Ecological memory of prolonged plasmid persistence after transient antibiotic exposure

Abstract

Plasmids play critical roles in spreading and maintaining antimicrobial resistance (AMR). They often exhibit prolonged persistence after antibiotic treatment, even when they impose a substantial burden on their hosts. This persistence has been primarily attributed to rapid horizontal transfer or low plasmid cost. However, these mechanisms cannot account for the slow decay of burdensome plasmids with poor mobility. Here, we show that the decoupling of time scales between slow segregation loss and fast growth competition leads to a slowing of plasmid-abundance decay at high initial plasmid abundance. Integrating theory, simulations, and quantitative experiments across clonal populations and multispecies bacterial communities, we demonstrate that a transient antibiotic pulse can eliminate plasmid-free cells and create a segregation-loss bottleneck that extends plasmid persistence from days to weeks. Our study reveals a generalizable mechanism for the prolonged ecological memory of antibiotic exposure and underscores the need for proactive strategies to curb the spread of AMR.

Zhou, Weiss and colleagues show that a brief antibiotic pulse can prolong plasmid persistence by removing plasmid-free competitors, creating an ecological memory that may sustain antimicrobial resistance.

Introduction

Plasmids are self-replicating genetic elements that play critical roles in bacterial evolution by enabling the acquisition, maintenance, and dissemination of adaptive traits such as antimicrobial resistance (AMR)^1–3, giving rise to many multidrug-resistant high-risk clones^4,5 and accelerating the global AMR crisis³. Although plasmid carriage is often assumed to be burdensome in the absence of selection, leading to rapid plasmid loss^6,7, empirical studies suggest otherwise. Resistance frequency often remains high for months after short-term antibiotic treatments (days to weeks), despite the absence of continued selective pressure thereafter^8–12. Given that many resistance determinants reside on plasmids^11,13–19, the slow resistance reversal suggests burdensome plasmids can still exhibit prolonged transients after antibiotic exposure.

Different mechanisms may account for this prolonged persistence after antibiotic treatment. For instance, positive selection can increase the frequency of compensatory mutations that reduce the cost of plasmid carriage and slow down the decay of plasmids²⁰. The long half-lives of certain antibiotics can sustain selection for resistance after treatment has ceased⁹. More broadly, the persistence or slow reversal of resistance or plasmid carriage can be associated with low or zero fitness cost^6,21–23, high conjugation rate^24–26, or both^24,27. Co-selection by antibiotics^6,28 or environmental contaminants²⁹ on the same mobile genetic element has also been proposed. However, these mechanisms cannot explain the slow decay of plasmids that are burdensome and poorly mobile, because such plasmids are expected to be lost rapidly from low-abundance states.

Previous theoretical work suggests another underappreciated possibility: the slow segregation loss of plasmids^30,31. Segregation has been incorporated in modeling frameworks of steady-state plasmid persistence^24,25,27, compensatory evolution²⁰, and plasmid copy number control³², yet is rarely the main focus^20,24,25,27. Over time, slow segregation loss can sustain plasmid carriage in clonal populations for prolonged periods^32–34, before high-burden plasmid cargo drives eventual plasmid loss. The ecological consequences of the timescale decoupling between segregation loss and growth competition have not been explored. We propose this timescale decoupling as the basis for prolonged plasmid carriage after antibiotic exposure: antibiotics eliminate plasmid-free competitors, so the population must first generate plasmid-free cells through slow segregation loss before those segregants drive rapid plasmid decline through growth competition.

We combine theory, numerical simulations, and quantitative experiments to demonstrate this principle in clonal populations, synthetic E. coli communities, and multispecies communities of hospital sink isolates. Across these systems, the separation of timescales between segregation loss and growth competition generates prolonged ecological memory of plasmid carriage. This principle underscores the need for active intervention strategies beyond standard antibiotic stewardship.

Results

We first establish how segregation loss and plasmid-associated fitness costs define distinct decay regimes in clonal populations. We then test how a transient antibiotic pulse resets populations into the segregation-limited regime and extend this framework to synthetic and multispecies communities, where either segregation bottlenecks or host shifts can sustain plasmid persistence.

Segregation loss and plasmid cost define two plasmid-decay regimes

We start from a simple ordinary differential equation (ODE) model depicting the dynamics of a non-mobilizable plasmid within a single strain population (Methods). Specifically, we modeled the density of two populations, plasmid-free cells $f$ and plasmid-carrying cells $p$:

$$\frac{df}{dt}=\alpha \mu \left(1-\frac{f+p}{c}\right)f-Df+\kappa p$$ 1

$$\frac{dp}{dt}=\mu \left(1-\frac{f+p}{c}\right)p-Dp-\kappa p$$ 2

Here $\mu$ is the growth rate of plasmid-carrying cells, $\alpha$ reflects the fitness effect of the plasmid: costly when $\alpha$ > 1 or beneficial when $\alpha$ < 1, $c$ is the carrying capacity, $D$ is the dilution rate, $\kappa$ is the segregation loss rate. The dynamics of plasmid loss in a population are mainly dictated by two processes: (1) segregation loss, where a daughter cell fails to inherit the plasmid during cell division (Fig. 1a, pink), and (2) growth competition, where plasmid-carrying cells are outcompeted by the plasmid-free population because of plasmid-imposed metabolic burden on their host (Fig. 1a, purple). In this model, segregation loss is taken as a slow yet continuous process, where different plasmids can deploy active partitioning, post-segregational killing, and high copy number to minimize their loss during cell division^35,36. Meanwhile, growth competition operates on a faster timescale.

Fig. 1. High plasmid abundance nonlinearly increases plasmid half-lives.

a Plasmid abundance decreases through two processes operating on different time scales, slow segregation loss of the plasmid (pink) and rapid out-competition of plasmid-carrying cells by plasmid-free cells (purple). b Simulated time series of plasmid abundance decay. When starting from high initial abundance P₀% (teal: 100%, orange: 99%), plasmid abundance exhibits biphasic decay, with the slow phase driven by segregation loss (pink arrow) followed by a faster phase driven by growth competition (purple arrow). When starting from lower P₀% values (dusty blue: 90%, pink: 50%, and lime green: 20%), the plasmid abundance follows exponential decay. c Dependence of plasmid half-life ${\tau }_{1/2}$ on the initial abundance P₀%. The scatter points represent simulation results; the curve shows the best fit to Eq. (3) in Methods. d Time series of average pSC101 plasmid abundance decay (mean ± SE, n = 3) in E. coli MG1655 across different P₀% (teal: 100%, orange: 99%, dusty blue: 90%, pink: 50%, and lime green: 20%). e Dependence of plasmid half-life ${\tau }_{1/2}$ (mean ± SE, delta method) on the initial abundance P₀% for non-mobilizable plasmids pSC101, colE1, and pUC. n = 3 biological replicates, shown individually in each panel. Curves connect the median half-lives at each given initial abundance P₀%. Right-censored data points (${\tau }_{1/2}$ > 12 days) are shown without error bars. f Dependence of plasmid half-life ${\tau }_{1/2}$ (mean ± SE, delta method) on the initial abundance P₀% for conjugative plasmids pCU1 and R6K. n = 3 biological replicates, shown individually in each panel. Curves connect the median half-lives at each initial abundance P₀%. Right-censored data points (${\tau }_{1/2}$ > 25 days) are shown without error bars. Source data are provided as a Source data file.

We used the ODE model to simulate plasmid abundance (fraction of plasmid-carrying cells in a population) over time (Fig. 1b), across different initial plasmid abundances of P₀% = 100%, 99%, 90%, 50%, and 20%. When starting from a pure plasmid-carrying population (P₀% = 100%), plasmid abundance decays in two phases, corresponding to the two processes: an initial slow decay driven by segregation loss, followed by a faster decay driven by growth competition once plasmid-free cells have emerged. This two-phase behavior is consistent with previous experimental observations^33,37, where the presence of post-segregational killing mechanisms results in a long plateau in which plasmid abundance remains ~100%. In the absence of active partitioning or post-segregational killing mechanisms, this plateau was significantly shortened or even eliminated. As the initial plasmid abundance decreases, growth competition becomes more prominent, driving the plasmid abundance decay exponentially. Based on previous work³⁰, we can analytically derive the half-life of plasmid-abundance decay (Methods), which increases nonlinearly with initial plasmid abundance (Fig. 1c). When P₀% $\ll$ 100%, the abundance decays exponentially with half-life ${\tau }_{1/2}=\frac{1}{\delta }\log 2$. When P₀% = 100%, plasmid half-life becomes ${\tau }_{1/2}\to \frac{1}{\delta }\log \left(\frac{\delta }{\kappa }\right)$, where $\delta$ is the rate of growth competition between the plasmid-free and plasmid-carrying populations, and $\kappa$ is the segregation loss rate (Methods). In most cases, $\kappa \ll \delta$²⁴, thus, the half-life becomes large, and the high-abundance state appears as the ghost of an equilibrium³⁸. This logarithmic scaling between $\kappa$ and ${\tau }_{1/2}$ resembles the ghost effect coined by Strogatz³⁸, originally referring to the loss of stability in a dynamical system during saddle-node bifurcation. The prolonged plasmid carriage in our case instead arises from the remnant of a saddle point (when $\kappa$ increases from 0 to a small number).

To test our prediction of diverging half-lives near pure population, we monitored the abundance of five different plasmids in two E. coli hosts over time (Figs. 1d and S1, S2). Specifically, we included three non-mobilizable, sfGFP-encoding plasmids in E. coli strain MG1655, and two conjugative plasmids in E. coli DA28102 (chloramphenicol resistant, Cm^R). The three sfGFP-encoding plasmids replicate under different replication origins, pSC101, colE1, and pUC, thus differing in their copy numbers, estimated as 10.90 ± 0.13, 19.86 ± 0.83, and 46.50 ± 1.38 (mean ± SE, n = 3) based on qPCR (Fig. S3, Methods). To facilitate high-throughput measurements, we used the GFP/OD ratio to quantify the GFP-encoding plasmid abundances (calibration details in Methods). The abundances of the two conjugative plasmids, pCU1 and R6K, were tracked over time through selective plating. Plasmid-carrying E. coli cells were mixed with plasmid-free ones at different volume ratios to generate populations in which plasmid-carrying cells constituted 100%, 99%, 90%, 50%, or 20% of the total population. MG1655 populations carrying GFP-encoding plasmids were cultured in LB and diluted 1:500 every 24 h, and DA28102 populations carrying the conjugative plasmids were cultured in LB supplemented with 25 μg/mL Cm and diluted 1:100 daily (Methods). Consistent with the model predictions, in populations with low initial abundances of 50% or 20%, plasmid abundances exponentially decayed over time. In contrast, with increasing initial plasmid abundance, a more pronounced plateau in plasmid abundance was observed before exponential decay (Figs. S1b and S2), significantly prolonging the half-life of plasmid abundance. This nonlinear increase in plasmid half-life with initial abundance was consistently observed across all five plasmids (Fig. 1e, f). Thus, the initial purity of the population with respect to plasmid carriage dictates the plasmid half-life (Fig. 1e, f).

In pure populations, when P₀% = 100%, plasmid persistence is controlled by both segregation loss and growth competition. Since the GFP-encoding plasmids lacked stability systems, segregation loss should mainly depend on plasmid copy number. Under a random-partitioning model, higher-copy plasmids are less likely to get completely lost during cell division, leading to a smaller effective $\kappa$ value and longer persistence. By assuming random partition of plasmid copies, no per-cell copy-number variability, and estimating $\delta$ values based on the half-lives of P₀% = 20% populations, we estimated pSC101, colE1, and pUC to have half-lives of 5.8 ± 0.5, 14.2 ± 1.5, and 80.8 ± 9.2 days (mean ± SE) (Methods). Experimentally, we observed increasing persistence with plasmid copy numbers (Fig. S1b), with plasmid half-lives of 5.1 ± 0.2 days (mean ± SD, n = 3) for pSC101, 6.8 ± 3.5 days (mean ± SD, n = 3) for colE1, and >12 days for pUC (Fig. 1e). The experimental half-lives were comparable but shorter than theoretical expectations, suggesting that copy-number heterogeneity, imperfect random partitioning, or decreasing GFP expression affected the measured dynamics.

For pCU1 and R6K, copy number alone is unlikely to capture segregation loss due to their stability mechanisms. Plasmid pCU1 contains a putative stbABC-like operon, homolog loci of which promote plasmid stability in various plasmids^39,40. Plasmid R6K contains an enhancer module adjacent to the gamma origin that confers stable inheritance⁴¹. These systems can reduce the birth rate of plasmid-free cells, thereby extending the plateau predicted by the model. As a result, at P₀% = 100%, pCU1 and R6K had long half-lives, with a pCU1 half-life of 10.4 ± 0.5 days (mean ± SD, n = 3) and an R6K half-life of 21.4 days for one replicate and >25 days for the other two. Because all five plasmids imposed substantial fitness costs on their hosts, as reflected in their growth curves (Supplementary Note, Fig. S4), they had half-lives between 0.8 and 2.3 days at the low initial abundance P₀% = 20%.

Transient antibiotic selection resets populations into the segregation-limited regime

Antibiotics can eliminate plasmid-free cells in a population and purify the population to enter the kinetic bottleneck (Fig. 2a). We therefore hypothesized that antibiotic selection would prolong plasmid persistence. We assembled populations of E. coli MG1655 with an equal mixture of plasmid-free and plasmid-carrying cells (P₀% = 50%). We used the non-mobilizable, sfGFP-encoding plasmids pSC101, colE1, and pUC for high-throughput measurements. We applied a pulse of antibiotic selection to the populations between days 2 and 3. After growing the population in LB supplemented with a gradient of antibiotics (kanamycin for pSC101 and colE1, and spectinomycin for pUC) for 24 h, we removed the selection and tracked the GFP/OD readout over time (Figs. 2b and S5b). The half-lives of the three plasmids all substantially increased with the strength of the pulsed antibiotic selection (Fig. 2c), and plateaued above critical concentrations. The critical concentrations were 3.125 µg/mL Kan for pSC101, 6.25 µg/mL Kan for colE1, and 12.5 µg/mL Spec for pUC.

Fig. 2. A transient antibiotic pulse prolongs plasmid persistence.

a In a mixed population containing both plasmid-carrying and plasmid-free cells (dusty blue), an antibiotic purifies the population by eliminating the plasmid-free cells, moving the population into the bottleneck regime (orange), and extending the half-life of plasmid decay (right). b Time series of average plasmid abundance (mean ± SE, n = 3) for non-mobilizable plasmids pSC101, colE1, and pUC under different selection strengths. Between days 2 and 3 (gray), antibiotics were applied at different concentrations to remove plasmid-free cells to different extents. Three representative conditions are shown here, corresponding to no antibiotic treatment (circle), 1/32 × 50 μg/mL (triangle), and 50 μg/mL (square). pSC101 and colE1 populations were treated with Kan, and pUC populations were treated with Spec. c Dose-dependence of plasmid half-lives ${\tau }_{1/2}$ (mean ± SE, delta method) on the pulsed antibiotic concentration for pSC101, colE1, and pUC. Right-censored data points (${\tau }_{1/2}$ > 17 days) are shown without error bars. n = 3 biological replicates, shown individually in each panel. d Plasmid half-lives at P₀% = 50% of nine clones (three clones per biological replicate) on day 3 and day 20. For the clones on day 3, the median half-lives were 2.6 days [IQR: 2.4–3.1; min – max: 1.8 – >4.0] for pSC101, 1.7 days [IQR: 1.6–1.8; min – max: 1.1–1.8] for colE1, and 2.2 days [IQR: 2.1–2.3; min – max: 2.0–2.4] for pUC. In one pSC101-carrying clone, plasmid abundance remained stable and did not show exponential decay; its ${\tau }_{1/2}$ was therefore >4 days. For the clones on day 20, the median half-lives were 2.0 days [IQR: 1.6–2.0; min – max: 1.4–2.8] for pSC101, 1.5 days [IQR: 1.4–1.5; min – max: 1.3–1.8] for colE1, and 1.9 days [IQR: 1.6–2.1; min – max: 0.9–3.7] for pUC. Whiskers extend to the smallest and largest values within 1.5 times the interquartile range. Source data are provided as a Source data file.

In addition to purifying the population to enter the kinetic bottleneck, antibiotics could reduce the effective plasmid burden by enriching compensatory mutations²⁰, contributing to increased stability after antibiotic removal. To test whether this mechanism contributed to the extended plasmid half-lives observed in our experiments, we cured plasmids from nine evolved clones (three clones from each of three biological replicates) per plasmid (Methods). Focusing on the populations treated with 50 µg/mL of Kan or Spec, we first quantified plasmid burden by comparing growth curves of plasmid-carrying clones with those of their cured counterparts (Fig. S6, Methods). We compared clones from the ancestral stock, immediately after antibiotic selection (day 3), and at the end of the experiment (day 20). Over time, the burdens of pSC101 and pUC were reduced, while colE1-carrying populations did not show statistically significant changes (Fig. S6 and Supplementary Note). We further tested whether the reduced burden translated into extended half-lives in P₀% = 50% populations (Fig. S7, Methods). Out of the 54 clones tested across the three plasmids and two days, 53 clones had half-lives below 4 days (Supplementary Note). pSC101 persisted in one clone on day 3, with GFP level remaining stable (Fig. S7a). In comparison, plasmid half-lives in P₀% = 100% populations were 5.1 ± 0.2 days (mean ± SD, n = 3) for pSC101, 6.8 ± 3.5 days for colE1, and >12 days for pUC (Fig. 1e). Therefore, these results suggest that plasmid burden was reduced after the antibiotic pulse, yet the ameliorated burden alone could not account for the prolonged persistence.

Antibiotic pulse induces prolonged plasmid carriage in synthetic E. coli communities

To test whether this effect extends to complex communities, we constructed five synthetic communities from E. coli Keio strains⁴² (Supplementary Data 1 and 2). Each strain carried a uniquely barcoded, non-mobilizable plasmid for NGS-based tracking of community structure (Methods). Comm87 comprised 86 background strains and one R388-carrying donor; four Comm57 communities comprised the same 56 background strains plus a donor carrying R6K, pCU1, R388, or RP4. Communities were assembled at a 3:1 background-to-donor ratio and passaged daily in LB containing carbenicillin (Comm87) or chloramphenicol (Comm57) to maintain the barcode plasmids. Target-plasmid abundance was tracked by selective plating (Figs. 3a and S8). Hereafter, plasmid-carrying and plasmid-free refer only to the target conjugative plasmid, not the barcode plasmid.

Fig. 3. Antibiotic pulse induces prolonged plasmid carriage in synthetic E. coli communities.

a Plasmid abundance over time for pCU1 and R6K in Comm57 and R388 in Comm87. Dusty blue, daily passage without target-plasmid selection; orange, a target-plasmid-selective pulse between days 2 and 3 (gray bar). Mean ± SE from technical triplicates are shown for each of n = 3 biological replicates. Growth media contained 25 μg/mL chloramphenicol (Comm57) or 100 μg/mL carbenicillin (Comm87); pulses used 100 μg/mL carbenicillin (pCU1), 100 μg/mL streptomycin (R6K), or 10 μg/mL trimethoprim (R388). b pCU1 half-life (individually shown as mean ± SE, delta method) increased from 1.0 ± 0.1 to 9.9 ± 0.9 days after carbenicillin treatment (mean ± SD, n = 3; t(2.01) = 16.50, p = 0.004; mean difference, 8.8 days [95% CI, 6.5–11.1]; Cohen’s d = 13.47). Invasion by pCU1-free strain #3 reduced it to 4.9 ± 1.7 days (t(3.04) = − 4.36, p = 0.022; difference, −5.0 days [95% CI, −8.6 to −1.4]; d = −3.56). c R6K half-life increased from 4.4 ± 0.1 to ≥17.0 days after streptomycin treatment (n = 3; t(2.00) = 190.56, p = 2.8 × 10⁻⁵; difference, 12.6 days [95% CI, 12.4–12.9]; d = 155.59). Invasion by R6K-free strain #4 reduced it to 7.2 ± 1.5 days (t(2.00) = − 11.17, p = 0.008; difference, −9.8 days [95% CI, −13.6 to −6.0]; d = −9.12). Right-censored values (>17 days) were set to 17 days for testing. (d). R388 half-life increased from 2.2 ± 0.4 to 27.3 ± 1.7 days after trimethoprim treatment (n = 3; t(2.21) = 24.91, p = 9.5 × 10⁻⁴; difference, 25.1 days [95% CI, 21.1–29.1]; d = 20.34). Invasion by R388-free strain #76 did not reduce plasmid abundance by half within 15 days. All tests were two-sided Welch’s t-tests. *p < 0.05, **p < 0.01, ***p < 0.001. Source data are provided as a Source data file.

Without additional target-plasmid selection, pCU1 and R6K in Comm57 and R388 in Comm87 decayed with half-lives of 1.0 ± 0.1, 4.4 ± 0.1, and 2.2 ± 0.4 days, respectively (mean ± SD, n = 3; Fig. 3a–d); R388 and RP4 persisted throughout the 20-day experiment in Comm57 (Fig. S8g, h). A target-plasmid-selective pulse between days 2 and 3 drove abundance toward 100% and extended half-lives to 9.9 ± 0.9 days for pCU1 (p = 0.004) and 27.3 ± 1.7 days for R388 in Comm87 (p = 9.5 × 10⁻⁴). In Comm57, R6K, R388, and RP4 remained near 100% for more than 17 days after selection (Figs. 3c and S8). Similar behavior was observed when Comm57 communities were passaged at a 1:500 dilution (Fig. S8).

NGS revealed that, without target-plasmid selection, donor abundances declined rapidly, consistent with inter-strain growth competition (Figs. 4a–c and S9). The pulse reorganized all communities, usually reducing diversity and increasing donor abundance (Figs. 4d and S10). Comm57 carrying RP4 was the exception: diversity increased, and donor abundance decreased, likely because RP4’s rapid conjugation^24,43 had transferred the plasmid to fitter community members before tetracycline selection.

Fig. 4. Community dynamics in synthetic E. coli communities.

Strain-level community dynamics (mean ± SE, n = 3) for Comm57 carrying pCU1 (a) or R6K (b), without (left) or with (right) a 1-day target-plasmid-selective pulse. The 12 most abundant strains are colored; donor strains were #3 and #4, respectively. c Strain-level dynamics (mean ± SE, n = 3) for Comm87 carrying R388, without (left) or with (right) a 1-day trimethoprim pulse. The five most abundant strains are colored; strain #1 was the donor. d Change in donor relative abundance from day 2 to day 3 (mean ± SE, n = 3). Two-sided Welch’s t-tests, no pulse versus +Ab: Comm87 + R388, t(3.2) = 31.4, p = 4.4 × 10⁻⁵, difference 0.78 [95% CI, 0.70–0.86], d = 25.6; Comm57 + R6K, t(2.0) = 26.9, p = 1.4 × 10^-3, difference 0.83 [0.70–0.96], d = 22.0; Comm57 + pCU1, t(2.6) = 25.7, p = 3.3 × 10⁻⁴, difference 0.53 [0.46–0.61], d = 21.0; Comm57 + R388, t(3.9) = 55.0, p = 7.8 × 10^-7, difference 0.64 [0.61–0.67], d = 44.9; Comm57 + RP4, t(3.7) = −5.6, p = 6.4 × 10⁻³, difference −0.29 [−0.44 to −0.14], d = −4.5. **p < 0.01; ***p < 0.001. e Dependence of plasmid half-life ${\tau }_{1/2}$ (mean ± SE, delta method) on the initial abundance P₀% for plasmids pCU1 in strain #3 (left) and R6K in strain #4 (right). Curves connect the median half-lives at each initial abundance P₀%. Right-censored data points (${\tau }_{1/2}$ > 14 days) are shown without error bars. n = 3 biological replicates, shown individually in each panel. f Dependence of plasmid half-life ${\tau }_{1/2}$ (mean ± SE, delta method) on the initial abundance P₀% for plasmid R388 on strains #1 (teal), #15 (lime green), and #76 (pink). Curves connect the median half-lives at each initial abundance P₀%. Right-censored data points (${\tau }_{1/2}$ > 28 days) are shown without error bars. n = 3 biological replicates. Source data are provided as a Source data file.

Segregation bottlenecks and host shifts maintain post-antibiotic plasmid persistence

We next asked whether the prolonged plasmid persistence after antibiotic selection resulted from slow segregation loss, as in the clonal scenario, or maintenance by a favorable host in which the plasmid imposes little burden. Under this alternative hypothesis, plasmid-free cells would have little or no competitive advantage, and prolonged plasmid persistence would not require a segregation bottleneck. For each community, we focused on dominant strains after antibiotic selection: strain #3 in Comm57 + pCU1 (Fig. 4a), #4 in Comm57 + R6K (Fig. 4b), and strains #1, #15, and #76 in Comm87 + R388 (Fig. 4c). Strains #1, #3, and #4 were the donor strains for R388, pCU1, and R6K in each community.

We first tested these mechanisms for pCU1 and R6K in Comm57 by measuring plasmid persistence in individual strains. Populations were initialized at P₀% = 100%, 90%, or 50% and tracked over time (Fig. S11). If persistence reflected negligible plasmid burden, lowering the initial abundance should have had little effect. Instead, half-life decreased from 11.94 ± 0.88 days at 100% to 3.63 ± 0.04 days at 50% for pCU1 (mean ± SD, n = 3; Fig. 4e). R6K remained above the half threshold for more than 14 days at 100% or 90%, but its half-life was 6.15 ± 0.13 days at 50%. Thus, at the single-strain level, both plasmids imposed measurable fitness costs on their hosts.

We next tested whether the same logic applies in the community context. If post-antibiotic plasmid persistence depends on the bottleneck of segregation loss, then directly adding plasmid-free competitors should bypass this bottleneck and accelerate plasmid abundance decay. In contrast, if persistence is caused by minimal burden on a favorable host, invasion by plasmid-free cells should have little effect, and plasmid abundance should remain stable or increase. We therefore invaded day 3 samples (immediately after antibiotic selection) of Comm57 + pCU1 with the pCU1-free strain #3, and Comm57 + R6K with R6K-free strain #4 (Methods). In both cases, plasmid abundance declined after invasion (Figs. S12a and S12b), with half-lives of 4.9 ± 1.7 days for pCU1 and 7.2 ± 1.5 days for R6K (mean ± SD, n = 3), significantly shorter than in the uninvaded community (two-sided Welch’s t-tests, p = 0.022 and 0.008, respectively). Together, these results show that plasmid-free cells have a competitive advantage once present, and that the persistence of pCU1 and R6K after antibiotic treatment is maintained by the segregation bottleneck.

R388 imposed less burden in Comm87: at P₀% = 50%, half-lives were 10.3 ± 2.2 days in strain #1 and 13.9 ± 2.1 days in strain #15 (mean ± SD, n = 3; Figs. 4f and S11b). In strain #76, R388 remained above the half threshold through day 28, identifying #76 as a favorable host. Because #76 dominated Comm87 after approximately day 10, host shift likely contributed to plasmid persistence later. Consistently, invading the treated community with R388-free #76 on day 12 did not destabilize plasmid abundance for more than 15 days (Figs. 3d and S12c). Thus, R388 persistence reflected both segregation bottlenecks in strains #1 and #15 and low burden after transfer to strain #76. Although #76 was also abundant without treatment, R388 was not maintained in the untreated community, indicating that antibiotic selection and conjugation together enabled this host shift. These results distinguish two complementary routes to persistence: a segregation bottleneck when plasmid-free competitors are absent, and transfer to a favorable host (Supplementary Data 4).

Prolonged plasmid persistence in multispecies, niche-partitioned communities

Natural communities can avoid competitive exclusion through resource or habitat partitioning (Fig. 5a). In the clonal and E. coli community experiments, antibiotic selection prolonged persistence after target-plasmid abundance approached 100%. Simulations showed that the same segregation bottleneck can occur without whole-community purification when niches are partitioned (Fig. 5b, c; Methods). In a two-member model, an antibiotic-sensitive host carried the plasmid while a nonhost occupied an independent niche and survived selection through moderate resistance. Although total plasmid abundance remained below 100%, the plasmid-carrying host population approached local purity and therefore had to generate plasmid-free competitors through segregation loss before rapid decay could begin. Thus, niche-specific purification is sufficient to create prolonged plasmid persistence in a compositionally mixed community.

Fig. 5. Prolonged plasmid persistence in niche-partitioned communities following antibiotic pulse.

a Niche partitioning allows microbes to coexist by using different resources or habitats. b Simulated composition of a two-strain community with complete niche partitioning. Green and dusty blue denote strains 1 and 2; orange shading denotes the plasmid-carrying fraction of strain 1; gray bar, plasmid-selective pulse between days 2 and 3. c Simulated plasmid abundance with (orange) or without (dusty blue) a 1-day pulse. After the pulse, plasmid abundance remained stable despite staying below 100%. d Experimental communities containing plasmid-free and plasmid-carrying E. coli MG1655 plus eight sink isolates, cultured in LB with sponge-imposed structure. Sink isolates comprised one Pseudomonas (P. aeruginosa), three Bacillus, one Citrobacter, and three Enterobacter isolates, including one E. cloacae. Colors denote genera as in panel f; pSC101, colE1, and pUC were tested in independent communities. e GFP/OD time series for pSC101, colE1, and pUC. Dusty blue, daily passage at a 2 × 10⁷-fold dilution in LB-filled sponges; orange, an additional pulse between days 1 and 2 (gray bar; 50 μg/mL kanamycin for pSC101 and colE1, 50 μg/mL spectinomycin for pUC); green bar, mean ± SE baseline for plasmid-carrying E. coli MG1655 (n = 3). Biological replicates (n = 3) are shown individually; error bars indicate propagated per-well measurement uncertainty (SD). f 16S-resolved community dynamics from day 0 to day 2 (n = 1 at day 0; mean ± SE, n = 3 otherwise). After the pulse, E. coli and P. aeruginosa dominated. Because 16S sequencing assigned E. coli to Escherichia-Shigella, it is labeled Escherichia. Taxa with a maximum relative abundance <1% are omitted. Source data are provided as a Source data file.

To test this prediction, we constructed three communities containing eight isolates from Duke Hospital sink p-traps⁴⁴, plasmid-free E. coli MG1655, and MG1655 carrying one of three burdensome, non-mobilizable sfGFP-encoding plasmids (pSC101, colE1, or pUC; Methods). The sink isolates comprised one Pseudomonas aeruginosa strain, three Bacillus strains, one Citrobacter strain, and three Enterobacter strains (Fig. 5d). We immersed cellulose sponges in LB to create porous, physically partitioned habitats (Fig. 5d), an approach previously shown to promote microbial community diversity⁴⁵.

Without antibiotic selection, GFP/OD fell rapidly below 10². A transient kanamycin or spectinomycin pulse between days 1 and 2 increased GFP/OD and slowed subsequent decay: pSC101 declined gradually, whereas colE1 and pUC persisted through day 10 (Fig. 5e), consistent with their clonal-population dynamics (Figs. 1 and S1). For colE1 and pUC, GFP/OD remained below the pure plasmid-carrying MG1655 baseline, showing that prolonged persistence did not require total community purification.

16S sequencing showed that, without the pulse, P. aeruginosa and Enterobacter dominated by day 2, while the other genera declined or became undetectable (Fig. 5f; Methods). The pulse enriched the resistant E. coli and P. aeruginosa populations and eliminated the remaining taxa. Survival of P. aeruginosa confirmed that the post-treatment communities were not composed exclusively of plasmid-carrying cells. In the pSC101 condition, P. aeruginosa was detected in two of three replicate communities, potentially reflecting stochasticity caused by the daily 2 × 10⁷-fold dilution.

In LB cultures without cellulose sponges, GFP/OD after antibiotic treatment approached the pure plasmid-carrying baselines and decayed more slowly than in antibiotic-free controls (Figure S13). Because the plasmids are non-mobilizable, the post-treatment communities likely consisted primarily of E. coli. By contrast, P. aeruginosa survived in the sponge-containing cultures, consistent with spatial structure that can mediate competition⁴⁵ and promote biofilm formation⁴⁶.

Chemical perturbations reduce plasmid persistence

Previous studies have used chemicals to reduce plasmid abundance and reverse persistence^24,47. Because plasmid abundance and half-life are jointly governed by conjugation, growth, and segregation²⁷, we hypothesized that plasmid-curing compounds could accelerate decay by shortening the segregation bottleneck or strengthening growth competition. We tested rifampicin⁴⁷, promethazine⁴⁷, maprotiline⁴⁸, and phenothiazine²⁴. Using pSC101 as a representative burdensome plasmid, we applied rifampicin alone or with other compounds (Methods). The treatments reduced pSC101 half-life by nearly half (Fig. S14b). A single-day combination produced an effect comparable to continuous treatment, suggesting that most of the effect occurred during the initial treatment phase. This result is consistent with accelerated segregation loss, although additional experiments are required to establish the mechanism.

Discussion

Resistance limits antibiotic efficacy and poses a major threat to human health. Many resistance genes are maintained and disseminated by mobile genetic elements, including plasmids^1–3. Understanding why plasmids persist is therefore important for strategies to contain AMR. Across clonal populations and community contexts, we found that plasmid half-life increased sharply when plasmid-carrying host populations approached purity, either during assembly or after a transient antibiotic pulse. This behavior is consistent with plateaus reported in earlier plasmid-stability assays^20,22,49. Our analyses identify the separation of timescales between slow segregation loss and faster growth competition as the mechanism underlying this initial-state dependence. Transient selection can therefore create a niche-specific segregation bottleneck, whereas community restructuring can additionally sustain conjugative plasmids through host shift. Together, these results show how antibiotic exposure can reorganize populations and communities to produce prolonged plasmid persistence.

Our results provide a mechanistic understanding of the slow reversal of plasmid-mediated resistance after antibiotic treatments, a phenomenon that occurs across microbial ecosystems ranging from the human oral, urinary, and gut microbiomes, to those of livestock. Macrolide antibiotics have been reported to increase the abundance of resistant streptococci (primarily carrying the mef gene or the macrolide-lincosamide-streptogramin-B resistance cassette erm(B)) in oral microbiomes to more than 80%; resistance abundance then required more than 6 months to return to baseline⁹. The authors attributed this prolonged resistance carriage to the long half-life of azithromycin, and the low burden of the erm(B) gene. Similarly, veterinary β-lactams strongly amplified the abundance of resistant bacteria in the pig intestinal microbiome, persisting for more than 3 weeks beyond the drug withdrawal. The extended-spectrum β-lactamase CTX-M-1 was carried by both an exogenous conjugative plasmid and indigenous gut microbiota. Consistent with our results (Figs. 3, 4 and S9), the authors observed significant amplification of resistance and donor abundances upon antibiotic selection, followed by rapid decline of the donor strain abundance after antibiotic withdrawal. Importantly, both the erm(B) cassette^13–17 and the bla_CTX-M-1 gene^18,19 are commonly found on conjugative plasmids and other mobile genetic elements. Yelin et al. reported that resistance in urinary tract infections shows autocorrelation lasting more than six months, with strong and persistent correlation between patients’ antibiotic purchase history and subsequent resistance¹². Anthony et al. observed increases in the abundance of β-lactam, macrolide, and tetracycline resistance following treatment with cephalosporin (β-lactam), azithromycin (macrolide), or both¹⁰. During the development of the infant gut microbiome, resistance genes on mobile genetic elements were found to persist longer than those encoded on chromosomes¹¹. The diversity of antibiotics, resistance genes, bacterial hosts, and ecosystems all point to a general population-dynamical mechanism. Since plasmids encode a variety of mechanisms against segregation error, including high copy number, stability systems, active partitioning, and post-segregational killing, the slow process of segregation loss likely serves as a common kinetic bottleneck to promote plasmid persistence.

The magnitude of this effect should depend on ecological context. First, our experiments revealed host shift as an alternative mechanism for plasmid persistence. Resolving the relative contributions of these mechanisms will require finer-resolution measurements within communities. Establishing host–plasmid associations would inform strain-level plasmid dynamics^50,51, and help determine the main driver of plasmid persistence in different ecological contexts. Second, our experiments approximated a relatively closed community with little or no influx of plasmid-free competitors. In natural and clinical settings, dispersal from the environment or untreated reservoirs can reintroduce these plasmid-free competitors. Their colonization of the plasmid carriers’ niche should shorten plasmid persistence by restoring growth competition. Conversely, as discussed in Fig. 5, plasmids can still persist at the community level if some host niches remain inaccessible to plasmid-free competitors. At the community level, selectively removing plasmid-free competitors has two opposing consequences. It can lengthen the persistence time of a plasmid within its current host background, but it may also reduce opportunities for transfer into other sensitive backgrounds. In our experiments, multiple strains were eliminated from the Keio communities after the antibiotic pulse (Fig. 4b,c), indicating they had not yet acquired the plasmid before selection. Therefore, the impact of an antibiotic pulse on the emergence of new plasmid-carrying strains should depend on the balance between plasmid residence time before selection and the timescale of transfer into additional host backgrounds. Thus, from a resistance epidemiology perspective, antibiotic-induced plasmid persistence could stabilize resistance within a lineage while not necessarily increasing, and in some cases potentially limiting, host-range expansion of the plasmid.

While our work focuses on plasmid population dynamics, the prolonged persistence also has strong evolutionary consequences. Prolonged plasmid carriage enables co-evolution of the plasmid and host, creating opportunities for compensatory mutations. Long-term co-evolution could ameliorate the burden of the plasmid²⁰, and may even turn the cost of plasmid carriage from deleterious to beneficial²². For instance, in our experiments, plasmid burden was reduced or even became negligible over time in some clones, as reflected by increases in maximum growth rate and area under the curve (AUC) (Fig. S6), and elevated clone-level plasmid half-lives in comparison to their ancestors (Fig. 2d). These results are consistent with previous experimental and modeling work showing that antibiotic selection and compensatory adaptation can jointly promote plasmid stability. San Millan et al. modeled this mechanism through segregation loss, plasmid cost, antibiotic selection, and compensatory evolution²⁰. In contrast, here we adopted a minimal model to isolate the early-stage ecological effect. The segregation bottleneck and compensatory mutations are complementary, with prolonged plasmid persistence increasing opportunities for compensatory mutations. Thus, while certain plasmids should not persist under theoretical models given their high burden and low conjugation rate^24,27, over time, the burden can be reduced to enable long-term persistence before eventual extinction.

The antibiotic-induced persistence also provides an environment-driven explanation for the survival of burdensome plasmids, addressing the so-called plasmid paradox⁷. Previous studies have attributed plasmid survival to biotic factors, such as fast conjugation^24,25,52, variable fitness effects across different hosts²³, spatial partitioning⁵³, and compensatory mutations that ameliorate the plasmid burden^20,22. These studies typically consider constant environments without positive selection for plasmids. However, plasmids in natural and clinical settings experience fluctuating selection forces^54–58. In such cases, burdensome plasmids can persist when the interval between episodes of positive selection is comparable to the plasmid half-life imposed by the segregation bottleneck. This mechanism further raises questions about the efficacy of simply reducing antibiotic use to contain AMR⁶—similar persistence can be achieved with far less antibiotic exposure with sporadic selection. Conversely, pulsed antibiotic treatments have been proposed as a strategy to constrain the development of resistance^59–63. Such strategies should be designed to avoid the bottleneck region to prevent slow population responses that disproportionately favor resistant bacteria.

The clinical, ecological, and evolutionary implications of the prolonged plasmid persistence after an antibiotic pulse underscore the need for strategies that counteract unintended antibiotic exposure. Here, we showed that plasmid-curing chemicals can accelerate plasmid abundance decay, suggesting a potential approach to mitigate persistence caused by transient antibiotic exposure. Future work identifying the mechanism of action of each curing compound could help guide the choice of intervention based on the dominant driver of plasmid persistence, such as segregation bottleneck, low plasmid burden, or fast conjugation.

Overall, our work established the timescale decoupling between segregation loss and growth competition for burdensome plasmids and demonstrated the disproportionate impact of transient antibiotic pulses on plasmid persistence. Our findings provide a generalizable mechanism for the prolonged ecological memory of plasmid carriage following antibiotic exposure and have implications for clinical management of antibiotic use and the ecology and evolution of plasmids.

Methods

Statistical analysis

No a priori sample-size calculation was performed. Unless otherwise stated, experiments used three independent biological replicates per condition. These sample sizes were selected to test the large, several-fold differences in plasmid persistence observed in this study and to evaluate reproducibility across independent cultures.

The archived analysis scripts were tested with Python 3.14.2, NumPy 2.4.1, SciPy 1.17.0, pandas 2.3.3, Matplotlib 3.10.8, and scikit-learn 1.8.0, as specified in the archived environment file.

For pairwise comparisons, we did not assume equal variances between groups. We therefore used two-sided Welch’s t-tests to assess differences in group means. Welch–Satterthwaite degrees of freedom were used for the test statistics. No formal tests of normality were performed because of the small sample sizes. Data are reported as mean ± SD or mean ± SE, as indicated in the text and figure legends. Mean differences and 95% confidence intervals are also reported where applicable.

Effect sizes were calculated as Cohen’s d, defined as the difference between group means divided by the pooled standard deviation $d=\left(x_1-x_2\right)/s_p$, and $s_p=\sqrt{\frac{\left(n_1-1\right)s_1^2+\left(n_2-1\right)s_2^2}{n_1+n_2-2}}$.

For right-censored half-life estimates where plasmid abundance did not decrease below the half-threshold during the experiment, the final observation time was used as a conservative lower-bound estimate for hypothesis testing, as indicated in the figure legends.

Mathematical modelling of plasmid dynamics

Clonal population

To investigate the plasmid dynamics in the clonal population, we used a previously published model of conjugative plasmid dynamics²⁴.

$$\frac{df}{dt}=\alpha \mu \left(1-\frac{f+p}{c}\right)f-Df+\kappa p-\eta fp$$

$$\frac{dp}{dt}=\mu \left(1-\frac{f+p}{c}\right)p-Dp-\kappa p+\eta fp$$

Here $f$ denotes the cell density of plasmid-free cells, $p$ denotes the cell density of plasmid-carrying cells. $\mu$ is the growth rate of plasmid-carrying cells, $\alpha$ reflects the fitness effect of the plasmid: costly when $\alpha$ > 1 or beneficial when $\alpha$ < 1, $c$ is the carrying capacity, $D$ is the dilution rate, $\kappa$ is the segregation loss rate, and $\eta$ is the conjugation efficiency of the plasmid from plasmid-carrying cells to plasmid-free cells.

Cooper et al.³⁰. obtained an analytical solution of non-mobilizable plasmid dynamics by assuming constant difference in the growth rates between plasmid-free and plasmid-carrying cells. Similar to this derivation, we extended the analytical solution to a conjugative plasmid (Supplementary Note). Under the approximately constant-total-biomass assumption described in the Supplementary Note, the temporal dynamics of plasmid abundance (fraction of plasmid-carrying cells) $P=p/\left(f+p\right)$ can be expressed as

$$P\left(t\right)=\frac{a{P}_0}{\left(a-b{P}_0\right)\exp \left(at\right)+b{P}_0},$$

where

$$a=\delta +\kappa -\eta \left(f+p\right),b=\delta -\eta \left(f+p\right)$$

$$\delta =\left(\alpha -1\right)\mu \left(1-\frac{f+p}{c}\right).$$

As the total cell density $f+p$ is not a constant, we can show

$$\left(1-\frac{1}{\alpha }\right)D\le \delta \le \left(\alpha -1\right)D.$$

The plasmid half-life can then be derived as

$${\tau }_{1/2}=\frac{1}{a}\log \left(\frac{2a-b{P}_0}{a-b{P}_0}\right)$$ 3

For a non-mobilizable plasmid ($\eta =0$), at low plasmid abundance, P₀% $\ll$ 100%, and when segregation loss is drastically slower than growth competition, $\kappa \ll \delta$, the plasmid half-life can be simplified as:

$${\tau }_{1/2}=\frac{1}{\delta +\kappa }\log \left(\frac{2\left(\delta +\kappa \right)-\delta P_0}{\delta +\kappa -\delta P_0}\right)\approx \frac{1}{\delta }\log \left(\frac{2\left(\delta +\kappa \right)}{\delta +\kappa }\right)\approx \frac{1}{\delta }\log 2$$

In contrast, when P₀% = 100%, the plasmid half-life can be approximated as:

$${\tau }_{1/2}=\frac{1}{\delta +\kappa }\log \left(\frac{2\left(\delta +\kappa \right)-\delta P_0}{\delta +\kappa -\delta P_0}\right)\approx \frac{1}{\delta }\log \left(\frac{\delta +2\kappa }{\kappa }\right)\approx \frac{1}{\delta }\log \frac{\delta }{\kappa }$$

To simulate the burdensome plasmid dynamics in clonal populations, we used the following parameters: $\mu =0.6h{r}^{-1}$, $\eta =0$, $\kappa =5\times {10}^{-5}h{r}^{-1}$, $D=0.2h{r}^{-1}$, $\alpha =1.1$, $c=1$. These parameters correspond to plasmid-carrying cells with a doubling time of 1.16 h and a non-mobilizable plasmid with a copy number of 15.08. Simulations were performed with initial population sizes of $p_0=1,0.99,0.9,0.5,0.2$, and $f_0=1-p_0$, using the scipy.integrate.odeint function with default parameters and time step of $0.1hr$.

As discussed in Supplementary Note, $\delta$ is a variable parameter. To fit Eq. (3) to the simulation data in Fig. 1c, we used the scipy.optimize.curve_fit function for nonlinear least-squares fitting by parameterizing $a$ and $b$ in Eq. (3), with appropriate initial conditions.

Estimation of GFP plasmid half-lives

To evaluate the plasmid half-lives theoretically, we need to obtain the rate of growth competition $\delta$ and segregation loss $\kappa$. We estimated $\delta$ based on the plasmid half-lives at P₀% = 20%, in which case the measured plasmid half-life should be ${\tau }_{1/2}\approx \frac{1}{\delta }\log 2$. To estimate segregation loss, we used plasmid copy number. We assumed that, during replication, plasmids are randomly partitioned into daughter cells, and there is no variability in per-cell copy number. In this case, the probability of creating a plasmid-free cell is 2^1-CN/generation. For pSC101, colE1, and pUC, which were passaged daily at a 1:500 dilution, the populations underwent log₂500 generations daily. Therefore, using a pSC101 copy-number estimate of 10.90 ± 0.13, a colE1 estimate of 19.86 ± 0.83, and a pUC estimate of 46.50 ± 1.38 (mean ± SE, n = 3), we can estimate the segregation loss $\kappa$ = 2^1-CN×log₂500. Substituting the estimated $\delta$ and $\kappa$ into Eq. (3), we can obtain the estimated half-lives of the GFP plasmids.

Niche-partitioned community

To simulate the effect of antibiotics on the burdensome plasmid dynamics in communities with niche partition, we extended the model above to include an additional population. Here we denote the two populations as

$$\frac{d{f}_1}{dt}=\alpha \mu \left(1-\frac{s_1+{\gamma }_{12}{s}_2}{c_1}\right)f_1-D{f}_1+\kappa p_1-\eta f_1\left(p_1+p_2\right)-A_1{f}_1$$

$$\frac{d{p}_1}{dt}=\mu \left(1-\frac{s_1+{\gamma }_{12}{s}_2}{c_1}\right)p_1-D{p}_1-\kappa p_1+\eta f_1\left(p_1+p_2\right)$$

$$\frac{d{f}_2}{dt}=\alpha \mu \left(1-\frac{s_2+{\gamma }_{21}{s}_1}{c_2}\right)f_2-D{f}_2+\kappa p_2-\eta f_2\left(p_1+p_2\right)-A_2{f}_2$$

$$\frac{d{p}_2}{dt}=\mu \left(1-\frac{s_2+{\gamma }_{21}{s}_1}{c_2}\right)p_2-D{p}_2-\kappa p_2+\eta f_2\left(p_1+p_2\right)$$

Here $s_1=f_1+p_1$, and $s_2=f_2+p_2$. The key additional assumptions of this model are: 1) populations 1 and 2 engage in interactions with coefficient ${\gamma }_{12}<1$ and ${\gamma }_{21}<1$, and 2) the plasmid-free populations are killed by antibiotics at rate $A_1{f}_1$ and $A_2{f}_2$. For simplicity, we take the two populations to have the same growth rate $\mu =0.6h{r}^{-1}$, same plasmid burden $\alpha =1.1$, segregation loss rate $\kappa =5\times {10}^{-5}h{r}^{-1}$, and no conjugation $\eta =0$. We assume complete niche partition ${\gamma }_{12}={\gamma }_{21}=0$, and $c_1=c_2=1$. Importantly, we assumed population 2 was killed more slowly by the antibiotic, with $A_1=1h{r}^{-1}$ and $A_2=0.3h{r}^{-1}$. Simulations were performed with initial conditions of $f_1=0.2$, $f_2=0.5$, $p_1=0.3$ and $p_2=0$. When population 2 is an incompatible host and is killed more slowly than population 1, a segregation bottleneck can occur without the plasmid abundance reaching ~ 100%.

Plasmids

The pSC101 plasmid was obtained from Addgene (plasmid #138942) and was previously described by Rosado et al.⁶⁴. The colE1 plasmid was obtained from laboratory stocks. The pUC plasmid was previously described by Yao et al.⁶⁵. Reference sequences for the conjugative plasmids are available under the following accession numbers: R388, NCBI RefSeq NC_028464.1; R6K, GenBank LT827129.1; pCU1, GenBank PQ358084.1; and RP4, GenBank MT868884.1.

Quantifying non-mobilizable plasmid dynamics in clonal populations with different initial plasmid abundances

We started the experiment by inoculating glycerol stocks into 500 µL LB Broth (Apex LB Broth Mix, Genesee Scientific, 11−120) for 24 h in deep well plates (Genesee, 27-413S) at 37 °C with shaking at 700 rpm and amplitude of 2 mm, with the corresponding plasmid selections: 50 µg/mL Kanamycin (Thermo Fisher, 11815032) for pSC101 and colE1, 50 µg/mL Spectinomycin (Sigma Aldrich, S4014) for pUC. A 1 µL aliquot of each culture was then inoculated into 500 µL LB in deep well plates with corresponding selections. After growing for another 24 h at 37 °C, the resulting cultures of plasmid-carrying and plasmid-free cells were diluted in fresh LB to yield optical density at 600 (9) nm (OD) ≃ 0.5, mixed at different volume ratios (P₀% = 100%, 99%, 90%, 50%, 20%, and 0%), and OD and GFP (excitation: 488 (9) nm, emission: 520 (20) nm) were measured (Tecan 200 Pro Plate Reader). After subtraction of the LB baseline, GFP/OD was linearly regressed to the plasmid abundance P₀%, yielding estimates of the slope (±SE) and intercept (±SE) later used to quantify the plasmid abundance in the culture based on OD and GFP measurements.

The mixed populations with different initial plasmid abundances were then cultured in LB in deep well plates, shaken at 700 rpm, 2 mm amplitude, at 37 °C, and diluted 1:500 every 24 h, for 12 days. Every 24 h, the cultures were diluted 4-fold (50 µL of culture into 150 µL LB) to measure OD (0.4–0.6) and corresponding GFP intensity. After subtracting the baseline LB readings, the GFP/OD ratios were used to calculate the plasmid abundance. On a single day, variances in OD and GFP measurements in blank controls (pure LB) were calculated to represent the measurement uncertainty in the GFP/OD readout. This measurement uncertainty, together with the uncertainties in the estimated slope and intercept in the calibration curve, was used to calculate the per-well measurement uncertainty (SD) in plasmid abundance using the delta method.

Quantifying non-mobilizable plasmid dynamics after antibiotic pulses of different concentrations

The GFP/OD calibration and experimental initialization were performed as described above. We started the experiment by inoculating glycerol stocks into 500 µL LB Broth for 24 h in deep well plates at 37 °C with shaking at 700 rpm and amplitude of 2 mm, with 50 µg/mL Kan for pSC101 and colE1, 50 µg/mL Spec for pUC. A 1 µL aliquot of each culture was then inoculated into 500 µL LB in deep well plates with corresponding selections. After growing for another 24 h at 37 °C, the resulting cultures of plasmid-carrying and plasmid-free cells were diluted in fresh LB to OD ≃ 0.5, mixed at different volume ratios (P₀% = 100%, 80%, 50%, 20%, and 0%), and OD and GFP were measured. After subtraction of the LB baseline, GFP/OD was linearly regressed (σ-weighted) to the plasmid abundance P₀%, yielding estimates of the slope m (±SE) and intercept b (±SE) later used to quantify the plasmid abundance in the culture based on OD and GFP measurements.

The equally mixed populations (P₀% = 50%) were then cultured in LB in deep well plates, shaken at 700 rpm, 2 mm amplitude, at 37 °C, and diluted 1:500 every 24 h. Between day 2 and 3, the populations were cultured in LB or LB with serially diluted antibiotics (with dilution rates of 2⁶, 2⁵, …, 2⁰ of 50 µg/mL Kan or 50 µg/mL Spec). After the 24 h antibiotic pulse, all populations were transferred to LB medium and cultured for another 17 days at a 1:500 dilution. Every 24 h, the cultures were diluted 4-fold (50 µL of culture into 150 µL LB) to measure OD (0.4–0.6) and corresponding GFP intensity. After subtracting the baseline LB readings, the GFP/OD ratios were used to calculate the plasmid abundance.

Kan (aminoglycoside) and Spec (aminocyclitol) both target the 30S ribosomal subunit and inhibit translation, thus reducing the expression of GFP. Under antibiotic selection, plasmid copy number increases, increasing plasmid encoded GFP level. We calibrated the GFP/OD to plasmid abundance using mixture standards prepared with Kan and Spec due to the potentially unknown plasmid abundance drift related to segregation error. Calibration based on mixtures prepared without antibiotics produces the same qualitative trends, indicating our conclusions do not depend on the calibration regime.

Quantifying conjugative plasmid dynamics in clonal populations using selective plating

E. coli DA28102 cells (chloramphenicol resistant, Cm^R) carrying either the pCU1 (carbenicillin resistant, Carb^R) or R6K (streptomycin resistant, Strp^R) plasmid and plasmid-free DA28102 cells were cultured overnight in 500 µL LB supplemented with 25 µg/mL Cm (Sigma Aldrich, C0378) and appropriate antibiotics for the plasmids: 100 µg/mL Carb (Thermo Fisher, 10177012) for pCU1 and 100 µg/mL Strp (Sigma Aldrich, S6501) for R6K. The overnight cultures were diluted to OD ≃ 0.5, and plasmid-carrying cells were mixed with plasmid-free cells at different volume ratios, yielding P₀% of 100%, 99%, 90%, 50%, or 20%.

E. coli Keio strains #3 and #4 in Comm57 (Cm^R, Supplementary Data 1) carrying pCU1 or R6K were cultured overnight in 500 µL LB supplemented with 25 µg/mL Cm and appropriate antibiotics for the plasmids (100 µg/mL Carb for pCU1 and 100 µg/mL Strp for R6K). Keio strains #1, #15, and #76 in Comm87 (Carb^R, Supplementary Data 2) carrying R388 (Trimethoprim resistant, Trim^R) were cultured overnight in 500 µL LB supplemented with 100 µg/mL Carb and 10 µg/mL Trim (Sigma-Aldrich, T7883). The overnight cultures were diluted to OD ≃ 0.5, and plasmid-carrying cells were mixed with plasmid-free cells at different volume ratios, yielding P₀% of 100%, 90%, or 50%.

Populations were cultured at 37 °C in LB with 25 µg/mL Cm or 100 µg/mL Carb to prevent contamination, shaken at 700 rpm with an amplitude of 2 mm, and diluted 1:100 (5 µL into 500 µL) every 24 h. The plasmid abundance was quantified using 100 µg/mL Carb, 100 µg/mL Strp, or 10 µg/mL Trim selective plating (LB agar, Genesee, 11−122) for pCU1, R6K, or R388, respectively, every 3–5 days using drip plating. Briefly, after making an OD measurement to estimate cell culture density, each culture was serially diluted 10⁴ to 4 × 10⁶-fold, such that a 10 µL sample would contain 30-150 CFUs on LB plates. A 10 µL droplet from the diluted culture was then placed on an LB or LB + selection plate, and the plate was placed at a 90° angle to allow the droplet to run down the plate. Plates were then allowed to dry flat on the lab bench before being placed at 37 °C for colony formation. Colonies were counted, and the percentage of plasmid-carrying cells was quantified after 24 h. When plasmid abundance was low, additional samples at lower dilution factors were plated on LB + selection plates to quantify their abundances.

Extraction of plasmid half-lives

The measured time series of plasmid abundance (mean ± SD for GFP/OD measurement or SE for selective plating) from each biological replicate were log-transformed and linearly interpolated to determine the half-life of plasmid abundance decay with respect to the initial mixture ratio. The SEs of the half-life estimates were calculated using the delta method. For samples in which plasmid abundance never dropped by half, the half-life estimates were right-censored at the final observation time. For experiments with an antibiotic pulse, the half-life was calculated as the time after the pulse at which plasmid abundance dropped below one-half of its day-3 value (measured immediately after the pulse).

Quantifying plasmid dynamics under chemical treatment

Populations of E. coli MG1655 carrying pSC101 were grown overnight from glycerol stock in 500 µL LB + 50 µg/mL Kan (24 h at 37 °C with shaking at 700 rpm, 2 mm) to ensure population purity. The populations were cultured in LB in black 96-well plates (Sigma Aldrich, CLS3603-48EA) with periodic shaking (every 5 min, orbital for 5 s and 2 mm), and diluted daily 1:200 (1 µL into 200 µL). When treated, the cell cultures were supplemented with 8 ng/mL Rifampicin (Sigma Aldrich, R3501), 80 µg/mL Promethazine (Sigma Aldrich, 46682), 16 µg/mL Maprotiline (Sigma Aldrich, M9651), and 320 µM Phenothiazine (Sigma Aldrich, 88580). Drip plating was performed to determine the plasmid abundance.

Quantifying plasmid fitness effects using a microplate reader

Growth curve measurement

Plasmid-free and plasmid-carrying strains were inoculated from glycerol stock in 500 µL LB Broth overnight in deep well plates at 37 °C with shaking at 700 rpm and amplitude of 2 mm, with appropriate antibiotic selection (MG1655, DA28102: no antibiotics, pSC101, colE1: 50 µg/mL Kan, pUC: 50 µg/mL Spec, pCU1: 100 µg/mL Carb, R6K: 100 µg/mL Strp). A 2 µL aliquot of the overnight culture was inoculated into 198 µL of LB with appropriate antibiotic selection. The edge wells were filled with water only, without LB or bacterial culture, and were not used as culture wells to minimize edge effects. To reduce evaporation, 50 µL of mineral oil (Sigma Aldrich, M5904) was added to each culture well. Cultures were incubated in black 96-well plates with periodic shaking (every 5 min, orbital for 5 s and 2 mm), with OD600 measured every 5 min for 24 h.

Estimation of plasmid fitness effects

OD values were blank-corrected for downstream analysis. We quantified the fitness effects of the plasmids using two metrics, area under the curve (AUC) and maximum growth rate. AUC was calculated from the OD time series over a 24 h period. To estimate maximum growth rate, OD values from the first 30 min were excluded, and the rest were floored at 0.01 to reduce amplification of measurement noise during the early growth phase. Local growth rates were then calculated using rolling-window linear regression on log-transformed OD values. For each window of consecutive time points, log(OD) was regressed against time, and the fitted slope was taken as the local growth rate. The maximum fitted slope was taken as the maximum growth rate. For each metric, plasmid-carriage fitness was quantified as the ratio for the plasmid-carrying strain relative to the plasmid-free strain.

Quantifying plasmid copy numbers through qPCR

Sample preparation

Plasmid-carrying strains were inoculated from glycerol stock in 500 µL LB Broth overnight in deep well plates at 37 °C with shaking at 700 rpm and amplitude of 2 mm, with appropriate antibiotic selection (pSC101, colE1: 50 µg/mL Kan; pUC: 50 µg/mL Spec). A 1 µL aliquot of each overnight culture was inoculated into 500 µL of LB + antibiotics and cultured for 24 h. These cultures were used for qPCR.

Calibration

The cultures were diluted by mixing 50 μL of culture with 150 μL of molecular-grade water. For cell lysis, 50 μL of the diluted culture was boiled at 98 °C for 20 min. Then cell lysate was serially diluted by 10-fold, and 4 µL of each diluted sample was used in qPCR reactions, with six technical replicates performed for each dilution rate. qPCR reactions were thermally cycled at 95 °C for 5 min, followed by 40 cycles of 95 °C for 1 s and 60 °C for 20 s, using TaqMan™ Fast Advanced Master Mix (Applied Biosystems, Cat. No. 4444965) on a QuantStudio™ 5 Real-Time PCR System (Applied Biosystems). Ct values were plotted against log-transformed dilution rates, followed by linear regression to obtain the slope m of the standard curve.

Plasmid copy number estimation

For each plasmid, we included two primer–probe sets targeting the chromosomal E. coli dxs gene and the plasmid, respectively (Supplementary Data 3). Plasmid copy numbers were estimated as PCN = 2^Ctp/mp / 2^Ctc/mc. Here, Ctp and Ctc denote the Ct values of the plasmid and chromosome targets in the measured samples, respectively, and mp and mc denote the slopes obtained from the calibration curves for the two probes. Based on long-read sequencing, plasmid pSC101 formed a concatemeric structure. Therefore, each plasmid copy carried two copies of sfGFP, the target used for qPCR quantification. The copy number estimated from the equation above was therefore divided by 2 to obtain the actual pSC101 copy number.

Plasmid curing of evolved clones

We employed a CRISPR/Cas9-based plasmid curing system to eliminate plasmids from the evolved clones. The pCas9 plasmids were constructed by assembling DNA fragments encoding Cas9, a P_BAD promoter-driven gRNA expression cassette, a chloramphenicol-resistance marker, a MOB sequence, and a temperature-sensitive pSC101 origin of replication. All fragments were amplified using 2 × Phanta Flash Master Mix (P520-02, Vazyme Biotech) and assembled by Gibson assembly using the ClonExpress Ultra One Step Cloning Kit V2 (C116-02, Vazyme Biotech). L-arabinose (0.2% w/v; Sigma-Aldrich, 1014920100) was used to activate the P_BAD promoter and induce transcription of gRNAs either targeting sfGFP or the origin of replication of pUC. The plasmids were individually transformed into E. coli BW29427 and cultured at 30 °C for 36 h. Correct transformants were validated by plasmid sequencing. Transformants were grown overnight in LB supplemented with 0.3 mM 2,6-Diaminopimelic acid to support BW29427 growth. BW29427 supplies RP4 conjugation machinery, allowing mobilization of pCas9 if the plasmid carries a compatible MOB sequence.

Overnight cultures of the transformant and evolved clones were pelleted by benchtop centrifugation for 3 min, the supernatants were discarded, and the pellets were resuspended in the same volume of LB. The resuspensions were mixed in equal volumes and incubated at room temperature for 1–2 h to allow pCas9 transfer into the target clones. The mixture was diluted and plated on double-selection LB agar plates with 25 μg/mL Cm and 50 μg/mL Kan, or 25 μg/mL Cm and 50 μg/mL Spec. The plates were incubated at 30 °C for 36 h. Single colonies were inoculated in 500 μL LB supplemented with 0.2% L-arabinose (w/v) and 25 μg/mL Cm to induce gRNA transcription. Cultures were grown at 30 °C overnight to cut the plasmid. Overnight cultures were inoculated at a 10⁴-fold dilution in 500 μL LB and shaken at 42 °C overnight to eliminate pCas9. The cultures were plated on LB agar plates. Colony PCR was used to identify and confirm colonies cured of both the target plasmid (pSC101, colE1, or pUC) and pCas9.

Competition assay to quantify plasmid half-lives in evolved clones

Glycerol stocks of plasmid-carrying strains and their cured counterparts were inoculated in 500 µL LB (or LB supplemented with 50 µg/mL Kan for pSC101 and colE1, 50 µg/mL Spec for pUC) for 24 h in deep well plates at 37 °C with shaking at 700 rpm and amplitude of 2 mm. A 1 µL aliquot of each culture was then inoculated into 500 µL LB in deep well plates, with corresponding selections for the plasmid-carrying strains. After growing for another 24 h at 37 °C, the subsequent cultures of plasmid-carrying cells and plasmid-free cells were diluted in fresh LB. OD and GFP intensity were measured for each diluted culture and for a 50% (v/v) mixture prepared at equal OD (≃ 0.35). The GFP/OD ratios were calibrated to plasmid abundance P₀% (100%, 50%, and 0%) as described above, with estimates of the slope (±SE) and intercept (±SE) used to quantify the plasmid abundance in the culture based on OD and GFP measurements.

The mixed populations (P₀% = 50%) were then cultured in 500 µL LB in deep well plates, shaken at 700 rpm, 2 mm amplitude, at 37 °C, and diluted 1:500 every 24 h, for 4 days. Every 24 h, the cultures were diluted 4-fold (50 µL of culture into 150 µL LB) to measure OD (0.4–0.6) and corresponding GFP intensity. After subtracting the baseline LB readings, the GFP/OD ratios were used to calculate the plasmid abundance.

Generation of barcoded strains

The Keio strains were obtained from Dr. Meta Kuehn at Duke University. Barcoded plasmids were generated following a previously published protocol⁴⁴. Briefly, DNA barcodes containing random base sequences were synthesized by an external source and introduced into a linearized plasmid backbone (p15A-GFP-Carb^R for Comm87, and p15A-mCherry-Cm^R for Comm57) by Gibson assembly cloning. Barcoded plasmid libraries were subsequently transformed into NEB® 5-alpha Electrocompetent E. coli cells (Cat. no. C2989), generating a library of barcoded plasmids that was recovered by plasmid miniprep after overnight expansion. Plasmid libraries were used to transform individual Keio strains by chemical transformation. Each strain was validated by Sanger sequencing. Each barcode contained two random 18-bp sequences surrounded by Illumina sequencing adapters.

Plasmid dynamics in synthetic E. coli communities

Passaging experiment

E. coli Keio strains with barcoded plasmids were cultured in deep well plates for 24 h at 37 °C with shaking at 700 rpm and amplitude of 2 mm from glycerol in 500 µL LB supplemented with 25 μg/mL Cm (for Comm57 strains, Supplementary Data 1) or 100 μg/mL Carb (for Comm87 strains, Supplementary Data 2) to select for the barcode plasmids. Donor strains were cultured for 24 h at 37 °C with shaking at 700 rpm and amplitude of 2 mm in 500 µL LB with double selection for both the barcode plasmid and the target plasmid (10 μg/mL Trim for R388, 10 μg/mL tetracycline (Sigma Aldrich, 58346-M) for RP4, 100 μg/mL Carb for pCU1, and 100 μg/mL Strp for R6K). The overnight cultures of the plasmid-free strains were mixed in equal proportions, and the mixed recipient community was further mixed with the overnight culture of the donor strain at a 3:1 ratio. By doing so, we constructed five communities, Comm87 + R388, Comm57 + R388, Comm57 + RP4, Comm57 + pCU1, and Comm57 + R6K. Each of the four Comm57 communities contained 57 strains: one donor and 56 plasmid-free strains. For example, Comm57 + R388, with Keio #1 as the donor, excluded Keio strains #2–#4, which served as donors for the other three plasmids. At an initial plasmid abundance of 25%, communities were cultured in LB with either 100 μg/mL Carb (for Comm87) or 25 µg/mL Cm (for Comm57) to prevent contamination and maintain the barcode plasmids, and diluted 1:100 (5 µL into 500 µL) every 24 h. The four Comm57 communities were also cultured at a daily dilution of 1:500 (1 µL into 500 µL) (Fig. S8). The plasmid abundance was quantified by selective plating. Cell cultures on the given sampling days were pelleted by benchtop centrifugation, the supernatants were discarded, and the pellets were stored at −20 °C before next-generation sequencing (NGS) library preparation.

Invasion experiment

E. coli Keio strains #3 and #4 from Comm57 (Supplementary Data 1), and #76 from Comm87 (Supplementary Data 2) were cultured in deep well plates for 24 h at 37 °C with shaking at 700 rpm and amplitude of 2 mm from glycerol in 500 µL LB supplemented with 25 μg/mL Cm (for #3 and #4) or 100 μg/mL Carb (for #76). A 10 µL aliquot of each thawed glycerol sample—Comm57 + pCU1 and Comm57 + R6K from day 3, and Comm87 + R388 from day 12—was inoculated in 500 µL LB with either 25 µg/mL Cm (for Comm57) or 100 μg/mL Carb (for Comm87) and cultured for 24 h at 37 °C with shaking at 700 rpm and amplitude of 2 mm. The overnight cultures of the invading strains were mixed in equal volumes with the corresponding community. The invaded communities were then diluted 1:100 (5 µL into 500 µL) every 24 h, and their plasmid abundances were quantified by selective plating. Plasmid half-lives were quantified as the time from the start of the invasion until plasmid abundance fell below 50% of its initial value.

Library preparation

To prepare the NGS library, nuclease-free water was added to cell pellets, and samples were boiled at 98 °C for 10 min. Barcodes were amplified by PCR (NEBNext Ultra II Q5 Master Mix, NEB, M0544) using primers containing unique 8-bp sample indexes. PCR products were pooled and separated on a 2% agarose gel. DNA was extracted and purified using the Zymoclean Gel DNA Recovery Kit (VWR, 77001-122) according to the manufacturer’s instructions. The purified PCR products were compatible with standard Illumina sequencing platforms. Libraries were denatured, diluted, and mixed with an equal amount of PhiX according to standard Illumina protocols for 16S library preparation on the MiniSeq. The final libraries contained 50% PhiX (Illumina, FC-110-3001) to increase sequence diversity and were sequenced using 300-cycle paired-end reads on an Illumina MiniSeq (Illumina, FC-420-1003 or FC-420-1004). The final library structure was:

5′-AATGATACGGCGACCACCGAGATCTACACXXXXXXXXAACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNNNNNNNNNNNNNCCTCAGGGTCACTAGGNNNNNNNNNNNNNNNNNNAGATCGGAAGAGCACACGTCTGAACTCCAGTCACXXXXXXXXATCTCGTATGCCGTCTTCTGCTTG-3′

Here Ns represent the barcode sequences, and Xs represent sample index.

NGS data analysis

A customized Galaxy workflow was used to demultiplex the samples based on the index sequences from the reads and to count the barcodes of each community. The following tools were used: Trim sequences (Galaxy Version 1.0.2+galaxy0), Barcode Splitter (Galaxy Version 1.0.1), FASTQ joiner (Galaxy Version 2.0.1.1+galaxy0).

Plasmid dynamics in multispecies communities

Passaging experiment

We utilized the previously reported SynkC community, consisting of eight isolates from a sink p-trap in a Duke University Hospital intensive care unit⁴⁴. Based on the 16S V3/V4 region, the community comprised one Pseudomonas aeruginosa strain, three Bacillus strains, one Citrobacter strain, and three Enterobacter strains. The eight isolates and E. coli MG1655 were cultured in deep well plates for 24 h at 37 °C with shaking at 225 rpm and amplitude of 2 mm from glycerol stocks in 1000 μL LB medium. E. coli MG1655 carrying one of the three plasmids (pSC101, colE1, and pUC) were cultured overnight from glycerol in LB medium supplemented with 50 μg/mL Kan for pSC101 and colE1, or 50 μg/mL Spec for pUC.

Since we aimed to achieve habitat partitioning in sponges, we performed extreme dilution to enable stochastic localization of bacterial cells of each strain within the sponge’s porous structure⁴⁵. The overnight cultures of all 10 strains were diluted to an OD of 0.3 and mixed together by volume so that plasmid-free and plasmid-carrying MG1655 each comprised 25% of the total population, and each of the eight sink isolates comprised 6.25%. The mixed communities were diluted and inoculated in sponges immersed in 1 mL LB, yielding a further 4 × 10⁶-fold dilution relative to OD = 0.3. To assess the baseline GFP/OD for each plasmid, the three plasmid-carrying MG1655 strains were diluted 1.6 × 10⁷-fold relative to OD = 0.3 and inoculated in sponges immersed in 1 mL LB supplemented with corresponding antibiotics (50 μg/mL Kan for pSC101 and colE1, or 50 μg/mL Spec for pUC). The communities were cultured at 37 °C with shaking at 700 rpm and amplitude of 2 mm. The cellulose sponges (ARCLIBER, TBWS-10006) were manually cut into 0.55 × 0.55 × 2 cm pieces, placed in deep 96-well plates, and autoclaved for 45 min.

Every 24 h, we squeezed the sponges using pipette tips and extracted the liquid cultures. As above, the cultures were diluted 4-fold (50 µL of culture into 150 µL LB) to measure OD and corresponding GFP intensity. The cultures were further diluted and inoculated in new sponges at a final dilution of 2 × 10⁷. Between days 1 and 2, LB supplemented with the corresponding antibiotics was used instead of LB alone (50 μg/mL Kan for pSC101 and colE1, or 50 μg/mL Spec for pUC). The same experiment was performed for communities in the absence of sponges (Fig. S13). The extracted liquid cultures were stored at −80 °C before 16S V3/V4 amplicon sequencing. 16S sample preparation, sequencing, and analyses were performed by SeqCenter, LLC.

Sample preparation and sequencing

Genomic DNA was extracted using the ZymoBIOMICS DNA Miniprep Kit (Zymo Research, #D4300). Samples were prepared using Zymo Research’s Quick-16S kit (Zymo Research, #D6400) with phased primers targeting the V3/V4 regions of the 16S gene. The forward primer was (341f) CCTACGGGDGGCWGCAG/CCTAYGGGGYGCWGCAG; the reverse primer was (806r) GACTACNVGGGTMTCTAATCC.

Following cleanup and normalization, samples were sequenced on a P1 or P2 600-cycle NextSeq 2000 flow cell to generate 2 × 301 bp paired-end reads. Quality control and adapter trimming were performed with Illumina bcl-convert (v4.2.4). Primer-dimer sequences identified as PCR artifacts were filtered from the generated FASTQ files based on the following criteria: read length >150 bp, PolyN strings <10 sequential Ns; PolyG strings <150 sequential Gs.

Sequence analysis

Sequences were imported into QIIME 2⁶⁶ for analysis. Primer sequences were removed using QIIME 2’s cutadapt⁶⁷ plugin using the following degenerate primer queries: region: V3/V4; forward trim sequence: CCTAYGGGNBGCWGCAG; reverse sequence: GACTACNVGGGTMTCTAATCC.

Sequences were then denoised using QIIME 2’s DADA2⁶⁸ plugin. Denoised sequences were placed into a feature table detailing which amplicon sequence variants (ASVs) were observed in which samples, and how many times each ASV was observed in each sample. Identified ASVs were taxonomically assessed using the SILVA 138 99% full-length sequence database⁶⁹ and the VSEARCH⁷⁰ utility within QIIME 2’s feature-classifier plugin. ASVs were then collapsed to their lowest taxonomic units (strain, species, genus, etc.), and their counts were converted to reflect their relative frequency within a sample.

Reporting summary

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

Author contributions

Conceptualization: Z.Z., A.W., and L.Y.; Methodology: Z.Z. and A.W.; Software: Z.Z.; Validation: Z.Z. and Z.Y.; Formal analysis: Z.Z. and A.W.; Investigation: Z.Z., A.W., Z.Y., X.C., J.Q., K.L., and G.H.; Resources: Z.Z., A.W., Z.Y., X.C., K.L., H.S., and L.Y.; Writing–original draft: Z.Z., A.W., and L.Y.; Writing–review and editing: all authors; Visualization: Z.Z.; Supervision: L.Y.; Funding acquisition: L.Y.

Peer review

Peer review information

Nature Communications thanks Jana Huisman, Alvaro San Millan and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

L.Y. discloses support for the research of this work from National Institutes of Health (R01AI125604, R01GM098642, R01EB031869) and the National Science Foundation (Cooperative Agreement No. EEC-2133504).

Data availability

Amplicon sequencing data for E. coli Keio communities and 16S sequencing data used in this study are available under NCBI BioProject accession PRJNA1360409: https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1360409/. Other data have been deposited at https://github.com/youlab/Plasmid-Ghost-Effect, with a Zenodo record⁷¹: https://doi.org/10.5281/zenodo.22184932. Source data are provided with this paper.

Code availability

All code has been deposited at https://github.com/youlab/Plasmid-Ghost-Effect, with a Zenodo record⁷¹: https://doi.org/10.5281/zenodo.22184932.

Competing interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT and Claude to improve the language of the manuscript. After using these tools, the authors reviewed and edited the content as needed; they take full responsibility for the content of the published article.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-78240-2.