Nutrient competition predicts gut microbiome restructuring under drug perturbations

SUMMARY

Human gut bacteria are routinely exposed to stresses, and community-level responses are difficult to predict. To interrogate these effects, we screened stool-derived in vitro communities with 707 clinically relevant drugs. Across ~5,000 community–drug conditions, compositional and metabolomic responses were shaped by nutrient competition, with certain species expanding due to suppression of competitors. Most compositional changes arose from strain extinction and were reversed by reintroducing extinct species, although certain drugs promoted alternative states long after treatment. Despite strong selection pressures, resistance emergence was infrequent. Responses to drugs were qualitatively conserved across communities, while nutrient competition quantitatively tuned species abundances, consistent with consumer-resource model predictions. Together, nutrient competition provides a predictive framework to anticipate and potentially mitigate drug side effects on the gut microbiota.

INTRODUCTION

The human gut microbiota is a dense ecosystem routinely exposed to perturbations. Many therapeutic drugs, particularly antibiotics, exert collateral damage on gut communities^1–3, causing loss of key commensals and unintended compositional shifts that compromise host health. Numerous non-antibiotic drugs also inhibit human gut commensals^4–7. While studies of isolates have revealed drug susceptibility patterns^3,4, it remains unclear how these susceptibilities translate to communities, in which interspecies interactions can modulate, buffer, or amplify drug perturbations.

In vitro culturing has enabled mechanistic studies of microbial communities, including nutrient competition^8,9, cross-feeding¹⁰, and responses to supplements¹¹ and drugs^12,13. These communities preserve many ecological interactions¹⁴ while enabling high-throughput screening. Prior work shows that a species becomes indirectly sensitized when its cross-feeding partner is inhibited¹⁵ or if another species alters pH¹⁶. A 32-species synthetic community revealed cross-protection via drug biotransformation and bioaccumulation¹³. These examples highlight that drug responses reflect both isolate-level susceptibility and community-level interactions, motivating a systematic, quantitative examination of drug–community interactions.

Nutrient competition is a key but underexplored factor in shaping community responses¹⁷. Competitive release, in which rare opportunistic pathogens expand due to suppression of dominant species, has been observed during antibiotic treatment¹⁸. Yet, it remains unclear how broadly such ecological mechanisms influence responses across drug classes and community compositions, or the extent to which these effects can be predicted from the growth and susceptibility of constituent species.

Collateral damage can have severe consequences. Antibiotics can shift the gut microbiota into a prolonged disrupted state^19,20, potentially driven by species extinction^9,21, which can be restored via fecal material transplant^22,23. However, mechanisms determining recovery versus long-term disruption remain poorly understood. A systematic and predictive understanding of how drugs reshape communities is therefore essential for anticipating side effects and improving therapeutic outcomes.

Here, we screened 707 small-molecule drugs against stool-derived in vitro communities (SICs) from nine donors. 141 drugs altered community growth, composition, and metabolome. Species-level response was collectively explained by isolate susceptibility and nutrient competition. Short-term drug treatments led to long-lasting compositional changes due to extinction, reversible by reintroduction of the extinct strains. In rare cases, a phylogenetically related species stably replaced an original member, suggesting higher-order interactions. Resistance emergence was infrequent, highlighting the interplay between selection and community interactions. Finally, nutrient competition alone caused non-monotonic, yet predictable dose-dependent responses, consistent with computational predictions. In summary, our study provides a quantitative framework to predict drug responses of microbial communities, emphasizing the pivotal role of nutrient competition.

RESULTS

Chemical screening of a stool-derived in vitro community

We screened a stool-derived in vitro community (SIC) against the National Center for Advancing Translational Sciences (NCATS) clinical collection of 707 small-molecule drugs targeting diverse organisms (Figure 1A, Table S1). Of the 479 drugs with annotated administration routes²⁴, 402 (84%) can be orally administered. The screen was performed via anaerobic culturing in BHI medium with 20 μM of each drug, as in a previous screen of isolates⁴ (STAR Methods). This SIC (hereafter, SIC-0) was derived from a germ-free mouse colonized with a human fecal sample¹⁴, and contains phylogenetically diverse amplicon sequencing variants (ASVs, a proxy for species) representing 14 bacterial families (Figure 1B,C). We acquired community growth, composition, and extracellular metabolite profiles (STAR Methods).

Figure 1: Drug treatment induces large-scale compositional changes in a stool-derived in vitro community.

A) The NCATS library consists of small-molecule chemicals targeting various organisms. Numbers in parentheses denote the total count of drugs in each category. B) Schematic of the chemical screen. C) Relative abundance of ASVs in SIC-0, colored by family. Dashed line: limit of detection at a typical sequencing depth. Some ASVs were identified despite having an average relative abundance below the limit of detection based on replicated sequencing of SIC-0. D) While antibiotics generally had large inhibitory effects, many other drugs also inhibited community growth. Dashed line: threshold for growth inhibition (STAR Methods). E) ASV-level absolute abundance fold change in drugs with growth inhibition, sorted by the extent of growth inhibition. Only fold changes >10-fold or <0.1-fold are shown with colors. Top: drug targets; bottom: relative growth AUC. Asterisks: selected drug–ASV combinations as shown in Figure S1D. See also Figure S1 and Table S1.

Many drugs inhibited SIC growth, measured by reductions in maximum OD₆₀₀ or the area under the growth curve (AUC), which were highly correlated (Figure S1A). Most antibiotics (51/69) inhibited growth (Figure 1D). Of the remainder, five target the obligate aerobe Mycobacterium tuberculosis, and six are beta-lactams, likely degraded by secreted beta-lactamases²⁵. Of the non-antibiotics, a smaller fraction (14%) inhibited growth, and to a lesser average extent (Figure 1D), consistent with previous findings in monocultures⁴. Antineoplastic agents were enriched for growth inhibition in human-target drugs (10/32, p=0.03, one-tailed Fisher’s exact test).

To connect growth inhibition to changes in species abundance, we estimated the colony forming units (CFU) of each ASV by scaling their relative abundance (from 16S rRNA sequencing) to the final OD₆₀₀ of the community and multiplying by 10⁹ CFU/mL. The 10⁹ CFUs/mL factor was validated across 34 communities (Figure S1B). Stronger growth inhibition (relative growth AUC <0.8) associated with larger inhibition of high-abundance ASVs like Escherichia fergusonii, Bacteroides thetaiotaomicron, and Enterococcus hirae (p<0.001, one-tailed Mann–Whitney U test). Drugs with stronger growth inhibition suppressed more ASVs (Figure S1C, left), and promoted the expansion of low-abundance ASVs, sometimes by >100-fold (Figure 1E, S1D). Many low-abundance ASVs grow to high yield as monocultures (Figure S1E), indicating they were inhibited competitors in SIC-0. Drugs causing intermediate inhibition allowed the most ASV expansions, presumably by selectively inhibiting high-abundance species while allowing others to grow (Figure S1C, right). Regardless of initial abundance or taxonomy, all ASVs displayed large abundance ranges under drug treatments (Figure S1F,G). Thus, drug perturbations generated a diverse range of communities.

Growth inhibition correlated with compositional and metabolomic changes

We next explored whether growth inhibition correlated with compositional changes. Alpha diversity (STAR Methods) showed weak or no correlations with growth (Figure S2A). However, the compositional differences to the vehicle (DMSO)-treated control correlated strongly with growth across all beta diversity metrics (STAR Methods, Figure 2A, S2B). No drugs altered community composition without impairing growth. 5-nonyloxytryptamine and minocycline inhibited growth but not composition, presumably because the inhibition was so strong that the communities did not grow and remained their initial composition. Such correlation is consistent with competition for nutrients playing a major role in community assembly⁹.

Figure 2: Growth inhibition in communities correlates with compositional and metabolomic changes.

A) Top: the relative growth AUC of drug-treated communities negatively correlated with the weighted UniFrac distance to the vehicle-treated control. Shaded area: mean±1 S.D. for vehicle-treated controls. Bottom: growth (left) and composition (right) of selected communities. B) Communities with large compositional changes also exhibited large metabolomic distances to the control. Colors represent the degree of growth inhibition. C) The number of consumed metabolite features positively correlated with growth yield. D,E) The distribution (D) and cumulative distribution function (CDF) (E) of Spearman’s ρ between metabolic features and E. fergusonii (Ef) absolute abundance. Compared to all features (gray), those depleted in Ef monoculture (purple) tended to have more negative ρ values. F) For other ASVs, the CDF of ρ for the monoculture-depleted subset (colored by family) was shifted to the left compared to all features (gray), except for B. thetaiotaomicron (Bt). Inset: in Ef spent media, the CDF for the Bt-depleted subset was also shifted to the left. G) (i) for a given species (blue), its total nutrient niche comprises nutrients exclusively available to itself and those shared with other species. (ii,iii) The relative ratios of exclusive and shared nutrients of a species affect the fraction of consumed metabolic features in communities with or without the species. When the species is absent, its exclusive nutrients are left unconsumed. H) In all five species, communities without the corresponding species had a small yet significant reduction in the fraction of consumed features. See also Figure S2, Table S2, S3.

Untargeted metabolomics revealed that SIC-0 depleted at least 137 annotated metabolites from fresh BHI, including sugars, amino acids, peptides, and nucleotides (Table S2). We further selected 90 drugs, assayed their concentrations before and after community growth, and found that most drugs remained stable (STAR Methods, Table S3). We assumed that each depleted metabolite feature represents a nutrient consumed by the community⁹, and calculated the metabolic distance between the vehicle control and each drug-treated community (STAR Methods). Metabolomic distance correlated with compositional change (Figure 2B). 5-nonyloxytryptamine and minocycline had metabolomes similar to fresh medium, again consistent with minimal growth. The total number of consumed features correlated with growth yield (Figure 2C), indicating that growth inhibition frees up nutrients.

We next assessed nutrient usage by individual species. Operationally, we refer to metabolites consumed by a focal species as its nutrient niche²⁶. We first focused on the most abundant ASV, E. fergusonii (Ef). For all metabolites, the Spearman’s ρ between metabolite level and Ef absolute abundance exhibited a large spread (Figure 2D, left). Metabolites with strongly negative ρ’s are likely uniquely consumed by Ef. Compared with a monoculture metabolomics dataset⁹ from the same human donor as SIC-0¹⁴, the metabolites depleted >100-fold by Ef monoculture (382/3599, ~11%) tended to exhibit negative ρ in drug-treated communities (Figure 2D, right), shifting the cumulative distribution function (CDF; Figure 2E). Thus, Ef largely consumes the same nutrients in isolation and community.

We extended this analysis to all other high-abundance ASVs (relative abundance >0.01) with monoculture metabolomics data⁹. E. hirae, Clostridium symbiosum, and Flavonifractor plautii exhibited similar shifts in the CDFs but not B. thetaiotaomicron (Bt; Figure 2F). Bt isolate grows poorly in BHI but is boosted by the supernatant of a saturated Ef culture (hereafter, spent medium), likely due to siderophore production⁹. Indeed, when cultured in Ef spent medium, Bt monoculture-consumed metabolites shifted toward negative ρ (Figure 2F, inset). This finding highlights that non-competitive interactions and/or environmental variables can impact species metabolism, and the monoculture–community comparisons are most relevant when the growth environments are closely matched. Taken together, high-abundance species maintain distinct and conserved nutrient niches regardless of community context or drug perturbations.

We further estimated the extent of niche overlap by comparing communities where the focal species is present or absent (Figure 2G). When the species is absent, its unique nutrient niche remains unused. For five species exhibiting variable presence/absence across drug-treated communities, communities lacking the focal species exhibited significant lower fractions of consumed features (Figure 2H), suggesting partial niche exclusivity. However, the exclusive niche fraction only accounted for 5–15% of metabolites consumed in monoculture, suggesting that competition for shared nutrients is strong for these species.

Drug susceptibility and nutrient availability collectively alter ASV abundance

To determine how drug susceptibility in community relates to susceptibility in monoculture, we assayed 14 isolates from SIC-0 against 43 drugs (Figure 3A). These isolates account for ~92% of abundance in SIC-0 and represent six families. Since some isolates grew poorly in BHI, we supplemented with heme for Bacteroidota members and mucin for Akkermansia muciniphila to accurately quantify growth inhibition (STAR Methods). The 43 drugs encompassed a range of growth inhibitions and contained drugs targeting different organisms (Table S4). Similar to a previous study²⁷, there were no strong trends between phylogenetic relatedness and drug responses (Figure S2C).

Figure 3: Growth inhibition in communities is collectively explained by isolate susceptibility and nutrient competition.

A) Schematic of drug susceptibility quantification for strain isolates. B) Across species, the fold change in monoculture after drug treatment was largely correlated with the fold change in the community. In some cases, the linear fit is close to y=x (left), while in others it deviates substantially (right). Solid lines and shaded areas: best linear fits and 95% confidence intervals. C) Top: Isolate susceptibility weakly correlated with absolute fold change in the community. Bottom: the rescaled growth of isolates correlated with fold change in the community. Blue: binned data. Bt: B. thetaiotaomicron; Bf: B. fragilis; bif: bifonazole. D) Schematic of nutrient competition resulting in apparent protection for the blue species. The blue and gray cells each exclusively consume some nutrients (blue and gray squares, respectively) and share some nutrients (squares with blue/gray gradients). Without drug treatment, the gray species outcompetes the blue by consuming all shared nutrients. For drug treatments that selectively inhibit the gray species, the blue species has access to all shared nutrients and increases in abundance. E) Schematic of growth rescaling strategy with four isolates. F) B. thetaiotaomicron was sensitized to bifonazole with low levels of external heme or in E. fergusonii spent media. G) Heme supplementation rescued Bacteroides species in a community during bifonazole treatment (n=4). H) Schematic of a 3-member community with unique and shared nutrient niches. Inhibiting the blue species provides a larger growth opportunity to the purple species since purple shares more nutrients with blue compared to red. I) In defined communities, the growth rescaling model (left) qualitatively predicted changes in abundances but tended to overestimate abundances. Incorporating community-level growth data (middle) or isolate metabolomics data (right) improved the accuracy of predictions. Blue: the most sensitized species in each community–drug pair; gray: all other species. Error bars in the right panel denote the abundance range due to potential nutrient competition among species. n=165. RMSE: root mean square error. See also Figure S2, Table S4, Table S5.

Our baseline hypothesis was that a drug induces the same fold change of absolute abundance in monoculture and community. This hypothesis held for drugs like enrofloxacin but not others, such as gatifloxacin (Figure 3B). Gatifloxacin strongly inhibited most isolates while A. muciniphila expanded by >1000-fold in community despite ~40% inhibition in its monoculture. We observed similar instances with other species and drug conditions (Figure 1E, S1D). Overall, monoculture and community fold changes were only weakly correlated (Figure 3C, top).

Since many species in SIC-0 have overlapping nutrient niches (Figure 2I), we postulated that relative, rather than absolute, susceptibility determines community-level outcome. A species less inhibited can have a growth advantage by accessing nutrients freed up from its competitors (Figure 3D). To approximate nutrient competition, we rescaled the growth of each isolate by the average growth of all isolates in that drug (Figure 3E). Across all species and drugs tested, rescaled growth correlated with the abundance fold change in communities (Figure 3C, bottom). Although the rescaling omits the details of niche overlaps or non-nutrient interactions, it suggests that nutrient competition, in addition to drug susceptibilities, shapes the community response.

We next focused on instances where the fold change deviated from the expected nutrient competition (STAR Methods). Of all 602 strain–drug pairs, 20 (3.3%) were classified as protected and 32 (5.3%) as sensitized. Protected cases all involved low-abundance species (<1% in SIC-0). For 8 of the sensitized conditions, drug treatment enabled the expansion of another ASV for which we lacked monoculture data; hence, the corresponding rescaled growth was likely an overestimate. The remaining sensitized conditions primarily involved Bacteroides species and nitroimidazole chemicals. For example, in the presence of bifonazole, B. thetaiotaomicron (Bt) and Bacteroides fragilis (Bf) both grew normally in monoculture, while their abundances dropped by >100-fold during community treatment (Figure 3C, bottom).

Across bifonazole concentrations, Bt was consistently sensitized in community with ≥20 μM bifonazole (Figure S2D,E). Bt monocultures were supplemented with 5 μM heme but not the community, as the production of siderophores by species like E. fergusonii (Ef) supports Bt growth^9,28. Since heme supplementation decreases metronidazole susceptibility in other bacteria^29,30, we assayed Bt monoculture susceptibility to bifonazole at varying heme concentrations. Bt growth was not affected by 20 μM bifonazole with 5 μM heme supplementation but was inhibited by ~60% with 0.16 μM heme (Figure 3F), although 0.16 μM heme still supported some Bt growth (Figure S2F). Ef spent medium (without external heme) also sensitized Bt to bifonazole (Figure 3F). Supplementation with 5 μM heme in SIC-0 cultures with bifonazole rescued both Bt and Bf (Figure 3G). Thus, heme scarcity in the community sensitizes Bt to bifonazole. Compounds like heme that affect both growth and drug sensitivity, although rare in our screen, highlight the interplay between media composition, cross-feeding, and drug susceptibility.

Metabolomics quantitatively predicts ASV abundance in defined communities

The growth rescaling model tended to overestimate fold changes for high abundance species because these species already access many available nutrients and gain little from additional released nutrients. Thus, we designed 9 defined communities containing 3–7 members for which we had monoculture metabolomics data to test whether nutrient niche information could improve predictions. If a drug eliminates one species, we infer that its nutrient niche becomes partially accessible to others in proportion to the pairwise shared niches (Figure 3H). Since higher-order niche sharing across three or more species is relatively small⁹, we focused on pairwise interactions.

We treated the defined communities with drugs inhibiting one members more strongly than the others. The rescaling model qualitatively predicted but tended to overestimate abundance changes, so that drug-treated communities had similar yields as controls (Figure 3I, left). We then multiplied these predictions by the relative reduction of OD₆₀₀ in community, which substantially improved the model’s predictions (Figure 3I, middle). Thus, adding limitation on community yield improves model accuracy for defined, low-diversity communities, although it had little benefit for the more complex SIC-0 (Figure S2G). However, this approach requires additional growth measurements in each community–drug condition.

We then tested an orthogonal approach, using monoculture metabolomics and isolate susceptibility data to directly infer nutrient competition (STAR Methods). We estimated how inhibition of one species redistributed its nutrient niche to others, based on pairwise overlaps in monoculture metabolomics. The resulting model achieved comparable or better performance than using community OD₆₀₀ data (Figure 3I, right), with lower prediction errors (p<10⁻⁵, two-sample F-test). Thus, incorporating niche overlap information can improve predictions of community responses without community-level growth assays.

Extinction restructures post-recovery community composition

To assess community robustness after drug perturbations, we passaged SIC-0 in drug-free media after treatment (Figure S3A). Most communities partially recovered in growth and composition, although their metabolomics did not exhibit trends of recovery (Figure S3B). We next tested 23 drugs with six recovery passages. In one set of recovery passages, the untreated SIC-0 was seeded at 1% relative abundance (v/v) after drug treatment (Figure 4A), mimicking recolonization from environmental reservoirs³¹. We assayed each condition in duplicate, and observed high reproducibility regardless of seeding (Figure S3C–E).

Figure 4: Recovery from drug treatment is hampered by extinction.

A) Schematic of multiple recovery passages with or without seeding. B) Both growth (top) and composition (bottom) required 2–3 recovery passages to equilibrate. Communities exhibited greater recovery with seeding. n=2 replicates. Gray: DMSO control. C) Left: definition of extinct (E), partially recovered (PR), or fully recovered (FR) ASVs. Middle: without seeding, most ASVs that were not detected during treatment remained extinct. Right: with seeding, many more ASVs fully recovered. D) Even with seeding, E. fergusonii (Ef) did not recover after fluoroquinolone treatment (left), and was replaced by K. pneumoniae (Kp; right). E) Before and after the swap from Ef to Kp, the abundances of Ef and Kp at equilibrium (R3–6) were inversely correlated. Dashed line denotes a linear inverse correlation. F) Strain swap did not affect the final OD₆₀₀ of the resulting communities. n=2 replicates in D-F. G-I) In an Ef-Kp pairwise co-culture with seeding of the pairwise mixture, fluoroquinolone treatment temporarily promoted Kp growth (G), and we observed similar inverse correlations as in (E) in abundance (H) as well as constant final OD₆₀₀ (I), except during treatment. n=4 replicates. J) Spent media from various cultures inhibited Ef growth. One-tailed Mann–Whitney U tests. K) Although spent media from low-Kp communities, Kp monoculture, and low-Kp communities + Kp dually spent media allowed for some Ef growth, the time required for Ef to double twice in the spent media was significantly longer than in fresh BHI medium. *: p<0.05, **: p<0.01, one-tailed Mann–Whitney U tests. L) In BHI mixed with the spent media from high-Kp communities, Ef growth was inhibited compared to BHI mixed with PBS. n=6, one-tailed Mann–Whitney U test. M) For the communities with low Ef, spiking in SIC-0 enabled Ef to quickly take over. Expansion occurred more rapidly without Kp. n=4. See also Figure S3, S4.

Communities required 2–3 recovery passages to equilibrate. Recovery without seeding was partial, such that many drugs resulted in long-lasting growth inhibition and compositional changes, while seeding promoted complete recovery (Figure 4B). To quantify the promotion of recovery, we identified 523 ASV–drug combinations in which an ASV present in SIC-0 was not detected after treatment. We classified these apparent extinct ASVs based on their equilibrated abundance after recovery (average of R3–R6): extinct (not detected), partially recovered (≤10% of pre-treatment), and fully recovered (>10% of pre-treatment) (Figure 4C, left). Without seeding, 86% (448/523) of the instances remained extinct (Figure 4C, middle). With seeding, 77% (401/523) fully recovered (Figure 4C, right). The no-seeding conditions were also accompanied by more blooming of initially undetected ASV during recovery. Some seeding communities also retained blooming ASVs, suggesting that drug treatment can lead to lasting community reconfiguration (Figure S3F). For the ASVs that persisted through treatment, their equilibrated abundance was similar to pre-treatment (slope=0.93±0.03 without seeding; 1.03±0.03 with seeding), even though in many instances (92/302) they decreased by >10-fold during treatment. Taken together, long-lasting compositional changes caused by drug treatment are mostly due to extinction, which are alleviated by resupplying the extinct ASVs via seeding.

Interactions beyond nutrient competition drive long-term strain swapping after drug perturbations

We next focused on three fluoroquinolones (levofloxacin, gatifloxacin, and moxifloxacin) that caused large compositional shifts. Without seeding, the Enterobacteriaceae members E. fergusonii (Ef) and Klebsiella pneumoniae (Kp) remained at low abundance, and community yield decreased (Figure S4A,B). With seeding, Ef was almost entirely replaced by Kp (Figure 4D), and their abundances were inversely correlated (Figure 4E). The resulting communities remained the same yield as pre-treatment (Figure 4F). The equilibrated seeding communities had similar compositions as the control after excluding Ef and Kp (Bray–Curtis distance 0.27±0.05, similar to the variation in the control passages, 0.32±0.09). We term this replacement between two species without substantively affecting other community members as a “strain swap”. A similar strain swap occurred in rifampicin treatment, in which Bt was replaced by Bacteroides intestinalis (Figure S4C–E). In Ef and Kp pairwise co-culture, Kp transiently dominated during fluoroquinolone treatment but was outcompeted by Ef after 3–4 recoveries (Figure 4G–I, S4F,G). In contrast, Ef remained low in four of the six seeded fluoroquinolone-treated communities (high-Kp communities) and four of the six unseeded communities (low-Kp communities), even after 13 recovery passages (Figure S4H,I). We did not identify any single ASV explaining such suppression (Figure S4J,K).

To test whether community members collectively outcompete Ef, we grew Ef in spent media from saturated low-Kp and high-Kp communities. Ef failed to grow in high-Kp spent media, and had low initial growth rate and yield in low-Kp spent media (Figure 4J,K). Ef growth was further suppressed in dually spent media from Kp grown in the low-Kp spent media (low-Kp + Kp spent). All spent media had neutral pH (~6.8–7.0), and we could not discern any metabolite features unique to the low-Ef communities that could explain their suppression of Ef. Mixing high-Kp spent media with fresh BHI reduced Ef growth compared to PBS (Figure 4L), indicating its inhibitory ability beyond nutrient depletion. Thus, Ef abundance is kept low both by nutrient competition, particularly its preferred nutrients, and by inhibitory compounds from other members.

During the passage of low-Ef communities, Ef expansion was dependent on a threshold: when Ef reached ~5×10⁶ cells/mL, it rapidly expanded. This threshold was not observed in monoculture (Figure S4L). We then focused on the low-Ef communities without Ef recovery in R8–R13 (Figure S4H,I) and spiked 1% of SIC-0 (v/v; introducing ~5×10⁶ Ef cells/mL) into the R8 communities. SIC-0 spike-in led to immediate Ef recovery, faster in low-Kp than high-Kp communities (Figure 4M). In contrast, when we mixed SIC-0 with Kp monoculture, SIC-0 resisted invasion even with a large spike-in (Figure S4M). Thus, Ef recovery depends on reaching a critical density, and Kp slows down but does not prohibit its expansion. This critical density likely allows Ef to outcompete other species for its preferred nutrients before they exert inhibitory effects. Such multibody dynamics represent a complex interplay between nutrient competition, inter-species chemical warfare, and priority effects, which can lead to long-lasting compositional changes over >100 generations of growth.

Resistance selection is rare in SICs

Antibacterial agents naturally drive resistance selection³², but the rate and mechanism of resistance are context-dependent³³. In communities, nutrient competition constrains population size of many species, which may limit the rate of evolution³⁴ and reduce the chance of fixation if resistance comes with a fitness tradeoff³⁵. After the drug-treated communities reached equilibrium, we re-treated with the same drugs (Figure 5A) to identify ASVs with potential resistance selection.

Figure 5: Selection for resistance is rare in communities.

A) Schematic of the second round of drug treatment (T2). B) Consumer-resource simulations of communities without (left) or with (right) resistant mutants during T2. Inset: receiver operating characteristic curve for identifying resistance. C) Experimentally, regardless of seeding condition, most ASVs exhibited similar fold changes across the two treatments, and potential resistance selection was rare. Shaded areas: 60% confidence interval of linear fit. D) Community growth AUC (left) and final OD₆₀₀ (right) were largely the same between the two treatments. E) All potential cases of resistance selection without seeding (C, left) involved F. plautii (Fp). Tet: tetracycline; cef: cefoxitin; rol: rolitetracycline. F) Tetracycline-resistant Fp mutants isolated from the community exhibited higher resistance to tetracycline than those isolated from Fp monocultures. Darker circles represent isolates whose genomes were sequenced to identify resistance mutations (Figure S4P). See also Figure S4.

We simulated our experimental procedure using a consumer-resource (CR) model³⁶ (STAR Methods). Without resistant mutations, fold changes after the two treatments were strongly correlated and similar in magnitude. Introducing a resistant species in treatment passage 2 (T2) led to a >2-fold change in abundance for that species, while others largely remained unchanged (Figure 5B). Intuitively, for resistance selection, the fold change during T1 should be smaller than that of T2. Therefore, we set the cutoff for fold change during T1 (FC1) also to be 2, corresponding to a false-discovery rate of 0.004 on a receiver operating characteristic (ROC) curve, with the area under ROC curve(AUROC) 0.99 (Figure 5B, inset).

Experimentally, ASV fold changes across the two treatments were largely the same (Figure 5C). Regardless of the seeding condition, growth AUC and final OD₆₀₀ across the two treatments were similar (Figure 5D), further confirming the lack of observable resistance effects at the population level. We then applied the model-inspired thresholds, and excluded cases located within the 60% confidence interval from the fit line to account for experimental noise. We identified 6 of 376 (1.6%) no-seeding instances and 14 of 698 (2.0%) seeding instances of likely resistance selection (Figure 5C). We further tested nine drugs with only one recovery passage between drug treatments to allow potential resistance mutants with growth disadvantages, and observed similar rates of resistance selection (Figure S4N, 6/123 (4.9%) instances, p=0.08, Fisher’s exact test). In monocultures, resistance selection occurred in ~11% of conditions (STAR Methods). Thus, resistance selection is comparatively rare in community.

All six resistance instances without seeding occurred in F. plautii (Fp), under tetracycline, cefoxitin, or rolitetracycline (Figure 5D), and a similar pattern occurred with seeding (Figure S4O). We isolated Fp mutants from tetracycline-treated monocultures or communities (49 and 11 isolates, respectively; STAR Methods). As expected, most isolates were more resistant to tetracycline than the ancestor (Figure 5E). Mutants isolated from the community grew significantly better than those from monocultures (p=0.02, one-tailed Mann–Whitney U test). Whole genome sequencing of select isolates (Figure 5E, circles) revealed distinct resistance mechanisms, with monoculture mutants carrying SNPs in efflux genes while community mutants tended to have gene duplications increasing the copy numbers of resistance-related genes (Figure S4P). Taken together, the different resistance phenotypes (Figure 5E) suggest that Fp may develop resistance through different mechanisms in monoculture versus community.

Drug treatment effects are quantitatively adjusted by nutrient competition

We extended our screen to two other SICs using a random subset (480 of 707) of the drugs. These two SICs were derived from mice colonized with the same human fecal sample as SIC-0. SIC-MD from a mouse with a diet deficient in microbiota-accessible carbohydrates, and SIC-cip from a mouse after five days of ciprofloxacin treatment^14,31. SIC-0, SIC-MD, and SIC-cip consisted of 37, 34, and 30 ASVs present at >0.01%. Of these ASVs, SIC-MD shared 23 with SIC-0, accounting for 95% of the total abundance in SIC-0 and 98% in SIC-MD. SIC-cip shared 14 ASVs with SIC-0, accounting for 76% (SIC-cip) and 72% (SIC-0) of total abundance (Figure 6A, S5A). These SICs represent a wide range of initial beta diversities.

Figure 6: SIC composition quantitatively tunes the effects of drug treatment.

A) UpSet plot of shared and unique ASVs in three SICs, using a cutoff of 10⁻⁴ relative abundance. B) Relative growth AUCs after drug treatment were largely correlated across the three SICs. C) Growth curves (left) and composition (right) of selected drug conditions. D) For ASVs shared between SICs, their abundances after drug treatment were correlated. Dashed line: mean Spearman’s ρ for random ASV pairs across SICs. E) For all ASV pairs in (D), their fold-changes in different SICs generally did not fall along the line y=x. Slopes of the linear fits deviated from 1 (left) and y-intercepts deviated from zero (right). F) In both experimental data and CR simulations, the majority of ASV pairs exhibited fit slopes deviating from 1 and non-zero y-intercepts, defined as the 95% confidence intervals of slopes and y-intercepts of the linear fits not containing 1 (for slopes) or 0 (for y-intercepts). Exp: experiment; Sim: CR simulations. G) Dose responses of high-abundance ASVs to metronidazole across SICs. In SIC-0 and SIC-MD, E. hirae (Eh) increased in abundance at higher metronidazole concentrations. Gray dashed lines: ASVs without monoculture susceptibility data. H) In the 3-member community, Eh increased its abundance at higher metronidazole concentrations. I) K. pneumoniae exhibited non-monotonic dynamics during treatment with rifampicin in SIC-0. J) Across the SICs and drugs tested, increasing/non-monotonic dose responses occurred, and were dependent on community context and drug conditions. Numbers denote ASV counts. N.D.: no data (no growth was observed in SIC-cip with 10 μM or higher rifampicin). K) Examples of predicted ASV dose responses in SIC-0. Right: SIC-0 with rifampicin, after introducing a fully resistant strain that mimics the effect of T. ramosa. L) Receiver operating characteristic (ROC) curve of the rescaled growth model (without accounting for the effect of T. ramosa in rifampicin). M) Model prediction and experimental measurements of high-abundance ASVs in SIC-0 with moxifloxacin, ampicillin, and triclosan titrations. Experimental data from 40 μM and 80 μM moxifloxacin exhibited clear signs of resistance emergence, and 80 μM triclosan strongly inhibited community growth. These conditions were excluded from analyses. N) ROC curve of the rescaled growth model for the drug conditions in (M). See also Figure S5, S6.

Upon drug treatment, relative growth AUC was correlated across all three SICs (Figure 6B). The outliers were largely due to compositional differences. For instance, SIC-cip was less inhibited by moxifloxacin, in which A. muciniphila and Phocaeicola vulgatus (~3% and ~13% in SIC-cip pre-treatment) collectively comprised ~82% of the moxifloxacin-treated community (Figure 6C). Both ASVs were at <0.01% in the other two SICs pre-treatment. We did not obtain an isolate of P. vulgatus, but the A. muciniphila isolate was largely resistant to moxifloxacin, while major ASVs in SIC-0 and SIC-MD were inhibited by >90% (Figure S5B, left). Oxiconazole inhibited all ASVs in SIC-0 and SIC-MD except for Ef. SIC-cip contained very low levels of Ef (~0.07%) and became dominated by Enterococcus after oxiconazole treatment (Figure 6B,C). Although the representative Enterococcus isolate from SIC-0 (E. hirae) was resistant to oxiconazole, we isolated two other Enterococcus species (E. faecium and E. faecalis) from SIC-0 with the same V4 16S sequence as E. hirae, and these species exhibited variable oxiconazole susceptibility (Figure S5B, right). Thus, the larger growth reduction of SIC-cip in oxiconazole may be due to the presence of a sensitive Enterococcus species. Together, the quantitative growth differences between SICs can reflect strain-level composition and sensitivity.

At the ASV level, drug-induced abundance changes were correlated for all overlapping ASVs (Figure 6D, FDR-adjusted p≤10^–6; Figure S5C), suggesting qualitatively conserved drug responses across SICs. Principal Coordinate Analysis also showed that drug treatment exerted some conserved effects across the SICs (Figure S5D). We extended our drug screen to eight other SICs derived from different human hosts⁸, and again found strong correlations in growth and compositional changes, with exceptions that were explained by strain-specific resistance in particular communities (Figure S5E–G).

Despite these qualitative correlations, the magnitude of responses differed substantially. For example, linear fit of Ef fold change in SIC-0 versus SIC-MD had a slope of 0.71±0.10. Eh further exhibited a non-zero y-intercept (Figure S5C). Such deviations were observed across almost all shared ASVs (Figure 6E). We queried whether the quantitative differences could emerge simply from nutrient competition using CR models (STAR Methods). Even in this simplified scenario where nutrient competition is the only interspecies interaction, many linear fits had slopes deviating from one and/or non-zero y-intercepts (Figure 6F). Thus, while drug responses are qualitatively predictable across communities, resource competition quantitatively adjusts each species’s response to drug perturbations.

Epistatic, context-dependent dose responses in communities arise from nutrient competition

Given the large range of isolate susceptibilities (Figure 3C), we further tested the dose-dependent responses of SIC-0, SIC-MD, and SIC-cip to four drugs (bifonazole, metronidazole, tetracycline, and rifampicin). In monocultures, bacterial growth typically decreases with increasing drug concentration or stays constant if the strain is resistant. In the SICs, during metronidazole treatment, most high-abundance ASVs (median abundance across drug concentrations >2×10⁶ cells/mL) also remained largely unaffected or monotonically decreased with increasing drug concentration (Figure 6G). The dose dependence of E. hirae (Eh) was more complex. While Eh abundance decreased monotonically in SIC-cip, in SIC-0 and SIC-MD its abundance increased 2- to 10-fold. This increase in Eh abundance coincided with a decrease in other high-abundance species such as B. thetaiotaomicron (Bt). We then performed metronidazole titration experiments in a defined community of E. fergusonii (Ef), Eh, and Bt, and found that Eh also increased in abundance at higher metronidazole concentrations (Figure 6H), suggesting that the increase in Eh abundance was community dependent.

We expanded our analyses to the three other drugs. In rifampicin, several ASVs, including K. pneumoniae, exhibited non-monotonicity, reaching their highest abundance at intermediate concentrations (Figure 6I). Similar behaviors were observed in tetracycline (Figure S6A). We then systematically classified the ASV-level dose responses into three categories (STAR Methods): monotonically decreasing (46/85 instances), no change (9/85), or increasing/non-monotonic (30/85). Increasing/non-monotonic ASV–drug combinations were more prominent in SIC-0 and SIC-MD, especially for rifampicin and tetracycline treatment (Figure 6J). These dose responses were reproducible in two more replicate experiments with SIC-0 (41/53 ASV–drug combinations exhibited the same dose response dynamics; Figure S6A,B). Thus, drug inhibition and species–species interactions potentiate ecologically epistatic, context-dependent dose responses. Here, we use the term ecological epistasis to refer to cases in which drug–species interactions depend on the presence or absence of other community members, analogous to gene-gene interactions in classical genetics.

We next evaluated the role of nutrient competition in epistatic dose responses. We applied the rescaled growth model using the monoculture dose-response data each drug concentrations. After accounting for a small, baseline level of growth in the monoculture data (STAR Methods), this model qualitatively captured the responses to metronidazole in SIC-0 for all ASVs that we had monoculture susceptibility data (Figure 6K, left). In rifampicin, this model reproduced the non-monotonic behavior of Kp, although it overestimated the abundances of Eh and Bt at high concentrations (Figure 6K, middle). Such overestimation could be due partly to the limited strain isolates, as a Thomasclavelia ramosa ASV dominated the community at high rifampicin concentrations (Figure 6I) but was absent in the monoculture data. After introducing the T. ramosa strain into our model with the assumption that it is completely resistant to rifampicin, our modified in silico model better recapitulated the experimental results (Figure 6K, right). Across all drug–community combinations, the AUROC of the model was 0.72 (Figure 6L), demonstrating successful qualitative predictions of community-dependent drug dose responses.

We extended our growth rescaling model to other drug conditions with monoculture susceptibility data (STAR Methods), and predicted that three drugs, moxifloxacin, ampicillin, and triclosan, to contain many epistatic dose responses. Follow-up experiments demonstrated qualitative agreements with model predictions (Figure 6M). Since we selected drug conditions that favored the growth of species with known monoculture susceptibility data, our model exhibited higher predictive power, yielding an AUROC of 0.89 (Figure 6N). Taken together, although we could not rule out the possibility of other, non-competitive interactions, such as certain drug concentrations promoting resistance selection and thereby resulting in the observed increasing/non-monotonic dose responses, epistatic dose responses can emerge from straightforward interplay of drug inhibition and nutrient competition.

DISCUSSION

This study assessed the effects of therapeutic drugs on the growth, composition, and metabolome of diverse communities of gut commensals. Community growth correlated with compositional and metabolomics profiles (Figure 2A–C), supporting its use as a rapid, high-throughput assay to identify impactful drug perturbations. ASVs behaved qualitatively similarly across community contexts (Figure 6C,D, S5G), suggesting broad applicability to other synthetic or stool-derived communities of gut microbiota, particularly if nutrient competition is conserved. In vivo responses will also depend on factors like host diet, which can directly modulate nutrient competition³¹ or alter drug pharmacokinetics³⁷. Nonetheless, our in vitro data provide a controlled baseline for comparison. Drug treatment outcomes were highly reproducible (Figure S3C,D). Collectively, our findings regarding growth-based filtering, generality across communities, and reproducibility provide a roadmap for streamlining and increasing the scale of future screens.

Although antibiotics generally caused more inhibition, many non-antibiotics also disrupted community structure (Figure 1D,E). Since many drugs from our screen can reach ~20 μM in the colon⁴, the effects we observed may be clinically relevant. Nutrient competition played a pivotal role in shaping community responses. Similar to a previous study with defined bacterial communities³⁸, growth inhibition correlated with a shifted metabolome and reduced nutrient consumption (Figure 2B,C), consistent with overlapping nutrient preferences across species (Figure 2H). We also found that bacterial species generally metabolize a similar set of nutrients in monoculture and community (Figure 2E,F), validating a common assumption made in many studies^9,39. The extinction of dominant species often involved competitive pressure (Figure 4D–F, S4A–E)⁹, as evidenced by the expansion of low-abundance species (“competitive release”, Figure 1E, 3B). In general, the potential for epistatic behaviors that emerge from nutrient competition highlights the utility of CR simulations for mechanistic interrogation of microbial ecology (Figure 5B, 6F). Rescaling isolate susceptibility by relative growth improved predictions of abundance changes in communities (Figure 3C), and integrating metabolomics data further enhanced model accuracy in defined communities (Figure 3I). While non-nutrient interactions may also play roles, nutrient competition offers a baseline framework that explains much of the variance in community response to drug treatment and enables predictive modeling (Figure 6K–N). Our results enable future studies to focus on outlier cases in which non-nutrient interactions such as drug sequestration¹³, chemical signaling⁴⁰, or environmental modification¹⁶ strongly shape the drug response landscape.

Post-drug recovery dynamics are pivotal microbiota stability and robustness. In mice, residual ciprofloxacin remained high two days after treatment³¹. In our in vitro study, the residual drugs during the first recovery passage could still inhibit some species (Figure S4F), emphasizing the benefits of multiple recovery passages (Figure 4B). Seeding with the untreated community enhanced recovery (Figure 4B,C), mirroring fecal microbiota transplantation in vivo²³. Most ASVs exhibited consistent responses upon re-treatment, suggesting limited resistance selection (Figure 5C,D). However, the few resistance mutations observed differed between community and monoculture contexts (Figure 5F, S4P), which aligns with emerging evidence that microbial interactions constrain evolutionary trajectories^41–43. Thus, community context influences ecological and evolutionary dynamics.

Although most extinct strains recovered with seeding, we observed occasional strain swapping (Figure 4D, S4C). Such collateral damage to the microbiota may be difficult to predict, especially if the emergent strain was undetectable prior to treatment. Although we only observed occasional strain swapping events, these cases serve as representative examples underscoring the ecological nuances underlying complex communities. Invasive strains such as K. pneumoniae could persist over long time scales if the invader(s) and existing community members collectively outcompete the original strain by exhausting its preferred nutrients, and persistence can be accentuated by non-competition-based growth inhibition (Figure 4D–L). The propensity for strain swapping is likely higher in undefined microbial communities such as SICs, since they tend to harbor a diverse reservoir of low-abundance species as compared with defined synthetic consortia. Thus, microbial behaviors in a community may be predominantly predictable via nutrient interactions, with strain swapping events as rare exceptions. Identifying and understanding strain swapping should provide clinical strategies for curing infections^44,45 and for anticipating long-term shifts in microbiome composition^20,46.

Epistatic, increasing/non-monotonic dose responses were common in drug–community contexts, and can be largely explained by nutrient competition (Figure 6G–N). A sensitive species may benefit from the inhibition of its competitors due to the additional nutrient availability. While obtaining ASV-level dose responses across drugs and communities is laborious, computational methods combined with isolate susceptibility data (Figure 6K–M) can identify potential emergent responses and interpret in vivo clinical studies in which chemical concentrations are difficult to quantify.

Moving forward, our approaches and findings provide general insight and principles for high-throughput screens, which are now extending beyond drugs to consider the effects of environmental toxins, dietary compounds, and many other molecules to which the gut microbiota may be exposed. Our findings identify nutrient competition as a fundamental driver of drug responses (Figure 3, 6). This framework enables baseline predictions and highlights specific higher-level interactions, such as cross-feeding molecules that modulate drug susceptibility (Figure 3F,G) and chemical inhibition that enhances strain swapping (Figure 4D–L). A general understanding of these effects will require broad profiling of many representative isolates (Figure 3C). Such efforts hold the promise of transforming the complex landscape of chemical–microbe interactions into a means of rationally engineering microbiome composition.

Limitations of the study

Our study highlights nutrient competition as a key mechanism governing gut microbiota responses to drug perturbations, but several limitations remain. First, other potential interspecies interactions beyond nutrient competition, such as cross-feeding, chemical inhibition, or environmental modification could modulate drug responses. Thus, our framework should be viewed as a foundation on which more complex models of community-level behaviors incorporating these other interactions can be built. Second, SICs are tractable and reproducible model systems but lack key in vivo features, such as spatial structure and host factors. For instance, our well-mixed liquid cultures are limited in population size and density and do not capture the spatial heterogeneity, peristaltic flow, and host immune factors that can shape microbial organization in the intestines. Third, the isolates used for metabolomics and monoculture susceptibility remain an incomplete representation of the SICs and of the human gut microbiota more generally. The unprofiled taxa can hamper community-level predictions, especially for conditions in which those taxa have substantial growth advantages. Lastly, while our modeling approach considers resource competition, we assumed for simplicity that biomass yield was uniform across metabolites and equivalent across species. The model also does not account for strain-specific metabolic regulation or resistance evolution, which may become relevant in long-term experiments or fluctuating drug conditions. Future studies incorporating spatial structure, host–microbe interactions, and higher-resolution metabolic models will be critical to extend our framework to the native human gut environment.

RESOURCE AVAILABILITY

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Kerwyn Casey Huang (kchuang@stanford.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• 16S amplicon sequencing reads of SICs and the assembled genomes of F. plautii mutants have been deposited at NCBI SRA as BioProject number PRJNA1297205 and are publicly available as of the date of publication.

• The untargeted metabolomics data have been deposited at Metabolomics Workbench as Study ID ST004087, doi: 10.21228/M8ZK1T and are publicly available as of the date of publication.

• All original code has been deposited at BitBucket at https://bitbucket.org/kchuanglab/chemicalscreen_figurescripts and https://bitbucket.org/kchuanglab/crsimulation, and are publicly available as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

The bacterial strains used in this study are listed in the Key resources table. The identity of each strain was verified using 16S rRNA sequencing.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and virus strains
Escherichia fergusonii	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Klebsiella pneumoniae	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Enterococcus hirae	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Blautia producta	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Blautia pseudococcoides	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Clostridium symbiosum	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Enterocloster clostridioformis	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Hungatella hathewayi	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Hungatella effluvii	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Bacteroides fragilis	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Bacteroides thetaiotaomicron	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Bacteroides uniformis	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Parabacteroides distasonis	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Akkermansia muciniphila	This study	N/A
Flavonifractor plautii	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Lachnoclostridium pacaense	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Bacteroides stercoris	KC Huang Lab, Stanford; Aranda-Diaz et al.14)	N/A
Chemicals, peptides, and recombinant proteins
Brain Heart Infusion	BD	Cat. #2237500
Chemicals for initial screen: see Table S1	NCATS	N/A
Chemicals for additional characterization: see Table S4	This paper	N/A
Critical commercial assays
DNeasy UltraClean 96 Microbial Kit	Qiagen	Cat. #10196-4
AccuStart II PCR SuperMix	Quantabio	Cat. #95137-100
Deposited data
Raw and analyzed data; MATLAB scripts to generate the figures	This paper	https://bitbucket.org/kchuanglab/chemicalscreen_figurescripts
Python scripts for consumer-resource simulations	This paper	https://bitbucket.org/kchuanglab/crsimulation
16S rRNA gene amplicon sequencing of SICs	This paper; NCBI SRA	NCBI BioProject PRJNA1297205
Whole genome sequencing of Flavonifractor plautii isolates	This paper; NCBI GenBank	NCBI BioProject PRJNA1297205
Untargeted metabolomics dataset	This paper; Ho et al.9; metabolomicsworkbench	http://dx.doi.org/10.21228/M8DB1F; http://dx.doi.org/10.21228/M8ZK1T
Oligonucleotides
27F: 5’-AGAGTTTGATCCTGGCTCAG-3’	IDT	N/A
1391R: 5’-GACGGGCGGTGWGTRCA-3’	IDT	N/A
Software and algorithms
MATLAB	The MathWorks, Inc.	R2024b (v24.2.0.2172019); https://matlab.mathworks.com
R	The R Foundation	v4.0.2; https://www.r-project.org
DADA2	Callahan et al.56	v1.16.0; https://benjjneb.github.io/dada2/
Python	Python Software Foundation	v3.8.5; https://www.python.org
Other
Thermo Q Exactive HF Hybrid Quadrupole-Orbitrap	Thermo	N/A

METHOD DETAILS

Chemical screening of SICs

SICs were revived from glycerol stocks by inoculating 5 μL of glycerol stock into 1 mL of fresh BHI medium, and incubated anaerobically for 48 h at 37 °C, which allowed the SICs to reach saturation. BHI was selected based on our previous finding that BHI retained the highest strain richness in SICs across several common rich media, and that the composition of BHI-derived SICs largely mimicked that of the stool from which the SICs were derived (mean weighted UniFrac distance ~0.4 between stool inocula and SIC compositions)^14,47. Drug stocks at concentration 2 mM were prepared in dimethylsulfoxide (DMSO). For the initial screen, the resulting cultures were diluted 1:200 into 96-well polystyrene microplates (Greiner Bio-One) containing 200 μL of fresh medium. Two microliters of drug stocks (final concentration 20 μM) or DMSO (for the vehicle-treated control) were added to the mix. The microplates were covered with optical film (Thermo Scientific) with a small hole poked at the edge of each well to allow for gas exchange. The microplate was incubated anaerobically in an Epoch2 plate reader (BioTek) with continuous shaking and OD₆₀₀ (optical density measured at 600 nm) readings every 7 min. After 48 h, the cultures were harvested for downstream analyses. For recovery passages, the 48-h culture was diluted 1:200 into fresh medium. After a 1:200 dilution into R1, the drug was carried over at up to ~0.1 μM. In a previous study quantifying the MICs of 63 antibiotics against 5 strains, there were only 13 instances (4%) with MIC close to or less than 0.1 μM⁴⁸. Thus, in most cases, drug carryover should not affect community growth even in R1, let alone R2–R6. For the R1 passage with seeding, an aliquot of the resulting 48-h cultures was mixed with saturated SIC-0 at 100:1 (v/v), then diluted 1:200 into fresh medium. For an ASV present at ~10⁻³ relative abundance in SIC-0, this strategy seeds ~10 cells into the next passage.

16S rRNA gene amplicon sequencing

16S sequencing was performed as previously described⁴⁹. Genomic DNA was extracted from 50–100 μL of cultures using a DNeasy UltraClean 96 Microbial Kit (Qiagen), and the 16S rRNA gene was amplified with 515F/806R primer pairs⁵⁰ using Platinum II HotStart PCR Master Mix (ThermoFisher) with the following thermocycler settings: 94 °C for 3 min, 35 cycles of [94 °C for 45 s, 50 °C for 60 s, and 72 °C for 90 s], then 72 °C for 10 min. The resulting PCR products were pooled at equal volume, concentrated, and cleaned up via agarose gel. Pooled libraries were sequenced with 250-bp paired-end reads on a MiSeq (Illumina).

Full 16S rRNA gene amplicon sequencing

For strain isolates, the full 16S rRNA gene was amplified with 8F/1391R primer pairs (8F: 5′-AGA GTT TGA TCC TGG CTC AG-3′, 1391R: 5′-GAC GGG CGG TGW GTR CA-3′) directly from saturated bacterial cultures following the same PCR protocol as the amplification of genomic DNA from communities. The resulting PCR products were sequenced via Sanger sequencing. Taxonomy was obtained from the full 16S sequences using the NCBI BLAST tool⁵¹.

Untargeted metabolomics

Bacterial culture supernatants were collected after centrifuging saturated cultures at 4,000g and 4 °C for 10 min, and immediately frozen at −80 °C. Before LC–MS/MS analysis, samples were mixed with extraction mix (47.5% acetonitrile, 47.5% methanol, 5% H₂O, and stable isotope-labelled internal standards) and passed through a 96-well filter microplate with a 0.2-μm polyvinylidene fluoride membrane (Agilent). Samples were injected onto a Waters Acquity UPLC BEH Amide column with an additional Waters Acquity VanGuard BEH Amide pre-column. Spectra were collected using a Thermo Q Exactive HF Hybrid Quadrupole-Orbitrap mass spectrometer in positive and negative mode ionization. Full MS-ddMS2 data were collected. Data were processed using MS-DIAL v. 4.60^52,53. Alignment retention time and mass tolerance were set to 0.05 min and 0.015 Da, respectively. Protocol details are as previously published⁹.

Isolation of Enterobacteriaceae strains

SICs were grown anaerobically to saturation in BHI. The resulting cultures were diluted, plated onto LB agar (10 g/L tryptone, 5 g/L yeast extract, 5 g/L NaCl, and 15 g/L agar) supplemented with 10 μg/mL vancomycin, and incubated aerobically for 24 h. The resulting colonies were re-streaked twice onto LB agar with vancomycin, and their identity was verified via 16S rRNA sequencing.

Drug sensitivity of strain isolates

Glycerol stocks of strain isolates were streaked onto BHI agar plates with 5% horse blood, and for each strain, two independent colonies were picked for downstream experiments. The colonies were grown to saturation in liquid, diluted 1:200 into fresh liquid medium, and arrayed into 384-well microplates. All drugs were dissolved in DMSO at 64 mM and arrayed in 96-well microplates, then added to the 384-well microplates at final concentrations of 160 μM, 80 μM, 40 μM, 20 μM, 10 μM, 5 μM, and 2.5 μM. The resulting microplates were incubated anaerobically at 37 °C for 24 h, and OD₆₀₀ was measured for the final cultures. Although our community-based screens were performed with 48-h passages (Figure 1B), the isolates all saturated within 24 h. Thus, for monoculture sensitivity experiments, we assayed their growth after 24 h to limit the opportunity for resistant subpopulations to expand.

All strains were grown in BHI-based liquid media. To promote growth and obtain reliable OD₆₀₀ readings, the media for all Bacteroidota members were supplemented with 5 μM heme, 2 mg/mL NaHCO₃, and 1 mg/mL l-cysteine, and the media for A. muciniphila were supplemented with 5 μg/mL of vitamin K3, 1 μM hemin, 0.2 mg/mL l-tryptophan, 1 mg/mL l-arginine, and 0.5 mg/mL l-cysteine.

Isolation of resistance mutants

To isolate tetracycline-resistant F. plautii mutants from monocultures, a previously isolated F. plautii strain from SIC-0¹⁴ was inoculated from a glycerol stock and grown overnight. To isolate from a community, SIC-0 was inoculated from a glycerol stock and grown overnight. The monoculture and community culture were diluted 1:200 into fresh BHI, treated with 20 μM of tetracycline for 48 h, then diluted 1:200 into fresh BHI and grown for another 48 h. The final cultures were diluted and plated onto BHI-blood plates and individual colonies were picked. F. plautii isolates exhibited a flower-like colony morphology, and for the community plates, we biased toward picking colonies with such morphologies and verified their taxonomy via 16S rRNA sequencing.

Whole genome sequencing of strain isolates

F. plautii strains were streaked onto BHI-blood agar plates from glycerol stocks and grown for 48 h. Colonies were scraped, washed in phosphate buffered saline (PBS), resuspended in DNA/RNA shield (Zymo Research), and sent for genomic DNA extraction and sequencing (Plasmidsaurus Inc., CA, USA).

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical tests, p-values, and sample sizes are provided in figures, figure legends, and text. Unless otherwise specified, r denotes Pearson’s correlation, ρ denotes Spearman’s correlation, and p values are from two-tailed Student’s t-tests. Unless otherwise specified, data points are mean±1 S.D. In violin plots, the center points represent the median of the distribution, thick bars represent the first and third quartiles, and whiskers represent 1.5 IQR. Significance was defined as p<0.05. *: p<0.05, **: p<0.01, N.S.: not significant.

For Figure 3C, data were binned along the log-transformed x-axis into 10 equal-width intervals. Each bin included 14–173 data points. No data were excluded.

Growth curve quantifications

A background OD₆₀₀ value measured from a medium-only well was subtracted from all samples. Afterward, OD₆₀₀ readings were corrected for non-linearity using a standard curve as previously reported⁵⁴.

Relative growth area under the curve (AUC) was calculated by adding all OD₆₀₀ values from 0 h to 40 h during a growth passage and normalizing to that of the vehicle-treated control. We choose to use AUC rather than maximum OD₆₀₀ to quantify growth inhibition because it represents a more comprehensive summary of growth dynamics (including changes in lag) and is more robust to noise in OD₆₀₀ measurements⁵⁵.

Any samples in which the growth AUC was lower than the 10^th percentile of the control replicates were defined as growth inhibited.

16S rRNA gene amplicon analyses

All experiments yielded similar sequencing depths, with a median of 17,000 reads per sample and 90% of samples having >10,000 reads. Sequencing reads were processed with DADA2⁵⁶. Forward and reverse reads were truncated at 240 and 180 bp, respectively. To account for the multiple 16S genes in certain strains⁵⁷, we curated the ASVs by combining those that differed by 1–2 bp and that exhibited strong linear correlations (FDR-adjusted p<0.05) across all drug conditions. Taxonomies of the resulting ASVs were assigned using the SILVA reference database⁵⁸. For ASVs at high relative abundance, their taxonomic annotation was manually curated by blasting the 16S sequence against the NCBI database⁵¹. For all downstream analyses, samples with >10,000 reads were rarefied to 10,000 reads.

Estimation of absolute abundances

The absolute abundance of an ASV was estimated by scaling its relative abundance by the final OD₆₀₀ of the community and then multiplying that number by 10⁹. The multiplier 10⁹ corresponds to a consensus (~10⁹ CFUs/mL/OD₆₀₀) from measurements of colony-forming units for 34 communities with varying compositions (Figure S1B). Thus, absolute abundances are estimates of total cells per mL of culture. Although variations in cell size due to growth conditions⁵⁹ or species-specific morphologies⁶⁰ could affect CFU counts, such effects are typically within ~5-fold, smaller than the typical changes in absolute abundances (~100-fold or larger) focused on in this work.

Alpha and beta diversities

All alpha diversity metrics were calculated at the ASV level, after rarefying all samples to 10,000 reads. Shannon diversity was calculated as $H=-\sum _i{p}_i\text{ln}{p}_i$, where $p_i$ is the relative abundance of ASV $i$. CHAO1 index was calculated as $C=\mathit{\text{obs}}+\raisebox{1ex}{$n_1^2$}\!\left/ \!\raisebox{-1ex}{$2n_2$}\right.$, or $C=\mathit{\text{obs}}$ if $n_2=0$, where $\mathit{\text{obs}}$ is the total number of observed ASVs and $n_1$ and $n_2$ are the number of ASVs present with 1 and 2 reads, respectively. Bray–Curtis distance between samples A and B was calculated as $d=\frac{\sum _i\left|A_i-B_i\right|}{\sum _i\left(A_i+B_i\right)}$, where $A_i$ and $B_i$ are the relative abundances of ASV $i$ in communities A and B. Weighted and unweighted UniFrac distances were calculated using the R function phyloseq::distance⁶¹.

Metabolomics analyses

Across all 707 drug conditions, 36 samples had poor quality metabolomics data in which most features were not detected, presumably because the samples were not injected properly; these were excluded from future analyses.

In the 671 samples with metabolomics data, 361 had annotated peaks corresponding to the drugs added. We selected 90 of these 361 drugs, individually added the drugs at 20 μM to fresh BHI medium, and measured the intensities of the corresponding drug peaks. We then compared the corresponding peak intensities with those from supernatant after 48 h of SIC-0 growth in the presence of the drug to assay for potential bioaccumulation and/or degradation of the drugs. We found that 62/90 (69%) of the drugs remained present at ≥50% concentrations after community growth, indicating at most limited bioaccumulation or degradation; only 6/90 (7%) decreased by ≥80% in the community supernatant. We did not observe any trend between the fraction of remaining drug and the extent of community growth inhibition (Table S3, p=0.99, one-sided Student’s t-test).

To calculate distances between metabolomes, we first quantified the changes in area of each MS peak relative to the media control, and defined features as consumed if their concentration was depleted >100-fold and as secreted if their concentration increased >100-fold. The metabolomic distance between samples A and B was defined as $d=\sum _i\left|A_i-B_i\right|$, where $A_i$ ($B_i$) is −1, 0, or 1 if feature $i$ was consumed, unchanged, or secreted, respectively, in sample A (B).

To identify shared metabolite features from this study and a previous dataset⁹, we first identified the features with identical annotations and m/z difference <0.002, and performed a linear fit to correct for retention time (RT) between the two data sets. Unannotated features were matched when the corrected RT difference was <0.21 min and the m/z difference was <0.002. The 0.21 min threshold for RT corresponds to the 95^th percentile of RT differences in the annotated features. Finally, features exhibiting >10-fold differences across the media controls of the two data sets were filtered out.

Drug sensitivity analyses for isolates

In each titration series, if the final OD₆₀₀ at a certain drug concentration was higher than that at a lower concentration, the OD₆₀₀ at the lower concentration was used instead, since the higher OD was likely due to the emergence of resistant mutants⁴. IC₅₀ values were calculated as the concentration that corresponded to 50% growth inhibition, as estimated by linear interpolation of the corrected titration series.

To identify potential cross-protection (Figure 3C, bottom), we defined conditions as protected if the rescaled growth was <0.75 and the fold change in the community was >10^0.5, and sensitized if the rescaled growth was >1/0.75 and the fold change in the community was <10^−0.5.

Drug-response model for defined communities

In a defined community without drug treatment, we denote the absolute abundance of all member species as $\left\{A_1,A_2,A_3,\dots \right\}$. To predict the response of each species to a drug (Figure 3I, right), we first estimated the abundance of the most sensitized species $x$, whose relative monoculture growth in the presence of the drug is $s_x$. In the community, its abundance after drug treatment is estimated to be $s_x{A}_x$. In the experiment, we selected community and drug combinations so that $x$ is the only species strongly inhibited by the drug, meaning $s_x\ll s_i$ for $i\ne x$. Thus, the direct effect of drug treatment is to inhibit the growth of species $x$, allowing other species to utilize its open niche.

The reduction in biomass from species $x$ corresponds to the availability of additional nutrients for all other species, in the amount $\left(1-s_x\right)A_x$. Of all nutrients consumed by $x$, we inferred from isolate metabolomics data that the fraction shared with another species $y$ is $r_{xy}=\frac{N_{xy}}{Nx}$, where $N_x$ is the number of metabolite peaks consumed in monoculture by $x$, and $N_{xy}$ is the number of metabolomics peaks consumed by both $x$ and $y$ in their respective monocultures. In the presence of a drug, the additional nutrients accessible to $y$ due to inhibition of $x$ is $r_{xy}\left(1-s_x\right)A_x$. Thus, the final abundance of $y$ is predicted to be ${a_y=s}_y\left({A_y+r}_{xy}\left(1-s_x\right)A_x\right)$, where $s_y$ is the relative monoculture growth of $y$ in the presence of the drug versus without.

Beyond the direct effect of inhibiting species $x$, drug treatment could also affect how the other species compete for nutrients. We therefore inferred a range of growth potential by considering the extreme conditions in which each species $y$ other than $x$ utilizes all versus none of the nutrients shared with all other species. We then calculated the number of metabolite peaks uniquely accessed by $y$, ${N_y}^{\text{unique}}$. The minimum abundance of $y$ is estimated to be ${{a_y}^{\text{min}}=a}_y\frac{{N_y}^{\text{unique}}}{N_y}$, and the maximum abundance is estimated to be ${{a_y}^{\text{max}}=a}_y\frac{N_y}{{N_y}^{\text{unique}}}$. We did not calculate the range of growth for the most susceptible species, $x$, since drug treatment strongly inhibited its growth and hence it is unlikely to outcompete others.

This model assumes that all species translate nutrients into biomass with the same efficiency, and that drug susceptibilities are not correlated with the nutrient-consumption profile.

Consumer–resource (CR) simulations

Species abundance dynamics during drug treatment were simulated using a CR model as previously described³⁶. For the simulations in Figure 5B, communities were randomly initialized with 30 species with distinct profiles for the consumption of 30 resources. We simulated these communities using the CR model with a constant flow of resources until the composition reached equilibrium. Communities with fewer than 30 co-existing species at equilibrium were discarded, and a total of 1,000 communities with 30 co-existing species were used for downstream simulations. Then, a random death rate sampled from a uniform distribution between 10% and 40% of the dilution rate was imposed for each species to mimic drug treatment, and growth was simulated until equilibrium. The resulting community was allowed to recover to equilibrium with all death rates reset to zero. The community was then exposed to the drug for a second time; for a single species selected to be resistant, the death rate was set to zero while the same death rate as the first treatment was used for the second treatment for all other, non-resistant species. In each round of simulations, species that decreased to relative abundance <10^–4 were considered extinct and removed from the community.

For the simulations in Figure 6F, communities were generated with 15 species shared across all communities and 15 unique species. A fixed amount of each resource was supplied, and when all resources were depleted, the resulting abundances were scaled down to mimic a 1:200 dilution, and fresh resources were re-supplied. This passaging was repeated until equilibrium was reached. Three communities with 30 species co-existing at equilibrium were used for downstream simulations. To mimic drug treatment, a random death rate sampled from a uniform distribution was imposed for each species, and the steady-state communities were diluted, supplied with fresh resources, and allowed to grow for a fixed time in the presence of the non-zero death rates. For each community, 200 sets of death rates were used to mimic distinct drug treatments.

Monoculture resistance

In monocultures (Figure 3A), if the final OD₆₀₀ of a strain in the presence of a drug at 20 μM concentration was higher than its final OD₆₀₀ at 10 μM by 1.1-fold, we defined the 20 μM condition as having selected resistance. The threshold of 1.1 was chosen because the coefficient of variation across vehicle-treated controls was ~10%.

Whole genome sequencing analyses

Single nucleotide polymorphism (SNP) analyses were performed using SNIPPY v. 4.6.0⁶² with default parameters. Genome alignment was performed with the progressive Mauve algorithm⁶³ using Geneious Prime v. 2023.2.1 to identify structural variations.

Classifying dose-response dynamics

In the drug titration experiments (Figure 6G–L, S6), we classified dose response dynamics of ASVs with a median abundance >2×10⁶ cells/mL across drug concentrations.

An ASV was classified as non-monotonic if its maximum (or minimum) abundance occurred at an intermediate concentration and was at least 10^0.5-fold higher (or lower) than its abundance in either the no-drug control or the highest drug concentration.

For ASVs not exhibiting non-monotonic behavior, we performed a linear fit of log₁₀-transformed abundance versus log₁₀-transformed drug concentration. If the slope of the fit was >0.167 (corresponding to a 2-fold increase from the no-drug control to the highest drug concentration), the ASV was classified as monotonically increasing. Similarly, if the slope was <-0.167, it was classified as monotonically decreasing. ASVs with intermediate slopes between −0.167 and 0.167 were classified as exhibiting no change.

To calculate the ROC curves in Figure 6L,N, the threshold used to determine non-monotonic changes was varied from 1-fold to 5-fold.

Rescaled growth model for dose response

The rescaled growth for each species at each drug concentration was calculated using monoculture dose-response data (Figure 3A) as shown in Figure 3E. We then predicted each species’s abundance in the community by multiplying its rescaled growth with its absolute abundance in the untreated SIC. We noted that in monocultures, even with very strong growth inhibition, all species grew to OD₆₀₀~0.05, corresponding to a ~50-fold reduction in growth given an untreated OD₆₀₀~2.5, much less than the extent of inhibition observed in a community (Figure 1E, 3C). The additional growth in monocultures may be due to the lack of competition, allowing for slow growth that would be inhibited in the community. This baseline growth in monoculture limits the range of fold changes in the rescaled growth model compared to experimental measurements. To account for this difference, we modified our rescaled growth model by subtracting a baseline level of growth from the monoculture data. The baseline growth, assumed to be constant across all drugs and species, was obtained by minimizing the difference of the fold change distributions across all conditions between the model and experiments. Mathematically, such a baseline growth level does not alter the qualitative dynamics of any dose-response curve but enables better comparison between model and experiment across similar ranges of growth change.

To predict additional drug conditions (Figure 6M) with epistatic dose responses, we first calculated dose responses based on monoculture susceptibility data as mentioned above, focusing on high-abundance species (abundance in SIC-0 >10⁶ cells/mL, and predicted to have >10⁶ cells/mL in at least two drug concentrations). We further excluded drugs that in our initial screen (Figure 1E) promoted the growth of ASVs other than the 14 for which we acquired monoculture susceptibility data.

Supplementary Material

Table S3

Table S3: Metabolomic measurements to evaluate bioaccumulation/degradation for 90 drugs. Related to Figure 2. The drugs were selected based on annotations and having high peak intensities.

Table S2

Table S2: List of annotated metabolomic features depleted by SIC-0 >100-fold relative to fresh BHI. Related to Figure 2. Only peaks with integrated area >10⁴ in fresh BHI were included for analyses.

Table S4

Table S4: List of drugs selected for monoculture susceptibility testing. Related to Figure 3.

Table S1

Table S1: List of NCATS drugs and their primary targets. Related to Figure 1.

Table S5

Table S5: List of defined communities, chemical conditions, and final abundances of each species. Related to Figure 3.

Figure S1

Figure S1: Drug treatment leads to large-scale abundance changes. Related to Figure 1. A) Left: across drug conditions, maximum OD₆₀₀ and area under the growth curve (AUC) were highly correlated, with the lone outlier being chloramphenicol. Right: selected growth curves with different maximum OD₆₀₀ and growth AUC. Drug treatment typically prolonged lag time and reduced the carrying capacity, except for chloramphenicol, which resulted in a longer lag time and hence a lower AUC but did not affect maximum OD₆₀₀. B) Colony-forming units (CFU) of 34 communities with varying OD₆₀₀. Despite the differences in OD₆₀₀, all communities exhibited similar CFU/mL/OD₆₀₀, with overall variation within ~3-fold of 10⁹. C) Left: degree of community growth inhibition was correlated with the number of inhibited ASVs. Right: intermediate growth inhibition led to more ASVs expanding in abundance. D) Two ASVs that increased in absolute abundance after drug treatment. Red: replicates of vehicle-treated control. E) Many of the ASVs present at 1% or lower relative abundance in SIC-0 were able to grow to much higher yields (equivalent to >20% of the total yield of the untreated SIC-0) during monoculture growth in BHI. n=35 for community-level data, and n=2 for monocultures. F,G) The range of absolute abundances for each ASV across drug conditions. Regardless of their initial abundance in SIC-0 (red dots), the abundance of each ASV (F) or family (G) varied >300-fold across drug conditions.

Figure S2

Figure S2: Growth inhibition in a community is related to compositional changes and inter-species interactions. Related to Figure 2 and 3. A) There was no systematic trend between degree of growth inhibition and alpha diversity metrics. B) Across treated communities, relative growth AUC was negatively correlated with compositional distance as measured by unweighted UniFrac, Bray–Curtis, and Jaccard distances to the vehicle-treated control. Shaded areas in (A) and (B) represent the mean±1 S.D. for vehicle-treated controls. C) Phylogeny of bacterial isolates (left) juxtaposed with their growth responses to 20 μM drug treatments (right). Phylogenic distances were calculated from multiple sequence alignments of the V4 region in the 16S rRNA gene. Chemical response distances were calculated as Euclidean distances of their monoculture growth yield across 43 drugs. Lines connect the same species to aid visualization. Overall, clustering based on phylogeny and chemical responses resulted in distinct outcomes (Mantel statistic r=0.09). D) Bt growth in monoculture was not impacted by bifonazole treatment at concentrations 20 μM or lower. E) In a community, Bt was strongly inhibited by ≥20 μM bifonazole. At higher bifonazole concentrations, the extent of inhibition in a community was higher than that in monocultures. F) At concentrations ≥0.16 μM, supplemented heme promoted Bt growth in BHI. Dashed line: Bt growth without additional heme. n=2 replicates (shown as individual dots). *: p<0.05, **: p<0.01, one-tailed Student’s t-test. G) Incorporating community OD₆₀₀ data into the growth rescaling model for SIC-0 led to little improvement compared to Figure 3C.

Figure S3

Figure S3: Community recovery from drug treatment is hampered by extinction of some species. Related to Figure 4. A) Schematic of the recovery passage after the initial screen in Figure 1B. B) In the 141 conditions with growth and compositional changes (Figure 1D,E), a single passage after treatment enabled partial recovery of community growth (left) and composition (middle), but metabolomic profiles did not show clear signs of recovery (right). C,D) After multiple recovery passages (Figure 4A), the replicate pairs exhibited more similar compositions compared to random pairs (C). Growth dynamics were also more similar between replicate pairs without seeding (D, left). With seeding (D, right), all communities at R6 recovered control-like growth dynamics, and hence all replicate pairs grew similarly. E) Left: pazufloxacin was an outlier among the no-seeding conditions, in which the two replicates exhibited different compositions (Bray–Curtis distance=0.72). The variation was due mainly to the differential abundance of certain high-abundance ASVs; nonetheless, there was still a positive correlation across all ASVs. Right: During rifampicin treatment, replicate community compositions were strongly correlated across all ASVs. Shaded areas represent 95% confidence intervals of linear fits. F) Left: for all ASVs that were not detected in SIC-0 but bloomed during treatment, they were defined as transient blooming (TB) or long-term blooming (LB) based on their behavior during recovery passages R3–6. Middle: without seeding, 44 (35%) of these instances were transient and the ASVs reverted to undetectable levels after recovery, while 80 (65%) persisted at detectable levels throughout all recovery passages. Right: with seeding, there were somewhat fewer instances of persistent blooming (p=0.03, two-tailed Fisher’s exact test), and those that persisted were at much lower abundances than those that bloomed without seeding (median abundance of 1×10⁷ cells/mL in unseeded communities versus 3×10⁶ cells/mL in seeded communities, p<10⁻⁵, one-sided Mann-Whitney U test).

Figure S4

Figure S4: Strain swapping in drug-treated communities. Related to Figure 4 and 5. A) Without seeding, fluoroquinolone treatment strongly suppressed both Ef and Kp. B) The no-seeding communities recovered to lower OD₆₀₀ compared to the DMSO-treated control. C) Rifampicin treatment with seeding led to a strain swap in which Bt was replaced by B. intestinalis (Bi). D) Bt and Bi relative abundances were inversely correlated when the communities reached equilibrium. The dashed line represents a conserved total abundance summed across the two species. E) The strain swap of Bt and Bi slightly reduced the OD₆₀₀ of the resulting communities. n=2 replicates in A-E. F) Kp was somewhat more resistant than Ef to fluoroquinolones, with much higher relative growth AUC at 0.1 μM concentration. The first recovery passage after a 1:200 dilution of the treated culture involved drug carryover at a concentration ~0.1 μM, which likely allowed Kp to outcompete Ef after seeding during R1. G) Regardless of the initial mixture ratio between Kp and Ef, Ef outcompeted Kp after 3–4 passages, consistent with the observation that fluoroquinolones only induced transient blooming of Kp in pairwise cultures. n=3 replicates. H, I) When the six communities with seeding (H) and six without (I) were passaged further, Ef eventually recovered in some of them. In the seeded communities, only the levofloxacin-treated communities recovered. We identified a Bacteroides stercoris (Bs) ASV that exhibited lower abundance in the levofloxacin-treated communities compared to the gatifloxacin- and moxifloxacin-treated communities (FDR-adjusted p=0.007, one-tailed Mann–Whitney U test). This ASV was present at ~10^–3 or lower relative abundance in all seeded communities. J) Spent medium from a Bs monoculture still allowed for substantial Ef growth. Thus, the low amount of Bs in the community is unlikely to fully inhibit Ef growth via nutrient competition or other non-competitive interactions such as toxin production. K) Across all communities (both no-seeding and seeding) with low Ef after R13, compositions at the ASV and family levels were highly different as measured by Bray–Curtis distance. Moxi: moxifloaxacin; levo: levofloxacin; gati: gatifloxacin. L) In monoculture, Ef growth yield was high regardless of its initial inoculum size. M) SIC-0 was resistant to challenge with Kp, always recovering to a low-Kp, high-Ef state regardless of the initial ratio of Kp spike-in. N) With only one recovery passage, the slope of fold changes across the two treatments deviated from one, since the community did not reach new equilibrium (Figure 4B). Thus, for a susceptible ASV, its abundance after the recovery passage was lower than its steady-state value, artificially increasing its fold change in T2. However, the frequency of potential resistance selection was still similar to the frequency with six recovery passages (Figure 5C). O) With seeding, F. plautii exhibited resistance selection under similar drug conditions as in the unseeded communities (Figure 5E). Rol: rolitetracycline; cef: cefoxitin; tet: tetracycline. P) To explore potential mechanisms of resistance, we sequenced the genome of three isolates from monoculture and three from a community (Figure 5E) as well as the ancestor. F. plautii isolated from tetracycline-treated monocultures or communities had distinct alleles likely to confer tetracycline resistance. The monoculture isolates contained SNPs in genes related to drug efflux, while all community isolates had gene duplications that increased the copy number of potential tetracycline-resistance genes.

Figure S5

Figure S5: Effects of drug treatment are largely conserved across SICs. Related to Figure 6. A) Relative abundances of ASVs detected in both SIC-MD and SIC-cip. The ASVs are sorted by their abundance in SIC-0 (Figure 1C). These SICs contain some of the same ASVs as SIC-0, but the relative abundances of many overlapping ASVs differed across SICs. B) Left: Among ASVs with representative isolates, A. muciniphila was the only major ASV in the SICs that exhibited moxifloxacin resistance. Right: Ef was resistant to oxiconazole treatment, while different Enterococcus isolates exhibited variable susceptibilities. C) Absolute fold-change of Ef (left) and Eh (right) ASVs across SICs. In both cases, the fold changes in two SICs were correlated but deviated from the line y=x. D) PCoA plots of the communities resulting from drug treatment of SIC-0, SIC-MD, and SIC-cip. The pre-treatment SIC (left) and drug condition (right) both affected community composition. Small dots are colored by pre-treatment SIC (left) or drug condition (right), and large dots are the corresponding centroids. Ellipses represent the 60% confidence intervals of the dots. E) Left: SICs derived from other human subjects exhibited similar growth dynamics as SIC-0 across all drugs except for three tetracycline-family drugs. r>0.81, p<10^–5 for all SICs. Right: there were large variations in growth across the SICs for three tetracycline-family drugs: oxytetracycline, rolitetracycline, and tetracycline. F) The variations in growth AUC for the three tetracycline-family drugs can be largely explained by the susceptibility of isolated Enterobacteriaceae ASVs in each SIC. <(≥)1/2: the fold change in Enterobacteriaceae family abundance under drug treatment is less (greater than or equal to) than 1/2 in monoculture. G) These eight SICs were derived from different subjects and the same ASV is highly likely to represent evolutionarily distinct strains across communities. Nonetheless, across virtually all ASVs shared between SIC-1-8 and SIC-0, we observed strong positive correlations in abundance change across all drugs (FDR-adjusted p≤0.02). The only exception was the Ef ASV in SIC-6 (FDR-adjusted p=0.4), likely because SIC-6 contains another Enterobacteriaceae ASV (a Proteus species) at much higher abundance (~70% relative abundance compared to ~10% relative abundance of Ef in the untreated SIC-6), which presumably affects Ef growth. N.D.: no data (due to the absence of the ASV in one or both SICs). N.S.: not significant.

Figure S6

Figure S6: Increasing/non-monotonic dose responses are reproducible. Related to Figure 6. A) Dose responses of high-abundance ASVs to bifonazole, metronidazole, tetracycline, and rifampicin treatment across three biological replicates of SIC-0. Gray lines represent ASVs for which we lacked isolate susceptibility data. To simplify visualization, in bifonazole and tetracycline, only the ten highest-abundance ASVs are shown. In replicate 3 of tetracycline treatment, data from 80 μM tetracycline exhibited clear signs of resistance emergence and were excluded from analyses. B) Across the three replicates of SIC-0, the types of dose responses of individual ASVs were largely reproducible. Numbers denote ASV counts.