i.p. injection (Fig. 3g). Thus, CompoRes identified that oral antibiotic treatment affects circadian gene expression in the gut via effects on the microbiota, but suggested that i.p. vancomycin treatment affects these genes independent of the microbiota. These results probably reflect the poor absorption of vancomycin when administered orally (Extended Data Fig. 1a,b).

Microbiota-independent disruption of colonic circadian rhythms

To investigate the impact of vancomycin on intestinal circadian rhythms, SPF and germ-free mice were treated via i.p. injection (as in Fig. 1). Transcriptional analysis of colons across a 24-h cycle revealed that systemic vancomycin altered core clock gene expression in both groups (Fig. 4a). These findings corroborate our CompoRes analysis, confirming that vancomycin-induced shifts in core clock genes are microbiota independent (Fig. 3g). Computational analysis identified genes that either gained or lost oscillatory expression patterns (Methods). At most time points, the overlap of affected genes between SPF and germ-free mice was significantly lower than expected (Extended Data Fig. 5). Moreover, although the peak transcriptional effect in SPF mice occurred at Zeitgeber time (ZT) −1 (Fig. 4b,c), vancomycin suppressed the greatest number of transcripts in germ-free mice at ZT 17 (Fig. 4d,e). Furthermore, the overlap of suppressed genes at ZT 17 was significantly lower than would be predicted by chance (P < 10⁻⁴; Extended Data Fig. 5). Collectively, these data demonstrate that vancomycin disrupts colonic diurnal gene expression through a microbiota-independent mechanism that is distinct from the microbiota-dependent mechanism.

Fig. 4. Vancomycin alters diurnal gene expression patterns in the gut in a microbiota-independent manner.

a, Expression levels (normalized reads) of transcripts over a 24-h period in colonic tissues of SPF or GF mice treated with i.p. PBS or vancomycin (n = 4). Statistical significance was determined by two-tailed student’s t-test. The numbers on the charts indicate P values and the data represent mean ± s.e.m. b–e, Heatmaps showing suppressed (b,d) and enhanced (c,e) groups of genes, determined based on statistical analyses across successive time points, in SPF (b,c) and GF (d,e) mice treated as in a. Each column represents a mouse. Specific lists of genes are provided in Supplementary Table 5.

Source data

Extended Data Fig. 5. Overlap of DEGs following vancomycin treatment in SPF and germ-free mice.

Venn diagrams depict the number of DEGs (unadjusted P ≤ 0.05; t-test) after i.p. administration of vancomycin versus PBS control at each indicated time point, comparing SPF and germ-free mice. Upper panels show genes with increased expression when compared with the SPF mice (enhanced genes), whereas lower panels show genes with decreased expression (suppressed genes). The intersection in each Venn diagram indicates genes significantly altered in both SPF and germ-free mice at that specific time point. Statistical significance of overlap size was assessed by Fisher’s exact test; arrows indicate whether the intersection is significantly larger (upward arrow) or smaller (downward arrow) than expected by chance. The significance of gene overlap between SPF and germ-free conditions was assessed using a one-sided empirical permutation test over 10,000 iterations (the P value is the proportion of permutations yielding an overlap equal to or greater than the observed overlap). Significance thresholds: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Source data

Vancomycin affects enteric neurons and induces hypophagia via neuromedin U in a microbiota-independent manner

Next, we wanted to determine the physiological consequences of the vancomycin-induced, microbiota-independent disruption to diurnal gene expression in the colon. Focusing on ZT 17, when transcriptional changes were most pronounced (Fig. 4d,e), Kyoto Encyclopedia of Genes and Genomes analysis revealed that these genes were primarily involved in the peroxisome proliferator-activated receptor (PPAR) signalling pathway and metabolism (Extended Data Fig. 6a,b). The PPAR family of nuclear receptors are nutrient-activated transcription factors⁴². Specifically, PPARɑ is induced by free fatty acids during fasting and displays diurnal expression^42–44. Whereas Ppara expression was lowest at ZT 17 in phosphate-buffered saline (PBS)-treated germ-free mice, vancomycin-treated germ-free mice exhibited a reversed pattern, with peak expression at ZT 17 (Fig. 5a).

Extended Data Fig. 6. Effects of vancomycin on the colon.

a,b, Kyoto Encyclopedia of Genes and Genomes (a) and Reactome (b) pathway analysis of transcripts with enhanced expression in vancomycin-treated germ-free mice at ZT 17. Each dot represents an enriched pathway. c, Microscopy of the myenteric plexus in mice treated orally with FITC-labelled vancomycin. Scale bar, 10 µm. Images represent similar results from three mice. d,e, Pathway analysis using MGI phenotype (d) and transcription factor (TF) prediction (e) of suppressed transcripts in vancomycin-treated germ-free mice at ZT 17. Each dot represents an enriched pathway.

Source data

Fig. 5. Vancomycin disrupts diurnal MHC class II expression by altering feeding behaviour via NMU.

a–d, Expression levels (normalized reads) of Ppara (a), Fabp4 (b), Cd36 (c) and Lpl (d) over a 24-h period in the colonic tissues of GF mice treated with vancomycin (n = 4). The data represent mean ± s.e.m. e,f, Change in food intake during 3 days of vancomycin treatment compared with 3 days before treatment in GF (e; n ≥ 9) and SPF (f; n ≤ 7) mice. g, Change in food intake after cessation of vancomycin treatment compared with 3 days before treatment (n ≥ 7). h,i, Expression levels of neuronal genes in the colons of vancomycin-treated SPF (h) and GF (i) mice. Each column represents a mouse. j, Microscopy images of the myenteric plexus in mice treated with fluorescein isothiocyanate (FITC)-labelled i.p. vancomycin. Scale bar, 10 µm. k, Change in food intake during 3 days of vancomycin treatment compared with 3 days before treatment in wild-type (WT) and Nmu^−/− mice (n = 3). l,m, Expression levels (normalized reads) of H2-Ab1 (l) and H2-Aa (m) in the colonic tissues of GF mice treated with vancomycin over a 24-h period (n = 4). The data represent mean ± s.e.m. n,o, Quantitative PCR analysis of H2-Ab1 transcription in the colons of fasting GF mice (n ≥ 5) treated orally with PBS versus glucose, but without vancomycin in either condition (n) or with both glucose and vancomycin versus vancomycin alone (o) (1 g kg⁻¹ body weight). p,q, Quantitative PCR analysis of H2-Ab1 transcription in MODE-K cells with or without starvation (p) and with or without glucose (q) (n ≥ 3). r, Microscopy image of neuronal-like differentiated PC12 cells treated with FITC-labelled vancomycin. Nuclear staining is shown in blue and the neuronal marker Tubb3 is shown in red. Scale bar, 10 µm. s–u, RNA-seq analysis of neuronal-like differentiated PC12 cells treated with vancomycin, showing DEGs (s), pathway analysis using genes downregulated in response to vancomycin (t) and the expression of ribosomal genes in response to vancomycin (u). v, Statistically significant changes in the expression levels of neuronal genes in the brains of flies reared on vancomycin. Each column represents 20 pooled fly head samples from an independent experiment. Statistical significance in a–g and k–q was determined by one-tailed student’s t-test. The numbers on the charts indicate P values. In t, each dot represents an enriched pathway. In e–g, k, n and o, each dot represents a mouse. In p and q, each dot represents a well. In s, each dot represents a gene. DAPI, 4′,6-diamidino-2-phenylindole; RQ, relative quantity.

Source data

As nocturnal animals, mice typically increase their expression of free fatty acid-scavenging genes (Cd36, Lpl and Fabp4) during the light phase as fasting triggers triglyceride hydrolysis; levels then drop during the dark phase as feeding commences^41,45. This pattern was observed in PBS-treated germ-free mice but reversed in vancomycin-treated mice, which reached peak expression at ZT 17 (Fig. 5b–d). This misalignment of metabolic gene expression suggested that vancomycin might alter feeding behaviour. Indeed, we found that vancomycin treatment reduced food intake in both germ-free and SPF mice (Fig. 5e,f), with feeding normalizing after treatment (Fig. 5g). Interestingly, a known side effect of i.v. vancomycin administration in humans is appetite loss⁴⁶.

We then wanted to determine how vancomycin induces altered feeding patterns. Transcriptional analysis showed that i.p. vancomycin uniquely altered enteric nervous system transcripts in both SPF and germ-free mice (Fig. 5h,i). Indeed, we detected vancomycin in myenteric plexus neurons following i.p. injection (Fig. 5j) but not oral administration (Extended Data Fig. 6c), potentially explaining these neuronal transcriptional changes. Notably, vancomycin induced expression of the neuropeptide neuromedin U (Nmu) and its receptor Nmur2 (Fig. 5h,i). Interestingly, the binding of Nmu to Nmur2 is known to suppress food consumption⁹. Indeed, we found that Nmu^−/− mice were resistant to vancomycin-induced hypophagia (Fig. 5k). Collectively, these data indicate that vancomycin disrupts diurnal metabolic gene expression by altering feeding behaviour via NMU signalling.

Diurnal MHC class II expression is dictated by feeding and is disrupted by vancomycin

We next focused on the consequences of the altered feeding patterns in vancomycin-treated mice. As the ZT 17 time point revealed enhanced expression of metabolism-related transcripts, we focused on suppressed transcripts at this time point. Pathway analysis suggested that vancomycin treatment suppressed the expression of genes transcribed by nuclear factor κB and genes involved in lymphocyte function (Extended Data Fig. 6d,e). Epithelial cells in the gut can affect lymphocyte function by the presentation of non-self antigens via MHC class II molecules¹⁴. We found that expression of the MHC class II representative genes H2-Aa and H2-Ab1 peaked at ZT 17 in the colons of PBS-treated germ-free mice, whereas this peak expression was completely suppressed in vancomycin-treated germ-free mice (Fig. 5l,m). This diurnal expression pattern in the colons of PBS-treated germ-free mice contrasts with MHC class II expression in the small intestine, which does not oscillate in the absence of gut microorganisms¹⁴. Given that vancomycin reduced feeding (Fig. 5e), we hypothesized that nutrients drive this diurnal MHC class II expression. Supporting this, glucose induced H2-Ab1 in fasting germ-free mice treated with either PBS or vancomycin (Fig. 5n,o). Furthermore, H2-Ab1 expression in MODE-K intestinal epithelial cells was glucose dependent in vitro (Fig. 5p,q). These data indicate that colonic epithelial MHC class II expression follows a diurnal, glucose-responsive pattern that is suppressed by vancomycin-induced hypophagia.

To understand the effects of vancomycin on neurons, we differentiated PC12 cells with nerve growth factor to produce neuronal-like cells⁴⁷ and treated them with vancomycin. Similar to our observations in mice (Fig. 5j), vancomycin was internalized into intracellular puncta in neuronal-like PC12 cells (Fig. 5r). RNA-seq identified 708 DEGs; most notably, there was suppressed expression of the gene encoding Ybx1, a regulator of neurogenesis and messenger RNA (mRNA) stability^48,49 and ribosomal proteins (Fig. 5s–u). This downregulation of translational machinery implies that vancomycin triggers a cellular stress response, consistent with the known role of NMU in endoplasmic reticulum stress and translation inhibition⁵⁰. This suggests that Nmu expression following vancomycin treatment might inhibit protein translation in response to vancomycin-induced stress.

Finally, we wanted to determine whether vancomycin-induced alteration of diurnal and feeding-related gene expression was conserved in evolutionarily distant animals. To test this, we reared flies (Drosophila melanogaster) in dishes containing vancomycin and sequenced adult fly brain RNA. We found that the expression of dClock (clk), the fly homologue of the mammalian Clock gene⁵¹, and its downstream target timeless (tim)⁵¹, both of which are part of the core clock machinery in flies, were altered in response to vancomycin (Fig. 5v). Also, the expression of klumpfuss (klu), which controls feeding behaviour in flies⁵², was altered following vancomycin treatment (Fig. 5v). Interestingly, klu was previously found to control the expression of hugin (hug)⁵², the fly homologue of Nmu⁵³. Accordingly, the expression of PK2-R1, the fly orthologue of the NMU receptor⁵⁴, was induced in fly brains in response to vancomycin (Fig. 5v), as was observed in mice for Nmur2 (Fig. 5i). Thus, the molecular response to vancomycin is, at least in part, preserved in evolutionarily distant organisms.

Discussion

Research in the past two centuries has identified the microbiota as a major player in host physiology and health⁶. For example, the microbiota’s roles in mediating the effects of non-antibiotic drugs such as metformin^55,56, immunotherapy^57–63 and psychotherapy drugs⁶⁴ are now well established. However, how exactly antibiotics shape host physiology (excluding toxicity) is not fully understood⁶⁵. Antibiotics are traditionally assumed to influence host physiology solely through microorganisms. Consequently, although tools such as FMT into germ-free models exist to test microbiota-mediated effects, tools to detect direct host–antibiotic interactions were lacking.

Addressing this gap, we developed CompoRes, distinguishing between host gene expression changes correlated with microbial shifts and those occurring independently. By comparing observed correlations against shuffled data, CompoRes identifies host genes that change in a microbiota-independent manner. With CompoRes, we demonstrated that vancomycin modulates host circadian and immune functions independent of its antimicrobial activity.

Investigating vancomycin-induced behavioural changes revealed two significant phenomena regarding MHC class II expression in the colon. First, unlike the small intestine, where MHC class II oscillations require microbial presence¹⁴, colonic expression remains diurnal in germ-free mice. Second, this expression is directly induced by feeding, specifically by glucose levels. These findings suggest an anticipatory immune mechanism. Since feeding is the primary route for microbial entry, the gut immune system may use glucose fluctuations to signal imminent antigen encounters⁶⁶. This link between metabolism and immunity offers a potential explanation for the rise of chronic inflammatory diseases in industrialized societies, where high glucose consumption and disrupted circadian feeding patterns are prevalent. Indeed, the link between diet-induced circadian expression of MHC class II molecules and inflammatory bowel diseases was demonstrated in mice¹⁴. In summary, our work challenges the microbiota-only model of antibiotic action and provides a framework for identifying direct effects on host physiology.

One limitation of this study is that we used antibiotics as a perturbation model in healthy animals, whereas the most routine use of antibiotics in humans is to treat individuals suffering from microbial infections. Because infection drastically alters host physiology, this limits the clinical interpretation of our results. However, our study explains how the host responds to antibiotics, which can affect infection resolution. Moreover, antibiotic use in healthy individuals is a common prophylactic in the clinic³⁷, and administering antibiotics to healthy livestock is common⁶⁷, making our findings relevant.

Methods

Ethics statement

All of the experiments using mice were conducted in compliance with the European Union directive regarding the protection of animals used for experimental and other scientific purposes. Experiments were approved by the institutional animal care and use committee of Bar-Ilan University (study ID 25-04-2019).

Mice

C57BL6/J and Swiss Webster mice were bred and maintained under a 12-h light/12-h dark cycle at 25 °C and fed standard chow under either SPF or germ-free conditions at the Azrieli Faculty of Medicine of Bar-Ilan University. Male SPF C57BL6/J mice were used for the experiments with multiple antibiotics (Figs. 1–3) and female SPF and germ-free Swiss Webster mice were used to study diurnal function (Figs. 4 and 5). Wild-type C57BL6/J mice were purchased from Envigo RMS Israel and Nmu^−/− mice⁶⁸ (Nmu^tm1Mko (RBRC04549); RIKEN Japan) were generated from frozen sperm by Bar-Ilan University’s Transgenic Unit. Mice aged 8–12 weeks were used for all of the experiments.

Antibiotic treatment

Mice were randomly divided into either control or antibiotic treatment groups (three mice per cage). Mice were treated twice a day for 3 days, and once more on the fourth day, via oral gavage, i.p. injection or i.v. injection. Mice received 2.5 mg of either ampicillin, metronidazole, neomycin or vancomycin, a mix of all four antibiotics or selected pairs of antibiotics, in 100 µl PBS. Control mice received 100 µl PBS. Four hours after the last treatment, mice were euthanized and tissues were extracted. For the treatment of germ-free mice, mice were removed from sterile isolators into Tecniplast ISOcages, where they were maintained for the duration of the treatment. For food intake measurements, food weight was measured daily.

For in vivo imaging of labelled vancomycin, mice were administered 750 µg fluorescein isothiocyanate-labelled vancomycin either orally or via i.p. injection. Three hours after administration, mice were euthanized for collection of the myenteric plexus. Isolated tissues were fixed in 4% paraformaldehyde for 30 min. For immunofluorescence, tissues were incubated in a blocking buffer (10% foetal bovine serum (FBS) and 0.1% Triton X-100) and then treated overnight at 4 °C with a mouse anti-β3-tubulin primary antibody (Abcam; 1:200). Following primary incubation, a secondary anti-mouse Alexa Fluor 594 antibody (1:300) was applied for 1 h at room temperature. Representative images were captured using a ZEISS Apotome microscope.

Flies

Fly stocks (Oregon R) were obtained from the Bloomington Drosophila Stock Centre. The flies were maintained in 50 ml Bio-Reaction tubes (RED BRT010051-ZX; Alex Red) to which ~7 ml cornmeal, molasses and yeast (CMY) growth media was added. The flies were raised in a light/dark cycle of 12 h light/12 h dark at 25 °C in an incubator. After reaching a population size of ~300 flies, most of the flies were transferred to a population cage. The flies in the population cage were maintained on CMY media in 10-cm petri dishes. The eggs and larvae were harvested from the petri dish every two days when the plate was changed and then transferred to the tubes with CMY media. The population in the cage was maintained by transferring new flies bi-weekly into the cage.

For antibiotic treatment, a fresh 10-cm petri dish with CMY media was kept in the Drosophila population cage for 1 h to collect eggs. This was done so that almost all of the larvae would hatch around the same time from the eggs. The collected eggs were then transferred to 50 ml reaction tubes with CMY media. When the larvae matured into the late pupae stage, they were collected and then transferred onto 10-cm petri dishes with CMY media + vancomycin (6 mg ml⁻¹). The adult flies that hatched from the pupae were reared on vancomycin-laced plates for 24 h. After 24 h, flies exposed to the antibiotic-laced food were flash frozen in liquid nitrogen and kept at −80 °C. Twenty fly heads were pooled per sample, and total RNA was extracted using TRIzol reagent. RNA-seq was performed by Procomcure NGS, Austria.

Microbial DNA extraction for 16S amplicon sequencing

DNA was extracted from mouse faecal samples using the PureLink Microbiome DNA Purification Kit (Invitrogen) according to the manufacturer’s protocol and following a 2-min bead-beating step. The V4 region of the bacterial 16S rRNA gene was amplified by polymerase chain reaction (PCR) using the 515F (AATGATACGGCGACCACCGAGATCTACACGCT) barcoded and 806R (TATGGTAATTGTGTGYCAGCMGCCGCGGTAA) primers⁶⁹. A reaction contains a final concentration of 0.04M of each primer and 0.50M of PrimeSTAR Max DNA Polymerase (Takara Bio) or Phanta Flash mix (Vazyme) in 50 μl total volume. PCR reactions were carried out via 30–35 cycles of denaturation (95 °C), annealing (55 °C) and extension (72 °C), with final elongation at 72 °C. PCR products were purified using AMPure magnetic beads (Beckman Coulter) or VAHTS DNA Clean Beads (Vazyme) and quantified using a DeNovix spectrophotometer. Samples were then pooled at equal amounts, loaded on 2% agarose E-Gel (Thermo Fisher Scientific), purified and sent for sequencing using an Illumina NextSeq platform (Genomic Technology Center, Azrieli Faculty of Medicine, Bar-Ilan University).

High-performance liquid chromatography analysis

Samples were analysed by injecting 5 µl of the solutions into an ultra-high-performance liquid chromatography instrument connected to a photodiode array detector (Dionex UltiMate 3000) with a reverse-phase column (Kinetex C18; length = 100 mm; internal diameter = 2.1 mm; particle size = 2.6 μm; pore size = 100 Å; pentafluorophenylpropyl). The mobile phase comprised distilled deionized water with 0.1% formic acid (A) and acetonitrile containing 0.1% formic acid (B). The gradient was started with 2% B and increased to 90% B in 7 min, then kept isocratic at 90% B for another 2 min. Phase A was returned to 2% B in 1 min and the column was allowed to equilibrate at 2% B for 4 min before the next injection. The flow rate was 0.3 ml min⁻¹.

Vancomycin liquid chromatography–tandem mass spectrometry analysis

Liquid chromatography–tandem mass spectrometry analysis was performed with a heated electrospray ionization (HESI-II) source connected to a Q Exactive Plus Hybrid Quadrupole-Orbitrap Mass Spectrometer (Thermo Fisher Scientific). The electrospray ionization capillary voltage was set to 3,500 V, the capillary temperature to 300 °C, the gas temperature to 350 °C and the gas flow to 10 ml min⁻¹. The mass spectra (m/z 680–1,500) were acquired in positive-ion mode (ESI+). Peak determination and peak area integration were performed with Quan Browser (Thermo Xcalibur, version 4.1.31.9). Auto-integration was manually inspected and corrected if necessary.

RNA extraction and sequencing

RNA from flash-frozen mouse colonic tissues was extracted using an RNeasy Plus Universal Mini Kit (Qiagen). The integrity of the isolated RNA was analysed using an Agilent High Sensitivity RNA ScreenTape Kit and TapeStation 4200 at the Genomic Technology Center of the Azrieli Faculty of Medicine at Bar-Ilan University. 1,000 ng total RNA was treated with a NEBNext Poly(A) mRNA Magnetic Isolation Module (E7490L; New England Biolabs). RNA-seq libraries were produced using the NEBNext Ultra II RNA Library Prep Kit for Illumina (E7770L; New England Biolabs). Quantification of the library was performed using a Qubit dsDNA High Sensitivity Assay Kit (Molecular Probes and Life Technologies) and qualification was done using an Agilent TS D1000 kit and TapeStation 4200. 250 nM of each library was pooled together and diluted to 4 nM according to the NextSeq manufacturer’s instructions. 1.6 pM was loaded onto the flow cell with 1% PhiX library control. Libraries were sequenced with an Illumina NextSeq 550 platform with single-end reads of 75 cycles, according to the manufacturer’s instructions.

Quantitative real-time PCR

Total RNA was extracted from colonic tissues, as above, for RNA-seq. Equal amounts of mRNA were then converted by reverse transcription to complementary DNA using a High-Capacity RNA-to-cDNA Kit from Thermo Fisher Scientific. Quantitative PCR was performed using PowerUp SYBR Green Master Mix (Applied Biosystems). Data were collected and analysed according to the 2^− ΔΔCT method using either a StepOnePlus Real-Time PCR System or the QuantStudio 5 Real-Time PCR System (Applied Biosystems, Thermo Fisher Scientific). Each sample was tested in triplicate at a final volume of 10 μl. Gene expression was calculated relative to mRNA levels of the housekeeping gene RPLP0 as a control. Primers were designed using Primer-BLAST (National Center for Biotechnology Information (NCBI), National Institutes of Health). The primers used were RPLP0 forward (AATCTCCAGAGGCACCATTG) and reverse (ACCCTCCAGAAAGCGAGAGT) and H2Ab1 forward (AGCCCCATCACTGTGGAGT) and reverse (GATGCCGCTCAACATCTTGC).

Cell culture

The mouse intestinal epithelial cell line MODE-K⁷⁰ was cultured in Dulbecco’s Modified Eagle Medium high-glucose medium (Gibco) supplemented with 10% FBS (Biological Industries), 1% penicillin–streptomycin (Biological Industries) and 1% GlutaMax (Gibco). Cells were cultured in a 37 °C humidified incubator with 5% CO₂. Cells were seeded in six-well plates at a density of 500,000 cells per well. After 24 h, the media was replaced with Dulbecco’s Modified Eagle Medium without glucose (starvation media; Gibco) for 16 h, after which the media was supplemented with 6 mM glucose for 4 h.

PC12 cells were seeded in six- or 12-well plates and maintained in RPMI 1640 medium supplemented with 2% FBS, 1% L-glutamine and 1% penicillin–streptomycin. To induce neuronal differentiation, 50 ng ml⁻¹ nerve growth factor was added every 48 h for 7 days. Following differentiation, cells were treated with 500 µg ml⁻¹ vancomycin for 24 h for RNA extraction or 100 µg ml⁻¹ fluorescein isothiocyanate-conjugated vancomycin for 3 h for immunofluorescence analysis. For imaging, cells were fixed in 4% paraformaldehyde for 10 min. After blocking in a buffer containing 10% FBS and 0.1% Triton X-100, cells were incubated overnight at 4°C with a mouse anti-β3-tubulin primary antibody (Abcam; 1:200). Secondary labelling was performed using an anti-mouse Alexa Fluor 594 antibody (1:300) for 1 h at room temperature. Images were acquired using a ZEISS Apotome microscope.

Metagenomic sequencing and analysis of mouse stools

DNA quality control and dilution

The DNA was isolated and quantified using Quant-iT 1X dsDNA Broad Range Assay Kits (Q33267; Life Technologies) in combination with a Varioskan LUX Multimode Microplate Reader (VL0000D0; Thermo Fisher Scientific). The DNA was then diluted in water for the following library preparation.

Library preparation and sequencing

The sequencing libraries were prepared with the Illumina DNA Prep (M) Tagmentation (96 Samples, IPB) Kit (20060059; Illumina) in combination with Illumina DNA/RNA Unique Dual Indexes (2009167/8/9; Illumina) and the amplified libraries were purified using the double-sided bead purification procedure, as described by the Illumina protocol. Then, library concentrations (ng µl⁻¹) were quantified using Quant-iT 1X dsDNA High Sensitivity Assay Kits (Q33232; Life Technologies) in combination with the Varioskan LUX Microplate Reader (VL0000D0; Thermo Fisher Scientific). In addition, the base pair length was evaluated using the D5000 ScreenTape Assay (5067-5588/9; Agilent Technologies) in combination with the TapeStation 4150 (G2992AA; Agilent Technologies). By knowing both the library concentration and the base pair length, it is possible to obtain the correct library volume to pool in the same tube in order to achieve optimal cluster density. The library pool was then quantified with the Qubit 1X dsDNA High Sensitivity kit (Q33231; Life Technologies) through the Qubit 3.0 Fluorometer (Q33216; Life Technologies) and the base pair length was evaluated as described before. Finally, the library pools were sequenced using the NovaSeq X Plus platform (Illumina) at an average depth of 7.5 Gb per sample.

Inferring community composition

Raw paired-end shotgun metagenomic reads were processed to remove host-derived sequences using KneadData (version 0.12; bioBakery) with the mouse C57BL/6NJ reference genome (Bowtie2 index). Each sample was processed independently with the --reorder flag to maintain read-pair synchronization. Paired reads that did not align to the host genome were retained. Taxonomic profiling of host-depleted reads was performed with MetaPhlAn 4.1 (mpa_vJan25_CHOCOPhlAnSGB_202503 database; downloaded February 2026) using the --unclassified_estimation flag to report the fraction of reads that could not be classified. Per-sample relative abundance profiles were merged into a single table using the merge_metaphlan_tables.py utility distributed with MetaPhlAn, filling absent taxa with zeros.

Analysis of the 16S amplicon sequencing data using QIIME 2

For dataset 1 (Gene Expression Omnibus (GEO) accession GSE293029), the 16S rRNA-seq data were analysed using QIIME 2 (ref. ⁷¹) (version 2023.9.2). Raw single-end sequences were imported, demultiplexed and processed with DADA2 (ref. ⁷²) for quality control and denoising, generating amplicon sequence variants (ASVs). Taxonomic classification of ASVs was performed using a pre-trained Naive Bayes classifier specific to the V4 region, based on the Greengenes2 (ref. ⁷³) database (version 2022.10). The feature table was exported in BIOM format for downstream analyses.

For dataset 2 (BioProject accession PRJNA1442090), the 16S rRNA-seq data were analysed using QIIME 2 (version 2024.10). Raw single-end sequences were imported and demultiplexed using qiime demux emp-single with reverse-complemented barcodes; the resulting per-sample FASTQ files were deposited with the accession PRJNA1442090. Demultiplexed reads were denoised with DADA2 (denoise-single), applying a truncation length of 270 base pairs (bp), no left trimming, a maximum expected error threshold of 2 and consensus-based chimera removal, generating ASVs. Taxonomic classification of ASVs was performed using a Naive Bayes classifier trained in house on the V4 hypervariable region (primers 515F (GTGCCAGCMGCCGCGGTAA) and 806R (GGACTACNVGGGTWTCTAAT) of the SILVA rRNA database (version 138; 99% operational taxonomic units). The feature table was collapsed to species level (level 7) and exported in BIOM format for downstream analyses.

Using ANCOM to detect statistically significant differences in relative microbial abundance

To identify bacterial genera significantly altered by antibiotic treatment, we applied the ANCOM-BC2 (refs. ^74,75) method to genus-level 16S rRNA-seq data. Analyses were conducted separately for each combination of antibiotic (ampicillin, vancomycin, neomycin, metronidazole or a mixture) and administration route (i.p., i.v. or oral), comparing antibiotic-treated groups with their respective PBS controls. Data were organized using the phyloseq⁷⁶ package, and family or genera with no variance within any group were excluded. The ANCOM-BC2 model was fitted with treatment category as the fixed effect, allowing estimation of log[fold changes] for each genus in the antibiotic-treated group relative to the PBS control. Statistical significance was determined using Benjamini–Hochberg adjusted P values, with a threshold of q < 0.05. The output is summarized in Supplementary Table 2.

Principal coordinates analysis of the taxonomic composition

Principal coordinates analysis was performed to visualize the dimensional reduction of microbial community composition data across experimental conditions. Aitchison distances were computed from the input abundance matrix by first applying the centred log-ratio (CLR) transformation. To prevent undefined values in log-based transformations, a pseudocount of 1 × 10⁻⁸ was added to all entries. The CLR-transformed data were then used to calculate pairwise Euclidean distances, forming the Aitchison distance matrix. Principal coordinates analysis was conducted using the PCA implementation from the Python scikit-learn library, retaining the first two principal components for visualization. To aid in assessing group-level variation and separation, ellipses representing 2-s.d. confidence regions based on the data’s covariance were drawn around each group’s centroid.

Using Adonis2 to infer significant changes between antibiotic and PBS treatments

To analyse differences in microbiome composition between treatment and control groups, we performed comparative analyses across multiple antibiotic conditions using permutational multivariate analysis of variance. Raw microbial abundance data were preprocessed by adding a small pseudocount (1 × 10⁻⁸) to prevent undefined values during log-ratio transformation. Data were subjected to CLR transformation, and Aitchison distances were calculated as Euclidean distances between the transformed samples. We conducted focused permutational multivariate analysis of variance tests (adonis2 function; vegan R package⁷⁷) with 10⁵ permutations to statistically assess differences between experimental groups. Our analytical framework systematically compared antibiotic-treated groups with corresponding PBS controls and assessed baseline variability by comparing PBS control groups with each other.

Mapping RNA sequences for mouse transcriptome analysis

The RNA-seq data were analysed using the RASflow pipeline⁷⁸. The reads were trimmed using Trim Galore version 0.6.2 and adapters were removed using Cutadapt version 3.5 (quality Phred score cutoff = 20; adaptor sequence = AGATCGGAAGAGC (Illumina TruSeq, Sanger inverse PCR; auto-detected); maximum trimming error rate = 0.1 (default); minimum required adaptor overlap (stringency) = 1 bp; minimum required sequence length before a sequence is removed = 20 bp). Reads were mapped to the transcriptome (Mus_musculus.GRCm39) using Salmon⁷⁹. The normalized values are estimated transcripts per million from Salmon and are scaled using the average transcript length over samples and then the library size using tximport. Any missing gene names (using mmusculus_gene_ensembl) are replaced with gene IDs. We filtered out sparse genes, which we defined as those with over 50% zero values, to concentrate on genes with meaningful expression levels and to decrease computational complexity. We started with 35,534 genes and removed 15,077 sparse genes, leaving us with 20,457 genes. Samples with less than 55% alignment (four mice) were excluded from the analysis. We normalized each sample’s reads to 1 million reads. Six mice were removed after being identified as obviously different from their group members and control mice in an a multidimensional scaling plot. We then normalized each gene by its PBS controls to obtain a z score in the following way. Missing values were imputed using the minimum non-zero value of the same gene from other samples sharing the same treatment and antibiotic condition. After imputation, the data were log₂ transformed. Expression levels were then standardized using z scores relative to the mean and standard deviation of the PBS-treated (control) group for each treatment. Standard deviations of zero were replaced by the smallest non-zero standard deviation to avoid division errors.

In vitro analysis of neuron-like PC12 cells

Sequencing

RNA-seq libraries were produced using the QuantSeq 3′ mRNA-Seq V2 Library Prep Kit (LX-191.96). Quantification of the library was performed using a Qubit dsDNA High Sensitivity Assay Kit (Molecular Probes and Life Technologies) and qualification was done using the Agilent TS D1000 kit and TapeStation 4200. Then, 200 nM of each library was pooled together and diluted to 2 nM, after which 750 pM was loaded onto the flow cell with 2% PhiX library control according to the NextSeq manufacturer’s instructions. Libraries were sequenced with the Illumina NextSeq 1000 platform, with P1 kit single-end reads of 100 bp.

Mapping

Reads were mapped to the transcriptome (Rattus_norvegicus.GRCr8) using Salmon⁷⁹. The normalized values are estimated transcripts per million from Salmon and are scaled using the average transcript length over samples and then the library size by tximport. Any missing gene names (using rnorvegicus_gene_ensembl) are replaced with gene IDs. The resulting feature table can be found in Supplementary Table 13. We filtered out sparse genes, which we defined as those with over 50% zero values, to concentrate on genes with meaningful expression levels and to decrease computational complexity. We started with 22,886 genes and removed 9,520 sparse genes and 13 mitochondrial genes, leaving us with 13,353 genes. We normalized each sample’s reads to 1 million reads. Outliers were defined using a leave-one-out approach (z score > 3 compared with the remaining group replicates), a log₂[fold change] > 2 and an expression level of >5 for both the individual sample and the group mean to filter out low-expression noise. Based on these criteria, two samples (C16 and V21) were excluded due to an excessive number of outliers (769 and 227, respectively), whereas all of the remaining samples exhibited between 60 and 125. We then normalized each gene by its PBS controls to obtain a z score in the following way: missing values were imputed using the minimum non-zero value of the same gene from other samples sharing the same treatment and antibiotic condition. After imputation, the data were log₂ transformed. Expression levels were then standardized using z scores relative to the mean and standard deviation of the PBS-treated (control) group for each treatment. Standard deviations of zero were replaced by the smallest non-zero standard deviation to avoid division errors.

GSEA

To obtain an overview of how the various antibiotics influence gut gene expression, we employed the Gene Ontology framework³¹, which organizes genes into hierarchically arranged clusters associated with biological functions ranging from broad categories (for example, immune system process) to specific functions (for example, positive regulation of T cell antigen processing and presentation). GSEA³² was used to identify statistically significant, coordinated differences in predefined gene sets between control and treated samples. GSEA assesses whether genes associated with a particular functional group are over-represented at the top or bottom of a ranked list—where the ranking is based on a signal-to-noise metric—with genes at the top considered enhanced (upregulated) and those at the bottom considered suppressed. After identifying differentially expressed Gene Ontology clusters, we aggregated them according to the highest Gene Ontology annotation level to determine the fraction of Gene Ontology terms within each top-level category that were differentially expressed (for example, the proportion of Gene Ontology clusters enriched in the immune system process category). Each Gene Ontology category was assigned an enrichment value between 0 and 1, with 1 indicating that all Gene Ontology terms in that category were either upregulated or downregulated. The aggregated GSEA results are shown in Fig. 1i and detailed enrichment tables are provided in Supplementary Tables 3 and 4.

Random forest classification

A random forest classifier from the sklearn.ensemble library was used to classify samples across different antibiotic treatment groups. To address imbalances in group sizes, each group was randomly subsampled to match the minimum number of samples across groups in every repetition. The classifier was trained and evaluated over 10,000 iterations, with performance assessed by averaging confusion matrices and feature importance scores across runs. The test size was 0.3 for comparisons of different antibiotics with the same administration route (Fig. 2a–c) and 0.25 for comparisons of different administration routes for the same antibiotic (Fig. 2g). The top 400 most important features were identified and displayed for further analysis. The gene importance rankings can be found in Supplementary Tables 6 and 7. Specific algorithm settings were: n_estimators=100, max_depth=None, max_features=‘sqrt’, criterion=‘gini’ (we did not use permutation importance).

CompoRes algorithm

Correlating host gene expression with microbial community composition poses two major challenges. First, 16S rRNA-seq data are compositional in nature, which can lead to spurious correlations if not appropriately handled. Second, the high diversity of microbial taxa results in a vast number of pairwise comparisons relative to the limited sample size, increasing the risk of false-positive associations. To address these issues, we developed CompoRes. Initially, pairwise correlations among taxa abundances were estimated using SPARCC^39,80, a method that accounts for the compositional constraints inherent to 16S rRNA amplicon data. These correlations were then used to define OCUs, which are coarse-grained aggregations of ASVs. Specifically, ASV–ASV correlations were hierarchically clustered such that, with increasing clustering thresholds, ASVs were progressively bundled into OCUs containing an ever-larger number of ASVs. This hierarchical aggregation permits the evaluation of correlations between OCU abundances and gene expression at varying levels of data coarseness: fewer OCUs reduce the number of comparisons and thus enhance statistical power, albeit at the potential cost of signal dilution due to the aggregation of diverse ASVs.

To leverage the advantages of different coarsening levels while mitigating signal dilution, we assessed gene expression–composition correlations across multiple clustering thresholds and averaged the resulting metrics (the mean of the log of the P values). At each level, correlation was quantified by performing linear regression on the log ratio of all possible OCU pairs (the so-called balance, as described by Selbal³⁸, and selecting the pair that exhibited the strongest association. Given the extensive number of comparisons, high correlation coefficients might occur by chance. Therefore, we generated a null distribution by shuffling the data at least 50 times and recalculating the correlation coefficients for each permutation. The significance of the observed correlation was then determined by computing the P value relative to this null distribution, and the logarithms of the P values across all OCUs were averaged.

To test the usefulness of our mean-log P value approach we sought to find whether it can classify between ground-truth correlated and uncorrelated data that mimic experimental data. As an example, we used the clock genes in orally administered vancomycin. The correlation between composition and clock gene response, for example, in Nfil3, gives a regression slope and mean squared error about this regression line. Taking the values for the slope and mean squared error from the experimental data, we generated 30 noisy realizations of synthetic data from the same microbiota composition but synthetically generated gene expression responses: 15 realizations with the slope as in the real data and 15 with zero slope. Each realization had the same noise level as in the experimental data regression. We then computed the CompoRes P values for each of the 30 realizations and plotted the P values split into the two groups (Extended Data Fig. 4a). By choosing increasing values of the P value cutoff (going up the y axis in Extended Data Fig. 4a), we plot the receiver operating characteristic (ROC) curve (Extended Data Fig. 4b) and compute the area under the ROC curve. We observed that CompoRes displayed very high area under the ROC curve values for the clock genes (Extended Data Fig. 4c–e), for a range of root mean squared error values corresponding to different OCU clusterings of the composition. Therefore, the CompoRes P values are demonstrably useful metrics to classify whether or not experimental data are correlated with microbiota changes.

Having established the usefulness of the CompoRes P value approach, we applied it on the gene lists described in the main text. We noticed that in some cases the correlation from the true data was well above the distribution of correlations from the shuffled datasets, yielding small P values indicative of statistically significant associations between gene expression and microbiota composition (for example, Extended Data Fig. 4f). In contrast, when the data fell within the range of the shuffled distribution, the P value was large, and the association was deemed non-significant (for example, Extended Data Fig. 4g). Note that in this example, the difference in conditions is between orally administered vancomycin (correlated with microbiota) or vancomycin administered by i.p. injection (uncorrelated) when considering the clock gene Ciart.

Finally, we sought to verify that CompoRes performs well on community compositions inferred from full metagenomic sequencing, as well as for 16S. For this purpose, we used our antibiotics dataset with mouse RNA-seq and both metagenomic and 16S sequencing of the same mouse. We inferred community composition from the metagenomic sequencing using MetaPhlAn 4.1 (ref. ⁸¹). We then ran CompoRes separately on the differentially expressed mouse genes versus either the 16S-inferred or MetaPhlAn-inferred community composition. For each gene, we plotted the CompoRes P value using the 16S community versus the P value using the MetaPhlAn community (Extended Data Fig. 4h). As is apparent, genes that are significantly correlated with the microbiota composition (P < 0.05) are deemed so by both sequencing methods. Furthermore, the CompoRes P value itself is highly correlated between the methods (Pearson r = 0.579; P < 1 × 10⁻¹⁰⁰). To validate the ability of CompoRes to identify transcripts whose expression levels correlate with changes to microbial composition, we focused on known circadian genes, as in Fig. 3. We found that running CompoRes using either 16S or metagenomic data both identified Per2 and Clock and they were both deemed correlated with the microbiota (Extended Data Fig. 4i).

Oscillation patterns

For each gene, we first performed a t-test at the initial time point (ZT −1) using a P value threshold of 0.05 and a log₂[fold change] threshold of 2. Subsequently, we iterated over pairs of successive time points, performing statistical analyses to classify gene expression changes as enhanced, suppressed or stable. Specifically, a gene was classified as enhanced at ZT −1 if the treated group exhibited significantly higher expression than the controls and as suppressed if the reverse was true. At later time points, enhancement was assigned when the treated group’s expression increased significantly relative to its previous time point while the control group remained stable or decreased, or when the control group’s expression decreased significantly while the treated group did not. Conversely, suppression was assigned when the control group’s expression increased significantly while the treated group remained stable or decreased, or when the treated group’s expression decreased while the control group remained stable. This method selects for the enhanced and suppressed genes at each time point relative to the control group, and as a result also often shows the opposite effect at the previous time point, although this opposite effect is not selected for.

Statistical analysis and illustrations

Analyses with Student’s t-test and analysis of variance were performed using GraphPad Prism. Scientific illustrations were created using BioRender or Python.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

A.E. and S.B. contributed equally. M.N.-O. and S.B. conceived of the study idea. L.Z., Y.L., O.K., A.E. and S.B. designed, performed and analysed the experiments and wrote the paper. J.S., M.N.-O., E.R., A.G., Z.M., A.Y., A.A., H.S., S.B.-S., S.M., M.N., S. Telpaz, M.Z., J.D., M.K., S.P., L.M., S. Turjeman, M.W. and S.K. performed the experiments and analyses. O.K., A.E. and S.B. supervised the study.

Peer review

Peer review information

Nature Microbiology thanks Samuel Nobs and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Funding

This project was supported in part by the Azrieli Foundation Early Career Faculty Fellowship (to S.B.), Israel Science Foundation (925/19 and 1851/19 to S.B.), European Research Council Starting (101039927 to S.B.) and Consolidator grants (101001355 to O.K.) and Israeli Ministry of Science and Technology (3011006447 to A.E.). Y.L. was partly funded by the Hebrew University’s Center for Interdisciplinary Data Science Research.

Data availability

Microbial compositional data and host transcriptional data were deposited in the NCBI’s GEO database under accessions GSE293364 (a superseries containing data on the RNA-seq of colonic tissue from mice treated with different antibiotics, as well as 16S amplicon sequences from faecal samples from these mice) and GSE293029 (RNA-seq data from SPF and germ-free mice treated with i.p. vancomycin and sampled every 6 h over 24 h). Data on the paired antibiotics with stool metagenomics and 16S amplicon sequencing have been deposited in the NCBI’s BioProject database under accession PRJNA1442090, with metadata provided in Supplementary Table 9. RNA-seq data for the PC12 cells have been deposited in the NCBI’s GEO database under accession GSE324684. Source data are provided with this paper.

Code availability

Code has been deposited at GitHub via https://github.com/AmirErez/CompoRes. The scripts used to analyse the data have been deposited at GitHub via https://github.com/AmirErez/Manuscript-Analysis-of-Host-and-Microbiome-Responses-to-Antibiotics.

Competing interests

The authors declare no competing interests.

Extended data

is available for this paper at https://doi.org/10.1038/s41564-026-02453-w.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41564-026-02453-w.